DETAILED RESEARCH & DEVELOPMENT PROJECT PROPOSAL

(For the Component Project)

(To be accomplished by the researcher)
(1) Title/Leader/Gender/Agency/Address/Telephone/Fax/Email
Program Title: 12Tech
Project Title: Motor Anomaly Detection and Lifetime Expectancy Estimation Through Deep Learning
of Harmonic Features
Leader/Gender: Emmanuel Bueta / Male
Agency/Address: OneWatt / #26 Greystone St., Greenview Village Sauyo, PH 1116
Telephone/Fax/Email: +639176203859 / em@onewatt.xyz
(2) Cooperating Agencies (3) Research & Development
OneWatt Station
OneWatt Labs
(4) Site of Implementation/Municipality/District/Province/Region
Quezon City
(5) Classification (6) Mode of Implementation
Research: Development:
__X_Single Agency
____Biotechnology __X_Alternative Energy
____ Multi Agency
__X_ICT ____Environment
____Health Products/Pharmaceutical
____Basic Research

(7) Sector Commodity (8) Discipline

ICT/Industrial Manufacturing AI / Motor Health Monitoring
(9) Significance

Problem
Motors are the backbone of the industrial world. Their continuous rotation drives the production of
manufactured goods, delivery of resources, and extraction of raw materials in productions plants. This
dependence, however, puts the factories at a huge risk, since it just takes one bad motor to stop a
whole production line.

This is not an uncommon scenario, in motor-heavy industries, one out of eight motors breakdown
every year. When production stops, a cascade of problems follow: raw materials spoil, employees
become unproductive, and clients do not get their orders. Industrial users claim that they lose upwards
of $250,000 for every hour of downtime, and that still does not include the costs of repair and non-
pecuniary costs such as damage to reputation [1]. At this stage, a motor repair can cost eight to sixteen
times more than if the fault was found early and averted using preventive maintenance.

Even with the trained maintenance engineers, 63% of the early signs of faults still go by undetected.
With some motors failing as early as two weeks from installation, the plant managers and maintenance
staff overseeing the health of hundreds of motors need to constantly check each one of them, which
is the typical scenario in an industrial facility [2]. More than 70% of the maintenance staff’s time is
devoted to data entry and inspection while only 30% goes into work that increases the reliability of the
facility. This also means periodic planned downtimes for specialized tests & unnecessary maintenance
that can only be done on controlled conditions, decreasing the productivity of the whole plant and even
worse may contribute to shortening the lifespan of the motors [3].
Proposed Solution
OneWatt’s Embedded Acoustic Recognition Sensor (EARS) is a predictive motor health management
system that identifies faults before they happen to prevent revenue losses, and remove the guesswork
from maintenance activities. By literally listening to motors, applying machine learning and frequency
analysis, OneWatt can identify early signs of faults and predict the time to failure. Plant managers can
then plan appropriate actions at the right time before the motor goes bad and completely shuts down.

Using acoustic sensing allows OneWatt to be as non-intrusive as possible. It works without even
touching the motor, ensuring we do not affect production. This also allows the system to work with
motors from different manufacturers, different sizes, and ratings without any complicated installation
& alterations needed.

History
The sensor and the algorithm have already undergone testing in controlled laboratory environments,
where the acoustic emission of functioning and faulty motors were sampled. These were pre-
processed using our algorithm and used to train and evaluate the neural network for classifying faults.
With the 126GB of acoustic data gathered, the classifier reached 95% accuracy in identifying the top
eight motor faults. With these promising results, OneWatt has been selected as one of the top
innovations in Accenture’s global Innovation Awards and Shell’s New Energy Challenge.

Impact
With the current population of industrial motors, there is a huge variance in the type, build, and
configuration in each installation [4]. With the help of the 12Tech, we plan to address this issue by
gathering a diverse set of motor acoustic data from industrial partners within the DOST network. With
this, OneWatt can identify faults earlier with a wider breadth of faults and machine types.

The team would also benefit from the business modelling assistance from 12Tech. There is a growing
number of request from industrial users to have their motors monitored. By September, the team will
be sampling acoustic noise from natural gas pumps of Gasunie and electric generators of Engie. The
team has also engaged with Aboitiz Power, MegaGlobal, and Fastech as potential pilot clients.

