-1196 796 STRTISTICAL SIMLITY CONTROL OF EMNIIE ENIAJUCHENTS 1/2

M
(U) NEXUS ASSOCIATES HAYLAMONA ACHER SEP697
MES/CRIL-07-2 DAN9-83-N-6967
UNWS c& ASSIFIEDi'132 UL

ma-mommhhhhul

mhmhhmhhmhIN
 WII1 1.08
jlfl 1.2 I I I:::: WhlI
i6

IIU

25
1-.. 11111.4 _L --- .-
 CONTRACT REPORT GL-87-2
STATISTICAL QUALITY CONTROL

.a -- OF ENGINEERED EMBANKMENTS
by
Gregory Baecher
CD-- NEXUS Associates "
Wayland, Massachusetts 01778-1401

OTIC
ELECTE
5':',- OCT26 1987

kh f-\ -.

September 1987
Final Report

k It

t,, ,,
&DEPARTMENT OF 1 HE ARMY

US Amyorp ofEngineers
Washington, DC 20314-1000
S
Conrat Nor oDACW39-83-M-0067
I ,Geotechnical Laboratory
LABORATORY US Atmy Engineer Waterways Experiment Station
PO Box 631, Vicksburg, Mississippi 39180-0631

S. D
 Destroy this report when no longer needed. Do not return
it to the originator.

The findings in this report are not to be construed as an official

Depatr ment of the Army position unless so designated

by other authorized documents.

The conttents of this report are r ot to he used for

advertising, pulLblication, or promotional purpoes.
Citation of trade names (oes rlot (:o'istitite all
official endorsemrent or ipproval of the iwr, )f
such cornmercial prodLcts

9t

,V,
v%
 ii.

UnClassified
SECURITY CLASSIFICATION OF THIS PAGE
AZ
Form Approvred
REPORT DOCUMENTATION PAGE OMS8No 0704-0188
la REPORT SECURITY CLASSIFICATION lb RESTRICTIVE MARKINGS
Unclassified
2a SECURITY CLASSIFICATION AUTHORITY 3 DISTRIBUTION /AVAILABILITY OF REPORT
* ______________________________________________ App roved foer piibl i c relIease;
Zb OECLASSIFICATION/DOWNGRAOING SCHEDULE distribution unliiited.

4 PERFORMING ORGANIZATION REPORT NUMBER(S) 5 MONITORING ORGANIZATION REPORT NUMBER(S)

Lontnrat Report (;J-87-2

6a NAME OF PERFORMING ORGANIZATION 6b OFFICE SYMBOL 7a NAME OF MONITORING ORGANIZATION

NNF;XUii Assoc i at, , (Ofapplicable) I'SAHWfrS
% Gee0tec)Ini Cal (dbe rat II
6c. ADDRESS (City. State. and ZIP Code) 7b ADDRESS (City, State, and ZIP Code)

WIvland, MAS I0i78-J/4Il L '' 3

Vi Cksburg, MS 39180-0631

8. NAME OF FUNDINGi SPONSORIN 8b OFFICE SYMBOL 9 PROCUREMENT INSTRUMENT IDENTIFICATION NUMBER

ORGANIZATION(iaplcbe

t'S Arm'- Iorp ,IIt)jttr ACW39-8 3-M'-00l,7

8c. ADDRESS (City. State. and ZIP Code) 10 SOURCE OF FUNDING NUMBERS
PROGRAM PROJECT I TAS ~ WORK UNIT
Wlslll lt't oll 1)( -11 i I - 1000)1 ELEMENT NO NO 0 ACCESSION NO
r 3I2
11 TITLE (include Security Classification)

3,Stift i:, t I, ,il i t\ (lnt il ot (igint-ct-d Flhankments

12 PERSONAL AUTHOR(S)

1 3a TYPE OF REPORT 1 3b TIME COVERED 14 DATE OF REPORT (Year, M~onth, Day) 15 PAGE COUNT
Fii II rctpl I FROM _ TO _ 1ell f ember 1987

16 SUPPLEMENTARY NOTATION
:%%,Ii I l cliI I iNat ionilIV li Ii I IntiiIIat iLenISv-rvit ct, )1283) Port Rovyal Read,
Spri i1141 it- Ill , Vt\ 12 10~ 1

17 COSATI CODES 18 SUBJECT TERMS (Continue on reverse if necessary and identify'by block number)
FIELD GROUP SUB-GROUP ~ 11111I i, ~
It

19 ABSTRACT (Continue on reverse if necessary and identify by block number)

tl ti i -,II t I t lit tt I I t rtc

t ilt t I tlltLI -i1 Wii
tll5 t1 it- iIt11,1rit Lci t ill t oil

1
I olr III it I It ll Iri-s
p iik II- l It 1 )1- I lIT I I i a r t I, W I IL I , .jt i st I t I Ill ttt

*20 DISTRIBUTION' AVAILABILITY OF ABSTRACT 121 ABSTRACT SECRiTY CLASSIFICATION

*0 UJNCLASSIFIED UNLIMITED [I SAME AS RPT C3 SERS
,~ I III .r i- 1 c
22a NAME OF RESPONSIBLE IFDIVDULI 22b TELEPHONE (include Area Code) ~HOFFICE SYMBOL

DO Form 1473, JUN 86 Pre vtous editions are obsolete SECLjRtTi CLASSIFICATON OF THS PAGE

%. % %
')
 * ~~- - w vFrwrwinw

PREFACE %

This report, prepared by Gregory B. Baecher of NEXUS Associates, Wayland,

Massachusetts, with assistance from D. DeGroot, C. Erikson, and A. Pais, under

Contract No. DACW39-83-M-0067, provides details for the statistical analysis of

geotechnical engineering aspects of new dam projects. It was part of work done

by the US Army Engineer Waterways Experiment Station (WES) in the US Army Civil

Works Investigation Study sponsored by the Office, Chief of Engineers, US Army.

This study was conducted during the period October 1983 to September 1985 under

CWIS Work Unit No. Civis 32221, entitled Probabilistic Methods in Soil

Mechanics. Mr. Richard Davidson was the OCE Technical Monitor. NP

The report presents an introduction to statistical quality control as

applied to the construction inspection of engineered embankments. It is

intended to be introduction to potential users who have little or no background 'p

in statistics. Examples in the report are drawn from actual construction

records of dam projects, and IBM-compatible microcomputer software supporting "J.

this report has been developed under separate funding. Two other instructional

reports were prepared under the same contract, "Statistical Ani)Iysis of

'
Geotechnical Data," and " Erro Analy:;is fi e;(: tec hnic ] Bgl necne"q, in

addition to a final report.

Ms. Mary Ellen Hynes-Griffin, Earthquake Engineering and Geophysics

Division (EEGD), Geotechnical Laboratory (GL), WES, was the Contracting

Officer's Representative and WES Principal Investigator for CWIS Work Unit

32221. General supervision was provided by Dr. A. G. Franklin, Chief, EFGD,

and Dr. W. F. Marcuson lII, Chief, GL. IV

01. 1). ':110 *. le, CV, wA.:s C' UIellt 'And fI)rect"I I WlA Iul i the j 1 1 it 1''11 -k

~ftils Yopj It. D1. Ph elt 1q. I

W111; ; Vlechnic~ I!
,1:w [110 t I.
%
•.. ]
 CONTENTS
%

Page

PREFACE . . .. .. .. .. .. . .. . . . . .. .. .. .. .. .* . . 1
1

LIST OF FIGURES .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. 4

LI1ST OF TABLES......... . . . . . . .. .. .. .. .. .. .. .. 7

LI1ST OF PLATES . . . . . . . . . . . . . . . . . .. .. .. .. .. .... 8

PART I: INTRODUC~TION. ........... .. .. .. .. .. ...... 9 j

Background . . . . . . . * . . .. .. .. .. .. .. .. .. .. .. 9p
Purpose . . .. .. .. .. .. ... . . . . .. .. .. .. .. .. 9
General Description of Statistical Quality Control . . . . . . . . . 10
organization of This Report . . . . . . . .. .. .. .. .. .. .... 11

PART II: FUNDAMENTALS ............ .. .. .. ..... . . . . 12el

Probability Theory . . . . . . . . . . . . . . . .. .. .. .. .. 12
Frequency . . . . . . . . . . . .. .. .. .. .. .. .. .. .. 12
Subjective Probability .. .. .. .. .. .. .. .. .. .. .. 12
Randomness .. .. .. .. .. .. .. .. .. .. .. .. .. .. 13
Conditional Probability and Independence...............14
Multiplication Theorem .. .. .. .. .. .. .. .. .. .. .. 16
Addition Theorem . . . . . .. .. .. .. .. .. .. .. .. . 16
Frequency Distributions.........................17
Discrete and Continuous Variables. . . . . . . . . . .17 'u

Histograms and Frequency Distributions . . . . .. .. .. .. .. 17%

Cumulative Distribution ............ .. .. .. .. .... 18
Importance of Frequency Distributions ........... . . . . 19
Summary Statistics .. .. .. .. .. .. . .. . .. .. .. .. .. 20
Central Tendency...........................20
Dispersion..............................22
Association Among Uncertain Variables. .............. 24
Quick Estimates . . . .. .. .. . .. .. .. .. .. .. .. .... 25
Shortcuts for Estimating the Mean . .. . . . . . . .. .. .. .. 25
Shortcuts for Estimating the Standard Deviation ............ 26
Shortcuts for Estimating the Correlation Coefficient .. ........ 27 .

Probability Distribution. ......... . . . .. .. .. .. .. .. 28

PART III: CONCEPTS OF STATISTICAL QUALITY CONTROL. ....... . . . . . . 49

Quality Assurance and Quality Control............. . . . 49

Sampling................ . . . . . .. .. .. .. .... 49
Scientific Sampling..........................50

2
 WWWK.i~ ww~ ~'

,.,.,
Page '.-

Random Sampling .... . . . . . . . . . . . . . . . . . . . 52

Sampling Plans .. .. .. . . .. . .. . . . . . . .. . . 52
Sampling Variation........... . . . . . ........ 54
Sampling Variability of the Mean ............ ........ . 55
Sampling Variability of the Standard Deviation .. . . .
. . . 57
Sampling Variability of the Range ................ . . . . 58
In-Control vs. Out-of-Control Processes .... . . . . . . . 58

PART IV: INSPECTION OF ENGINEERED FILLS . . . . .............. 67

Objectives of Sampling and Testing ................. . 68

Target Properties vs. Sampled Properties . . . . . . . . . . 69
Tests for Water Content and Dry Density .... .............. 69
Compaction Specifications ....... .............. . . . . . 70

PART V: QUALITY CONTROL CHARTS ...... ............... . . . . . . 71

Theory of Control Charts.............. ........ 71

Behavior of Chance Variation ...................... 72
Control Charts and Control Limits . . .............. 74
Control Chart for the Sample Range r x ...... ............... . 76
Cumulative Reject and Related Charts ..... ................ 77
Cumulative Reject Chart ......................... ........ 78
Moving Average, Standard Deviation and Range Charts ... ...... 80
Shape of the Frequency Distribution of Test Data . ........ 81
Cumulative Sum (CUISUM) Chart . . . . ................. 83

PART VI: QUALITY ASSURANCE BY ACCEPTANCE SAMPLING ... ........... 100

Structure of An Acceptance Sampling Plan ..... .............. 100

Buyer's Risk and Seller's Risk ...... ................. 101
Inspecting for Fraction Defective vs. Inspecting for the Mean 102
Operating Characeristic Curves ....... ................. 103
Acceptance Sampling to Give Assurance on the Mean . ......... 104
Single Limit, Standard Deviation Known . ... ............. 104 %
Are Compaction Data Normally Distributed ... ............ 108
Single Limit, Standard Deviation Unknown ... ............ 109
Double Specification Limits, Standard Deviation Known ... ...... 111
Double Specification Limits, Standard Deviation Unknown ..... . 116
Acceptance Sampling for Fraction Defective .... ............. 118
Operating Characteristic for Fraction Defective Sampling ....... 120".
Single Limit with Standard Deviation Known ... ........... 122
Single Limit with Standard Deviation Unknown .. .......... 12 7 q
Double Specification Limits ....... ................... 129
Double Limits, Standard Deviation Known ...... ............. 130 U1
Double Limits, Standar-J Deviation Unknown ... ............ 133

REFERENCES o........... .. ............................... 168

I 1a.

COP 01'
;%
. 1 . .... . '% -'' .' '-1 - : %-%-"- -"' v % . " ; - '- -' '. -
% 1, / / ' x€ € : ,.,. . € • . 'v.,:2
%. a% V 'a . V% P~% ~...... V_
.
',
, _a.. .. .. ).
 LIST OF FIGURES

Page

Figure 1 -- Venn diagram showing relations among probabilities of simple 37 %

events.

Figure 2 -- Histogram of SPT data. 38

Figure 3 -- Histogram of R-test data. 39

Figure 4 -- Frequency distributions for the data of Figure 2. 40

Figure 5 -- Frequency distributions for the data of Figure 3. 41

Figure 6 -- Cumulative distribution of the SPT data of Table 1. 42 -

Figure 7 -- Cumulative distribution of the R-test data of Table 2. 43

Figure 8 -- Symmetric and skewed frequency distributions. 44 A

Figure 9 -- Summary statistics of a frequency distribution. 45

Figure 10 -- Example application of balloon method for estimating the 46

correlation coefficient of experimental data. 4
Figure 11 -- Normal or bell-shaped frequency distribution. 47

Figure 12 -- Typical control chart for the output of a process operating 48

in a random manner.

Figure 13 -- Populations of interest in sampling. 64

Figure 14 -- Typical spatial sampling patterns. 65

Figure 15 -- Histogram of the means of n=5 SPT blow counts randomly 66.,'* -
sampled from a large data set.

Figure 16 -- Compaction control data for a construction process operating 87

in-control.

Figure 17 -- Shewart control chart (m-chart) for compaction control on a 88

dam project.

Figure 18 -- Shewart control chart (R-chart) for compaction control on a 89

dam project.

Figure 19 -- Cumulative reject chart for compaction water content and dry 90
density inspection data from the compacted clay core of a
rock fill dam.

2. N2,
 Page

Figure 20a-- Change in construction operation as reflected in cumulative 91 1.

reject chart (schematic) .

Figure 20b-- Change in construction operation as reflected in actual 92

cumulative reject chart.

Figure 21 -- Moving average water content chart. 93

Figure 22 -- Moving standard deviation chart of water content. 94

Figure 23 -- Cumulative reject (out-of-specification) chart for compaction 95

water content.

Figure 24 -- Distribution of compaction water content data for tests 1 96

through 459.

Figure 25 -- Distribution ot compaction water content data for tests 460 97 .

through 884.

Figure 26 -- Distribution of compaction water content data for tests 885 98

through 1175.

Figure 27 -- Cumulative sum (cusum) chart of the compaction water content 99

data of Fig. 23. ,

Figure 28 -- Frequency distribution of test results taken from an 151

acceptable fill (schematic).

Figure 29 -- Operating characteristic (OC) curve for an acceptance 152

sampling plan to assure the value of mean compacted dry
density of an engineered fill (see Plate 3).

Figure 30 -- Examples of empirical compaction control data showing 153

closeness to Normal distributions of frequency. %

Figure 31 -- Student's-t distribution. 154

Figure 32 -- Operating characteristic curves corresponding to various 155 -

sample sizes.

Figure 33 -- Frequency distribution of the sample average m. for a sample 156 .'
of size n.

Figure 34 -- Sampling variahility ot two litts of material, one with true 157 -
mean equal to U%1,1, the o)tht-r with true mpan equ~al to 17)l 1 .

Figure 35 -- Operatin'g :hira'ter it- ,:urvf, for in .t',:dptan-, ianFilnq 158

plan to ,1; t h!. v
,t ltui. of the mc',al.

"'I

r-]I
.+,,.,;,.,., ,,
..
...,..,.
+.....z,:+:.; .,
.,.,, ...,........_, .,, ,. _.. J
I,
%,,,,
 Page

Figure 36 -- Fraction defective for a Normal frequency distribution of 159

soil properties.

Figure 37 --Operating characteristic for fraction defective sampling, 160

standard deviation known.

Figure 38 -- Operating characteristic for acceptance sampling plan derived 161

in Plate 7.

Figure 39 -- Approximate and exact OC curves for single limit sampling 162
plan with standard deviation unknown (Plate 8).

Figure 40 -- Operating cahracteristic curve for sampling plan derived in 163

Plate 9.

Figure 41 -- Chart for determining fraction defective from z, standard 164

deviation unknown.

Figure 42 -- Chart for determining maximum allowable fractive defective M. 165

Figure 43 -- Operating characteristic curve for sampling plan derived in 166

Plate 10.

%6

• %

6
 LIST OF TABLES

Page

Table 1 -- Summary statistics. 31

Table 2 -- Ratio of average range to standard deviation for samples 32

from a Normal frequency distribution.

Table 3 -- Common Probability distributions. 33 Iv

Table 4 -- Cumulative frequencies of the Normal distribution. 34 I

Table 5 -- Random numbers. 61

Table 6 -- Percentage points of the Student-t distribution. 62

Table 7 -- Percentage points of the distribution of relative ronge 63

wx=rx/sx for small samples from Normal distributions.

Table 8 -- Fraction defective for double sampling limits. 135

.-

,.

%,%
 LIST OF PLATES
p"

Page -"

PLATE 1 -- Example calculations of Summary statistics for (R-test) data. 35

Shortcut estimates of summary statistics.

PLATE 2 -- Shortcut estimates of summary statistics. 36 %

PLATE 3 -- Acceptance sampling plan to assure mean value of compacted dry 136
density; standard deviation known.

PLATE 4 -- Acceptance sampling plan to assure mean value of compacted dry 137
density; standard deviation unknown.

PLATE-- Design acceptance sampling plan to assure average compaction 139

water content within double specification limits; standard
deviation is known. .

PLATE 6 Operating characteristic (OC) curve for fraction defective 141

sampling, typical calculation; standard deviation known.

PLATE 7 -- Acceptance sampling plan for fraction defective; standard 142

deviation known.

PLATE 8 -- Acceptance sampling plan for fraction defective; standard 145

deviation unknown.

PLATE 9 -- Acceptance sampling plan to assure fraction defective with 147

double specification limits; standard deviation known.

PLATE 10 -- Acceptance sampling plan to assure fraction defective with 149

double specification limits; standard deviation unknown

or

V
.S

0:

S:
 C.4

Dp.

STATISTICAL QUALITY CONTROL OF ENGINEERED EMBANKMENTS P

PART I: INTRODUCTION P

Ba ckg round .,
.p

Concern with quality and the control of manufacturing or construction

processes to assure quality underlie modern engineering and production.

Indeed, quality ontrol is as old as engineering itself. On the other hand,

statistical quality control is relatively recent. In the United States,

statistical quality control first came into its own with the wartime production

effort of 1939-1945, the main impetus of this push having been Army Ordinance

and the War Production Board. The military influence has been important to the

introduction of statistical quality control to American industry.

Quality control in construction has characteristics which are both similar

to and different from quality control in manufacturing. The control of quality -

C..

in dam projects, especially concerning the placement and compaction of

engineered embankments, is critical to the safety and performance of the entire

project. Consequently, a well planned inspection program is considered

essential on any moderately larje project. Current CE guidance for quality

control of engineered embankments is contained in EM 1110-2-1911, dated 17

January 1977. It is not statistically based, but is experience based.

Purpose-

The purpose of this report is to provide potential users of statistical

quality control of engineered fills with an introduction to practical concepts,

definitins, and techniquos;. '1he relxrt presents simple techniques which are

interiel f .r use by readersi havinq limited ftmiliarity with statistical theory.

-. . -. *. ., 'aA

*......
The report does not attempt to survey the literature of statistical quality

control, but concentrates on a few chosen techniques that fill the needs of

geotechnical engineering practice.

General Description of Statistical Quality Control

The placement of compacted fills, like any manufacturing or construction I

process, varies with time. The physical properties of soils being placed

varies in moisture content, gradation, plasticity and other ways; and the

process of placing soils varies, for example, in lift thickness, compactive

effort, and climatic conditions. These variations cause physical properties of

a resulting fill to differ from one point to another, A field inspection

program intends to ensure that--to an acceptible level of confidence--the

completed fill conforms to specified standards and thus will perform its

function acceptably.

Ideally, an inspection program could non-destructively screen all soil

placed in a fill and reject those materials with engineering properties not

conforming to specified standards of strength, deformability or permeability.

Such a program would guarantee perfection by detecting all parts of a fill

which were flawed. Unfortunately, cost and the lack of reliable testing

technology precludes this ideal program. Instead, typical inspection programs

consist of limited numbers of small-scale tests spread thinly throughout a

fill. The properties measured by most of these tests--for example, moisture

content and dry density--are merely surrogates for the engineering properties

of actual interest, although some engineering properties are also measured

10
 Statistical quality control uses simple probalility theory to develop
I.

inspection sampling plans. These plans make efficient use of resources, and

can be related to a quantitative confidence in the quality of a finished

product.
Organization of This Report

This report is organized in six parts. After the introduction, Part II

summarizes fundamentals of probability and statistics which are necessary for

later presentations. Part III presents basic concepts of statistical quality

control including sampling theory. Part IV briefly reviews field control of

4.-

compaction operations. Part V presents quality control chart techniques.

Finally, Part VI discusses the design (i.e., planning) of sampling schemes for
%-

field use. Following each chapter are tables and figures, and plates
N

presenting example calculations. .4

.4.

.
,.

.4

.4

.4"

.'4 ", "'" ' ,"2' '' """':"""":. - ""'."" 'i- --. - "•. "." """ -"""""" - - -"""". '-:' --
 PART II: FUNDAMENTALS

This section briefly reviews mathematical concepts underlying statistical

quality control.

Probability Theory

Probability theory is a branch of pure mathematics. It is logical and

internally consistent in the sense that all the mathematics of probability

theory can be derived from a small set of axioms. In essence, the axioms

specify properties that "probability" must have, for example probability is a

real number between zero and one. Yet, nowehere do the axioms say what the

concept of probability means. As a result many interpretations of what

probability means are in common use.

Frequency

In statistical quality control, probability is usually interpreted to be

the frequency of occurrence of some event in a long series of similar trials.

A trial is an individual occurence producing an outcome of some sort. For

example, each individual lift of soil placed in a compacted embankment might be

considered a trial. The frequency of soils, having low moisture content among

these lifts (i.e., among the trials) would be the probability of soil with low

moisture content.

