FRP PDF

Download as pdf or txt
Download as pdf or txt
You are on page 1of 19

Identication of Fiber-reinforced Plastic Failure Mechanisms From Acoustic Emission Signals Using Self-organizing Maps and Fuzzy C-Means

XuQin Li March 17, 2009

Introduction

A number of Non-destructive testing (NDT) methods have been used to identify the type of failure mechanisms whin ber-reinforced plastic (FRP) material. Among several NDT techniques, acoustic emission (AE) has been frequently adopted because it is well suited for detection of microfailures in all kinds of materials. The AE signal is an ultrasonic wave emitted when materials are strained. It is a transient wave resulting from sudden release of stored energy during a damage process. So any AE signal contains useful information on the damage mechanisms. The major issue in the use of AE technique is how to discriminate these failures using AE signatures. Usually, Fiber-reinforced plastics consist of more than one material. So its quite natural that the failure mechanism may occur in one material or in a combination of materials. The typical damage mechanisms of FRP are matrix cracking, ber breakage, ber-matrix debonding, delamination, and ber pullout. Due to the dierent characteristics of these sources, these eect will result in dierent AE responses. Many standard AE parameters such as peak amplitude, rise time, energy, duration, etc have been been focused on identifying these failure mechanisms. Among which the frequency is the most-widely used parameters to discriminate the dierent failures. But in previous publications where the frequency is employed to dierentiate between events, only one fre1

quency is considered and this frequency was not enough to thoroughly describe the behavior of the composite material. So we introduced the second frequency in a hope to denote a single event more precisely. A Fast Fourier Transform (FFT) is then applied to the signals resulting from the two frequencies. After that the frequencies are used as the features to recognize the dierent patterns among the events thus to discriminate dierent failure mechanisms. The pattern recognition work was done by means of two inelligent computing techniques: self-organizing map (SOM) and Fuzzy C-means (FCM).

Theoretical Background to Self-organizing Map (SOM) and Fuzzy C-means (FCM)

2.1

Introduction to Self-organizing Map

Self-organizing Map (SOM) is one of the most well-known and widely used Neural Network models based on unsupervised learning algorithm. It was proposed and developed by Kohonen to transform a data set of arbitrary dimensions into a one or more dimensional discrete map. In the process of this transforming, the input patterns that are located close to each other in the input space will also be located closely in the output space and dissimilar patterns will be mapped on opposite regions of the map. Suppose wij are the components of the weight vector Wj , connecting the inputs i to output node j ; the xi are components of the input vector X , the output of the neuron is the quadratic (Euclidean) distance dj between these two vectors. The description of each step for using the data to train the SOM is summarized as follows: Initialization Initialize all the weight vectors to random values between 0 and 1. Sampling Randomly select an input vector X ; present it to all the neurons of the network and evaluate the corresponding quadratic distance output dj according to the following equation:
n

dj = ||X Wj ||2 =
i=1

(wij xi )2

(1)

wm1

w11 x1

w12 w1n

wm 2

wmn

x2

xn

Figure 1: An architecture of a typical structure of SOM

where n is the number of input vector components. Similarity Matching Select the neuron with the minimum output dj as the wining neuron, i.e. the nearest vector to the input vector (using the minimum Euclidean distance criterion). Let j denote the index of the winner, the minimum output will be: dj =
j [1,,m]

min

||X Wj ||2

(2)

where m is the number of neurons. Updating Adjust the weight of the wining neuron according to Eq. (3). wij (k + 1) = wij (k ) + (k )[xi (k ) wij (k )] where k is the learning iteration count and is the gain term. The weights of the neighbors of the wining neuron, dened by the neighborhood function N (j ) (the neighborhood function N (j ) denes how many neurons in the neighborhood of the winning one will be updated for each learning input) are also updated following Eq.(4) and (5). wij (k + 1) = wij (k ) + (k )[xi (k ) wij (k )] if j N (j ),neighborhood of j wij (k + 1) = wij (k ) 3 (5) (4) (3)

if j N (j ),neighborhood of j Continuation Repeat steps 2 to 5 until no noticeable changes in the output map for all the input vectors.