Industries contributes almost 30% to the country’s GDP, employing roughly 6M Filipinos, as of 2015.
OneWatt aims to help small and medium enterprises with relatively aging motors, especially
businesses that are operating near breakeven. By ensuring their production does not stop due to motor
failure, they can produce the required orders reliably while preventing losses associated to motor
repairs.

(10) Objectives

General Objectives:
Develop a robust acoustic motor health monitoring system that can be used in industrial
environments.

Specific Objectives:
 To increase the prediction accuracy of the motor acoustic model above 95%
 To adapt the sensor to real industrial environment in the Philippines
  To develop a scalable and profitable business model suitable for SMEs in the Philippines

(11) Review of Literature

Current Methodologies for Motor Health Monitoring

There is no clear estimate on the lifetime expectancy for a motor. Even with proper operating
conditions, motors can break down in as early as two weeks, while some last for centuries. With all
the electronic and mechanical parts, there are a lot of factors affecting its operation and the
expected lifespan. Typically, a plant estimates a motors lifetime at 10 years, but expects several
scenarios where they fail and require servicing throughout this period. This unpredictability plagues
industrial users who require predictability and reliability in their everyday manufacturing processes. A
failure in just one motor can halt a factory’s production line leading to lost revenue and bad
reputation for the business [5].

Most industrial facilities employ reactive maintenance where they literally fix things only when they
break. While this sounds normal, this technique is very inefficient since it causes unplanned
downtimes in the industrial setting. This means that while the maintenance staff is finding,
diagnosing and repairing the fault, the whole sequential production line stops. Moreover, the cost of
repairs can easily spike up as the initially small fault has become huge that it brought down the
whole motor [6].

Plant managers who wish to protect their facility from unplanned downtimes sometimes subscribe to
preventive maintenance. Here, they spend a fair amount of time checking on the condition of their
motors on a regular basis. Inspection involves several criteria like electrical parameters, lubrication,
mechanical vibrations and analyzing patterns. This process requires multiple equipment for each
test, a planned downtime to inspect the motor, and a lot of time specially if there are hundreds of
motors running in the facility [7]. Moreover, some of the tests might lead to faults as they try to push
the motor to run at their limits [8]. These preventive maintenance checks are usually scheduled such
that the frequency of the checks catch the most number of developing faults between them.
However, some faults can happen in between maintenance checks, some even almost
instantaneously, and this methodology is unable to detect those.

Applications of Industrial IoT

With the advent of IoT, industries are actively adopting this technology into their systems to increase
operational efficiency of their production lines. Early adopters of these systems claim that they have
seen improvements by as much as 30%. By having IoT systems installed, operational teams can
gain access to real-time data, analytics, and insights into the production. This gives them the ability
to respond quickly and even anticipate the problem, allowing them to do corrective measures [10].

Motor maintenance is one of the major areas where these IoT systems are entering since it is the
most common industrial fixture. With over 500,000,000 in motors running, their number is almost
double the population of America, and still continuing to grow. At the same time, the types of IoT
sensors available in the market is also growing, with differences in installation process, cost, and
accuracy.

One of the most common type of sensor monitors the electrical parameters, like current and voltage,
going into the motor. Installation usually involves placing the sensor in between the ports of each
phase in the motor. Depending on the maximum power rating of the monitored equipment, the cost
of the sensor could grow exponentially to use more robust and higher tolerance components. This
type of sensor works best in identifying faults that are electrical in nature, like shedding winding
laminates, short circuits, and power imbalance. This solution, however, does not give a complete
picture into the motor’s mechanical domain.

To complement the electrical parameter sensor, mechanical operation is usually monitored using
accelerometers. This sensor measures the amount of force transmitted to it. Using this data, experts
can infer whether a motor is working perfectly or producing anomalous tremors. This type of sensor
requires that it is attached directly to the motor at specific sites, since the position and orientation of
the device will significantly vary the measured force. The cost of this sensor is quite invariant as it
does not change components based on the build and specifications of the motor, although can have
a huge base cost since high precision instrumentation is required. With proper use and analysis, this
can identify physical defects in motors such as misaligned shafts and deformations in the motor
body.