Subjective Probability

An alternative interpretation, common in geotechnical engineering, holds %

thdt probability is a rational degree of belief. The probability that an

12

'%N.e
important solution cavity exists in a limestone dam abutment is typical of

geotechnical problems which cannot be easily approached using the frequency

definition of probability. Such probabilities have to do with

one-time events, past experience, and amounts of information. They are

personal and subjective. They are not related to frequencies, actual or

conceptual.

In this report the frequency definition of probability is used, for it is

appropriate to quality control problems.

Randomness

A key concept of the frequency approach to probability is randomness.

There are two places where the concept of randomness is important. one is the

description of a construction process as operating in a random manner; the

other is the design of a random sampling plan.

A process is operating in a random manner when any part of the output may

be viewed as typical of the output as a whole. That is, when perturbations

show no discernable pattern. Usually it is not possible to demonstrate that a

process is operating randomly; rather, it is only possible to do the reverse,

to demonstrate that a process is not random. This is done by showing that the

output of the process does in fact have a pattern to it.

A process that is operating in a random manner has elements or events with

definite probabilites of occurrence. These probabilities may not be known, or

may be known only to the extent that data are available from which to draw

estimates. Because the elements or events have associated probabilities,

statistical theory and methods can be used to characterize them.

13

5%
 A-

The other place randomness is important is in the design of random

sampling plans. A random sampling plan is one in which sampled elements are

chosen with definite probabilities, but without a predictable pattern. In

:-a
large samples the relative frequency with which elements are sampled should

approach those probabilities; however, the collection of elements which make up

any specific sample reflect a chance distribution.

Conditional Probability and Independence

In quality control, probability is commonly defined as the relative

frequency with which a certain event occurs in a long series of similar trials.

For example, if there are N elements in a large set, of which na share a common

property A, then the probability of an element within the set having property A

is,
n
P(A) - , --
N

If some of the elements also share a common property B, and if the number of

these elements is nb, then the probability of property B within the set is,

nb
P(B) = -2-
N

Consider now that some elements in the large set possess both property A

and property B. Let the number of such elements be nab. Graphically, the

number of elements possessing property A, property B, or both can be depicted

14 b N

,
-.$ -
",*"N*."A. ,,*',, '' ' -,J 2" .'".""' -
%% ," "r' " '" "'I../ - ."z'-:. -r'= "" -""A --.
i " - " . ;/ , "
 p.,
"

as in Fig. 1. If we consider only those nb elements having property B, the

fraction of these also having property A is nab/nb. This relative frequency is

called the conditional probability of an element having property A, given that

it is known to have property B. The conditional probability is denoted,

nab
P(AJB) = -3-
nb

By analogy, the conditional probability of property B given property A would

be,

nnab
P(BIA) n
n -4-
a
The event that an element in the population possesses property A is said

to be independent of the event that the element possesses property B when the

probability of A is unchanged by knowing that an element possesses property B.

Mathematically, A is independent of B if

P(AIB) = P(A) -5-

Knowing that the element possesses property B in no way influences the

probability that it also possess property A. If the event that the element

possesses property A is also independent of the event that it possesses

property B, then properties A and B are said to be mutually independent.

15

[ 4%
 Multiplication Theorem

Two theorems of probability theory are basic and often encountered in

statistical quality control. These have to do with the relationships among the

probabilities of distinct events. The first is the multiplication theorem.

The multiplication theorem states that the probability of two mutually

independent events occuring simultaneously is the product of their individual

probabilities. If elements possessing properties A and B within some large set

are mutually exclusive events, then the probability of an element possessing

both property A and property B is, -

P(A and B) = P(A) P(B) . -6-

If A and B are not mutually independent, the more general form of the multipli-

cation theory states that the probability of them occuring simultaneously

depends on the conditional probabilities.

4.,

P(A and B) = P(A) P(BIA)

-7-
= P(B) P(AIB)

Addition Theorem

The addition theorem states that the probability of either one or the

other of two events A and B occuring equals the sum of the individual

probabilities of their occurrence, minus the probability that they occur

together.

P(A or B) P(A) + P(B) - P(A and B) -8-

1.

16 .,
 In Fig. 1, if area is taken to represent probability, the addition theorem can

be considered simply a statement of geometry. The area contained by the

combination of events A and B equals the sum of their individual areas, less

one times the area of their overlap (i.e., P(A and B) ) which would otherwise

be double counted.

Frequency Distributions

The variability of data on production output, soil properties, or other

variables is conN -iiently summarized in a frequency distribution, the

fundamental tool used by statisticians.

Discrete and Continuous Variables

Fig. 2 shows the variability of standard penetration test blow counts

measured in 40 borings in a silty sand deposit at a dam site. Blow counts can

only assume interger values, and therefore are said to be discrete variables.

Fig. 3 shows variability of water content measured in R-tests on 73 specimens

of a compacted clay. These strength data may assume any real number value

within a broad range, and are therefore said to be continuous variables.

Quality control in geotechnical engineering must deal with both discrete and

continuous variables, and many methods of statistical quality control apply to

each in a similar way.

Histograms and Frequency Distributions

A convenient way to graphically represent scattered data is in a

histogram. A histogram graphs the number of measurements falling within

specific intervals of value as a vertical bar. Fig. 2 shows a histogram of SPT

17

....

Aw*,
 5,'

data. For obvious reasons, a histogram is sometimes called a bar chart.

height of the bar above each interval shows the number of measured values
The
I.
within the interval, and the sum of the heights of the bars equals the total

number of measurements. Fig. 3 shows a histogram of R-test data.

The histogram of Fig. 3 divides the data into intervals of 1%. The choice

of intervals is arbitrary, but the intervals should be of uniform width and

have convenient end points. If too many intervals are chosen the general

picture of relative frequencies will not be obtained, while conversely, if too

few intervals are chosen the general picture will be blurred. A common

rule-of-thumb is to use about 10 intervals. More detailed

discussion is presented in the report "Data analysis for geotechnical

engineering" (January 1986).

A frequency distribution is constructed from a histogram by dividing each

vertical bar by the total number of measurements. This gives the relative

frequency of observed value in each interval as a decimal fraction. The sum of

the heights of the bars in a frequency distribution is 1.0. Fig. 4 shows the

frequency distribution (right had side scale) corresponding to the histogram of

Fig. 2. Fig. 5 shows the frequency distribution corresponding to the histogram

of Fig. 3.

Cummulative Distribution

A cummulative distribution of discrete or continuous data is constructed

by summing relative frequencies starting at the lower-value end of the data

and proceeding toward the upper value end. The cummulative distribution
N.

•'p .

18'

@1l"I
 denoted F(x) gives the fraction of measurtmnerit- l tan ;r equal to a

particular value,

F(x) = fraction of measurements 4 x. -9-

Cummulative frequencies for the data of Figs. 2 and 3 are shown in Figs. 6

and 7. The cummulative distribution has the properties that,

For x = lower limit (or -co) + F(x) = 0

-10-
For x = upper limit (or +oo) + F(x) = 1.0
h

For discrete data the cummulative distribution is a step function increasing to

the right. For continuous data the cummulative distribution is typically a

smooth S-shaped curve.

Importance of Frequency Distributions 1

Frequency distributions give a summary view of the variation in a set of

data. The shape of the distribution suggests whether the data have any central •

tendency, and if so, where along the x-axis the data are concentrated. The

width of the distribution indicates the dispersion or scale of variation of the

data.

Some frequency distributions have one point of concentration and are

. thus called unimodal. Others have more than one and are called multimodal.

Usually, soils data are unimodal. Multimodal distributions may indicate a

mixture of data from different soil types or different construction procedures,

that is, nonhomogeneous data.

.4-

4.

19

" "" ,"

g". "4 ,, ",, " " ",, . -
-.".' .,"., . . ."W". ."-" .,,,"
. . , 4 ,' - . .-, 4 4 ,",,
 The frequency distribution also shows whether the variation in data is

symmetric or asymmetric, that is, whether high and low variations are evenly

balanced. For data that are asymmetrically distributed, large variation from

the central tendency of the data set are more frequent on one side of the

center than on the other. This is illustrated in Fig. 8.

Summary Statistics

Frequency distributions are convenient representations of data for visual

inspection, but often numerical measures of distribution characteristics are

useful for calculation or for setting standards. Numerical measures are

essential for developing quality control criteria and quality control charts.

The most important numerical measures pertain to the central tendency of data

and to dispersion.

The term "statistic" refers to any mathematical function of a set of

measured data. For example, given the measurements xl,..., Xu, any function y

= T(xl,...,xu) is said to be a statistic of the data. The arithmetical average

is such a function, the largest value Xmax or the smallest value Xmin is such a

function, and so on. Any of these ways of summarizing the data would be called

a statistic. Obviously, there are an infinite number of statistics which could

be calculated from any set of data, but the most useful have either to do with

the central tendency of the data along the x-axis or to the dispersion of the

data.

Central Tendency

The most common measures of central tendency are the mean, median, and

mode. The mean is the arithmetic average of a set of data. The median is the

20

%"
 -r7f-r WV -7-"' RV VV Vt '

"!

Apo

.or

I,,

value for which half the observations are smaller and half larger. The mode is

the most frequent value (Table 1). '

The mean of a set of n data x = {x1 , ... , xn}, denoted mx, is defined as

the arithmetical average,

1 n
m =- x. -11-
x n 1 1

The mean is the center of gravity of the frequency distribution along the

x-axis, as shown in Fig. 9.

The median of the set of data x = {xj, ... , Xn}, denoted x0 .5 , is the "

value of xn which half the data are less than and half more than. The

* cummulative distribution evaluated at the median is 0.5,

F(xo. 5 ) = 0.5 -12-

The median is the midpoint of the data, when listed in increasing or decreasing

%
-. order. Common practice in the case of an even number of data is to define the

median as half way between the two middle data, that is, those of rank (n/ 2 ) P.

. and (n/ 2 + 1).

The mode of the set of data x = [xl, ... , xn}, denoted xo, is the most -,

often observed value. This is the value of x having the highest ordinate on

the frequency distribution.

421

I..

21O
W
I.. V

. . " " " . . . . " . '''-4

"0" * - .' ° . . . . .
" " - " ."" "
#" "". " -" "
%-.
,| ".".%,' .,, . '-,'',-.,
. - .. "% .', %"'% , ....--. , -,%-.,--%%-%'-,% '%% '%S
 N
.:

4.I

Dispersion

The most common measures of dispersion are the standard deviation, ranqe,

and inner quartiles.

The Standard deviation of a set of data x = {x 1 , ... , xn}, denoted sx, is

defined as the root mean square variability of the data,

/
1 n2
(x.- m 2 -13-
x V n-1 ( 1 x)
i=1.

in which m x the mean of the data. The denominator (n-1) rather than (n) is

used to correct d statistical bias. In estimatin(; the standard deviation from

data, the mean is usually also unknown. Thus, the mean must be estimated from

the same data as the standard deviation. This causes the averaqe squared

variability about m. to be smaller than it should be. On averaqe, it is smaller

by a factor (n-1)/n. Correcting for this error gives Bqn. 13.

The coefficient of variation of a set of data is the standard deviation

divided by the mean,

, x/mx
xx -1 4-

The coefticent of variation is used to express relative dispersion.

The variance of a set of data, denoted V x, is the square of the sta-ndard

devia tion,

0,..

22

%-if---
 U.

a' n

V = S 2 = (x.- m )2 -15-
'a'.
x x n-1 i=1 1 x

In many statistical calculations the variance is a more convenient term

than the standard deviation, and is thus widely encountered in statistical

quality control. The variance is the moment of inertia of the frequency

distribution about mx.

The range of a set of data, denoted r, is the difference between the

largest and smallest values,

r x = Xmax - xmin -16-

The range has poor statistical properties in that it is sensitive to extreme

values in a data set, however, it is easily evaluated and therefore often

useful.

The inner quartiles of a set of data, denoted x 0 . 25 and x 0 . 75 , are the

data values for which one-quarter of the data are smaller and one-quarter

larger, respectively. The quartiles may be found from the cummulative

distribution as

F(xo. 25 ) = 0.25 -17a-

F(x1.7 5 ) = 0.75. -17b-

The intorquartile ranqe, denoted rO.5,

tO. 5 (x0.7 5 - x0.25) -18-

23

%, %
is less influenced by extreme values than is the range itself, but it is

correspondingly more troublesome to compute. Various summary statistics

applied to the R-test data of Fig. 3 are evaluated in Plate 1.

5'..
Association Among Uncertain Variables *

When dealing with two or more soil properties the uncertainties in -

estimates may be associated with one another. That is, the uncertainty in one

property estimate may not be independent of the uncertainty in the other

estimate. Consider the problem of estimating 'cohesion' and 'friction'

parameters of a Mohr-Coulomb strength envelope. If the slope of the envelope

to the Mohr circles is mistakenly estimated too steeply, then for the line to

fit the data the intercept will be too low. The reverse is true if the slope k
is estimated too flat. Thus, uncertainties about the slope and about the .

intercept are not independent, they are associated with one another.
-. 5'
The correlation coefficient for paired data x,y = {(x 1 ,yl ),...,(xn,yn)} is

denoted pxy, and defined as,

xi-m y my

Pxy En ( ) m -19-

In effect, the correlation coefficient is equivalent to a normalized product

moment of inertia in solid mechanics. It expresses the degree to which two

parameters vary together. The costrelation coefficient is non-dimensional

because deviations of x and y are measured in the same units as their .

%",

24

%
.5~5% .. . . .. . % - 5' % 5 .
 rb

respective means. The value of Pxy may vary from +1 to -1. Px,y=+l implies a
1
strict linear relation with a positive slope; Px,y=- implies a strict linear r%

relation with a negative slope; px,y=O implies no association at all (i.e.,

independence) .

Quick Estimates

Often one wants quick, approximate estimates of means, standard -'

deviations, or correlation coefficients from limited numbers of data. Some

shortcut techniques are available for this purpose. These provide economies of

time and effort while causing sometimes only minor losses of accuracy or

precision. -V

L-

Shortcuts for Estimating the Mean

Rather than using Equation 11, a quick and often good estimate of the mean

can be obtained from the median. The median is the middle value of a data set.

For example, if, say, five data are listed in ascending order xj, x2 , x3 , x4 ,

x 5 , the median is x3 . For an even number of data, say n=6, the difference

between the two middle data is halved to give the median, that is (x3 +x 4 )/2.

For data scatter which is symmetric about its central value and for small

numbers of data, the sample median is a good estimate of the true mean. On the

other hand, if the data scatter is asymmetric--for example, if there are many

small values and a few large values--the sample median i., not such a g7ood

estimator of the mean.

A second shortcut for estimating the mean is taking one-half the sum of

the largest and smallest measured values, (1/2)(xmax + Xmin). This estimator

25
 p-

is sensitive to the extreme values in a set of measurements, and thus

fluctuates considerably. It is not a good shortcut estimator and should only

be used with caution.

Shortcuts for Estimating the Standard Deviation

Rather than using Equation 13, a quicker estimate of the standard

deviation from small numbers of tests can be made from the sample range

rx=(xmax-xmin). The range is the span of data from largest to smallest. Like

the standard deviation, the range is a measure of dispersion in a set of data.

However, the relationship between the standard deviation and the sample range,

on average, depends on how many tests are made. Th obtain a best estimate of

sx from the range of data r. a multiplier Nn is used which depends on sample

size (Table 2). The best estimate of the standard deviation is sx Nnrx (see .

Plate 2).

As for the sample median, the range is a good estimator of the standard

deviation for small n and symmetric data scatter. Even for modest n it remains

fairly good. However, for asymmetric data scatter the range, which is strongly

affected by outliers, is not a good estimator of the standard deviation.

Fortunately, with the notable exception of hydraulic parameters such as

permeability, most geotechnical data display symmetric scatter. In the case of

%'
hydraulic data a logarithmic transformation (Lee, et al., 1983) usually makes

the data scatter symmetric, and again the median and range become convenient

estimators.
S." %

26

~ -" " .. *...\..*'-

"1*"-. .~* .": ., . . , -' . ,."
 V% -I :W %W
-I TI WXW1O

Shortcuts for Estimating the Correlation Coefficient

0" Calculation of correlation coefficients by Eqn. 19 can be time consuming

and tedious. A simple and quick approximation is obtained graphically from the

shape of the scatter plot of y vs. x. The method works well whenever the

outline of the scatter plot is approximately ellipical, and works even with

small numbers of observations. Using Chatillon's (1984) term and prodedure,

this is called the balloon method:

STEP 1: Plot a scatter diagram of y vs. x.

STEP 2: Draw an ellipse (balloon) surrounding all or most of
the points on the plot.
STEP 3: Measure the vertical height of the ellipse at its
center, h, and the vertical height of the ellipse at its
extremes, H.
STEP 4: Approximate the correlation coefficient as r = V1 -(h/H)

An example of the method is shown in Fig. 10. The balloon method gives a

correlation coefficient of 0.81, whereas the correlation coefficient

calculated by Eqn. 19 is 0.83. Empirically, the method works well for r>0.5.

Shilling (1984) suggests a similar balloon method for approximately

estimating the correlation coefficient:

STEP 1: Plot a scatter diagram of (y-my)/Sx vs. (X-mx)/sy.

STEP 2: Draw an ellipse surrounding all or most of the points
on the plot.
STEP 3: Measure the length of the principal axis of the ellipse Y
having positive slope, D, and the length of the
principal axis of the ellipse having negative slope, d.
STEP 4: Approximate the correlation coefficient as r (D2 -d2 )/(D 2 +d').

This methods works about as well as Chatillon's. For the data of Fig. 10

Shilling's method gives r 0.80.

27

a-
 4"%,

Probability Distribution

For many problems in statistical quality control it is convenient to

approximate the empirical frequency distribution for some category of data by a

mathematical function. Surprisingly, a comparatively small set of mathematical

functions can be used to fit a broad range of frequency distributions

encountered in the field. By far the most important of these is the Normal or

bell-shaped distribution. Among other useful distributions are the log Normal,

Exponential, and 4-parameter Beta distribution, although many others exist.

The Normal distribution is discussed here, while parallel properties of the

other forms are given in Table 3. a-,

The Normal distribution is represented by the equation

x-m
1 x2
f(x) 1 e x -20-
V27 s

in which m x = the mean of x and s x = the standard deviation. The distribution

is unimodal at m x and symmetric (Fig. 11). The cummulative distribution of x

using the Normal equation is found from the area under the frequency

distribution up to x,

F(x) = (x f(x) dx. -21-

The normal distribution is defined for -- <x<+-, but the area under the

28
.4%

2 ~.4".
i -II I" iil~
1F -i . . . - ' I ""l ' -" " 1 '- " 1 -- -
 distribution beyond 3 to 4 standard deviations from the mean is neglible.

The area under the Normal distribution expressed as a function of the

standardized variable

x-m -

Z =-22- x
S
x %

and calculated by Eqn. 21 are given in Table 4. Benjamin and Cornell (1970) W

give examples. Numerically, these areas can be approximated by the series '

expansion (Abramowitz and Segun, 1964), .4

F(z) = fN(x) (blt + b 2 t 2 + b 3 t 3 + b 4 t+ bst5) + e -23-

in which,

bI = 0.319381530 t = (1+px) - l
b2 = -0.356563782 p = 0.2316419
b3 = 1.781477937 lel < 7.5 x 10-8 -24-
b4 = -1.821255978
b5 = 1.330274429

.4

and fN(x) is the ordinate of the Normal distribution function evaluated at x.

." The series expansion is generally more convenient than Table 4 for use with

computers.

If a construction process is operating in a random manner, and if good

estimates of the mean and standard deviation are available, and if the fre-

quency of data are observed to be well modelled by a representable distribu-

tion, then forecasts can be confidently made about the future performance of

29

%
that process. This is the basis for statistical quality control. For example,

if the process is observed to produce Normally distributed output, then a chart

such as Fig. 12 can be constructed which shows the process mean and envelopes

+3s x about the mean. As long as the process continues to operate in a random

manner, and the mean, standard deviation, and frequency distributions remain

unchanged, then a confident forecast can be made that 99.7% of the output

measurements to be made in the future will lie within the +3s x bouni (Figure

12). This forecast of 99.7% comes from Table 4. Such forecasts are considered

in greater detail in Parts V and VI.

[..'
 Table I

Summary Measures of Frequency Distributions

Measure Symbol Formula Comments

Central Tendency

Mean mx 1/n I x i center of gravity

Median F(x 0 5 )=0.5 middle value

Mode xo xo=max f(xi) most frequent value

",

,Dispersion

S 1)2
Standard Deviation sx 1
/ (xi-m root mean square
V xx variation

2
Variance Vx sx moment of inertia
about mx

Range rx Xmax-Xmin

Interquartile r 0 .5 x0.75-x0.25
Range

31

"'"., PA-"
 P.

Table 2
4 -I."

Ratio of average range to standard deviation for samples from a

Normal frequency distribution.
:..

n Multiplier N n n Multiplier Nn V.

2 0.886 12 0.815
3 0.591 13 0.300
4 0.486 14 0.294
5 0.430 15 0.288
6 0.395 16 0.283
7 0.370 17 0.279
8 0.351 18 0.275
9 0.337 19 0.271
10 0.325 20 0.268
11 0.315

from Snedecor and Cochran (1980)

.4%

32

r' P

S& 1 V
IL
%,.I.
P'i
 Table 3

Common Probability Distributions

(after Lee, et al (1983)

Type Formula Shape Comments

Uniform I Mean - '/a + b); Variance = '/,2(b - a)r - .

fx) = (b - a) Used when no reason to give other than
for a s; x :s b equal likelihoods to possible values of x.
Normal Mean = A; Variance =
N x) -2 7r ] Most common distribution. Used unless
a).'2i another distribution is more applicable.
for -r. x :r
Lognormal Inx- The random variable y = In x is normally
exp 2 a, J - distributed.

where y - In xo|i
forOs x f. w
Exponential fQx) = X exp(-k x] Mean = I/A; Variance =I/X'%
for 0 s x n; Used for particular physical situations when
positive values required, e.g., lengths of
joints in a rock mass. Also used to describe
the time between incidents of events which
can be described by a Poisson distribution
(such as earthquakes and floods). See
Benjamin and Cornell (1970).
Beta Bet
W (x -a)*-I(b - x)O - - ' Mean = a +2(b-a)
f~)= B(b - aW-1

where B (beta function) a(B - ,)

nap- a) '(/3 + I)
( J) - Extremely versatile distribution for matching
fuctdata
Fgam over the range [a.b]. Variation of
r gparameters a and 0 gives wide variety of
for a s x s b shapes. Contains as special cases the uniform
and normal distributions. Can be symmetrical
or skewed right or left. See Benjamin and
Cornell (1970). and Harr (1977).