2.2
2.2.1

Fuzzy C-means algorithm


Introduction to the Fuzzy C-means algorithm

Dierent kinds of denitions of a cluster can be formulated, depending on the objective of clustering. The objective of cluster analysis is to classify objects according to similarities among them, and organize a data set into subsets. Usually, clustering techniques are considered to be among the unsupervised methods, they do not use prior information about groupings before clustering. It can detect the underlying structure in a data set, thus it can be used for classication and pattern recognition. Usually we regard a cluster as a group of objects which bear more similarities to one another than to the members of other clusters. We can dene similarity by means of a distance norm. It can be measured among the data vectors themselves, or as a distance from a data vector to some prototypical object (perspective center) of the cluster. The prototypes may be vectors of the same dimension as the data objects and are usually not known beforehand, and usually they are sought by the clustering algorithms heuristically with the partitioning of the data. Fuzzy C-means (FCM) algorithm, also known as Fuzzy ISODATA, is a data clustering technique wherein each data point belongs to a cluster to some degree that is specied by a membership grade. This technique was originally introduced by Jim Bezdek in 1981 as an improvement on earlier clustering methods. It provides a method that shows how to group data points that populate some multidimensional space into a specic number of dierent clusters. The FCM-based algorithms are in practice the most widely used fuzzy clustering algorithms. The objective of the FCM clustering algorithm is to minimize an objective function called C-means functional, which is formulated as:
C N

J (U, V ) =
j =1 i=1

(ij )m ||xi vj ||2

(6)

Where X = {x1 , x2 , , xN }, xi Rn , represent a given set of feature data. V = {v1 , v2 , , vC } are the cluster centers. U = (ij )N C is a fuzzy partition matrix, in which each member indicates the degree of membership between the data vector and the cluster j. The values of matrix U should satisfy the following conditions: ij = [0, 1], 1 i N, 1 j C
C

(7) (8)

ij = 1, 1 i N
j =1

The exponent m [1, ] is the weighting exponent, which determines the fuzziness of the clusters. The most commonly used distance norm is the standard Euclidean distance dij = ||xi vj ||, although Babuska suggests that other distance norms could produce better results. Minimization of the cost function J (U, V ) is a nonlinear optimization problem, which can be minimized with the following iterative algorithm: 1) Initialize the membership matrix U with random values so that the conditions in Eq.(7) and (8) are satised.Choose appropriate exponent m and the termination criteria. 2) Calculate the cluster centers V according to the equation: vj =
N m i=1 (ij ) xi , j N m i=1 (ij )

= 1, , C

(9)

3) Calculate the new distance norms: dij = ||xi vj ||, i = 1, , N, j = 1, , C (10)

4) Update the fuzzy partition matrix U : if dij > 0 (indicating that xi = vj ) ij = else ij = 1 5) if the termination criterion has been met, stop. Else go to step 2. 1
C dij m2 1 k=1 ( dik )

(11)

2.2.2

Fuzzy clustering validity indexes

The goodness of the clustering depends heavily on the choice of the number of expected clusters (Babuska, 2001). A number of clustering indices exist which can be used as the guidance to decide how many clusters should be presented in a data set to make a meaningful clustering. We can calculate the validity measure indexes to estimate the success of an algorithm or to help to nd the optimal number of clusters for a data set. A description of the indexes used here has been described by Balasko et al (2005). The indexes can be grouped and interpreted as follows: i) For the Partition Coecient (PC) and Classication Entropy (CE), the indices naturally decay and grow with cluster number respectively and so local minima and maxima respectively should be used to nd good values. ii) For the Partition Index (SC), Separation Index (S) and Xie and Benn index (XB), a low value indicates a good choice, and where two or more cluster numbers have similar index values, the smaller number of clusters should be chosen. iii) The Dunn (DI) and Alternative Dunn (ADI) indexes should also be minimised.