Using accelerometers might not always be the best option due to its inherent need for calibration
and dependence on both position and orientation. The force transmitted outside the motor, causing
detectable vibrations can also be detected as propagated sound. This sound can be considered as a
derivative of the force measurements and analysed the same way as accelerometer data. In terms
of installation, this methodology does not require the sensor to be attached to the motor. Sound
waves can propagate through air and can be received at a reasonable distance from the source. The
cost can be much lower compared to accelerometers since the sensor is driven by either piezo
transducers or pressure sensitive microphones.

Acoustic Based Motor Monitoring

The noise emitted by a motor is largely due to the rotational movement of its moving parts and the
interactions with the system. This data can be analyzed both in the time and frequency domains to
get a picture of the motor in question. For example, a healthy motor should produce acoustic
emissions whose main frequency is related to the rotational speed. However, a fault like broken gear
teeth, would contribute a sound similar to clicking or impulse every time the broken teeth is passed
[11].

The most common motor fault is caused by broken bearings. This can be caused by intrinsic
irregularities in the bearing themselves, lack of lubrication, or external particles that accelerate wear
and tear. There are several interactions in this system, which includes the rotor, the outer race, inner
race and the rolling elements all contributing to the final noise emission. All these components
should contribute a specific frequency and deviations from those can be considered as developing
faults. Moreover, the severity of the damage can be estimated by the bandwidth and energy
contributed by the anomalous sound [12].

Machine Learning

With the vast diversity of motors, it would be difficult to custom tune a sensor to listen to the
frequencies that characterize their status. However, with the emergence of machine learning, this
need to identify those characteristics are no longer required. Acoustic emissions and sound patterns
can be sampled by a generalized training algorithm that can identify the relationships between each
frequency band and produce a classifier based on the fault and condition of the subject.

This is evident in the current developments like Alexa, Siri, and Google Assistant. These systems
use trainable models that take different samples of a person’s voice and use that model to increase
the accuracy of the conversion of the sound to words. This technique is translatable to the motor
acoustic domain and has been performed by several researchers [13]. OneWatt’s methodology
follows along these lines, construct a model of the motor and then optimizing the mathematical
model using the collected acoustic data to identify faults.

(12) Scientific Basis/Theoretical Framework

The core of OneWatt’s intelligence is based on a neural network. This technique allows the system
to run and learn about the subject motor as time progresses and producing an optimized predictor
for its state. Although it sounds easy to implement, this algorithm is heavily dependent on data.
Usually, this requires gigabytes of data and significant computing power to arrive at a robust model
of the target motor. So far, the system has been trained with 126GB of acoustic data containing
various states and faults of motors.

During the collection phase, it is important to acquire acoustic data from different states of motors.
Partnering with industrial facilities will allow us to sample working, repaired, and faulty motors in their
use. To supplement this, the team will acquire motors of different states from repair shops and
service providers of industrial facilities.

Typically, collected data is divided into training data and test data. The training data is iteratively
presented to the model as it optimizes itself. This is done by calculating the model’s predicted output
against the know states of the training data once it is entered into the system. An error function is
defined as the amount of deviation of expected results against the predicted results. Once this has
been calculated for the iteration, the network will perform a back-propagation wherein a neuron’s
weight is varied using the rate of change of its value against the final output. This boosts the system
to converge at the minimum of the error function, thus arriving at the model faster than most other
techniques. Once the system has reached an acceptable error rate, the results from the model’s
prediction are compared to the known state of the test data set producing the final accuracy rate.

This process requires continuous feedback as the number of the installations and acoustic data
increases. Because of this requirement, the best platform to perform this is based on Graphical
Processing Units or GPUs. Without the aid of these devices, a typical iteration of the system runs for
a week, while th GPU assisted processing completes an iteration in one minute.

The other branch of the development is the hardware, which allows the system to collect acoustic
data from the environment. Currently, the system uses a microphone array to gather motor acoustic
at different positions. With the use of digital signal processing, the hardware can estimate the source
of the sound and at the same time eliminate unnecessary noise.