'.

% %
Table 4 -- Cuilnil, l\.
t reduil1i- ')t
o th, ll-z)a 1 i-; tribution
tf l itLjri j dUi i in a rd Co rtie I1I, 19 7i t)

Cumulative Probabilities of the Normal Probability Distribution* (areas under the

normal curve from - to z)

z .00 .01 .02 .03 .04 05 .06 .07 .08 .09

.0 .5000 .5040 .5080 .5120 .5160 .5199 .5239 .5279 .5319 .5359
.1 .5398 .5438 .5478 .5517 .5557 .5596 .5636 .5675 .5714 .5753
.2 .5793 .5832 .5871 .5910 .5948 .5987 .6026 .6064 .6103 .6141
.3 .6179 .6217 .6255 .6293 .6331 .6368 .6406 .6443 .6480 .6517
.4 .6554 .6591 .6628 .6664 .6700 .6736 .6772 .6808 .6844 .6879

.5 .6915 .6950 .6985 .7019 .7054 .7088 .7123 .7157 .7190 .7224
.6 .7257 .7291 .7324 .7357 .7389 .7422 .7454 .7486 .7517 .7549
.7 .7580 .7611 .7642 .7673 .7704 .7734 .7764 .7794 .7823 .7852
.8 .7881 .7910 .7939 .7967 .7995 .8023 .8051 .8078 .8106 .8133
.9 .8159 .8186 .8212 .8238 .8264 .8289 .8315 .8340 .8365 .8389

1.0 .8413 .8438 .8461 .8485 .8508 .8531 .8554 .8577 .8599 .8621
1.1 .8643 .8665 .8686 .8708 .8729 .8749 .8770 .8790 .8810 .8830
1.2 .8849 .8869 .8888 .8907 .8925 .8944 .8962 .8980 .8997 .9015 "" ""
1.3 .9032 .9049 .9066 .9082 .9099 .9115 .9131 .9147 .9162 .9177
1.4 .9192 .9207 .9222 .9236 .9251 .9265 .9279 .9492 .9306 .9319

1.5 .9332 .9345 .9357 .9370 .9382 .9394 .9406 .9418 .9429 .9441 0 "
1.6 .9452 .9463 .9474 .9484 .9495 .9505 .9515 .9525 .9535 .9545
1.7 .9554 .9564 .9573 .9582 .9591 .9599 .9608 .9616 .9625 .9633
1.8 .9641 .9649 .9656 .9664 .9671 .9678 .9686 .9693 9699 .9706
1.9 .9713 .9719 .9726 .9732 .9738 .9744 .9750 .9756 .9761 .9767

2.0 .9772 .9778 .9783 .9788 .9793 .9798 .9803 .9808 .6812 .9817
2.1 .9821 .9826 .9830 .9834 .9838 .9842 .9846 .9850 .9854 .9857
2.2 .9861 .9864 .9868 .9871 .9875 .9878 .9881 .9884 .9887 .9890
2.3 .9893 .9896 .9898 .99Q1 .9904 .9906 .9909 .9911 .9913 .9916
2.4 .9918 .9920 .9922 .9925 .9927 .9929 .9931 .9932 .9934 .9936
IN
2.5 .9938 .9940 .9941 .9943 .9945 .9946 .9948 .9949 .9951 .9952 "N
2.6 .9953 .9955 .9956 .9957 9959 .9960 .9961 .9962 9963 .9964 %
2.7 .9965 .9966 .9967 .9968 .9969 .9970 .9971 .9972 .9973 .9974 ,%
2.8 .9974 .9975 .9976 .9977 .9977 .9978 .9979 .9979 .9980 .9981
2.9 .9981 .9982 .9982 .9983 .9984 .9984 .9985 .9985 .9986 .9986

3.0 .9987 .9987 .9987 .9988 .9988 .9989 .9989 .9989 .9990 .9990
3.1 .9990 .9991 .9991 .9991 .9992 .9992 9992 .9992 .9993 .9993
3.2 .9993 .9993 .9994 .9994 9994 .9994 9994 9995 9995 .9995
3.3 .9995 .9995 .9995 .99961.9996 .9996 .9996 .9996 .9996 .9997
3.4 .9997 ,9997 9997 .99971.9997 .9997 9997 .9997 9997 .9998

S_1.2829
645[1.60 2 326 2.576 3 o9013 .891 i4417-
Wf zT - .90 .95 .975 .99 .95 1999 .9995' .999951 .999995 "".
2 - Fz .20 • 05 .02 .01 .002 .001 001 .00001-

,....
F',,0

'O0r
 IIL

~.

PLATE 1 .--

SUBJECT: Example calculations of Summary statistics for (R-test) data.

I. PROBLEM:

Calculate summary statistics from experimental R-test data on

soil strength.

II. SOLUTION:

1. Measured data:

Measured values of R-test data as shown in Figure 3.

. 2. Measures of central tendency:

mean -- mx = (1/n) Exi = 32.6%

median -- x0 .5 = 32.1%
mode -- x0 = 28.5%

3. Measures of dispersion:

standard deviation -- s = I E(x. - m )2 6.5% €

X -1 1 x .

variance -- Vx = Sx 2 = 42.8%

range -- r = (Xmax - xmin) = 29% %

fractiles -- N0. 25 = 27.4%

xO .50 = 32.1%
xo. 75 = 36%

interquartile range -- r0 .5 = (x0 . 7 5 - x0 . 2 5 ) = 8.6%

35

"4. -e e
 PLATE 2

SUBJECT: Shortcut estimates of summary statistics. 9

I. PROBLEM:

Estimate summary statistics using shortcut methods and compare to accurate

calculations.

II. DATA:e

Measured Strength (kPa): 38, 51, 43, 39, 48, 45, 42, 45, 49.

III. ESTIMATE MEAN:

Shortcut Method Using Median

mx ftmedian of xi
= 45 kPa

By Equation 2

Mx E xi - (400 kPa) =44.4 kPa

IV. ESTIMATE STANDARD DEVIATION:

Shortcut Method Using Range

W (xmax - Xmin)
-51 - 38 kPa
lkPa
13

N from Table 1 (for n=9): 0.337

sx.(0 .337) (13)

= 4.4 kPa

By Equation 3 ..

s (X1 -m, 4.2 kPa

36

0Z.
 'p

'S

'S p
*4 p
J.

.4

V.

p
P

A + B
B

0
'S
-I

4' 1
0
-I

4.

p.
H~ qure 1 -- Venji d~aq ran sflow~ nq r*-' I a ti ons arnonq prohab~ 1 ties of s~ ople
P
OVUr1 ta

5'

'S
7,
S. p.

- .....................
 :%,e

TOTAL NUMBER OF DATA 235 .

MEAN 9.902] STD DEv 4.,283
NUMBER OF OUTLIERS BEYOND LOUER LIMIT OF PLOT - 0
NUMBER OF OUTLIERS BEYOND uPPER LIMIT OF PLOT - 0

30

25 --

< 20
<V

L.L
o 15

e:: - ._-.

0 5 10 15 20 25

N VALUES -UNCORRECTED

Figure 2 -- Histogram of SPT data.

38

% A., -. A .7
 .4
%

.%

TOTAL NUMBER OF DATA 73

MEAN 32.58] STO DEv 6.5427
NUMBER OF OUTLIERS BEYOND LOUER LIMIT OF PLOT - 0
NUMBER OF OUTLIERS BEYOND UPPER LIMIT OF PLOT - 0 .

10 ON.II

,..
0 5

3
4A
-.

oI .1

25 20 25 30 35 40 45 50 55 60

UATER CONTENI @ FAILURE

RE-REO DAM FOUND. SOILS "R" TESTS

Fquiret- Hfis toq rdl (t R- t S t lidt'i

a.%

.7*
~ .3
F S5 A-. 5P --
P F5 A A t. t~ .. S S S - S -
 - - .-- -- -V IT S 1- C.770
S T 7 ZVI *7 .Z 4

TOTAL NU'lISER DOF D"ATA

15.932]
235
SID DEv 4.4283
MEAN

NUMBER OF OUTLIERS BEYOND LOUER LIM]T OF PLOT - 0 Frequency

NUMBER OF OUTLIERS BEYOND UPPER LIMIT OF PLOT 0I

30 0.128

25 - 0.106
-.

< 20 0.085
P--.4

S_ _0.064

LUJ

z K 0.043

0.021

I %000

0 5 10 1t 20 25 ;

N VALUES UNCORRECTED

F.;kure 4 -- Frequency d: str ibulti )ti; t it.i of Fig. 2.

p.

40

I
 TOTAL NU' BER OF DATA 73
NEAN 32.58) STO DEv 6.5427

I
NUMBER OF OUTLIERS BEY OND LOUE-R L2IMIT OF PLOT - 0
NUMBER OF OUTLIERS BEYOtD uPPER LIMIT OF PLOT - 0
Frequency
SI0.137

10.110

< 0.082

4 0.055

Wo
2 0.027

-%TE TOTAL @ FAILURE

I0

- 5 D c~ K- ,.~*~~

M RE-RLER DAM FOUND. SOILS "R" TESTS

5 FiLure
Frequency UT
distributions for the data of Fig. 3.

.141

2JI II0 1
 '.

TOL4L ? Uj,1R D- DaTA 235

MEAN STD DEv 28 I
NUM
LR OF OUTLIERS REYQOD LO IER LIMIT PLOT OF 00 Cumulative
Frequency 1
NUMBER OF OUTLIERS BEYOND UPPER LIMIT OF PLOT w*

3B

cumulative distribution
25 1001Z

< 20

0 15 ANN-

rD 50*, .

z 10

525/

I0
.

0 10 5 20 25

N VALUES UNCORRECTED

RE-PEG DAM STA 2500 - STA 3200"

Figure 6 Cumulative distribution of the SPT data. .0

42

:I
 TOTAL NUMBER OF DATA ?3
NEAN 32.581 STD DEv 6.5427
NUMIBER OF OUTLIERS BEYOND LOUER LIMIT OF PLOT - 0 Cumulative
NUMIBER OF OUTLIERS BEYONJD UPPER LIMIT OF PLOT - 0 Frequency
10

7 7A 100%

distribu-
.
cumulative
C-
UMS tion
0L 75%

50%

225

15 20 25 30 35 40 45 50 55 60

UATER CONTENI @ FAILURE

R~E-REG DAM FOUND. SOILS "R" TESTS

Fi~jure 7 -- cumulative distribution of the R-test iiti.

4M

1 01
 (a) &ymetri c

IL

(b) skewed
• ."
.5.,

Figure 8 -- Syame:(ic and skewed frequency distributions.

%
,.t

_*(5

44

*".'*'*5***.".*.•-.'.-*- * - %**" %.- •.. . . .. .----. " • - - - '% - % lrj

'1
 TOIAL NUMBlEER OF DATA25
0 MEAN - .902) STD [)-v -4 .,263 '30
NUMB~ER OF OUTL IERS BEYOD Lb,_'ER L )M] OF PLOT - 0
NUMBER OF OUTLIERS B8EYO,'D UPPER LIMIT OF PLOT - 0

range

25 ea
5
I N N

< 20-

LL
0 15
LU Mp

z 1

iA
01
0 5 10 t) 20 25

N VALUES UNCORRECTED

RE-REG DAM STA 2't00 - STA 32t00

i
iA

S.F.qtire j summanry stat~ stIcs ot ai treoueiicy distr hultJon.

.0

p JJ...-.' .JC*- - Z~

-, . , . " . -.. .-
- -- - .. -_ . .r .- ., .-. -. '_ '. .'_ -. '_. .. . . .. . . . . , .'. ' . . . . , . . . " ; " ,-.. ', 4 ¢ ' . z . ' x ; ' ,
 10 I C-IHC

//o /

t0o0~~~ /
0 ILN
4~~ 0/0

0 %'

2 1
x. ~

Figue 1 Exapleappicaton
f baloo mehod or stimtin th

corrlaton ofexpeimetalceffcien ata

46A

0 I I
 Density
0.0040

0.0030

0.0020 r

0 -
a,200 300 400 500 600 700 800
,- 3a, j-2a u- a, A s+a ~+2a,;
A u+ 3a

Cumulative
- 1.0

0.2
a0.6,
'a VI

x~ 0.4

% 0.2

a0,

a, - a d o + a '-2,
a+01s+

aa

Fiqure 11 -- Normal or bell-shaped frequency distribution.

V, (from Chernoff and Moses, 1950, Elementary
Decision Theory, John Willey and Sons, NY)
%
47

a-.0
 P
%

"9,

%40

30 I I T ',

++,

4 2
25 + + + +
+ + + 4 + + 4

Lz + ++++ + + + + +4 +4 ++ + "
1 +_ + 4 + + +

+ + + -+ , +4 + -+
++. + + +1 4 4 -+ +
S20+ 44 4+ + + +
* ++ +~~ '44 ~ 4
#4* ,41 4 34+ t
,t
4* +
+
# +
*+41'
+
04~+$ _+
++ -f+ -+ -t++
+ *LJ ; 4*1+ f+ 4 4 +41

w* + 4 4 + + r
+ +++

10+ +t
+ +

CUMMULAT IVE NUMFBER Qfi IL S;I:.

x0A +44
6 f++

Fiqure 12 -- Typical control chart for the output of a process operdt~nI iin ,
random manner. _...

48 .1'
--. ,....

1..,..-_,0
 Part III: CONCEPTS OF STATISTICAL QUALITY CONTROL

Quality Assurance and Quality Control

The terms quality assurance (QA) and quality control (QC) are used in

special and differing ways by different organizations. In this report,

Quality assurance means an inspection program aimed at assuring that soils

placed in a fill meet specifications.

Quality control means an inspection program aimed at monitoring

construction performance to give early warning of changes that affect
quality and thus to provide a basis for controlling the process.

Quality assurance programs prescribe a procedure which when consistently

applied to inspection data yield a specified risk of accepting lifts of given

.449

quality. A QA program provides a decision procedure. Quality control, on the .

other hand, provides a way of estimating lift properties and the changes in ,

those properties
% %4 with time. A QC program provides a monitoring scheme. QA
%€
provides a rulea m by
• which the owner's risk of accepting poor quality as.
% .'./ - .€.-'.€.
.',2,..2 " "., -... •'.." '°€. . " ."- "-'. . . "• • . ," " " • " -"-'
construction is guaranteed and balanced against the contractor's risk of having

good quality construction rejected. QC provides a tool by which owner and

contractor alike can make efforts to maintain a uniformly high quality

product.

Sampling

Measurements are made on a set of soil specimens or at a set of locations

in order to estimate the properties of a soil deposit or an engineered

structure as a whole. Statisticians call this set of measurements a sample.

Pa-rI
 An individual piece of soil is called a specimen to distinguish it from the

concept of a statistical sample.

The soil deposit or structure whose properties are of interest is called

the target population (Fig. 13). A population in statistics is simply a large

(or infinite) collection of elements. Not all of the elements in the targjet

population may be accessible for samplinq. Those that are accessible are said

to compose the sampled population. Prom this sampled population a finite

number of elements are selected for testing and this set is called a sample.

If the way this sample is chosen satisfies certain rules, the samplo is said to

be a probability sample. Statistical methods can the be used ti) quantify the

uncertainty in estimates from the sample about properties of the sanpled

population. Statistics is powerless to say anything about the correspondence

between sampled and target populations, however, as this is a geological or

engineering question.

Scientific Sampling

The concept of scientific sampling, or probability sampling, is central to

quality assurance and control. A scientific sample is planned according to

statistical principles. The importance of scientific sampling is that it

allows quantitative statements about the uncertainty in parameter estimates .

which result from sampling. Other sampling schemes--as for example,

instructing an inspector to purposely seek out areas in a fill that appear .

poorly compacted--certainly have merit in special circumstances, hut thf.y do

not allow the quantitative analysis which has come to underlie modern

engineering practice.

5 0-

-.
:..--..v
:...:...> ......
-- - ~~................ .. ...... . . - .. . . . .. .. . .. . . . . . .- . .. . -:: :
 To be a probability sample, three criteria must be satisfied: (1) sample

points must be chosen randomly, (2) all elements in the sampled population must

have a non-zero chance of being sampled, and (3) different probabilities of

each element being sampled must be compensated by weighting.

If these two criteria are satisfied--and only if they are satisfied--

statistical methods can be used to determine uncertainties properly associated

with parameter estimates. This means that for statistical methods to be used,

some form of random sampling is necessary. Purposive sampling, by which an

inspector consciously selects for testing those elements that appear of poor
d?
quality, is intuitively appealing and can provide important information, but it

cannot form the basis for statistical quality control. From a purposive sample

there is, (a) no way to assign quantitative confidences to estimates of soil

parameters, (b) no way to explicitly review an inspection program after the

fact, and (c) no way to establish a defendable level of quality assurance.

People also talk about having an inspector seek out a 'representative'

sample. This, too, may have merit in special circumstances, but it does not

produce a sample from which quantitative conclusions can be drawn. No

individual sample is representative of a sampled population. A sample contains

!;pecific measurements which can never precisely mirror the subtlety of

variations in the sampled population. On the other hand, a sampling plan can

be made representative, if designed by scientific principles, in that it

affords every element within the sampled population an equal chance to

influence etimates thit are made.

- ] ..

*5... -.
 Random Sampling

Scientific sampling requires that every element in the sampled population

have a non-zero chan,-e of appearing in the sample. It does not require these

chances to A!, be the same, only that the relative probabilities are known.

This con lition requires that elements be selected from the sampled population

in a random way. Lacking a random procedure, the assumption that each element
-
has a non-zero chance of being sampled cannot be made, and the relative

probabilities of different elements being sampled cannot be assessed. Tne use

of a random procedure attempts to avoid any form of association between the -N

selection of elements for the sample and the properties of the elements that

are being sampled. Such association is called a bias.

Randomization means selecting elements of a sample in such a way that the

two conditions of probability sampling are satisfied. Randomization can be

accomplished many ways. A conceptually simple but operationally clumsy way is

to pick sampling locations by a random number generator or table of random

numbers (Table 5). If performed faithfully this scheme gives each element in 0

the sampled population an equal chance of being sampled. Another way to .

provide randomization is to layout measurements on a fixed grid and then

randomly locate the first point.

Sampling Plans

An essentially infinite number of sampling plans for quality assurance or

quality control satisfy the properties of probability sampling. These are all

randomized sampling plans in the sense that the exact elements whi-h are .%

sampled depend on the outcome of some chance event. 5

52

-
, ,.- -5,
, .' *..%
-. , • - e • ..
• " ", ,, ." . " " . .. .. •,i .4.," " "% " . " ' "'• • " "-" ".• • •.• "•. .
 'Jb

A more convenient sampling plan is to layout sample points on a grid, and

then locate the grid in the field by randomly selecting its first point (Fig.

14b). Only one pair of random numbers needs to be chosen from which all of the

sample points are determined. The disadvantage of a grid pla,I compared to-

purely random plans is that any spatial periodicity in the compaction process

may bias the outcome. An advantage compared to the purely randn plan,

especially with small sample sizes, is that uniform coverage of the site is

assured.

To provide coverage while at the -te time limiting the possible effects

of periodicities, stratified random sampling plans are sometimes usoi. 'Isinq a

stratified plan the sampled area is first ividd into a re,11lar Irray ,'f

squares or rectangles (Fig. 14c) and then a sample po>int is randm y loiat,1 in

each.

Another common plan is nested sampiing. Nested sampl ing u:es a pre-fixel

" grid of sample points with varying spa,-inqs (Fi ;. 141) T'he first 1 )int i -

located randomly as in grii samplinj ari fromn that poi'it iii tti r ,, .ir

specified. The principal se of nstJ samplinq is f)r -stimatniq, s;'- Ia,

.
J.aspects of the spatial strioctoir )t 01 ils 1ota, namely t tio t.)r. o

function or variograln, (se the report, "S atistioji in iy-, - i:r,

data", Instructional Report (;[,- 7). The, 1o )f ruist-i in) il' I' ) i ' I t7

control or quaility alwsiiroin,- )t *o lt- ti 1 I i m);*1v t r ip t

a sses' !1'1 m ,
111 nI r r r-, ri II Ii
1' r- I 'll 11 1

R n+isn
R* Ill.) j- 1 I mf 1, 1 1! r. I V'......c i i i
 -w~ -u IV - -

%.

operation is large compared to the incremental cost of testing. Clumped

sampling involves two stages. In the first stage a number of seed points are

randomly chosen. In the second stage a number of sampling points are chosen in

the vicinity of each seed point. At both the first and second stage the

sampling plan can be purely random, gridded, stratified, nested, or so forth.

The simpling plans reviewed here are typical of the very large number of

possible sanpling places. In practical situations the constraints of a parti-

cuiar project may dictate that a specialized plan be developed. This is

ac,:epte practice as long as the principles of probability sampling are adhered

t". These principles dictate three thinqs, (1) that sample points be chosen

a''oring to s,)me random process and not be affected by the intuition of an

inspe't,)r, (2) that all elements within the population to be sampled have a

nmn-zero ichrinc,
of being sampled, and (3) that if the prohabilities of each

element hetrig s mupled are not ill the same, these differences in probability he

a~plropr tely COflplSi t~eH for by weighting when the data are analyzed.

F'or o t ial i ty control and guali ty assurance sampl inq in qeotechnical

n;ne r nj the )ribaL rIitie n of elements within the sampled population beinq

,unpld are .111 the iame. Therefore, for these samplinq plans the problem of

,h t
tii; unp ,. !it <,ms ni- ldo-m of concern. For those cas,,i' where.

0 ill irr i r--in iry , (,)(.:hrarl

i t . (1964)
;.
provide,;
I I;
techniques tini prai t i-al

In ~~~ 1 T-tJ~ .-. ,~

"% .

IF
i-r .5.- % '%

.- %

, z... . "'. . . ."" .

.''z' " ' . ''. " .{ .' . 4-
. . • .. ; "." , ".' ,","."
 a single test may yield .ore than one numerical result--for example, water

content, dry density, plasticity index, and so forth--but for now nothing is

lost by considering only a single scalar outcome.

If another sample of n specifications is now taken from the same lift of

soil, however with the specimens taken at slightly different places, another

set of n numerical data will result. Each of these will differ somewhat from

their counterparts in the first sample, because the soil itself varies from one

spot to another and because there are a number of instrument or operator

effects which influence test results. This variation in numerical results from

one sample to another is called sample variation. Statistical techniques allow

such sample variation to be predicted and dealt with in a quantitative way.

The sample mean (Eqn. 11), sample standard deviation (Eqn. 13), and other

summary measures calculated from the test results x1,...,xn are simply

mathematical function of the data. If the data vary from one sample to

another, so will the summary measures.