3
3.1

Test Apparatus and Experimental Procedure


Acoustic emission equipment

Due to the requirement to have a exible and feasible process for manufacturing the test coupons during the current research hand lay up was the most suitable solution. The ber glass volume fraction that is normally achieved by this manufacturing process is 57%. All the Glass/Polyester plates were manufactured in-house in order to have a closer control of the parameters that inuence the mechanical properties; and ultimately the acoustic response of the materials. These parameters included ber orientation and coupon dimensions. The coupon dimensions for the current research were 20 mm x 200 mm x 1.8 mm with three layers of unidirectional glass tape. Each plate was cut into dierent

Figure 2: Testing Apparatus, the sensor was coupled to the specimen using a G clamp

orientations in 15 steps against the direction of the load. A broadband WD piezoelectric sensor manufactured by Physical Acoustic Corporation was used to capture the stress waves (AE events). After capturing the event, the piezoelectric sensor transforms the mechanical deformation of the material surface into an electrical signal. This signal is transferred to a sound card with a high sampling rate, which processes it and displays it as a wave. The software used with this sound card can display on screen the transient waveform for each event. Each event can then be individually exported into an ASCII le, which is then processed by a program written in the R programming language (open source code language). With the use of this program it is possible to perform, and visualize a fast Fourier transformation (FFT) for each event and select the most dominant frequency, which is to be used as the descriptor of the event. The distribution of these dominant frequencies as the test progresses shows the trends of the events occurring in the test specimen. The acoustic emissions software was the TRA (transient recorder package) from Physical Acoustics Corporation 2001. The hardware was a sound card AEDSP32 (Acoustic Emissions Digital Signal Processor) with a digital signal processor

by Texas Instruments (TMS320C40). It has a maximum sampling speed of 8 MHz and 16 bits resolution, between 20 and 100 dB. From the literature review, it appears that there are several dierent approaches to the task of event recognition analysis, from simple hit analysis to amplitudebased and frequency-based studies. The amplitude of an acoustic wave is attenuated as the wave passes through the material. Thus the amplitude recorded at a sensor depends upon the distance the sensor is from the source of the emission. Based on that, for the present study the dominant frequency of the signals was considered to be the most relevant characteristic.

3.2

Test Procedure

According to Qi 90% of AE activities for glass ber reinforced (GFR) composite materials are concentrated in the frequency range 10-550 kHz. One wide band sensor model PAC WD with an operational range of frequencies between 100 and 1000 kHz with a resonant frequency at 125 kHz was used. The sensor was attached to the specimen by means of a G aluminum clamp with a plastic screw. The surface of the sensor was covered with silicon grease in order to provide good acoustic coupling between the specimen and the sensor. The signal was detected by the sensor and enhanced by a 2/4/6-AST pre-amplier. The gain selector of the preamplier was set to 40 dB, as the material used was not considered loud. Figure ?? shows the test conguration, with the sensor clamped to the test specimen and the INSTRON grips where used for applying the tensile force.

Data Analysis

Initially, the MISTRAS software that records the counts from each event was used. It was observed that as each test progressed, the number of counts increases signicantly. Nevertheless, the data do not provide any information about the type of failure or the location of the failure. The TRA software, although using the same hardware, acquires the information in a dierent way; it records the complete waveform and stores the event. A transformation is then applied to dene the frequencies related with these waveforms, which facilitates their analysis. According to Qi these frequencies are useful to distinguish

Figure 3: (a), (top) transient event waveform; (b) (bottom) FFT power spectrum

dierent AE signals from various possible failure modes in ber reinforced composites. A waveform event contains several component frequencies, and it is not possible to determine them visually, therefore a program was needed that calculates the FFT for each waveform. The procedure produces a power vs. frequency graph of a waveform (power spectrum). Higher powers contained at particular frequencies are then shown as peaks on this representation of the waveform. If more than one frequency is super-imposed on the waveform each component appears as the secondary frequency and so on. The program that generates the FFT identies by default the ve frequencies from each signal with the highest power peaks as shown in Figure ??: (a) transient event waveform; and (b) FFT power spectrum. Sometimes the power dierence between the highest power frequency and the subsequent lower power frequencies is so distinct that it can be assumed that the single frequency constitutes the denition of that waveform. Based on our analysis of the data, it was evident that we can assume that the event primary frequency (1stEF) alone was enough to characterize each acoustic event. Figure ?? shows a typical AE event recorded by the system. The FFT power spectrum in this specic case shows several