Since the sensor is meant to run unsupervised in an industrial facility, there are cases where the
conditions requires safety compliances to be met. Most common ones are IP compliance for dust
and water proofing and EX compliance for chemical volatility. This is important in the product
lifecycle to increase the reliability and applicability of the sensor in the targeted environments.
(13) Methodology

1. Collect Motor Acoustics

a. Number of motor conditions sampled
i. Bearing fault
ii. Outer race defect
iii. Inner race defect
iv. Shaft misalignment
v. Rotor bar deformity
 vi. Uneven air gap
vii. Stator deformation
viii. Loose footing
ix. Soft Foot
b. At least 200 motors monitored
c. At least 60,000 hours of acoustic data recorded
2. Retrain the neural network with the data collected
a. At least 95% accuracy of classification
b. 10seconds maximum response time to identify one second of audio
3. Enhancement of existing circuit and enclosure
a. Number of industry standards met (EX, IP, Cole-Parmer, EMI, RF)
b. Electro-magnetic interference shielding
c. Tolerance to 80db noise level
4. Install the system in a real industrial environment
a. At least 25 dashboard requests/views per month
b. At least 80% accuracy of alerts (including false positives and false negatives)
c. At least 60% conversion rate from trial to paying client
(14) Major Activities/Workplan Please refer to Form 2B-1
(15) Expected Output

 A neural network model, customized for the Philippine industrial environment, that works on a
huge breadth of motors and an accuracy greater than 95%
 A circuit and enclosure that passed industrial environment accreditations
 Installation manual and case documentations on the effects of OneWatt for industrial facilities
 Business model and running installations of OneWatt

(16) Target Beneficiaries

Industrial facilities/utilities with at least five AC motors in their production line looking to increase
reliability of their production.
(17) Personnel Requirement

Type Skill Percent Devoted to

Project
Project Leader (1) Project Officer 100%
Others (1) Information Technology Officer II 100%
(1) Senior Science Research Specialist 100%
(1) Project Development Officer I 100%
(1) University Research Associate I 100%
(1) Project Assistant III 100%

(18) Budget (See form 2B-2)

(19) Literature Cited
 [1] “The Total Cost of Donwntime in Industrial Automation.”
https://www.stratus.com/assets/Infographic-Total-Cost-of-Downtime-Industrial-
Automation.pdf.
[2] “Reducing operations & maintenance costs.”
http://www2.emersonprocess.com/siteadmincenter/PM%20Central%20Web%20Documents/p
lantweb-ops-maint.pdf. September 2003.
[3] “ABB Ability Smart Sensor.”
http://search.abb.com/library/Download.aspx?DocumentID=9AKK106713A5345&LanguageC
ode=en&DocumentPartId=&Action=Launch. 2017
[4] “Industrial IoT Market Worth195.47 Billion USD by 2022.”
http://www.marketsandmarkets.com/PressReleases/industrial-internet-of-things.asp
[5] Nailen, R. “How Long Should a Motor Last.” www.brithinee.com/technical-articles/how-long-
should-a-motor-last/.
[6] Dabbs, T. “The True Cost of Maintenance.” Pumps & Systems. April 2004
[7] “Preventive Maintenance of Motors and Controls.” Electrical Construction and Maintenance
Magazine. October 01, 1995.
[8] “Standard for an Electrical Preventive Maintenance (EPM) Program.” The Hartford Steam
Boiler Inspection and Insurance Company. April 2013
[9] “Industry 4.0: Huge potential for value creation waiting to be tapped,” Deutsche Bank
Research, May 23, 2014.
[10] Press release, “Accenture to Help Thames Water Prove the Benefits of Smart Monitoring
Capabilities,” March 6, 2014.
[11] Singh, SK. “Acoustic Based Condition Monitoring.”
[12] Li, B., Chow, M., Tipsuwan, Y., Hung, J.“Neural-Network-Based Motor Rolling Bearing Fault
Diagnosis.” IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS. VOL. 47, NO. 5.
October 2000.
[13] Chow, M., Sharpe, R., Hung, J. “On the Application of Artificial Neural Networks for Motor
Fault Detection – Part II.” IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS. VOL.
40, NO. 2. April 1993.
(20) Capsule Curriculum Vitae (One-page brief C.V. only)

See attached appendix-CV

(21) Certificate of incorporation (if applicable)