Sampling Variability of the Mean

The sample mean m x is calculated by Eqn. 11. If many tests are made

(i.e., if n is large), variations in one test result within an average will be

offset by variations in others, and as a result m x should be fairly close to

the actual mean of the sampled population mx. In this report the actual

sampled populations mean is denoted by a prime, m' , as compared to the sample

% mean which is denoted without a prime, mx. On the other hand, if few tests are

made (i.e., n is small), variation in test results will not have as much

opportunity to averaqe out, and as a result the sample mean may deviate

considerably from mx'. This samplinq variability is the critical factor in

55

.,,0
deciding how many tests must be made in a quality control or quality aisurance

program.

If the standard deviation of the sampled population is known and if the

individual measurements are independent of one another, then the means of

individual samples each of size n will vary with a standard deviation of

Sn= sx / /n. -25-

For example, Fig. 15 shows a histogram of sample means, each corresponding to a

different sample of size n = 5 taken randomly from the SPT blow count data in

Fig. 2. The standard deviation of the sampled set of data is 4.4 bpf, while

the standard deviation of the variability of the sample means is 2 bpf 4.4 V-p

bpf/ n. If plotted as a frequency distribution, the variability of the sample

means will be approximately Normally distributed, almost without regard to the

shape of the frequency distribution of the sampled population. The V

approximation to the Normal distribution becomes better as n becomes larger.

In the more common case the true standard deviation of the sampled

population, Sx' , is not known, and thus the sample standard deviation, sx, is

used in equation 25 to approximate the variability of m x about mx'. Usinq s x

rather than s x ' underestimates the variability in mx, however, because the

estimate of s x differs somewhat from the true standard deviation Sx' lb

overcome this limitation a standardized mean is used,

m - m
x x -6
x

56

K,
 in which m x and s x are the sample mean and sample standard deviation, and m x %"

is the true mean of the sampled population. If the value t is estimated

separately from a large number of samples, each sample containing the same

number of observations, the frequency distribution of t over these many

separate samples will have a standard deviation of 1.0 and a shape known as

Student's-t distribution. The Student's-t distribution looks much like a

Normal distribution, but with thicker tails and a higher mode. That is, the

Student t has somewhat more of what statisticians call kurtosis than a Normal

distribution does. Areas under the Student curve are given in Table 6, and may

be approximated by series expansions, as given by (Abramowitz and Segun (1964). J-

The shape of the Student's-t distribution and thus the areas beneath it depend

on the number of measurements within a sample, n. This enters Table 6 as the

degrees-of-freedom parameter u = n-1.

Sampling Variability of the Standard Deviation

Just as the sample mean varies from one sample to another, so do other

summary measures such as the sample standard deviation or sample variance.

Unfortunately, the statistical results for the variability if the sample

standard deviation and variance are not as simple as those for the sample mean.

For samples taken from Normally distributed data, the sample standard deviation

varies approximately with a standard deviation of,

s s
S V2n ,-27-.'

- 57

N N-
 %°

in which n = the sample size (Snedecor and Cochran, 1980). The sample

variance, sx2 , varies approximately with a standard deviation of,

s s 2/2 -28-
SX x / n-1

.Similar results are available for non-Normally distributed data, but are more

complicated. Most basic statistics textbooks discuss these results (e.g., .6%
%

Snedecor and Cochran, 1980).

Sampling Variability of the Range

Because the sample range r x =IXmax Xminl is easier to calculate than the

standard deviation, it is often preferred as a measure of variablity in

prograis of quality control. For a sample taken from Normally distributed data

the frequency distribution of the relative range

wx = rx/Sx -29-

across samples of size n is tabulated in Table 7.

In-Control vs. Out-of-Control Processes

The above discussions are based on the concept of a construction process

operating in a random manner. When a process is operating in a random manner

any part of its output may be viewed as typical of the output as a whole and

perturbations 'n the process show no discernable pattern. A construction

process .perating in a random manner and producing few sample ont-oms which

deviate substantially from its average output is said to be "in-control." F1 q.

A. . . ,P . ..
.-.- .. - .. .A. , . - -.
%
q
B

,--
lo.p, -p . % .%.
I
l .g$ . u i-..
C J . - .*Tyj'.- .: i .o V' . .
-==

16 shows a chart of compaction data for a process operating in-control. The

variations in these data appear to behave randomly without trend or pattern. V

Few construction processes operate in-control for significant lengths of

time, and even when a process is in-control minor deviations from randomness

have to be overlooked. The statistical theory of quality control and quality

assurance is based on the idea of randomness in process output, and thus every

effort should be made to assure that non-random factors are not present.

Whether a process is in-control also depends on the level of detail with P -

which the output is scrutinized. Obviously, variations in the output of a.

construction process are not truly random. At some level of detail are all

explainable by physical arguments. The notions of randomness and a process

P
.-

being in-control have to do with engineering decisions and the cost

effective-ness of further reduction in output variability. Attempts are made

to identify and eliminate all major sources of variability, and what remains

and is not cost effective to further reduce is operationally handled as if it

were random variation. As long as our statistical models can be successfully

used to portray this residual variability and to characterize uncertainties

which arise from it, then the process is for engineering purposes

"in-control."

When the variability in the output of a construction process deviates from

randnness, that is, when significant trends or patterns begin to appear in the

output, statistical models no longer adequately capture the important features -

of the vdriations ind the construction process is said to he out-of-control.

The pr nci pat us(- of the control charts of Section 5 is to obtain early warning

that a pr,. is going out-ot-control, and to identify ste:ps that might he

% % % %
taken to bring the process back in-control. Examples of changes that could

cause a fill and compaction process to go out-of-control include a change in

borrow materials; change in rainfall; change in construction superintendent,

operator, or equipment; and so on.e!e

60

% . . . . . .
 p.-

%
Table 5 -- Table of uniform [0,1] random numbers
(from Cochran, 1977).
';I'

10 27 53 96 23 71 50 54 36 23 54 31 04 82 98 04 14 12 15 09 26 78 25 47 47
28 41 50 61 88 64 85 27 20 18 83 36 36 05 56 39 71 65 09 62 94 76 62 11 89
34 21 42 57 02 59 19 18 97 48 80 30 03 30 98 05 24 67 70 07 84 97 50 87 46
61 81 77 23 23 82 82 11 54 08 53 28 70 58 96 44 07 39 55 43 42 34 43 39 28
61 15 18 13 54 16 86 20 26 88 90 74 80 55 09 14 53 90 51 17 52 01 63 01 59

91 76 21 64 64 44 91 13 32 97 75 31 62 66 54 84 80 32 75 77 56 08 25 70 29
00 97 79 08 06 37 30 28 59 85 53 56 68 53 40 01 74 39 59 73 30 19 99 85 48
36 46 18 34 94 75 20 80 27 77 78 91 69 16 00 08 43 18 73 68 67 69 61 34 25
88 98 99 60 50 65 95 79 42 94 93 62 40 89 96 43 56 47 71 66 46 76 29 67 02
04 37 59 87 21 05 02 03 24 17 47 97 81 56 51 92 34 86 01 82 55 51 33 12 91

63 62 06 34 41 94 21 78 55 09 72 76 45 16 94 29 95 81 83 83 79 88 01 97 30
78 47 23 53 90 34 41 92 45 71 09 23 70 70 07 12 38 92 79 43 14 85 11 47 23
87 68 62 15 43 53 14 36 59 25 54 47 33 70 15 59 24 48 40 35 50 03 42 99 36
47 60 92 10 77 88 59 53 11 52 66 25 69 07 04 48 68 64 71 06 61 65 70 22 12
56 88 87 59 41 65 28 04 67 53 95 79 88 37 31 50 41 06 94 76 81 83 17 16 33

02 57 45 86 67 73 43 07 34 48 44 26 87 93 29 77 09 61 67 84 06 69 44 77 75
31 54 14 13 17 48 62 11 90 60 68 12 93 64 28 46 24 79 16 76 14 60 25 51 01
28 50 16 43 36 28 97 85 58 99 67 22 52 76 23 24 70 36 54 4 59 28 61 71 96
63 29 62 66 50 02 63 45 52 38 67 63 47 54 75 83 24 78 43 20 92 63 13 47 48
45 65 58 26 51 76 96 59 38 72 86 57 45 71 46 44 67 76 14 55 44 88 01 62 12

39 65 36 63 70 77 45 85 50 51 74 13 39 35 22 30 53 36 02 95 49 34 88 73 61
73 71 98 16 04 29 18 94 51 23 76 51 94 84 86 79 93 96 38 63 08 58 25 58 94
72 20 56 20 11 72 65 71 08 86 79 57 95 13 91 97 48 72 66 48 09 71 17 24 89
75 17 26 99 76 89 37 20 70 01 77 31 61 95 46 26 97 05 73 51 53 33 18 72 87
37 48 60 82 29 81 30 15 39 14 48 38 75 93 29 06 87 37 78 48 45 56 00 84 47

68 08 02 80 72 83 71 46 30 49 89 17 95 88 29 02 39 56 03 46 97 74 06 56 17
14 23 98 61 67 70 52 85 01 50 01 84 02 78 43 10 62 98 19 41 18 83 99 47 99
49 08 96 21 44 25 27 99 41 28 07 41 08 34 66 19 42 74 39 91 41 96 53 78 72
78 37 06 08 43 63 61 62 42 29 39 68 95 10 96 09 24 23 00 62 56 12 80 73 16 .0
37 21 34 17 68 68 96 83 23 56 32 84 60 15 31 44 73 67 34 77 91 15 79 74 58 . 5,

14 29 09 34 04 87 83 07 55 07 76 58 30 83 64 87 29 25 58 84 86 50 60 00 25
58 43 28 06 36 49 52 83 51 14 47 56 91 29 34 05 87 31 06 95 12 45 57 09 09
10 43 67 29 70 80 62 80 03 42 10 80 21 38 84 90 56 35 03 09 43 12 74 49 14
44 38 88 39 54 86 97 37 44 22 00 95 01 31 76 17 16 29 56 63 38 78 94 49 81
90 69 59 19 51 85 39 52 85 13 07 28 37 07 61 11 16 36 27 03 78 86 72 04 95

41 47 10 25 62 97 05 31 03 61 20 26 36 31 62 68 69 86 95 44 84 95 48 46 45
91 94 14 63 19 75 89 11 47 11 31 56 34 19 09 79 57 92 36 59 14 93 87 81 40
80 06 54 18 66 09 18 94 06 19 98 40 07 17 81 22 45 44 84 11 24 62 20 42 31
67 72 77 63 48 84 08 31 55 58 24 33 45 77 58 80 45 67 93 82 75 70 16 08 24
59 40 24 13 27 79 26 88 86 30 01 31 60 10 39 53 58 47 70 93 85 81 56 39 38

05 90 35 89 95 01 61 16 96 94 50 78 13 69 36 37 68 53 37 31 71 26 35 03 71
44 43 80 69 98 46 68 05 14 82 90 78 50 05 62 77 79 13 57 44 59 60 10 39 66
61 81 31 96 82 00 57 25 60 59 46 72 60 18 77 55 66 12 62 11 08 99 55 64 57
42 88 07 10 05 24 98 65 63 21 47 21 61 88 32 27 80 30 21 60 10 92 35 36 12
77 94 30 05 39 28 10 99 00 27 12 73 73 99 12 49 99 57 94 82 96 88 57 17 91

5.

..

.,S.

"*%.-<
.55PS :- .- ' ~ : a ~ ~ . K Y V .
Table 6 -- Percentage points (i.e., double tail areas) of the Student-t
distribution. For areas under a single tail, divide by two.
From Duncan, 1974.

Probability (P). e,

"0 "98 "7 '6 "5 "4 "3 .2 "1 05 '02 01 '001_
-
1 .158 .325 .510 -727 1.000 1.376 1.963 3.078 6.314 12.706 31.821 63.657 636 619 i"-_
2 .142 .289 "445 .617 .816 1.061 1.386 1.886 2-920 4.303 6.965 9.925 31.598 "-'
3 .137 .277 .424 -584 .765 .978 1.250 1.638 2,353 3' 182 4.541 5-841 12 941 "
4 -134 271 -414 -569 741 941 1 190 1-533 2-132 2 776 3.747 4. G04 8 610 "'
5 132 .267 .408 559 .727 .920 1156 1476 2.015 20571 30365 4032 6859

6 .131 .265 404 .553 1718 1906 1134 1440 1.943 2.447 3142 3707 5 959
2 .130 .263 .402 .549
* 711 •896 1.119 1.411 -895 2365 2. 992 3598099
5-
8 .130 .262 .399 .546 .706 .889 1.108 1397 1860 2306 2.896 3355 5 041

9 .129 .261 .398 *543 .703 .883 1100 1.383 1.833 2.262 2.821 3.250 4.781
10 .129 .260 .397 .542 '700 .879 1-093 1.372 1.812 2.228 2.764 3.169 4.587

11 .129 .260 .396 .540 *697 .876 1.088 1.363 1.796 2201 2.718 3.106 4.437
12 .128 .259 .395 -539 .695 -873 1.083 1.356 1.782 2.179 2.681 3-055 4 318
13 .128 '259 .394 .538 -694 .870 1.079 1.350 1.771 2.160 2.650 3-012 4-221
14 .128 .258 -393 .537 *692 '868 1.076 1.345 1.761 2.145 2 624 2.977 4 140
15 .128 -258 .393 *536 *691 .866 1.074 1.341 1.753 2,131 2.602 2.947 4 073

16 -128 *258 .392 -535 '690 '865 1.071 1.337 1.746 2.120 2.583 2.921 4-015
17 .128 .257 .392 "534 *689 '863 1.069 1-333 1.740 2,110 2.567 2.898 3.965
18 .127 -257 .392 .534 *688 '862 1.067 1.330 1'734 2.101 2'552 2.878 3.922
19 '127 -257 .391 .533 '688 .861 1'066 1-328 1.729 2.093 2.539 2-861 3-883
20 -127 .257 3.91 .533 *687 .860 1.064 1.325 1-725 2.086 2.528 2-845 3-850 ,

21 .127 .257 .391 -532 *.686 '859 1.063 1-323 1:721 2-080 2518 2-831 3819
22 .127 .256 -390 -532 -686 .858 1.061 1.321 1.717 2.074 2-508 2.819 3.792
23 .127 .256 -390 .532 .685 .858 1.060 1.319 1.714 2.069 2.500 2 807 3.767
24 '127 .256 .390 .531 .685 '857 12059 1318 1.711 2.064 2-492 2.797 3.745
25 .127 .256 .390 .531 -684 .856 1.058 1.316 1.708 2.060 2.485 2.787 3.725

26 .127 -256 .390 .531 .684 .856 1.058 1.315 1.706 2 056 2-479 2 779 3.707
27 .127 .256 -389 .531 .684 '855 1 057 1-314 1.703 2.052 2.473 2 771 3.690
28 .127 .256 "389 "530 -683 855 1.056 1.313 1.701 2.048 2.467 2-763 3 674
29 .127 .256 .389 .530 .683 -854 1 055 1.311 1.699 2.045 2.462 2.756 3.659
30 -127 •256 .389 .530 .683 854 1 055 1.310 1.697 2.042 2.457 2.750 3-646

40 .126 -255 388 -529 *6'81 .851 1 050 1 303 1 .684 2.021 2 423 2.704 3-551
60 .126 -254 387 .527 679 -848 1-046 1.296 1.671 2.000 2-390 2.660 3460
120 -126 254 386 .526 -677 -845 1-041 1-289 1-658 1-980 2-358 2-617 3.373
.126 253 385 524 674 842 1 036 1 282 1 645 1.960 2.326 2.576 3.291 -

=%1

%'

%
%.% %%
 -a

Table 7 -- Percentage points of the distribution of relative range

w =r /s for small samples from Normal distributions
(from Duncan, 1974).

II

Mean w Probability That w Is Less than or Equal to Tabular Entry

n or or
di ds 0.001 0.005 0.010 0.025 0.050 0.950 0.975 0.990 0.995 0.999

2 1.128 0.8525 0.00 0.01 0.02 0.04 0.09 2.77 3.17 3.64 3.97 4.65
3 1.693 0.8884 0.06 0.13 0.19 0.30 0.43 3.31 3.68 4.12 4.42 5.06
4 2.059 0.8798 0.20 0.34 0.43 0.59 0.76 3.63 3.98 4.40 4.69 5.31
5 2.326 0.8641 0.37 0.55 0.66 0.85 1.03 3.86 4.20 4.60 4.89 5.48

6 2.534 0.8480 0.54 0.75 0.87 1.06 1.25 4.03 4.36 4.76 5.03 5.62
7 2.704 0.833 0.69 0.92 1.05 1.25 1.44 4.17 4.49 4.88 5.15 5.73
8 2.347 0.820 0.83 1.08 1.20 1.41 1.60 4.29 4.61 4.99 5.26 5.82
9 2.970 0.808 0.96 1.21 1.34 1.55 1.74 4.39 4.70 5.08 5.34 5.90

10 3.078 0.797 1.08 1.33 1.47 1.67 1.86 4.47 4.79 5.16 5.42 5.97
11 3.173 0.787 1.20 1.45 1.58 1.78 1.97 4.55 4.86 5.23 5.49 6.04
12 3.258 0.778 1.30 1.55 1.68 1.88 2.07 4.62 4.92 5.29 5.54 6.09
-- [.

5.1

",''-5.
 1

SAMPLPED

Fiqjure 13 -- Populations of interest in sampli * .l

C V - c 0 C, --

C
c 0 C C C C C

0 0
C, 0 CC 0 C I

C, 0

l - in Cor C CorC na e

I~~ 0 0

cc C C C CC
E

~11
C vC 0 C

(c c (d

- a 0s --- DC 0n 0 ,-
 ~- -
p-s.,
F

'5
S.'...,

.5.
p.,

.5.

5'.'
.5%

p1: 9.(~ .5-5'

~f.

6
-5..

4 *55*
*5-5
S.
'5%
'5%

z
-. p

2
/

4 ~ 6. 7 ~ ~j ,'o ," ii ...

81CV~~' (' ~IN

Figure 15 -- Hi stoqram ot the ~1Pd~IS )~ I 2 T . .

from a larJe data set. I~ iii * .' -.. r

camp I ed.

'V ".~ i. -
S
.'2.
5. * -~I. ... ** .~ / *
 - h - - - . - - 'vw q-w-w-w-w----~- ~'-w-~-

*0~
0~

It
U-

P
-S
*4~
PART IV -- INSPECTION OF ENGINEERF -.

4.

The ronstruc tiori of enqi neerei embankmentci or f i 1 Is mist he ~- ii *~ 1.fl i -K.

It-
order to ensure that zones within the fill are satisfactorily bomvw~i#'i.; w~ 44
I,-

t4
tflat average prpertles 4 rrm~ f 111 7 )nf)rm S peo if ici ret1u1r~en*;. ;4j ;i
4
.1
constru'-ti in - )ntrJ1 1.s ~i~ )411F)I vt.-1 by i -vi'sl irIsf~-ti Mi, rs;.lerns-..'&..I fly I It

t
systematic pr'.'qr-tn H sampi in; nil ~4r I 1121.-I it ,nrify:n~ 'iC fl~ Id. i~/

I-;1 ;ri

I--i, i S

.4

1.1 1 4 ... 4

I. * , . . . 4
- . I- -,

4.'
.4
4%-

I .~. .

4-.

.4

.4,
4',
.4,
-a
d

N
"4

4.
4.
S.

S.
S.

.4
.4.

S.. p
.4.
4,'

:-* 4.:v-~:-. L C ~tc~.x-.t1 / tC.JK V1C&.Kc - - . . .4- .- .- - .4 .4 - j .4

*W U-W W W'NU W - ~ W V d V ~ W , .

Th'-" Io t -)dpli )11',a1'i

m Te-tlin

m iitiilv pr mun 4 )nparion ins~w-K' )n) i nvolvi n. ±iinplinq aml

i I r: ZY rm
7t ;-Mf)-1 r.'qirvn-n1q rvqjariih maqv.np'1ty ml .J

. .,.

inin -"'A! *o p.''? s f-'< -11 jjt'rn

ahi~~~~~ "h 1LM!KV

-,il O nM i r n f~

1* . I ja p- 110*- **I. m r w 4 i

t ;-. . . - w ws .
 .)

Thus, a sampling inspection program has two goals: to control homogeneity

and to control average properties. Which, if either, of these goals is more

important depends on the specific situation.

Target Properties vs. Sampled Properties

The properties of greatest interest to the engineering performance of an

..'
embankment or fill are strength, deformability, and permeability. However,

these target properties are cumbersome or expensive to measure directly, so

other more easily or quickly measured properties are used in their place. By

far the most commonly sampled properties in construction inspection of fills

are compaction water content and dry density. The fact that these are

correlated to strength, deformability, and permeability makes them useful

surrogates.

Tests for Water Content and Dry Density

A variety of tests are available for measuring water content and dry

J ensity. Water content can be measured directly by oven drying a specimen and

determining change in weight. Dry density can be measured directly by

ascertaining the weight and volume of a specimen, as for example, with a sand

cone Jensity test. •

Water :ontent and dry density can also be measured indirectly using %

vyo ios Aevi-es, for example by nuclear gage. These indirect tests are

typi- ally less expensive than direct testing but also less accurate. In

:rtr, in c',is, e'onom1; can be gained by combining a small number of direct

S4 it ti lirer macher o indirec t tests.

i,,. ) t fi. i (en,;ity an. wter content tests are given in Lambe

1 ),, . 'rar J, ,t .I. 1')f)i), F:nqin f'r Mnnkia EM1 1 10-2-1911, 1IS BR (1960), and

Li %. ~~.
Jil
"
AASHO and ASTM standard test procedures.

Compaction Specifications

Specifications for compaction quality are most often placed on water

content, dry density, or both. Fbr example, water content of the fill at time

of compaction might be specified to be within ±2% of standard or modified

Proctor optimum. Dry density might be specified to be at least 95% of standard

or modified Proctor optimum. These are performance specifications.

Specifications are also placed on coaipaction equipment and procedure. Fbr

example, a specified number of passes with equipment of specified minimum

capacity may be required in addition to some specified water content range. As

an example, on the USAE Carters Dam Project, Georgia, compaction specifications Z'

for the imperious core required as-placed water contents to be ±2% standar or

modified Proctor optimum and the fill to be compacted by a minimum number of

passes using specified equipment. If placed materials were found to have water

contents more than ±2% from optimum, the cost fell to the contractor to

moisten, dryout, or remove the material. If placed materials were within water

content specifications and had been properly compacted, but were less than 95%

standard or modified Proctor optimum density, then the cost fell to the owner

to undertake additional compaction or to remove the material. These are

compliance specifications. ..