peaks, the highest one being 280 kHz and having a second higher peak at 236 kHz. The FFT power spectrum can be used as the ngerprint of each event and therefore may be used as a means of distinguishing between them. A possible interpretation is that the waveform is formed from several individual sources associated with an event and each one has its own frequency. In previous publications where the frequency is employed to dierentiate between events, only one frequency is mentioned. As a rst approach the frequency with the highest power value was used as an identier for the events. However, this frequency was not enough to thoroughly describe the behavior of the composite material. Previous tests have found that the 1stEF contributes more than 50% of the total power of the waveform; the remaining highest power frequencies contribute the remaining power. Some researchers (e,g. Papadakis, 2004) attempted to nd a relationship between the rst and the second frequency, through a direct plot of the two of them and analyzing any relationship by two approaches (connecting vectors and kernel density estimations). Figure ?? shows a plot of the 1st frequency against the 2nd frequency from one set of FFTs calculated from waveforms recorded on a tensile test of 0 Carbon/Polyester. It can be seen that there is some clusterings of values where a relationship could be drawn, however according to Papadakis this representation requires more information for it to be used as a sensible tool to dierentiate the events that originated each waveform. Although these preliminary results supported the hypothesis that frequency could be used as an identier for dierent failure events in composite materials, nevertheless subsequent analysis showed that using only the rst frequency as a descriptor is not enough to produce optimal results. It was therefore decided to add the second frequency to analyze the data obtained from dierent ber orientations using neural networks.

Testing Results

The clustering process can be supervised or unsupervised. A supervised clustering process expects certain output, for example a xed number of classes, whereas the unsupervised process will result on the best solution after several

10

Carbon/Polyester 0
700000

600000

500000
2nd Frequency (Hz)

400000

300000

200000

100000

0 0 100000 200000 300000 400000 500000 600000 700000 1st Frequency (Hz)

Figure 4: Relationship between the 1st and 2nd highest power frequencies.

repetitions. Based on the process explained in previous sections, the data from the 7 dierent bres orientations shown in Figure ?? was passed through an unsupervised SOM and FCM. Here we take 0 degree test for example. Figure ?? shows the SOM output map with the 9 neurons positioned on their correspondent locations. Notice that some of the neurons are closer to each other (i.e. neurons 5, 6, 7 and 8), these neurons are eventually grouped into a single class by the neural network. Therefore the values that had been related to each neuron by the network will be considered to belong to that class. Once those neurons that are in the same vicinity are identied by the network as belonging

Figure 5: Diagram of ber orientations.

11

SOM Weight Positions 7

4 W(i,2) 3

Neuron 1 Neuron 2 Neuron 3 Neuron 4 Neuron 5 Neuron 6 Neuron 7 Neuron 8 Neuron 9 0 1 2 3 W(i,1) 4 5 6 7

Figure 6: SOM output for 0 Glass/Polyester

to the same class, then a plot with the number of events corresponding to each neuron is created. Apart from that we have some useful visualized gures which enable us to interpret the results from a more straightforward perspective. Figure ?? shows the neighbor distances within the SOM map, which indicates the distances between neighboring neurons. The squares represent the neurons and the lines connect neighboring neurons. The color containing the lines indicates the distance between neurons with dark colors representing larger distance and light colors representing smaller distances. As we can see, a group of light segments appear in the upper-right region, bounded by some darker segments. This grouping exactly indicates that the data relate to the neuron 5, 6, 7 and 8 should be clustered into a single group. This group can also be seen from Figure ??. The lower-left region of this gure contains several tightly clustered neurons. The corresponding weights are closer together and the positions are tighter in this region, which is indicated by the lighter colors in the neighbor distance gure. The segments in the lower-left region of the neighbor distance gure are darker than those in the upper left, since the distance from neuron 1 to neuron 5, the distance from neuron 2 to neuron 5 and the distance from neuron 4 to neuron 5 are all larger. This color dierence indicates that data points in this region are