..m

,%. .%

V V
70,

-S:
 .

PART V: QUALITY CONTROL CHARTS

Quality control techniques are used by both the contractor and the owner

to monitor the progress of construction, and thus to quickly identify changes

in soils or operational procedures before these changes adversely affect

construction quality.

Quality control differs from acceptance sampling in that " has the

principal purpose of identifying changes in construction materials or

procedures before those changes adversely affect construction quality. When a

change is detected, efforts are made to find assignable causes .in, fix them.

Acceptance sampling, in constrast, ha- the principal purp() .' f as)f r nq that

n
soils placed in a fill meet s5 Jwifira tions. on aI"
o-,s*1 t"iii,-o siml

results soils are either ac-eptod or rei ;ts part t a P1f1i y 1 ;'ir t '

program .

,Phoory t
f mtr )I "halrt

T'Fit in 11 t n W- )t 1 1- r y
Ir i - 1 -hti ' t- t , ' e.

' ca- k o; )f va r I t t 1 11i 'l ItI'" . * t I i' I I ;

- simply tlir, r,...,Il t 't ; "t v it i I iI t,r.

1
)r itm in T " If I
i , ,l; ; - I . , * - .

* )rI t I 7 t I -[ ; T I I * ' 7

[ . . . ~ . .
.-
I
*
*
.- .
.%-
I" --
many chancwe caios.- MOM I proluct. var ration in soil Pttsivy., man tir4- ntonto

or atner proner ties are sinri 'ir to the many forces which caulse- a t-os-wi coin to

IinAi up ht';i -r tails. su-h var iations; fol loYw prelii -t-ahl- laws of 4

pr )baii Ii

)n W' trier han i, o tter var ia tions 1 n quaIi ty a*ra- MeP to aq,~ I nahlot

'io,4~;. lIn ;nc a1 '-; a1 q; Yini f Lcan t f rdat ton of the va r iahi I i ty can hie traopel

nI I"~44 -Jj;. 'Hy i. iont i t yiti ti s; nssiq naiflo 'cte .ini takini stops t

r- i qI ' t 14-4 Ani t irmi ty A tne construct 1 )n pr )nfl-sp -an he improve I.

4%%

I'y i.tI r i'ls.-; i -ia I. ri te.ne mt;*i nM if'rn~

r I t

I Si iiiW, t ~
7 I 'It'. ' I 1. ~t'~w' it 40, 01 n~ i~e ''1171;.'5 i Wiy 'noi

1 '1 1 v I4 I
 regularity. Therefore, when the process is out-of-control an effort is made

to find assignable causes.

Suppose that samples of fixed size n (= number of tests) are taken from

lifts being placed in a compacted fill. From each of the n tests a measured

value of some soil property results. Prom these n values certain statistics

are calculated, for example the sample mean mx=(i/n)Zxi, standard deviation

sx=V(1/nI) (ni-mx) , or range rx=(xmax-xmin). Being sample results, these

statistics will be subject to fluctuation from one sample of n to another.

However, if the variations are due to chance causes, the frequency

distributions of the sample mean, standard deviation, range, or other

statistics are known to follow the reqk.lar distributional forms discussed in

Part I[. Eor example, the sample mean m× is known to have a frequency

li triut,)n In repeati-e s;implinq whi h is approximately normal (exactly normal

if tie ioil pr,)pert is reinq; tested art- themselves Normally distributed). The

.vI'r iT,. vi . it te 'mple me.nr in, -,u Is the real mean mx' , while the

It n r it 1-'Vm 'vi It lia M,

mI l n. Y(xn Table 4, o)nly ().2* of the

' u 'ii'-i, 's' ;~ it t' III ;'mpie jut., , t. t r -xinpl-, the ;.im1lP.

T , a;- . i I ' I IIi TI, I T " 1 ho rl W h It I im * r

44q
r '
.5 i
"''' ] i"',' I l ) e 1' t ' , 'I | '"I e*T t ' ] ''[ ' W " ' 1 'i i T I: , 'r~ - d

5-
, 'S.,,,, ' "5, ' . , ' I ,' i " t ' ' * * ! . , l . . } ) I * , ] f .'
I
 Control Charts and Control Limits

A control chart is a device by which the state of statistical control

(i.e., that a process is in-control) is operationally defined. It is used to

attain control in a new process, and check that control is maintained in an

ongoing process.

A control chart is constructed by plotting values of mix, sx rx or other

sample statistics as a function of time or of some other dimension for ordering

sample results (e.g., lift sequence number). The sample statistics are plotted

against the vertical axis, time or other dimension against the horizontal axis.

A horizontal line is drawn through the actual mean mix' , which could be fixed by

specification or calculated from dat.. Two other horizontal lines are drawn,

one above mx and one below nix' , showing limits which are highly likely to

contain the sample results. These are the control limi ts: the upper control

limit (tCL) and the lower control limit (LCL). Fig. 16 shows a typical control

chart for individual compaction data.

If sample values are plottedi foi a substantial range of produ:'tion and

timfe, .Ind if all these valties fal1. within the interval tormeli by the ULK21and

LCl-,, "nd if tnit, la L show nio cycles or runs, then it is conc llded that the -

-i)ns5triict in pr. cess is in-control for that particuilar It tr ihote. if the ia ta

A,, n't -7i)nfo rmtn tb s pa t torn, then the -onc lusion i s irawri tha t %var
iahr iity

i~ ~ iitit.~~~rAith t 1 -i not xpkiiaile- by chacet(-rs a! no iri an

A I P1 l . ',l I .i If 1t..;1,0 ;rri .' ~n w i* 4 I] f '

i t* I IT 1 !1, ia ri -o t
fa r ;it i ihi I i ty , t he~i t 'ite 1 t h'.;1 I

7% % 1
 ',,

random variation is reasonable and accepted. If the results would be

improbable based on the hypothesis of random variation, then that hypothesis is %

rejected. The choice of control limits, and the associated probability of

their being exceeded, is arbitrary. For example, the probabililty could be set
.4

at 0.01, implying control limits of ± 1.65 standard deviations; or at 0.002,

implying control limits of ± 3.0 standard deviations. Narrowing the control

limits means increasing the risk that the hypothesis of random variation will
j-,

be rejected when actually it is true. For example, with limits set at ± 3

deviations there is a chance of 0.002 that a process actually in control will

fixed at (mx'+3Sm), while a lower control limit (LCL) is fixed at (mx -3sm)

Usually the standard deviation of mx, that is, s , is estimated from the data
mx

as s =sx/Vn. Sometimes sx itself is specified as a target homogeneity.

Fig 17 shows an m-chart for compaction control data on a dam project.

To control current production, a sample of size n is taken periodically

from material placed in the fill and the average m x of the n test results is

plotted on the control chart. If all the mx lie within the UCL and LCL, the

construction process is concluded to be in-control. If any m x falls outside

either the UCL or LCL the process is deemed to be out-of-control. When the

deviation outside one of the control limits is adverse, for example, when mean
.A
compacted density falls below the LCL, specific cause for the variations are

looked for with the intent of improving the construction proces- and thus the

product of that process. When the deviation beyond a control limit occurs on

the favorible side, for example, when the mean compacted density ex-eeis the

I,.' 4,
 X. O aVll _V V -V -. -' 7 -. _. ly I-13 3

, %,

UCL, either no action is taken or the causes of this unusually high quality are

searched for in order to learn how to permanently improve quality.

Probabilities of individual sample means exceeding the UCL or LCL can be

found by reference to Table 4. For soil property data which are themselves

mx - mx
Normally distributed the probabilities from Table 4 are exact for ,
x

For soil property data that are not Normally distributed--presuming that the

distributions are not bizarre--the probabilities of Table 4 are still

approximately correct even for sample sizes as small as 3.

'.4.

Control Chart for Sample Range r.

A control chart on the sample range,

rx = Xmax - Xmin -30-

shows variation in the range as a function of time. The central line on an R-

chart is fixed at the empirical average range in past production, or in special

circumstances is set by specification on acceptable variability of the

compacted fill. The control limits are usually set at t 3s., in which sr is

the standard deviation of the sample range. Both the average range and the

standard deviation of the range can also be related to the standard deviation

of the soil properties being samples s x . Fig. 18 shows an R-chart for

ompaction control.

%:!
70o"
* - . N .,.
L

If data fall inside the UCL and LCL on an R-chart, the construction

process is deemed to be in-control with respect to homogeneity. When a single

data point falls outside the UCL or LCL the process is deemed to be out-of-

In the latter case actions are taken to -

control with respect to homogeneity.

find assignable causes. A sample result above the UCL is usually considered
.J,

adverse and efforts should be made to find out the cause of the variability and

fix it. A sample result below the LCL's usually considered favorable and

efforts can be made to find out what is being done so well so that the

construction process can be improved.

Because an m-chart and an R-chart control for different aspects of

quality, a process may be in-control on one but out-of-control on the other.

An m-chart controls for the mean or average quality of the compacted fill. An

R-chart controls for the uniformity with which compacted materials

are being placed. Compacted soils may be on average sufficiently dense, but

unacceptabely heterogeneous. On the other hand, the soils may be on average

sufficiently uniform but unacceptably loose.

Cumulative Reject and Related Charts

Unlike most industrial applications of quality control charts;, construc-

- tion involves a single project with a clearly identified beqinninq avi fli. As

a result, certain quality control charts are very useful in constr1i tin even

though they are not wi eIy usedl in the fa(ctory. Ono of these,; iz; the ciimi lativo

.1 reject 'hiart.

*5 I € , , - . - .. : - ' , , ., . . - ' . ., - . , - . " . ' . ' . . ' .. . ,o .- . . .

ed.
- - V

Cumulative Reject Chart

The cumulative reject chart plots the cumulative numhers of t- ;t- 1 ,1 "n

results outsiie speci tied limits, against time test sequence number, "r a

similar indicat-r of test order. Fig. 19 snows cumulativo reje 't itit)r i

compaction inspection progrn on the imperious core of a r:)ck fill lam. T-'"

upper figure (a) shows cumulative rejects due to inadequat- den; ties. Thl e-

middle (b) and bottom (c) figures show numerical values f wat,r -7" tent jii,

dry density, respectively, for the rejected tests. These are plot t,- I I)n I
J K-
with the cumu a tive reject test so that the cause of ri :'t ,r i,' 11 r-% ti

the cause can he readily seen.

In Fig. 19 cuimilative reject is plotti ajainst te-t s-iw-n - n-i1ber.

a result, the slope of the curve qives the rate of rele.,-ts at i- i, int !liarl a

the project. in all, 1175 inspection tests we.re mad.e -)r th- irper! 14 - ., ,

which 38 were re] ecteA -ither for being outsi,e 3 Pr W't.er ;,t-
1T1aA - ,-!

content or t )r hba-r< iry P,'na:;i ty lessi than 9"t* Pro,-tor r)t 1mi-u. Io 't

rejects f t tui ntire pr-)'o--'t _ 38/1175 3%.

While the a]ve'rt..- r it,? ,) re]ject tests was: , f ,r the onlt ]2' -c *i"" ,

"A I1
A( 1, 1 - dly sh,1-,':-
)t Cj str ttjn
ic the rite, w; 1,,,i ±1 1!

I~~, %11ri,
a*a' %
-er ' It t . +~I1
wit~s-)5 v I S-* .

r I ,''. li:. ' -2.' r +'r, *"1;',

a- "r--li;' n 1'' .:+ . ' +'
- '' S

"'p

• " .-.
. . - .. .,."
- . -- . - . .- -- . -. , . . . . . . . - . . . . .. . . .- . . . . - ,
 Ir

control. The mean quality was low and there was considerable variability in

compaction properties. As construction progressed the process was brouqht into

tighter control. The mean quality was better maintained and the variability

reduced.

The cumiilative reject chart can be used to monitor a number of subtle

changjes in the construction process. Fig. 20 shows schemacically the effect )f

a ma'o.-r -han(ge in the construction process, for example, a change of

* ontrac:t r, Chadngef of ?quixnerit, or change of borrow material. Th e change

cau.se s i reaik in the smuth progression of the learning curve, usually

stiriugin t-~rlealrning cycle.

.i~tiIn~-t1 t Fiqg. 119 showci two such oreaks. The first oc-curs at

~L~.9r . Te rec rate for tests 1 to 50 is about q*. Pr,)

a: "J:J-rknect ra.~increas;es sharply to ahout 0* In f ,t,

t t rni- rit., of re 1,,,ts shoul he oxpocteoi I,- -ereae

I
'*.i'::~i .. 11: Pr )"~I",i. k ,tjj~.t~ analysi-s -,hl-z *'ia

-1 - I J. 1 n, r-~r i'wr i r' w

cl ~ ~ ~ ~ ~ 1 -; or 1Virxr
n

%6
 -.

e.
Moving Average, StandJarl DeviatioDn andi Ranqe Chart-,

Convenient accesso)ries to the cumulative reject chart are t- -oo:

average, moving standard deviation and moving range charts. These pr PO

smoothed information on changes in constr-iction outplut fro whih tr;.is'h

more easily identified.

In a standard m-chart the averages of samples of n tests are plotted as a

function of time or some other ordering index. A different set of n tests is

used for each point, and the assixnption is made that each test is independent"

of every other. Thus, each plotted sample mean m x is also independent -f the .-

:sample means of adjacent to it, presuming that the construction process is

in-control and operating in a random manner (Note: in practice the problem of

serial correlation in the construction process itself sometimes arises, but

-1h autocorrelation is beyond the scope of the present report). The use of

- :,rts typically presumes that many data are being collected and that

: r : ti )n output is fairly high.

-"r :-'ty cases in construction the rate of testing is more modest or

.. r>eeis more slowly. Often a considerable time is required to

-viiL lifts that are to be tested. In these cases a moving

* ,r" m- more convenient than the standard m-chart. The moving S

;trw ted in the same way as an m-chart, but it provides a

n w the construction process may be changing.

, r-i7olat-i) over windows of fixed size n. For a lono •

,...:k), an averaeg is calculated for the first n

" -"i is; calcnulated for the next n iati

%f- e .

0i
• °* * Qo ." *°- . *o• * .~ °- . . o ° . • ° j • . ° . • ° %
 I.-.

S..

, .
"
P! + 'I1''
. **** .; "II+
*+i
-.

, - - ij* , £
*

•+ + "
. +, .
tr ' '

#
: .
. ..
. ., 7.+++.
!

. . . ... o
. . .
7.-io-
' r

. . ,- - .+ ..
v +++ '++ . .,'.
. >_- -.-
.. -... .., ...., ,..-_.-.+. . , ,+-,
+.-.,
_..--.-....
. -..-.. + .+
+ ....-..-....-.
+-..

" I
•t" 4 " +' -
"
-- "1ft +r " I "IJlI 2
 r.

S.
4..

4..

4.
9. . . . 4, .4 . . * 4 '3 .3 3'

4..
9.

9.
* 4 3 4 3
-l .4. -

'C 4'3 .. '' T .4 .. 32 C

- ~4 4 4~ 3'4~ 3 '44 3444' . '. f'S]"

- 34 .~ :.

4/ 4.4
3 .4.

.3
Nd .* - 4 *- *- - .'.~ !4-;I*'..' 4 ' - 4
4.1
4. .4
434' 4.

3 44

'4.

- i . r' , . . 4 44~34 44 - . * 3 . 4.4

*344~~ 444 -~ ~4 4 -, 1

-'5

.4.4. 4~4~ 4*: ,' '. -

* - , 1 4444 'j I $ '4 4 '.' -3~ 1..'

34

.4.;
*- *2~1 4 .. - - .r. r

* 4 -44 4 4 .. 4 4 .1* 44 * I,: ,'.~, 33 4*3 , .' 4'

* . 4.. .' 4 3 4 . . 'W'~' '~ 4 44. 44 4

- , . .. 4 - 2 ~4434444~444 . . 4 . , 4

* . , . .. . I 1, 44 , 4

.34

3*4, ' 4 -i. .2 *~'' ., .4

.4
t4
.4 4 . 2 . *. B 4 4 . 4

*34

S
1

4,
.44

* 43~

- ~ :j.. . . . - 4 .444% 444

w~f~tttt4SV../. ' /% . f...-A..r.... 2~..X..%.A....2...'.S§ *. .-. - *-v.-*.. -- .. . .. .a .. a..a .a.a
 distribution, as illustrated in the above example, may indicate a change in the

construction process which needs to be monitored or controlled.

Cumulative Sum (CUSUM) Chart a,

Changes in the output of a construction process are sometimes more quickly

detec---A by monitoring the change from one test to the next rather than the

absolute value ot the test. For example the changed conditions which appear in

-F',s. I,) and 23 become apparent earlier when increments of test results are

pi.)ttedi.

he sos ti ,ifnmon way til moni t,)r in-r(emnts is by the cumulative sum or

cusn :htrt. Tie cusutm ,artit is-es tr.nd: tn the (,I data to) identify process
.

Sharj 's, ri tier titian tireatirig the dita frixn each lift by themse 1v# 'Te
i. ma or

I.aivirt1 i'_i f' :. t chart; ;v,,r m- )r r-charts (sornetimpi Sntwhart ('harts after

- ti-, itt it'ii- , wh,)

qi t irst- ,r, st ttirien) ire th, i 'h,inqes ire- ieontiffi-, more '

" lii :kly, ;,,irtr:ulrly m,,,i.';t -hiip ;, .n,i that the tir,, )r loatin of thf"

* ~~ ~ ' 'l Po. :III I).' Tq -. !r'ij ?

1% it -1 - •I < x1 ;1 . . h

' " a 1* 1i ' I t I,., ' < :1, J,. r t'• "I "X I WI I

L* l I
1,i ' . ," ,Iaf *i . * P ' I' .1,
i 'T} ! .' A 'i ,1 *' ' ' 1 inn it

1
t :* t. " >,* , , ,

%• %

". - , " . . " ". -- " • . . ."'- - . - . . .' ",,;. --.-. \ 'V. .'-' '.a',,*
 As iorn ; .ii the co s tic t ion )r,_)C-,; i-I n -n t r -

should lie on a straiqht Iline. - ome indivi-itir 1 1XI 4L I

in wh i : h ca se the Ii,; tini 11r .'t- w I II r e in-i S'hfl 1 iL)

,,v I Ci x, 4'

M , ln whi,-h -,Ise tnIp -uI-irl -tir-vo wi 1 t 1! 1. 7*1,, . . '

*wand'er Up mln i'mwn Ihoyjt neZ.rj ) xi S.

i.inre mxLsi Il-eiilY fl')t- known In I'iti,.-'~

I- t I ii ' I i ht~y ~ - th , , njfI- i*.

- x -v'!:- in tr it -~l m n -ii,i vitr~e m h y Iti 1 t " '*i,,

enr.tto It t _r r i

Iit I I II it ti I t t r.rw wI tw.I ir' ~I.A T i. I I .

I-~~~~ t * ,' in I"

It~ J. I . I tr t~r*

r- On

t :w 'fl 4

i v r71 1,x 1), jv-.11

iS -

WIt SIin,.r 1t -- t -t '1 ~ 1

/-. -. ~
S* . * :* . * ,,-; * . I.. .* . t .
 ..

I
,!1 l* i ', i . i t --. i ' i

:-.4

e aq

- , ., ..

. * . . . * * , .. , .l I

.4-

I - I . 4 . , .

I t .'

* .- ,, , , I *I I I

.- , i - . - I

-. o4

-- . -..
, ... , ..
. ..- ., ... • • .. -,'- - 4.. .- .- .-.. • . -.. .. . -
 -S.
4.

-a

'Rh
sq..
-'5

*4,
'4

-4-

-5-

-p

~~
C

"p
'I

* A.

< h&S~&.t. Crt-(&ACCt~,'~xK' * - A -~ -- . -- -. -. A. - -

~fl2cflA~t.S.S ;,m%~. . A -~ \
 'a

*
, -I

i 4. . .. . . 4: ' - -, .t .,,- . ., 4 4-.k

I. - - " i -.

'"c I">.C' .

'.. , C ,.- , i

L." *°.4

V'e 4 a. * .. 4

- - . -,, .,
-. ,- , -. . . .,-,, . . . - ++, '+. . . • . - . ,, .. , - .. .. .. + ", , ' ,+ ' . - ,. ,. -. _ . . .. - . .+.-. . . +---
°
,,P
+ .. ".. .
" i",
. ,, , ,,,,,
_ ',, ,i"
+ ,, . i" +" ' d _. -" . '_ '_ ,,e _.l _,," . _ . .' "." +" . -',,o' - ., ' , . ,". . .,,i' ,,-
 " L

,
I.t?
w, t

%% + PU 14'+
. jii @ +++- --
S.

withit rt ' IAw

.
,%I

Fi~ur 1 har
Sipwrtcotrl (-(-Iir fr omacion contrl o i d"iI

-+%

F Iqure 17 She+war t cotitro I chart (rn-thar t) for ,:ompact ion Gont rol OnP a dam
*, --
project.

d l l m l i m m m
5'."J., "W't P."*€£' +~ ':+l .. + h~~~~mmllll~ldllll~
 .1', S

SII

5,.j

4 4'

-u I , ,

+I*'+ .' + +
+ +
I. I +' + I I 1

-II* , , .'

*,t"i,, I r2' ,

-A

44"l? t " ^ I,1 I l r - i

,oI.

lr+ i o+¢ t..

Ia

1 + °
 T - %

20A

4 -

2/

.4

aT 1T I N '!IV: ( I I -S

4 __0

t t r- ! u_ __ __t_

-4%
 %
i

'.-S

e-

|I

*
second cycle curve

change
'"
Cumulative %
Number of
* Rejects

--..
tyia leanin curve
a.°

Test Sequency Number

PN
-,%

4":

Fiqure 2Q,-- Change in construction operation as reflected in cunulitiv(e reject

chart (schematic). """

i~.
I
 4.-

40

+t

%10~

Q 4" +4

+++

+
20
* -

0g
10"

0tII I I

0 400
4400O 800 1000 20.-,

TOTAL NUMBER OF TEST RESLLTS

F i gure 201) (Aconin 11 s t ruct I I o'p t Ii.h 4'! It& t tI, i :I i

'ICtul LI III l t i VC rt. ('Cvt lrt

I * . '.,.
.4
4'
.4

a.

-4
4%
-4
4%

N.

,44

'I

-- 4
N

I 4 - ' 4 4' ~.s2LaAahaSa\kYAak

 Al"4 794 STATISTICAL OtONLITY CONTRO OF ENfEDD ENSMKNTS 2
(u) mXUS NSSOCIRTES IMLN M 0 0 DECHE* SEP607
AES/CVOL-9?-2 DAC3-3-H-S947
LOCLAS S SIFIEPFA13/2 L
 I .'I IIIIII
1111IL2111111.4 HI~i

Lot-
. .. % 4
. J,-%,
P. .%.%% .. . ,_. - ... \ .% * *.. ... *.% -. * ... % %*.~ .- % . ,"% . o %% .% -% % . , -
'., . %
' . . ,,,".
• '. ,,,".
-,. . . .- . ,'. ', .% .". ,% .' . ". .% . . ,,- *,. .. . ,,. . ,,. , .. ,, - .. % . % % " .. ',. ,, ', ' % ,,, '.,, %. • %.V., %,
4 ,
V) 2

litt

'I
.4

>

44

CUMMULAT]VE NUMBER CF TESTS llz

Figure 22 -- Moving standard deviation chart of water content.

. 94 .
 + 0

0
44
++4)

(S))

C)1'

+ tz 4-'

a) 4-)
.,A

4 J 0

+* CD (
++11* O-4 u 3
+

0- CD 0

95,,.

.r. %le
 Ih1 AL N br f)D'IA
L

* O'LIER SI '.
C OLALIER F
'i~fLr ;01iG.1 *. Ljr~I A 0
6 -- I()

,',

60

20-

0 •

0.9 j ILI
e 9 Va F,, 4
D /

9
09

0 7
bi
0r /

>J 05

1 %e
_ __
0 3

0 2

-,.,.

4
6

WATER CONTENT (% Opt)

Figure 24 -- Distribution of compaction water content data for tests 1 through

459.

96

%l I

,
.-
 T O T A L r ;u-
ER OF DA 'I - 40

8 I
IN O26 L19
0. -10 DEV .0997

i
' - R O OLTL IERS BE OND LOLER LJNIW ].OF PLt)T - E cTs
N',t'BLR OF OUTLIERS BEiOND LEPER LIMIT OF PLOT - 0

'7 I
-

,qb

0 4-

03 -+

088
r3 I A.
075

4 09

%6~/I +-PERCENT WATER CONTENT

%
L K0N.
Figure 25 -Distribution of compaction water content data
for tests 460",
through 884.

•.
97
,

',.
 T
OT7AL :FOATA

rEAN ~Ec.2 JK .'-

4 %0

440

LU I

0
P,%

05

4, -0

0 2k

0 L

0O1
+
+'t*/
~ '
j '

0.00"

Figure 26 -- Distribution of compaction water content data for tests 885

through 1175.

98

%*.

P.4
 JJJ

-4

C~dU
a,.~%~ - * ~ %~ C . . .*'a.
. . .-. . . .4 M ~ - -. . . . . . . . . . . . . . . .-
d* * a. a
4 P
 PART VI: QUALITY ASSURANCE BY ACCEPTANCE SAMPLING -

The purpose of quality assurance is to test fill as it is placed and to

make decisions on whether to accept or reject the fill as conforming to

standards. If the fill is rejected, further compaction could be made, the fill

could be removed, or some other course of action could be followed pursuant to

contractual arrangements between contractor and owner. Acceptance sampling is

the quantitative tool used to make the accept/reject decision. The objective

of acceptance sampling is not to control quality, but to make decisions.

Structure of An Acceptance Sampling Plan

A simple acceptance sampling plan is structured in the following way:

I. A random sample of size n is taken from the materials being

tested.

II. The results of the n measurements (xl,...,xn) are summarized

statistically in an index z. For example, z might be the sample
average (1/n) xi .

III. The index z is compared to a critical value z*, and if z lies on

the correct side of z* the materials are accepted as
satisfactory.

The questions in designing an acceptance sampling plan are how large to make

the sample size n, how to summarize the resulting data in an index z, and how

to select a critical value z* such that quality is assured without unduely

increasing the cost of construction.

The more stringent the acceptance criteria become, the greater the

likelihood of rejecting fill which is in fact satisfactory. The less -

stringent, the greater the likelihood of accepting fill that is in fact not

satisfactory. The problem of acceptance sampling is that, for a given qize

100
%0. .',,'-[
.:"
 sample, reducing the likelihood of accepting poor material usually means |

increasing the likelihood of rejecting good materials, and vice versa. To

simultaneously reduce both the likelihood of accepting poor materials and the

likelihood of rejecting good materials, the sampling plan must be made more

discriminating. This usually increases inspection cost. N.

Buyer's Risk and Seller's Risk

In specifying an acceptance sampling scheme two risks are balanced, 4

-q (a) The owner's (buyer's) risk of accepting material of poor quality,

and,

(b) The contractor's (seller's) risk of rejecting material of good

quality.

J Decreasing one of these risks typically increases the other.

Test results from an acceptance sampling program are variable whether the

fill is truly of acceptable quality or not. Because of this variability, it

may be, for example, that the lowest compaction test results on an acceptable

fill give lower dry densities than the highest test results on an unacceptable
fill.
I
"
The top of Figure 28 shows a hypothetical frequency distribution of test

results taken from an acceptable fill. Suppose that the criterion for

accepting the fill as meeting specification is that test results be above yd*.
I
Because test results are always variable, some fraction of the tests results

will always fall below the acceptance criterion and thus lead to rejection,

even though the fill might in fact be acceptable.

a
This fraction is

. proportional to the area under the frequency distribution to the left of yd•.

The probability of the test result lying beneath Yd*, and therefore the

101

Nr
 1

probability of improperly rejecting an acceptable fill, is called the

seller's risk.

The bottom of Figure 28 shows a hypothetical frequency distribution of

test results taken fron an unacceptable fill. Some fraction of these test
results will always fall above the acceptance criterion Yd* and thus lead to

the fill being accepted when in fact it should be rejected. This fraction is

proportional to the area under the frequency distribution to the right of Yd •

The probability of the test result lying above Yd and therefore leading to
>'•
acceptance of an unacceptable fill is called the buyer's risk.

For a fixed sampling plan there is an explicit trade off between the

buyer's risk and the seller's risk in selecting the acceptance criterion d•.

Higher values of Yd reduce the buyer's risk but raise the seller's risk; lower

value of Yd raise the buyer's risk but lower the seller's risk. This trade

off can be seen in Fig. 28. %%%

The buyer's risk and the seller's risk can be controlled simultaneously

only by making changes in the sampling plan, not just in the acceptance

criterion. The purpose of statistical acceptance sampling is to allow the e.,

buyer's risk and seller's risk to be quantitatively determined for a given

sampling plan and to be appropriately balanced by designing the sampling plan.

Inspecting for Fraction Defective vs. Inspecting for the Mean

Acceptance sampling typically addresses one or both of two aspects of

quali ty:

(a) The average property of the fill, that is the mean; or,

(b) The fraction of individual values within a fill which are below some
standard, that is, the fraction defective.

102

%
Each aspect of quality may not have the same importance in a particular

application. For example, the potential for internal erosion of a fill depends

on soil densities at the least compacted places. Conversely, the strength of a

fill to resist large slope instabilities more often depends on average soil

densities. Acceptance sampling plans differ depending on which aspect of

quality is to be assured.

Operating Characteristic Curves

The functional properties of an acceptance sampling plan are usually

summarized by an operating characteristic or OC curve. The operating

characteristic relates the quality of the fill being sampled--for example its

mean density or the fraction of the fill with out-of-specification water

content--to the frequency with which the sampling plan leads to a decision to

accept. As in Fig. 29, the horizontal axis usually shows the actual fill

quality, while the vertical axis shows the probability of acceptance. The

Buyer's risk and Seller's risk are read directly from the OC curve

corresponding to the definition of good quality and poor quality materials.

For example, the probabilities corresponding to the two frequency distributions

of Fig. 28 are shown as the Buyer's and Seller's risk, respectively, on Fig.

29. In principle, the better the acceptance sampling plan, the steeper the OC

curve in the vicinity of the contract-specified quality of the fill. A steep

OC curve reduces both the Buyer's risk and the Seller's risk.

The shape of the 00 curve depends on the desiqn of the acceptance sampling

plan, and can be used to make economic decisions about the reasonable extent

and cost of sampling. Usually, the easiest wdy t,) steepen the 0(' is by
!S

)1 3
.. "

%! %%
 N, lvN
%

increasing the sample size and thus sampling cost. The remainder of Part VI

discusses the relation between a sampling plan and its corresponding operating

characteristic, and how a sampling plan can be designed to achieve a desired -e

OC curve.

Acceptance Sampling to Give Assurance on the Mean

This section considers acceptance sampling plans the intent of which is to

assure that average properties of placed materials meet specification. Two

types of specification are considered, single limits and double limits. Using.

single Limits the concern is that the average properties are, for example,

greater than some specified value. For instance, average compacted dry density '

is to be greater than 95% standard or modified Proctor optimum. rising double

limits the concern is that the average properties are between two values. For

example, average compaction water content is to be within ±2% standard or

modified Procter optimum.

The sub-sections first consider the case of known or specified material

variability, that is, known standard deviation. This case is mathematically

easier than the more general case of unknown variability, and does sometimes

occur in practice. The more general case of unknown variability is treated

afterward.

Single Limi t, Standard Deviation Known

Suppose specifications (-all for soUi with an averagle or mean compacted dry 5

density of mx = 120 pcf. Suppo,,te ilso that th, dry density of the compacted

fill is known to have a constant stani ri dviation of s.' 15 pcf. An

, ,,W

%0
 . acceptance sampling plan to give assurance regarding the mean is constructed "

' such that material actually having a mean of at least 120 pcf (i.e., good
...
material) will be rejected no more frequently than some fixed value a. As

before, a is the Seller's risk. Simultaneously the sampling plan is

constructed such that material whose mean is substantially less than 120 pcf -o

" (i.e., poor material) will be accepted no more frequently than some other fixed

-2value B.As before, 8is the Buyer's risk. For the acceptance sampling plan

,' to be operational, a specific definition of what is meant by "substantially

less" must be adopted. In Fig. 29, poor material is defined as being an

average density less than 110 pcf.

The procedure for acceptance sampling with one fixed limit on the mean is

the following

i. Take a random sample of n tests

2. From the results xi . . .x n calculate the mean

mx=(I/n)Ex i .
3. Compare m x with a specified acceptance value m*;

if m A m*, then accept

if m < m*, then reject.

The OC curve for a sampling plan regarding the mean shows the probability

of acceptance as a funiction of the true mean value of the material. Mx' as in

,,--",Fig.
,%. #- ". 29.
., Thee
. Ta a curve
OC r4 nd m is
s , constructed
-l by using
of" n tests, the . standarized
."•"•"-"- . , .. o -".. variable
. - Zm, . ",,"@ "

m m
x
z -•-31-
m
s / n
x

1W5
 The denomination in Eqn. 31 is the standard deviation of the sample mean

(1/n) Zxi over repeated samples (cf., Eqn. 25). That is, the denomination

expresses the variability one naturally expects among different sets of tests.

The numerator is the separation between the acceptance criterion in * and the

true average quality of the soil m x . e variable Zm is the number of

standard deviation separating m* for m x ', and thus can be used to calculate the

fraction of samples in which the deviation of the sample mean for m. is given

than m*-mx.

When the property being tested has a Normal frequency distribution, the

frequency distribution of zm over multiple samples is exactly Normal. Yet,

even when the property being tested does not have a Normal frequency

distribution, the frequency distribution of zm is still approximately Normal.

The probability of accepting material with actual mean m.' is found by

comparing zm with Table 4 to find the corresponding frequency with which a

standard Normal variable exceeds zm .

Consider a sampling plan with an acceptance mean m*=105 pcf and sample

size n=6. Under this plan, n=6 tests are made, the mean mx of the results is

calculated, and if m x ; 105 pcf the material is accepted. If mx < 105 pcf the

material is rejected.

The OC curve for this plan is calculated by computing the quantity zm in

Eqn. 31 and looking in Table 4 to find the probability of a standard Normal

variable having an absolute value larger than Zm . Because the standard Normal

distribution is symmetric about Z=0, the area under the distribution above +Z

is the same as the area under the curve below -Z. Fbr example, if the true

'4
* 106
•S.'

D7
 mean were m x ' = 120 pcf and sx ' = 15 pcf, then zm = (105 pcf -120 pcf)/(15

pcf//6) = -2.4. Thus, the probability of accepting good material with the

specified mean density 120 pcf equals the probability of a standard Normal

variable being algebraically greater than -2.4, that is, about 0.01. Other

points on an OC curve such as that in Fig. 29 are evaluated by substituting

corresponding values of mx' into Eqn. 31.

An acceptance sampling plan with regard to the mean is designed by

specifying a Seller's risk a and a Buyer's risk 8. The Seller's risk is the

probability of rejecting fill which in fact is of better quality than some

decided upon acceptable quality level (AQL), or "good" material. The Buyer's

risk is the probability of accepting fill which is in fact of poorer quality

than some decided upon unacceptable quality level (UQL), or "poor" material.

The AQL and UQL are engineering decisions and must be quantitatively specified

to give meaning to the notions of good and poor quality material. The sampling

plan is defined by a sample size n and an acceptance level m*. The procedure

to find (n,m*) is:

i. Specify

c = Seller's risk %
= Buyer's risk
ma = Acceptable quality level of mean (AQL)
mu = Unacceptable quality level of mean (IJQL)

2. Find standard Normal variables (Table 4) with frequencies of not

being exceeded equal to (1-a) and 8,

zl_, = standard Normal variable with frequency of not being

exceeded (1-a).

z = standard Normal variable with frequency of not being

exceeded 3.

107 1.

% %
"% % "
 J..
I

3. Write the two equations

4.

." m- m a

= -z1_ a + Sets Seller's risk -32-

~m -m u
4.4. u
=-z + Sets Buyer's risk -33-
sx Ip

4. Solve simultaneously for n and m*.

-i An example is shown in Plate 3 and Figure 29.

a. Are Compaction Data Normally Distributed?

,., Experience has shown that empirical data on water content and dry density

for compacted soils are often well approximated by Normal distributions.

Examples are shown in Fig. 30. Specific experimental data may on occasion be

better fit by distributions other than the Normal, but this is uncommon.

Actually, the empirical fact that the variability of soil properties is

often well approximated by the Normal frequency distribution is not surprising.

The Central Limit Theorem, one of the cornerstone of statistics (Benjamin and

Cornell, 1970), shows that when variability among data is caused by the

cumulative effect of a large number of small pertubations or errors, the

resulting frequencies of observations should exhibit a Normal distribution.

108

\ .9-
..
, .-"..
2

* 0*
 Ail- . - 77 47 W. V. C

A',

Presumably for this reason, Normal distributions are common across the broad

spectrum of experimental science. In Part V of this report, deviations of

observed frequency distributions from Normality were used to identify changes

in construction process and inspection procedures.

Single Limit, Standard Deviation Unknown

The development of an acceptance sampling plan to assure the mean when the
i%%

,%'
.4' standard deviation is unknown is similar to the case when the standard

deviation is known, except that the index zm involving the known standard

deviation sx' is replaced by an index t involving the sample standard deviation

4,. x•
sx*

The inspection sampling procedure is

1. Take a random sample of n tests.

2. Calculate the mean and standard deviation of the test

results,

mx = (1/n) Exi -34-

2
sx = (1/n-I) E(xi - mx) -35-

3. Evaluate the sample statistic

ma - mx
-36-
* t=
x/

,.
in which ma = AQL.

4. Fix an acceptance criterion t*;

if t > t* then accept,

4 if t < t* then reject.

eI
'109
 actual standard d(-viation. The vertical axis of the OC curve is
II the

probability of accepting the tested material. The horizontal axis is the

non-dimensional quantity X = (ma'-m x )/Sx' , which involves the specified AQL

and both the real mean and the real standard deviation cf., plot in S.W.. V

The OC curve is calculated in a manner similar to that when the standard

deviation is known, but whereas the variation of z m across different samples

can be approximated by a Normal frequency distribution, the variation of t

across different samples -- at least for small n (Say n > 20) -- is wider than

for z m and ust be approximated by a so-called student -t requency SW

distribution. The variation in t is wider than in zm because the sample 5.

standard deviation varies somewhat from sample to sample. As n larger,sets

. ~the variation of s x about sx' becomes smaller, and the student t distribution .'

bapproaches a Normal distribution., v a o

acrTo design an acceptance sampling plan for the mean when the standard tha

,.. deviation is unknown the procedure is: ie

,, 1. Specify, '-
to = Seller's risk s
= Buyer's nrisk

ma = Acceptable (mean) Quality Level (AQL) ..

Mu = Unacceptable (mean) Quality Level (UQL) sadr

2. Make a rough estimate of sx

=Byr srisk
m - m

a -S
3. Compute AX -38-
5
x

110

%-,%'..- ''..;.--. ... <...;. . *. ......-. '.

 I I
This is the value of A when the actual mean m x equals the UQL mu '.

4. Make a rough estimate of n from Fig. 32.

Material having the value of A from step 3 should be accepted
only a fraction of the time. Find n in Fig. 32 providing ,

probability of accepting material of quality X

5. Find the acceptance criterion t* corresponding to a frequency of

not being exceeded (1-a) from Table 6, using v = n-i degrees of
freedom.

6. Specify sampling plan by

n = sample size
t* = acceptance criterion

M
= x - Ma test statistic. -39-
sx I-

Plate 4 shows the design of a sampling plan for the same condition as in

Plate 3, but that the standard deviation is not known. The effect of not

knowing the standard deviation in this case is that the sample size must be

increased by one test, from 9 to 10, to obtain the same precision in the OC

curve.

Double Specification Limits, Standard Deviation Known

Certain material properties, as for example compaction water content, have

specification limits both above and below their target value. Soil moisture

should be within some ± interval of optimum, say, no wetter than +2% of

standard or modified optimum Proctor and no dryer than -2%. An acceptance

sampling plan with double specification limits intends to assure that a

1%1
1 N %
-.% I I*%
N -VL" VA VW W- V-.
WLXNX N:- J IT %W.W kW' XW--1.rVW* 1r wV-r1 W IF WW ','W ~~VW ~;~ViJ'~S* '

material property is within the defined interval.

An acceptance sampling plan with double limits is designed by specifying

"a
two acceptance bounds. If the sample average lies between these bounds, the

lift is accepted. If the sample average lies outside, the lift is rejected.

The bounds are chosen to conform to specific values of the Buyer's risk and the

Seller's risk.

If the variability of the soil properties as measured by the standard ,

deviation is known, then the variability of the sample average of n tests from

one sample to another is also known (i.e., sm=sx'//n). As before, if the soil

properties are assumed to have a Normal frequency distribution, the variation

of the sample average also has a Normal distribution. Even if the soil

properties are not Normally distributed, the distribution of the sample average

is usually still approximately Normal. .

Let the target value or acceptable quality level of the average soil

A
properties be ma. If indeed the average soil property is ma, the sample

average of n test results will vary about ma as shown in Fig. 33. This

sample-to-sample variability of mx is centered on ma and has standard deviation

Let mu* and mL* be the upper and lower acceptance limits on the sample

mean mx. If mx is greater than mu* or less than mL* the lift is rejected. The

Seller's risk a is the frequency with which the sample mean mx lies outside mu*

and mL* when in fact the true mean is ma. That is, the Seller's risk is the

shaded areas in under the frequency distribution of m x in Fig. 33. Each tail.4

area has frequency (i.e., probability) a/2,

112

%J
 Let UQLU and UQLL be the upper and lower unacceptable quality levels. If

the actual average soil property lies just outside the UQLU or UQLL, there is

still a chance that sample variability will allow the measured sample mean to

lie inside the range (mL*, mU*), and thus lead to the lift being improperly

accepted. This frequency is the Buyer's risk B. The sampling variability of

m x for two lifts which have true means equal to UQLU and UQLL are shown in Fig.

34. The frequency (i.e., probability) with which the sample mean from these

soils lies within hie interval (mL*, mu*) is shown by the shaded areas under
L5%

the respective frequency distributions. Each tail area equals 8, the Buyer's

risk.

To design an acceptance sampling plan on the mean with double

specification limits, two constraints must be satisfied, the Seller's risk and

the Buyer's risk. Two parameters can be adjusted, the sample size n and the

location of the acceptance limits mu* and mL*. The sample size controls

the width of the frequency distribution of mx, in that the standard deviation
of m x equals Sx'/ 1 ; while the limits mU* and mL* control where the frequency

distributions are cutoff to yield a and 3.

From Table 4, the tail area under a Normal frequency distribution can be

related to numbers of standard deviation on either side of the mean. Let zp be

the number of standard deviations below which the area under the Normal

frequency distribution is (1-p) (i.e., Zp is the standard Normal variable which

has probability p of not being exceeded). For example, from Table 4, zO.975 =

+ 1.96, and z 0 . 0 25 = -1.96. Then, Figs. 33 and 34 lead to four relationships

from which an acceptance sampling plan can be designed:

A

113

-. .~ . . . . . . ... .
 7W--V~,.~
V W VV v UVWWVW JW vvv V WIPI' U-V1W ULW~j* . ' t~W V J Y 'w w

in -m

mU m
a z1 -a/2 -40-

sx
) *
m -im
L az/2 -41-
a

m - UQLu m s
o ,
z-4 -42-
S //n
x

*r

mL -UQL z -43-

x V'

As an example, consider an inspection plan for compaction water content in

.

which the Seller's risk and Buyer's risk were set at a=0.05 and B=0.10,

respectively. The target value of average water content is Proctor optimum,

and intolerable deviation from the target has been decided to be t3% water

content. Assume that from project records the standard deviation were known to

be about 1.5%. For these conditions, Eqns. 40 to 43 become

114

N "ee'.F "P "

 m
U
-o0
-44-
1 .5/lV .96

mL - 0

-1.96 -45-
1 .5//n
m U* --- 3% 1.282 -46- a.8

,1. =5//n

mL (-3%)
1.5//n -47-
1.282
.41 .5/7

Solving the first two equations simultaneously gives,

a.mU u = - mL
ML • 48
-48-

Solving the first and third equation simultaneously gives,

n 1)(1.5) (1.96 + 1.282)] 2 = 2.62, -49-

or rounding off, n=3. Putting n=3 into the equation for Seller's risk gives

mu 1.7%, mL = -1.7%. Putting n=3 into the equation for Buyers risk gives

mu = 1.9% and mL = -1.9%. Choosing ±1.8% as the acceptance limits gives a

Seller's risk of a=0.38 (i.e., less than 5%) and a Buyer's risk of B=0.08

'p...."
"-',.A
v";.
.. ,",a." .".".".v ".. . ,'....." v . .. .' . . ,,. . . . ". ". .,,.".,".". ", -'

%
* (i.e., less than 10%). The OC curve for this plan is shown in Fig. 35.