12

SOM Neighbor Weight Distances 3 2.5 2 1.5 1 0.5 0 0.5 1 1

Figure 7: SOM Neighbor Weight Distances

farther apart, which is conrmed in the weight positions gure. In addition, we have also produced another gure which can show directly that how many data points are associated with each neuron. It can be seen from Figure ?? that this data set is concentrated a little more in the middle of the upper-left region. The above mentioned results can also be justied using another technique: Fuzzy-c Means (FCM). The FCM algorithm starts with an initial guess for the cluster centers, which is intended to mark the mean location of each cluster. The initial guess for these cluster centers is most likely to be incorrect. Additionally, FCM assigns every data point a membership grade for each cluster. By iteratively updating the cluster centers and the membership grades for each data point, FCM iteratively moves the cluster centers to the right location within this dataset. This process is based on minimizing an objective function that represents the distance from any given data point to a cluster center weighted by that data points membership grade. As we have mentioned, the success of the clustering depends heavily on the choice of the prospective number of the classes. Fortunately, a number of clustering indices exist which can be used as guidance as to how many clusters or classes are present in a data set although 13

Hits 3 2.5 2 1.5 1 0.5 0 0.5 1 1


750 277 195 138 608 847 3232 956 49

Figure 8: SOM samples hit for each neuron

they are not the absolute indicators for deciding how many clusters should be classied. Figure ?? shows a possible clustering indices for 0 degree data. For many of the indices, mainly the Partition Coecient (PC), Classication Entropy (CE), Partition Index (PI), Xie and Beni Index (XB) and Dunn Index (DI), there appears to be an strong implication at 3 clusters, justifying the conclusion that the data set should be classied as three clusters. The FCM algorithm converges after 13 iterations. Figure ?? shows three clusters in the map which were all represented by a cluster center. The rst cluster lies around 2.5k Hz, while the second and the third lie around 4.5k Hz and 5.5k Hz respectively. Each cluster consists of some events (with the rst cluster 5761 events, the second and the third cluster 457 and 834 events respectively) which amount to 7052 events in total. In order to compare the consistency of these two algorithms, we plotted a vertical bar graph showing the percentage of the events in each class for each algorithm. Figure ?? shows the number of events per class. The X axis accounts for the 5 dierent classes found during the training process whilst the Y axis has been normalized against total number of events for that test. This is done so that it is easier to compare the relative population for each class per test 14

Partition Coefficient (PC) 0.9 0.85 0.8 0.4 0.2 0

Classification Entropy (CE)

3 4 Partition Index (SC)

3 4 Separation Index (S)

0.4 0.2 0 15 10 5 0.4 0.2 0

0.2 0.15 0.1 0.02 0.01 0 5 2 Alternative Dunn Index (ADI)

3 4 Xie and Beni Index (XB)

3 4 Dunn Index (DI)

2.5

3.5 Number of Clusters

4.5

Figure 9: Clustering indexes showing the No. of Classes versus the index

A:Test for Zero degree (total events:7052) 7

Cluster Center

5 Second Frequency

3 4 First Frequency

Figure 10: FCM output for 0 Glass/Polyester

15

Figure 11: Normalized events per class for 0 Glass/Polyester

across all the tests. After carrying out all the tests, we have presented all the results in Figure ??. For the 0 degree test, we have detected three dierent classes in which class 2 dominates all the event. As for the 15 degree test, we have found again Class 2 accounts for the majority of the events while the other three classes (newly discovered Class 1, Class 4 and Class 5) account for about 15% each for both algorithms. The normalized number of events corresponding to 30 shows that in both algorithms class 3 has no events and class 2 contributes with almost 70% of the total events. The other classes contribute less than 40%. 45 ber orientation shows a more even distribution than the previous ones. As we can see that between 15 and 45 ber orientations, all the classes were represented except for Class 3. For the 0 ber orientated material Class 1 was also missing. when it comes to 60 ber conguration, it is noticeable that there are no events that could be related to either Class 4 or Class 5. Class 1 does not show any obvious pattern in terms of the change of its size as the ber orientation turn around away from the direction parallel to the load. Class 3 continues presenting an increase in the number of events for 75 ber conguration compared

16

Total events

A:Test for Zero degree (total events:7052) 1 0.5 0 1 2 3 4 Classes B:Test for Fifteen degree (total events:5451) 5