*" Another example is given in Plate 5.

Double Specification Limits, Standard Deviation Unknown

When the standard deviation is unknown the procedure for specifying an

acceptance sampling plan is much the same as when the standard deviation is

known, except that the sample standard deviation sx replaces the known standard

deviation sx' in Eqns. 40 to 43, and the Student-t distribution (Table 6)

replaces the Normal distribution (Table 4).

As for the case of a single specification limit, the test statistic is,

m - m
x a
S /1 *
-50- ?

For sample sizes above about n=20 these modifications are unnecessary because

the sample standard deviation sx is sufficiently close to the actual standard

deviation s x

The inspection sampling procedure is,

1. Take a sample of size n

2. Calculate the mean and standard deviation of the test

results,

mx = (1/n)j x i -51-

(1/n-1 2
Sx= (xi- mx) 52

3. Evaluate the sample statistic

116

J%~~~~~~~~~~ * %%C. J, .%% ' -

 -. "
I

m - m
x a 5
sx/In -53-

4. Fix an acceptance criterion t*;

if Itl 4 t* then accept

if t > t* then reject.

As in the case of a single specification limit, an inspection sampling plan

with double limits is designed by specifying a sample size n and a criterion

t*. An initial guess at s x is made, and Fig. 32 is used to estimate a sample

size n based on the quantity,

JUQL - ma i -54-
s
x
in which ma is the target soil property and UQL is either the upper or lower

unacceptable quality level. This assumes that UQL[I and UQLL are symmetrically

placed about the target ma. See Duncan (1974) or Grant and Leavenworth (1972)

for asymmetric cases. Unacceptable materials at either the UQLU or UQLL should

be accepted only with frequency B. Thus, knowing the Buyer's risk 8 and the

number of standard deviations X separating the UQL from ma, an initial sample

size can be chosen from Fig. 32. Using UQL=±3%, ma=0%, and sx 1.5%, as

before, Fig. 32 leads to n 4.

The acceptance criterion t* is found from a table of the Student's-t

frequency distribution (Table 6). This table provides the frequencies with

which given values of the test statistic of Eqn. 50 are exceeded due to random

117

s,,
 sampling variations when in fact the soils being inspected are of target %

quality ma . Because both unacceptably high and unacceptably low values will be

rejected, the Seller's risk is the sum of the frequencies with which the test

index of Eqn. 50 lies above +t* and below -t*. Thus, t* is set so that the

tail areas on either side each have probability a/2. For the Student's-t

frequency distribution these tail areas depend on the sample size taken, though

the so-called degrees of freedom \)=n-1. For these conditions, Table 6 gives a

t* value of 3.25.

5. Specify Sampling Plan:

a) Take sample of 4 tests

b) Calculate sample mean m x and sample standard deviation s x
c) Calculate the test index

mx -0 -55-
s /14
x

d) If -3.18 < t < +3.18, then accept lift.

If t < -3.18 or t > +3.18, then reject.

S< Acceptance Sampling for Fraction Defective

The following section considers the case in which an inspection sampling

plan is employed to assure that the fraction of out-of-specification material

.9 in a compacted fill is within tolerable limits. Such plans are generally

' called acceptance sampling for fraction defective.

-A
If the soils data display a Normal or bell-shaped frequenc: iistrihution,

there is an exact mathematical relationship between the mean and rI

de%)iation of the data on the one hand, and the fraction defective on the other.

This is shown schematically in Fig. 36.

. - I"d : ' "II f

 Setting the lower limit of acceptable values or the UQL at

L = lower limit of acceptability

7
a standardized deviate ZL is defined as the number of standard deviations sx '

separating the mean of the data m x from the lower limit L,

m' - L
z = -56-
L
so
x

If the data are normally distributed, zL is uniquely related to the fraction

defective, as shown in Fig. 36. Numerical values of this relationship are

found in Table 4 or can be approximated by Eqn. 24. Fig. 36 illustrates that

the higher the mean and the lower the standard deviation, the lower the

fraction defective. An acceptance sampling plan for fraction defective is

structured in the following way:

1. Test a random sample of size n to obtain the data x1,...,xn.

2. From the results, calculate a sample mean mx, sample standard
deviation sx, and test index I
m -L
z -57-
5
x 0

Depending on the specific problem, the formula for z may vary

somewhat.

3. Compare the computed value of z with a critical value z*:

.,....

119 7S .

' 6
 If z ; z*, then accept.
If z < z*, then reject.

The choice of n and z* defines the performance properties of the sampling plan.

These parameters are usually chosen to satisfy specified levels of Buyer's risk

and Seller's risk.

Operating Characteristic for Fraction Defective Sampling

The operating characteristic or OC curve summarizes the discriminatory %

power of an acceptance sampling plan. The OC curve shows how the probability

of accepting a lift or other quantity of material varies as a function of the %

quality of the material being inspected. For plans aimed at fraction defective

the OC curve relates probability of acceptance to the fraction defective in the

lift.

Consider an acceptance sampling plan for percent compaction specified by:

n =5
L 95% maximum Proctor density
z* =1.645 (i.e., 5% of the soil less than 95% max. density).

Presume the standard deviation is known to be s.' = 2%. Under this plan 5

tests are made. The average of the tests m. is compared to L through z =

(mx - L)/Sx'. If z > 1.645 the material is accepted; if z < 1.645 the material

is rejected. The OC curve for this plan relates the probability of accepting

the material to the actual fraction of the lift compacted to less than 95%

Procter maximum.

For Normally distributed material with known standard deviation there is a

unique relation between the fraction defective and the mean. Fbr s x ' = 2% and

L = 95% Table 4 is used to find the following relations: '-

'1 20
! %
 Fraction defective

Mean
p'

mx '
0.05

98.3
0.10

97.6
0.15

97.1
0.20

96.7
0.25

96.4 j
7
Thus, the horizontal axis of the OC curve can be expressed either as actual

fraction defective or as actual mean.

For a given fraction defective or given mean, the probability of accepting

the material equals the probability that the test result z is greater than z =

1.645. This probability can be determined by noting that z is itself Normally

distributed. With L and sx' fixed, z depends only on the mean m x of the test %

results. When sampling from a Normally distributed population, the frequency

distribution of the sample mean is also Normal (Part II). Thus, the %

probability that z)1.645 is found by calculating the mean and standard

deviation of z and referring to Table 4.

The mean and standard deviation of z are found by the method described in

Part II,

m L
x -58- %
z
s
x%

sm sx //n
. .. 1/1n -59-
z I I

S ., ,-"

° %

- 1.21'.
:..

:..

.% . ., . .'. . •
 a .- -•

Table 4 is entered by calculating the number of standard deviation of z

separating mz from the acceptance criterion z*=1 .645. The corresponding number

on the vertical axis is the probability of rejection (i.e., the tail area of 6

the Normal curve, or Pr{z<z*}). The probability of accepting the material is

the complement of this number,

Pr{acceptingl = 1 - Pr{rejectingl. -60-

This procedure is illustrated in Plate 6.

The entire OC curve is found by calculating the probability of accepting

the material for various values of actual fraction defective. For the sampling

plan above, the full OC curve is shown in Fig. 37. If none of the material is

* defective the probability of accepting is 1.0, and as the actual fraction

defective increases (i.e., as the mean of the material decreases) the i

probability of accepting goes down.

Single Limit with Known Standard Deviation

The main question in designing an acceptance sampling plan is to decide

upon a sample size n and an acceptance criterion z*. These choices dictate how

the plan performs with respect to Buyer's risk and Seller's risk. Let the

probability of improperly accepting unsatisfactory material, the Buyer's risk

be,

Buyer's risk = -61- I

let the probability of improperly rejecting satisfactory material, the

Seller's risk be,-%

122"; ,,
 Seller's risk = e. -62-

An OC curve is defined by specifying two points through which it passes.

For this purpose the Buyer must specify a maximum fraction defective that he

considers tolerable and which would be accepted under the plan only some

fraction 8 of the time. This poor quality material as a fraction defective is

denoted Pu'• At the same time, a target or desired quality level is specified
" ~which would be accepted at least (1-a) fraction of the time. This good (i.e., ,.

acceptable) quality as a fraction defective is denoted Pa'- The OC curve can

be made to pass through the two points (Pa',1-a) and (pu ,8) by adjusting the

sample size n and acceptance criterion z*.

For example, consider that acceptable material has Pa =0. 01

fraction

defective and an unacceptable material has p '=0.10 fraction defective. To fix

the two points of the OC curve specified by the Buyer's risk and the Seller's

risk, the first task is to calculate the corresponding averages ma' and mu

which would give fractions defective of Pa'= 0 .0 1 and pu'=0.10, respectively.

From Table 4, the area under the Normal curve below -2.33 standard deviations

*from the mean equals 0.01, and the area below -1.28 standard deviation equals

0.10. Tius, an acceptable soil having pa'=0. 0 1 and standard deviation sx =2%

would ha;e a mean,

ma = L + 2.33 sx '
= 95% + 2.33(2%) -63-

""99.7%;

123

V
 1
%k,

10
and an unacceptable soil having pu'=0. would have a mean,

m u = L + 0.84 s x '
x-,
= 95% + 1.28(2%) -64-

= 97.6%. %

The test index z is calculated from Eqn. 56. Due to random sampling ,

variability, the value of z varies from one sample of n tests to another even

for the same soil. This sampling variability can be characterized by a mean m.

and standard deviation sz for each of the soils above. Specifically, for the

acceptable quality soil,

%
m - L
a 99.7-95
mz - = 2.35 -65-
sx 2

1/n
/ . -66-

For the unacceptable quality soil, m z = 1.28 and sz = 1n. These are the

means and standard deviations that the test statistic z would have if the

actual soil being tested were just at the edge of being acceptable or just at

the edge of being unacceptable, respectively.

p-

The Buyer's risk and Seller's risk specify target probabilities of

accepting the two types of soil above when using the acceptance sampling plan.