Total events

1 0.5 0 1 2 3 4 Classes C:Test for Thirty degree (total events:1251) 5

Total events

1 0.5 0 1 2 3 4 Classes D:Test for Fortyfive degree (total events:1268) 5

Total events

1 0.5 0 1 2 3 4 Classes E:Test for Sixty degree (total events:1434) 5

Total events

1 0.5 0 1 2 3 4 Classes F:Test for Seventyfive degree (total events:534) 5

Total events

1 0.5 0 1 2 3 4 Classes G:Test for Ninety degree (total events:967) 5

Total events

1 0.5 0 1 2 3 Classes 4 5

Figure 12: Comparison of all the results for the 7 tests

to previous ber congurations. Class 4 and Class 5 remain empty and the number in Class 1 continues to decrease as the bers keep rotating away from the direction parallel to the load. When the direction is perpendicular to the load, that is for 90 degree test, class 3 presents the largest contribution to the total number of events in the test with nearly 60%, and classes 4 and 5 are not present; Class 2 and Class 1 both contribute around 20% of the total number of events. Figure ?? also shows the overall transition in both SOM and FCM of the relative contribution per class for each test. Class 2 shows a progressive decrease in its contribution to the total number of events. Class 3 on the other hand appears from the 60 ber orientation and increases towards the transverse orientation. Classes 4 and 5 exist only for congurations below 45; whereas they are not present for 60, 75 and 90. Class 1 appears in all the tests with a variable

17

Test Classes SOM FCM Similarities Test Classes SOM FCM Similarities Test Classes SOM FCM Similarities Test Classes SOM FCM Similarities 1 20.17% 20.79% 1 50.63% 49.72% 1 18.07% 18.55% 2 2 81.13% 81.69%

0 4 6.14% 6.48% 0.9224 30 4 4.40% 5.28% 5 11.35% 9.59% 1 17.19% 18.22% 2 5 12.73% 11.83% 2 25.85% 27.10%

15 4 55.42% 57.84% 0.8914 45 4 23.34% 20.35% 5 18.77% 21.45% 5 18.73% 15.06%

66.19% 66.59%

40.69% 39.98%

0.9888 60 2 20.71% 20.42% 0.9616 90 2 25.96% 22.75% 0.8821 Table 1: Similarities for the seven tests 3 53.88% 56.46% 3 28.66% 29.85% 1 27.90% 28.28%

0.9450 75 2 21.54% 22.66% 0.9172 3 50.56% 49.06%

contribution to the total events except for the 0 conguration where it does not appear at all. In addition to the bar chart which presents a straightforward comparison, we also compared the similarities between the two methods for all the seven test (see Table ??). Since we have labeled all the testing events using these two different techniques, its quite natural we calculate the similarities (it is dened as the ratio of the number of events which have the same labels for both SOM and FCM to the total number of events). It is noted that the results generated by SOM and FCM separately bear a high similarity (from 88.2% to 98.9%) which again justied our conclusion.

18

Discussion ad Conclusion

After carrying out an analysis on only the 1st frequencies it was observed that more information was required for a proper description of the failure mechanism phenomenon. This information came from the same FFT previously obtained from all the AE recorded from the tensile tests. After a careful characterization of the recording system in order to assure the quality of the recorded data the clustering process was handled by intelligent system techniques. This classication process has resulted in a new set of graphs where the events were classied into 5 dierent classes by their relative closeness in both SOM and FCM. After analysis from a material property perspective, Class 1, Class 2 and Class 3 can be related to matrix cracking, ber/matrix debonding and ber/matrix pull out respectively, and Class 4 and 5 can be related to ber breakage. Conclusions can be drawn as follows: (1) Frequency based analysis from AE technologies can be used to correlate micro-mechanical failure events in composite materials to their corresponding AE signature. (2) Two intelligent system engineering techniques methods, SOM and FCM, have been used as a reliable classication and clustering tool for AE events recorded from damage in composite materials. (3) From the plots obtained from the SOM and FCM classication of AE events, Class 1, Class 2 and Class 3 have been found to be related to matrix cracking, ber/matrix debonding and ber/matrix pull out respectively, and Class 4 and 5 have been found to be related to ber breakage.

19

You might also like