For acceptable soil Pr{z<z*} = a; for unacceptable soil Pr{z>z*}=B. This gives

two equations. Again from Table 4, for Pr{z<z*} = 0.05 the mean of z must

be 1.645 standard deviation larger than z*,

..''
12 4
1-4 e

~~~~~~~A
-P ~ -' ~ .- .
 mza - 1.645 Sza = -67-
-67- 6
2.33 - 1.645/Vn = z*

For Prfz)z*} = 0.10, the mean of z must be 1.28 standard deviations smaller 4...

than z*,

mzu + 1.28 szu = z*

-68-
1.28 + 1.28I/n = z*

Eqns. 67 and 68 are solved simultaneously to give,

n = 7.79 say, 8 -69-

z= 1.75 -70-

An example of the acceptance sampling plan is specified as shown in Plate 7.

The design of an acceptance sampling scheme may be accomplished more %

quickly by algebraically solving for the sample size n and acceptance criterion

z* Define,

z1_ = standardized Normal variable for which the

probability of not being exceeded (Table 6) is 1-a. -%

z1_- = standardized Normal variable for which the eS

probability of not being exceeded (Table 6) is 1- B. %
.4r

Za = standardized Normal variable for which the

probability of not being exceeded (Table 6) is 1-Pa'

Zu = standardized Normal variable for which the

probability of not being exceeded (Table 6) is 1-pu'

For a single criterion acceptance sampling plan having parameters (Pa',a) and

(Pu',B) the sample size and acceptance criterion are,

125

%0%
 z +z
z -z
a U

z zI +z z
, a u 1-a
1- -72-
z z +z
1
-a 1-B

To summarize, the procedure for designing a single limit acceptance

sampling plan is:

1. Select a Seller's risk a, and a Buyer's risk 9.

2. Select acceptable quality level Pa' and unacceptable quality level

%
Pu'"

3. Find values for standard Normal variables correspondinq to 1-a, 1-8,

(1-pa'), and ( 1 -Pu') probabilities of not being exceeded (zl_,, zia,
)
Za, zu

4. Calculate the sample size by

z _+ zi_% 2
n = ( 2 -73-
a
u

5. Calculate the acceptance criterion by

'z* za z 1-8 + z u z 1-a %.

= 74-"'
z~ -74z
,1 I-a 1-B

6. Plot the OC curve

,.. -.

0
'5%.1"

%~~~~~~~' .. .... ...

*-?

iSi
 Single Limit with Unknown Standard Deviation

Usually the standard deviation of the material property being tested is

unknown. The only information about the standard deviation comes from the data

themselves in the form of the sample standard deviation,

s = - (x. - m 2 75-
x n-i x

If the sample size is large (n>20), the sample standard deviation will be close

to the real standard deviation and the assumption of known standard deviation

can be made with neglible error. If the sample size is not large, a slight

modification to the foregoing procedure must be made.

When the standard deviation is unknown the quantity z is calculated using

the sample standard deviation sx

m - L
z = -76-
x 4%

",..

Whereas, when the standard deviation is known the quantity z has a Normal

distribution, when the sample standard deviation is substituted for the real S

standard deviation the calculated value of z has more variability. Now the

denominator as well as the numerator will vary from one sample to another. Th e

frequency distribution of z takes on the sliqhtly broader shape of the

. Student -t distribution.

~ .* .. *~ *.>. ... *- .* *
%
 ..

The procedure for designing sampling plans and calculating OC curves when

the standard deviation is unknown is the same as when the standard deviation is

known, with the exception that tables of the Student -t frequency distribution

rather than the Normal distribution are used. Areas under the Student -t

distribution for the standardized case of zero-mean and unit standard deviation

are given in Table 6. Note, unlike the Normal distribution, Student -t

depends on the sample size n. As n becomes large the shape of the Student -t

approaches the Normal distribution.

%f
Convenient approximations for sample size and acceptance criterion when

the standard deviation is unknown are (Wallis, 1947),

z z + z z

1-B u 1-a -77-

Zz +

n + Z* 2 / 2 ) z1
z 1-t +zI-B 2 -78-

a u %
.

Thus, when the standard deviation is not known a larger sample must be taken to
2
get the same OC curve. The sample size must be larger by the factor (1+z* /2).

The example of the previous section is recalculated in Plate 8, now

relaxing the assumption that the standard deviation is known. The OC curve can

be calculated approximately but acceptably by assuming z to be Normally

2 /2
distributed with a standard deviation equal to sx (1/n + (z* /2n)1 .• The

% %,

-° S

,.,-. .:f..
 approximate OC curve is shown in Fig. 39. Thus, to calculate the real fraction
defective p' corresponding to a given probability of acceptance q (i.e, to plot

the OC curve), first the corresponding standardized Normal deviation Zq is

taken from Table 4. Next, Zq is increased by the factor (1/n + (z*2 /2n))1/ 2
.

Then a corresponding Zp' is calculated as

*= z* - Zq (1/n + (z*2 /2n))l/ 2

-79-

Then Table 4 is used to determine p'. For example, in Plate 7, z* = 1.63 and

n =6.

Double Specification Limits

The preceding plan pertains to the case of one specification limit. For

example, dry density should be at least 95% standard or modified Procter

maximum. When deviations in either direction are important the plan must be

modified. For a lower limit of acceptability L and an upper limit U, the

minimum fraction defective occurs when m x lies halfway between L and U; that

is, when the limits are symmetric about the mean. In this case,

m x-,-80-
- L
Z=x
L

U -m
zU - - and .
xS

S.S

],_9 4."

'. 4

','-.
 li.

z - z U-L
z = -- 8 2-
2s
2 x

The fraction defective equals the area under the Normal curve outside ±z, or

twice the fraction defective read from Table 4. Note that z depends only on

the upper and lower limits U, L and on s x It does not depend on m x . Thus,

if s x is known, the first step is to assume that the acceptable fraction

defective is greater than the area under the Normal curve outside z. This may '
be done without sampling, and indicates whether the variability of the

construction process reif)-t,1 n is so large as to preclude any

possibility of the tpstod s,)il beinq founi acceptable.

Presuming that z from 91 . 4) is sufficiently large that rejection is not

inescapable, the fraction defective will depend on both zL and zu, and the

acceptance criterion must 6e based on both. In concept, this is done by

summing the fraction defective beneath L and the fraction defective above U and

comparing that su t the criterion M. However, a simpler procedure can be

eveloped by considering the operating characteristic curve of the sampling

plan.

Double Limits, Standard Deviation Known

The fraction defective for double specification limits is that proportion

of the area under the frequency distribution of the material property which

lies either below a lower specification limit,

L = lower specification limit,

,I,°t)
13 )
~.%.-
• %
 or above an upper specification limit,

U = upper specification limit.

For constant standard deviation the fraction defective is minimized when the

mean m' lies halfway between L and U. in this case,

-(L-m x ' )/sx ' (U-mx')s x ' = (U-L)/2sx' -83-

Thus, a quick check should be made to see whether a material can possibly meet

the fraction defective double specification standard by finding the area under

the Normal frequency distribution outside ± z = (U-L)/2Sx.

greater than the acceptable fraction defective Pa' no sampling plan alone will

assure quality. The construction process must be changed

If this area is

to make the material

I
more uniform and thus reduce sx'.

In the general case for double specification limits, an acceptance

sampling plan to assure fraction defective follows the following procedure:

1. Take a sample of size n

2. From the results, calculate the sample mean

mx = (1/u)Exi

3. Compute the quantities

m -L -84-
L
S
x
J-mx
-85-
x

F-
W"
 <A
1*

4. Specify an acceptance criterion z*: If zL z* an z z*,

then accePpt otherwise, reject.

The problem with double specification limits is determining z*. In the

case of single specification limit z* was determined from areas under the__
Normal frequency distribution to one side of a specification limit. In the

double specification case z* must he determined from the sun of the areas above .

U and below L. 1.7

Consider the problem of acceptance sampling for compaction water content.

The target value is Proctor optimum water content. Tne upper specification

limit is U = +2% optimum; the lower specification limit is L = -2%. Presuming

the standard deviation of water content to be 1%, the limits are udrt

U-L +2 - (-2) "''

±z= + - = = +2 -86-.
2s2(1)•.

That is L and U are 4 (i.e., +2) standard oeviations apart. From Table 4 the

area under the Normal curve beyond z=2 is 0.02. hus, the lowest possible

fraction defective would be twice 0.02 or about 4%. e fraction defective for

values of the mean other than that halfway between U and L are shown in Table-'-

8.

A i ,. -. %
Presume for sake of example that the acceptable quality level or AQL
U-L +2 (-2)
expressed as a fraction defectime were conte. That is, the lift would be

I ° ", °o
- . ". , "- %. -, ° ° . •. . . . ) ±2 . -86-•.
.. ° .- .
• °•", . _" " ._ '%
-
• " -,W " o , " • , . """ " •"'-)t of-"" the." soil
o • "had
. " a compaction
"" . " , " water
" , -" " ° . "2"•
'%' considered .", % - acceptable
, '._'j,,o " % % ,- .•.'it at% lpast
° ." ,' ' - , % ' . •* ,, - • • %,- - . % % %%x' %

is between
content U -2% Proopitmmr )t m and i o +2%. rom Table 8 (by
interpolating m.' value ) any lift with an ben
averasdt water content bet -0.7%ets
 lmw Ar lv a . V ' - T- . 1r r(VW'Vw r -W -A -r I
and +0.7% would be acceptable, for the fraction ofxJany of tliese lifts with

water contents outside ±2% would be less than 0.10. At m'

x -0.7% the

fraction below -2% is 0.096 and the fraction above +2% is 0.004. The sum isI

0.10. Similarly but in reverse at m~ +0.7%, the fraction above +2% is 0.096

and the fraction below -2% is 0.004. The double limit specification can thus

be met by combining two single limit tests designed such that the acceptable

fraction defective in each is reduced from Pa' = 0.10 to Pa 0.096. One

applies on the upper limit side, the other applies on the lower limit side.

The design for these two plans is exactly as discussed before, and is carried

out in Plate 9.

Double Limits, Standard Deviation unknown

The problem of designing an acceptance sampling plan for fraction

Sdefective with double specification limits and unknown standard deviation is

less easily solved than the single limit problem. In particular, with double

limits the shape of the D ,urve d-penls on how the fraction defective is split

between the upper and lower tail the distribution.

pf However, the

availability of statistical l ; ancy graphs designed expressly for the

Spurpose (US DOD Military tandiari 41J) simplifis the task. For the -ra

purpose of acceptannce sampling of engineered fills, the frrcphs

f Fio. 41 and

Fig. 42 providie sufficient cray

beweSh
case
The procedur
pe n
begins
oe ~l
Diyr' s and Se 11er' s rink

unaccept b le frac t on;hi d,-f-cti

a for i thoifi ;1noi
ftedsrbto.Hwvr
ai de

1),'
robe•

ani 1)
imi t,

w th
i
unknown

fr7
and ac

ain
stan ar deviation
h
epta

a r- t
lt and

ip t I
%

133
 - , '- - -- . - - -. '-

estimate a sample size n and an acceptance criterion z*. From the estimates of

n and z* the quantity

-87-
Sn-I
2 -

is calculated and used to enter the absicca of Fig. 41. On the ordinate, and

corresponding to the appropriate value of n, an allowable fraction defective M

is read.

The test procedure is implemented by taking a sample size n, calculating

the sample mean m. and sample standard deviation Sx, and then computing the

test indices

m -L
ZL
ZL =S x -88-
x
m
U- x -89-
zu
ZU - 5s'
s
x

From Fig. 42, for the appropriate value of n, estimated fractions defective

corresponding to ZL and zU are read of as PL and PU, respectively. These are %

summed to obtain an estimate of the total fraction defective, which is in turn

compared to M to decide whether to accept or reject the lift. If PL + PU 4 M,

then the lift is accepted; otherwise, the lift is rejected. An example is

given in Plate 10. The OC curve for this procedure is approximately the same

as in the single limit case, using the same values of c,8, Pa', and pu .

134 .I

,, .;. ,,
?,w ., +.,',.><'.
a.Y.; +
I-,-
. ,&'.. . . . ,+.A:., ,...,.I: ...-...% % , .- %-,....+
 Table 8 -- Fraction defective for double sampling limits.

L-m U-r
mx x xPL UP=PL+PU

+2.0 4.0 0.0 -0.50 0.50

+1.5 3.5 0.5 - 0.31 0.31
+1.0 3.0 1.0 0.001 0.16 0.17
d+0.5 2.5 1.5 0.006 0.07 0.086 .

0.0 2.0 2.0 0.02 0.02 0.04

-0.5 1.5 2.5 0.07 0.006 0.086
-1.0 1.0 3.0 0.16 0.001 0.17
-1.5 0.5 3.5 0.31 - 0.31
-2.0 0 4 0.50 -0.50

Note: range U-L is kept constant.

IP

4.Z
%I 1 35
 N.
PLATE 3

SUBJECT: Acceptance sampling plan to assure mean value of

compacted dry density; standard deviation known.

I. PROBLEM:

Design an acceptance sampling plan to assure the value of mean

compacted dry density of an engineered fill.

II. SOLUTION:

1. Parameters:

Seller's risk a = 0.05

Buyer's risk 8 = 0.10 _
AQL ma = 120 pcf
UQL mu = 110 pcf
Std. Dev. sx ' = 10 pcf

2. Find standard Normal variable corresponding to 1-a and 8:

z1_ a = z0.95 = +1.65

z0.10
__= = -1.28

3. Set Seller's risk and Buyer's risk:

m m
m - ma_ m-120 pcf Z = -. 65
10 pcf//n
s /Vn
x

u M-110 pcf 1.
s //n = 10 pcf/Vn 1..8
s /°'m

4. Solve simultaneously to obtain:

n = 8.6 + 9
m = 114 + 114 pcf. .

5. OC curve shown as Fig. 29.

N"% N.

%',
P°

136
 PLATE 4 SHEET 1/2

SUBJECT: Acceptance sampling plan to assure mean value of compacted dry

density; standard deviation unknown.

I. PROBLEM: Design acceptance sampling plan to assure mean value of compacted

dry density.

II. SOLUTION:

1. Specify,
" a = 0 .05
8 = 0.10
ma = AQL = 120 pcf
mu = UQL = 110 pcf

2. Estimate Standard Deviation

sx ' 10 pcf.

3. Estimate Sample Size

m -M
a u 120 - 110 pcf. =
' 10 pcf
x

For X=1.0 the probability of accepting should be 80.10. Fig. 32

shows that n=10 is the approximate sample size providing this
probability of accepting.

4. Find t*,

From Table 6 the value of t which is exceeded (1-a)= 0.95 fraction

of the time is t*=-1.83.

01
4..

R ,,,,-,,U.,:.. . .:...,....
.... . . ... . . .- . . . . . . . . 'o

'I. -S,. ',• ,'

"," .' " '.9 ,2 ' .•,", ...',. . ,, .-i .".-... - ' ' '. ."F .. "' '''''' ' . '
"

[p .
i, ,
I " " - '' '' #' '+ ' -
'
:" " j ' + ' ' ' ' *,
", P.. ' ,

•~
%]
 PLATE 4 SHEET 2/2

SUBJECT: Acceptance sampling plan to assure mean value of compacted dry

density; standard deviation unknown.

5. Specify acceptance sampling,

a) Take a sample of size n=10

b) Calculate mean of sample mx=(1/n)Yxi
Calculate standard deviation of

2
sample sx = (1/n-i) I(xi-mx)
c) compute quantity

m x- ma
t /n

-..

m 1.20Opcf

~J.
d) if t > -1 .83, then accept
t < -1.83, then reject

Q
b.

138

:%
 PLATE 5 SHEET 1/2 .
4'

SUBJECT: Design acceptance sampling plan to assure average compaction water

content within double specification limits; standard deviation is
known.
0-4
4'%'

I. PROBLEM:

Design acceptance sampling plan to assure average compaction

water content is within ±2% Procter optimum. Standard deviation is known
s
to be x 2%.

II. SOLUTION: -d

1. Specify:

a=0.05 Target water content = Procter optimum.

B=0.05 UQLU = +2% Procter optimum.
4' sx' = 2% UQLL = -2% Procter optimum.

2. Write equations for Seller's risk and Buyer's risk:

m* - m m* - o
U a zU-.9 2
+1.96
z1-a/
s //n 2/n
x

L a z L
= /2 + -1.96
s //n 2//n
x
M* UQL *-2
U U- z U
_ _, __ +--= - 1.645
S //n 2//n 'S
x*

m*- UOL m* - (-2%) -. '.645

L L z - 1.U645

s //n 2//n
x

X 5

4V
% V..

I%

"°
 PLATE 5 SHEET 1/2

SUBJECT: Design acceptance sampling plan to assure average compaction water

content within double specification limits; standard deviation is U

known. %

3. Solve equations simultaneously: V

a) from Fans. 1 and 2,

mu -mL*

b) from Eqns. I and 3,

n . 13

c) from Eqns. 1,2,3,4,

mu* = -mL* = 1.09

4. Specify sampling plan:

a) Take sample of size n=13 water contents

b) Calculate mean of sample m x = (1/n)Exi
c) If -1.09% 4 m x < +1.09% Procter optimum, then accept lift.
d) If m x > 1.09% or m x < 1.09% Procter optimum, then reject
lift.

NOTE: The relatively large sample size and tight acceptance limits for this
inspection plan are caused by the large variability of the soil relative to the
unacceptable quality limits of ±2% Proctor optimum water content. 8"

-.-

140

*%V %
Z..

%\ *V
 '-4

PLATE 6

SUBJECT: Operating characteristic (OC) curve for fraction defective sampling,

typical calculation; standard deviation known.

1. PROBLEM: For an acceptance sampling plan specified by

n =5,
L = 95% optimum Proctor,
sif' =2%,
z 1.645;

Find the probability of accepting material actually having 5% defective

(i.e., 5% of the material compacted to less than L).

2. SOLUTION:

1 Find actual mean if s,' = 2% and fraction defective is P' = 5%.

From Table 4, 5% defective corresponds to a mean 1.645 standard

deviations greater than the lower limit of acceptable material.

mx' = L + 1.645 s x '

= 95% + 1.645 (2%) 98.3%

2 Find mean and standard deviation of z for n=5, m x ' = 98.3, L = 95%,
and sx ' 2%

mz = (mx' - L)/s x I
= (98.3%-95%)/2% = 1.645

sz = 1//n
= I//5 = 0.45

3 Find number of standard deviations of z separating mz from z*

1 .645 P%
|

number standard deviations = (mz - z*)/s z

= (1.65 - 1.645)/0.45
- 0.00

4 Find probability of accepting material if m, is 0 standard

F deviations above z.

From Table 4, z 0.00 + Pr{rejectinq = 0.50

Pr{acceptinql 1 - Pr{rejecting"'
-= 1 .50.50
- 0.50 -

'I".

p.e

-. ".... . '. , '-.*-, ,.. . ~. -. A W _ , ...

"""-.'" -- '% " "-" ,"" "" ' ," .
 PLATE 7 SHEET 1/3

SUBJECT: Acceptance sampling plan for fraction defective; standard deviation

known.

1. PROBLEM:

Design an acceptance sampling plan for fraction defective °-

%-compaction, when the standard deviation is known.

2. SOLUTION:

1. Specify: Buyer's Risk a = 0.05

Seller's Risk 8 = 0.10
Acceptable fraction defective Pa = 0.01
Unacceptable fraction defective pu' = 0.10
Standard deviation sx = 2% Procter optimum 1
Lower Limit of acceptable compaction L=95% Procter
optimum

2. Calculate average percent compaction corresponding to pa' and Pu :

m - 95%
Acceptable compaction:
a
= z = 2.33 + m = 99.7%
2% 0.99 a

m -95%
Unacceptable compaction; u 1.28 m 97.6%
2% 0.90 u

3. Define test index:

m -L
x r -.
5
x

%. ..
..'S-;.
 !WJW ~ WK r - -. F .-. 7..Fir Y. -Y . Vy iic' % MmK Wr *.-
IkV..-W.11.' " %rr W

PLATE 7 SHEET 2/3

SUBJECT: Acceptance sampling plan for fraction defective; standard deviation _

"
known.*

4. Calculate mean and standard deviation of z for acceptable and

unacceptable compaction:
m - 95%
a
Acceptable compaction m = = 2.33
za 2%
s =1//n
za
S m - 95%
= 1 I//n_
Unacceptable compaction m1 = - .28
zu 2%
szu 1V

5. Fix Buyer's and Seller's risk:

Seller's risk (acceptable compaction): Pr{z<z*} = a; = 5%

-.
Z*- m
1//n
= 6.05 1.4
z- 2.33
1.05 -1.645

Buyer's risk (unacceptable compaction): Pr{z>z*}=B=0.10

s
Z- m
1-0. I0,%
zu
z +1.28
= +1 . 28
z* - 0 .84 2 -

J 11

I %

[..

,,'

*
' -

|,..J if-_
_' -.
. ".
.,. _'.%
. . - ",-.-.
% .,,,%''
e
,,, " j'.,,'',''%'
,' .%,_,'%
%',%,,.'
.,",,'.-.'
'.'.'-"
% .'%'
,.
,- ,# .
 PLATE 7 SHEET 3/3

SUBJECT: Acceptance sampling plan for fraction defective; standard deviation

known.

Solve simultaneously to obtain:

n =7.8 +8
1.74
*=

6. Specify acceptance sampling plan:

a. Perform n =8 density tests

b. Calculate the mean of the four results, mx. ~
c. Calculate the quantity z,

m -L m - 95%
x x

s 2%
x
d. if,

z ;s 1.74, then accept

* < 1.74, then reject

The OC curve for the sampling plan (n=8, z*=1.74) is shown in Fig.
38. .

144
%.

N
 PLATE 8 SHEET 1/2

SUBJECT: Acceptance sampling plan for fraction defective; standard deviation

unknown. :

PROBLEM: Design an acceptance sampling plan for soil compaction with the
properties I
i = 0.05 Pa' = 0.01
= 0.10 Pu' = 0.20
L = 95% optimum Procter
sx = Unknown.

SOLUTION:

1. Find standard Normal variables corresponding to a,$,pa' ,Pu'.

From Table 4: a + z 1 _a = 1.645

+ z-a = 1.28
Pa' + Za = 2.33
PUl + Zu = 0.84

2. Calculate sample size n and acceptance criterion Z*.

z z + z z
. a 1- u 1-a
Z 1-a 1-8

= (2.33) (1.28) + ( 0.84) ( 1.645)

1.28 + 1.645

= 1.5 1.
2 1- 1-
( *1 z +z 2
n =(+
2
) z - z

1.645 + 1.28
2 2.33 - 0.84

= 8.13 8

-4'-

4A

]4 , .'0"

%" .
% o
 PLATE 8 SHEET 2/2

SUBJECT: Acceptance sampling plan for fraction defective; standard deviation

unknown.

3. Specify Sampling Plan

a. Take random sample of sizen=8

b. Compute sample mean mx, standard deviation sx and the test

m - L
xI
c. Compare zL with z* 1.5,
If zL ;0 1.5, then accept lift.
If zL < 1.5, then reject lift.

4. The OC curve for this plan is shown in Fig. 39.

%* 1
 PLATE 9 SHEET 1/2

SUBJECT: Acceptance sampling plan to assure fraction defective with double

specification limits; standard deviation known.

I. PROBLEM:.

Design an acceptance sampling plan to assure fraction defective on

the basis of compacted water content. Assume SD known to be 1% optimum ,
Proctor.

II. SOLUTION:

1. -Specify,

CE = 0.05 sx ' = 1%
= 0.10 U = +2% Procter optimum
Pa' = 0.10 L = -2% Procter optimum
Pu' = 0.30

2. Determine values of mx' such that the sum of fraction defective above
U and fraction defective below L equals the AQL, Pa'= 0 . 1 0

from Table 8 (by interpolation):

mean PU PL P TOTAL

Mx ' = +0.7 0.096 0.004 0.10

mx ' = -0.7 0.004 0.096 0.10

'J 3. Determine standard Normal

variables corresponding to
(1-a), B,
"A maximum of (pU,pL), and p,'

a = 0.05 Z I_, = 1.65

= 0.10 ZI-9 = 1.28
max(pu',P') 0.096 Zpa = 1.30
pu 0.30 zpu = 0.53

pl

0-

% .f.P . .. . .. .. . ...
. .-
 . -. . - ~ .... U :.,

PLATE 9 SHEET 2/2

SUBJECT: Acceptance sampling plan to assure fraction defective with double

specification limits; standard deviation known.

4. Evaluate n and z*,

z + z
From Eqn. 71 n = 1-a 162
a u
S+1.65+1.28 2
= (13-. .3 14.5 + 15

z z +z z
Prom Fln. 72 z* = a z +z :
1-a 1-s

(1.30)(1.28)+(0.53)(1.65) =0.87
(1.65)+( 1.28)

5. Specify acceptance sampling plan

a) Take random sample of size 15
b) Calculate sample mean mx, and test indices,
m
U x
zU =

m L"
ZL Sx
x

c) Compare z u and ZL with z* = 0.87,

If zu and ZL > z*, then accept ""
If z u and ZL < z*, then reject

6. OC Curve is shown in Fig. 40.

,-

148

%'
 PLATE 10 SHEET 1/2

SUBJECT: Acceptance sampling plan to assure fraction defective with double

specification limits; standard deviation unknown
*0*

I• PROBLEM:

Design an acceptance sampling plan to assure fraction defective on

the basis of compaction water content. Standard deviation is unknown

II. SOLUTION:

1. Specify,

= 0.10 U = +2% Procter optimum

B = 0.10 L = -2% Procter optimum
Pa' = 0.05
Pu' = 0.30

2. Determine standard Normal deviates corresponding to a,B,Pa' and Pu

a = 0.10 ZI- e = +1.28

8 0.10 ZI-B = +1.28
Pa i 0.05 Za = 1.65
Pu' 0.30 zu = +0.53

3. Use Eqns. 77 and 78 to estimate n and z*,

z + z 2
2
n = (1+z* /2) i- 1-B 2 (+1.02 +1.28+1.28 2 7.94 + 8
2 ~1.65-0.53 -

z z + z z
a 1-8
z, + zu 1-a (1.65)(1.28)+(0.53)(1.28)
(1.
•28)+(1.•28) 1.09 -

4. Estimate allowable fraction defective M

Compute- quantity

.I 1 - z* Vn
y:'
=- (n-1) - 1-(1 .'9) /8 -(8-1 - =0 2
2 2

Rel M frowm Ficl. 41 for y0.281, n=8 M = 0.15

.. . . . . .. . . . . .. . . . .

'1%(
 PLATE 10 SHEET 2/2 :

SUBJECT: Acceptance sampling plan to assure fraction defective with double

specification limits; standard deviation unknown

5. Specify test procedure%

a. Take a random sample of size n=8

b. Compute sample mean mx, and standard deviation sx, calculate
test indices,

U mx
z
U s
x 5

m -L
x
zL s
x p-

c. From Fig. 42 read values of Pa and Pu corresponding to zU and

ZL

d. Compare pL + PU to M,%A

If PL + PU 4 M, then accept.
If PL + PU < M, then reject.

6. The OC curve for this plan is shown in Fig. 43

% C;

% % %'S%
 wI

Acceptable Fill

% Frequency

an ik
Unacceptable, FillhoiqSllrsardHyr

J- %.
 1.

0..

1".
r-S. 0.6

-4-

1.0I ". %

- 0.6'
/

0.2 -

o ~~~~~?
woI
'
100 110 115 120 125 130 "
Dry Density
o "1"
.-- .

25

"'.
":"." "'. .'.. . . -" . : " . 2..
' ." -"v v '' -""" - ' ',- . . . . - . .-- ". ." ' '

Figure 29 --Operating characteristic (OC)

curve for an acceptance
sampling plan to assure the value "''%
of mean compacted dry density of
an engineered - ,..
fill (see Plate 3).
.:
U 0.6
* ww M~r ~wMUv qw
-mU WMwWM r WM W~M ~W WI1Wv -M U - - ~u i-w i --

EnI
96

4J

-4
0

41

E '4-

.4 0

.1 b.rn -

Ve
VIV Jo eit-

Zma
 ,,
w..

I,'.

j) = .o

The t cuvs o...5,ad

ditibto o

-3 -2 -1 0 1 2 3 "'
r ,

The t distribution curves for v = 2, 5, and cc.

w
*6l

S
~ -. . - - .* *
Fi gure 31 S-tudent 's -t ist ribiut i oil

154
-V ~ : . ~l.i.,.ur 6 .. 1 - tudn t ~rb tin
J
 Kr -71"16 110111 1 O -- WIT1 - WW-1~VVVW -.

FIG;UREi 15.9%
Operating Characteristic Curves for Single-Limit Sampling Plans Based on the Statistic

with a=O.OS*

0.90 - .

0.80.

0.70.

0.-60
I3

a, ~050- - - ---- - ----- _ -- _

30lRI lit ~X

*In acceptance sampling. A the A; of the plan and A - an) other lot or process quality. n suze of sample 1 he
lot or process is assumed to he normally distributed or approximately normally distrihiated. I he sampling plan has only
one acceptance limit. Source oif original data: ). Neyman and 11 -1obarska. - Friiur of the Second Kind in t esting 'Stu-
dent'% flypothesis." Joiurnal of the Anierii an Stiatisml .4jimii oni~, Vol. XX X1, pp. 3IS- 26.

Ao

a
5*
 W-.:;_,.]

AQL-

':.,!%IO
J

'*. %
156% _
,, , "
156
 .%

LI
-

,?. -•

UQLg

A,:::rpr 26E-,-7 , .-.,-

%7, %Q,

Figure 34 -Sarplin; variabhlity of two lifts of nateriaI, one wi th tr ue e--en

equal to IJVI,[, , the other with true inearn equdl to )I .- "

157
 ".-*'.

Z-0
4&.,

Water content relative to Proctor opitmum -' ,

Figure 35 -- Operating characteristic curve for an acceptance sampling , .

plan to assure the value of the mean, with double . .
sampling limits. £'

'..'N,

0AA

°
 II

Lower Limitx

* I

I frequency distribution of x
A

fraction
Vol..

Soil Property, x

Figure 36 -- Lwproperties.
Fraction defective tor a Normal frequency distribution of soil -

I 51 . .

"..S"
 /0
0 6-.'.V

Percet Defctiv

Fiu e 3 p r t n h
r c e i t c o
wit sigesmln r c i n d f c i e s m l n
Smstn

C.-

-41

% N" N, %'V-.1
 U 0.6

I.q
C,

13=
/0 z' 0-5
FrcinDeetv

Fiu e 3 O e a i g n r ~ er4 t c f r c e t nc a p :n , p a o rvU-

Plte7

J.A

0.%
-e % %
1'10
i-OS

pu0
4.j.
C-r

t.".

o.4,.

~ls -

O.s 0-I . O4z

-ecie Frcio

Fiur 39 ApO-4t o jj II~li 'j j

S~ t I(,1rd deviiti. I tiiikn(I'T
 n~s I

.6

00

0.2

Fract
10.

DeectIv
oJ

%-
%
 0.01

0.05
0.1 .

2.4%

,2.0
100 P

15
1.0
202

40

so __

1 2 3 4 5 6 7 ag91 15 20 30 40 500 60800 n

Fiqure 41 -- Chart for determining fraction defective from z (from Duncan,

1974).

% 00
 0 .10.2030 .0 .5
1 T 1 1 1 1 1

mI

.300.

.200

.060
.040
0..0 .5 .0 .5 3 3 4 4 5

.0n20 I
For~1 SWnadDvainPan
aeasis
,1719k 21

.5.0
i.00

.002

Fo
(rom

tnadDvainPasTk
1974).

bcsaJv/n
0.0
25.10
30
.15 .0.4.i
.5.20 1
1ucn

2
vrg ag lnsIaeasis /q
Fo
p2

Fique 4 Chrt
eteriniq mximm
or alowble raciondefctie,
 I.

100
C.K

<'%

0.

C3
4' ,0 -'

4,2 004 V.o0-s

16

-, t'.2

4..

0 o ,r- O./o D~YO. 2. .Z, O.3 o":'

Fraction Defective .
.. -

..

Figure 43 -- Operating cahnracteristic curve for samplingj plan derived in Plate-1

% "'..J

"a
'°
°
 1. Abramowitz, M. and I.A. Segun 1964). Handbook 3f Mathematical Functions.
Dover Publications, N.Y.

2. American Association of State Highway and Transportation Officials (1984).

"Acceptance sampling plans for highway construction," AASHTO R9-84I, in
Interim Specifications and Methods of Sampling and Testing Adopted by the
AASHTO subcommittee on materials, AASHT)), New York.

3. Benjamin, J. and C.A. Cornell (1970). Probability, Statistics, and

Decision for Civil Engineers. McGraw-Hi'1 Book Company, N.Y.

4. Cochran, W.G. (1977). Sampling Techniques (3rd Pd.). John Wiley and
Sons, New York.

5. Duncan, A.J. (1974). Quality control and industrial statistics (4th Ed.).
Richard D. Irwin, Inc., Homewood, Illinois, 1047.

6. Grant, E.L., and R.S. Leavenworth (1972). Statistical Quality Control

(4th Ed.). McGraw-Hill Book Co., New York.

7. Juran, J.M. (Ed.) (1974). Quality Control Handbook (3rd Ed.).

McGraw-Hill Book Co., New York.

8. Kotzias, P.C. and A.C. Stamatopoulis (1975). "Statistical quality

control at Kastraki Earth Dam," Journal of the Geotechnical Engineering
Division, ASCE, v.101(GT9): 837-853.

9. Kuhn, S.H., (1972). "Quality control in highway construction," CSIR

Research Report 294, NIRR Bulletin 11, pp. 1-40, National Institute for
Road Research, Pretoria, South Africa.

10. Lambe, T.W. (1950). Soil Testing for Englineers. John Wiley and Sons, New
York.

11. Lee, I.K., W. White, an] O.G. Ingle; (1983). Geotechnical Engineering.
5-.
Pitman Publishing Inc, Marshfield, Mas;sachusetts.

12. Sherard, J.L., R.J. Woodwari, s.F. Gizienski, and W.A. Clevenjer (1963).
Earth and Farth-Rock Dains. John Wiley an] ons, New York.

13. Snedecor, G.W. and W.G. Cochran (19R)). Stiti,3tical Method:i (7th E1.).
The Iowa State Univtrsity Press, Amp; low.

14. Turnbul L, W.J. and C. P. Mar, 1 (1 96o) . "Qua lity contr,) ,t compute-
earthworks, "Journal of the toiL
M-haniic ,n irindtion Divi sion, ASCE,
V. 92(0MI).

15. USBR (1960). Thet.Earth M nil. 1D}par nent )f the, ltri or, Washinqton,
D.C.
41

g -
%.'

11

. .. w[
 lflS ~.fl. r..ar~AM7JIP9AJ'.Xrt.IJtm a. na. t<w Lu I~~Wsw .3 t. ~ .- ' c- trw .r'. L .fWl- .7W u.n trw Is a-~ .~..s '.r.u-, JCLWUWL C ISV

IV
I'

I
r

p7

I. ~ ~