Content uploaded by Kam K.H. Ng
Author content
All content in this area was uploaded by Kam K.H. Ng on Jun 29, 2017
Content may be subject to copyright.
Copyright ©2017, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
DOI: 10.4018/978-1-5225-0956-1.ch004
Chapter 4
50
Big Data Analytics for
Predictive Maintenance
Strategies
ABSTRACT
Maintenance aims to reduce and eliminate the number of failures occurred during
production as any breakdown of machine or equipment may lead to disruption for
the supply chain. Maintenance policy is set to provide the guidance for selecting the
most cost-effective maintenance approach and system to achieve operational safety.
For example, predictive maintenance is most recommended for crucial components
whose failure will cause severe function loss and safety risk. Recent utilization of
big data and related techniques in predictive maintenance greatly improves the
transparency for system health condition and boosts the speed and accuracy in the
maintenance decision making. In this chapter, a Maintenance Policies Management
framework under Big Data Platform is designed and the process of maintenance
decision support system is simulated for a sensor-monitored semiconductor manu-
facturing plant. Artificial Intelligence is applied to classify the likely failure patterns
and estimate the machine condition for the faulty component.
C. K. M. Lee
The Hong Kong Polytechnic University, China
Yi Cao
The Hong Kong Polytechnic University, China
Kam Hung Ng
The Hong Kong Polytechnic University, China
Big Data Analytics for Predictive Maintenance Strategies
51
INTRODUCTION
Maintenance can be defined as all actions which are necessary to retain or restore a
system and a unit to a state, which is necessary to fulfill its intended function. The
main objective of maintenance is to preserve the capability and the functionality
of the system while controlling the cost induced by maintenance activities and the
potential production loss. Correspondingly, failures can be defined as any change
or anomaly in the system causing an unsatisfactory level of performance. Although
only certain failures will cause severe risk in productivity and safety, most failures
lead to disruptive, inconvenient, and expensive breakdowns and loss of quality.
Maintenance plans are designed to reduce or eliminate the number of failures and
the costs related to them.
There are two broadly accepted methodologies aiming at continuously enhanc-
ing maintenance excellence, with different focuses. As a human factor manage-
ment oriented policy, total productive maintenance (TPM) involves all employees,
especially the operators, in the maintenance program in order to achieve optimality
in overall effectiveness and zero breakdowns. Through the operators’ participation
in maintenance, such as through inspections, cleaning, lubricating and adjusting,
early detection of hidden defects, before service breakdown. TPM aims to diminish
and eliminate six significant losses of equipment effectiveness – i.e. breakdowns,
setup and adjustment, idling and stoppages, reduced speed, defects in process, and
reduced yield (Jardine & Tsang, 2013).
Reliability-centered maintenance (RCM) is another approach to strengthening
the system’s reliability, availability and efficiency which focuses on design and
technology. RCM program is based on systematic assessment of maintenance needs
after a complete understanding of the system function and the types of failure caus-
ing function losses.
Types of Maintenance
Maintenance activities can be categorized into three types:
1. Reactive or corrective maintenance,
2. Preventive maintenance (PvM), and
3. Predictive maintenance (PdM).
The following terms are also respectively used for the above three categories
interchangeably as:
1. Breakdown maintenance or unplanned maintenance,
Big Data Analytics for Predictive Maintenance Strategies
52
2. Planned maintenance, and
3. Condition based maintenance (CBM) or prognostic and health management
(PHM).
Reactive or corrective maintenance follows the run-to-failure methodology, which
is the repair and/or replacement work after an equipment outage has occurred. This
primitive maintenance approach, which has been applied in industry for decades,
and is still considered the best maintenance policy for non-critical components with
short repairing time in the system. However, in most cases, an equipment failure can
lead to unexpected production delay and lower the production efficacy rate, or more
seriously, cause severe damage to other components and/or injury to people. One
goal of a proactive maintenance plan is to reduce the overall requirement for reac-
tive maintenance and to apply PvM and/or PdM strategies on any feasible occasion.
Preventive maintenance is performed based on a certain periodic interval to
prevent and correct problems before breakdown without considering the actual
health condition of a system. Basic preventive maintenance, including inspections,
lubrication, cleaning and adjustment is the first step to be undertaken. After that,
rectification or replacement can be undertaken only for components identified with
defects and/or considerable risk of failure. Generally, most PvM actions can be
implemented by operators with basic training.
Predictive maintenance is a trend-oriented policy that begins with identifying the
states of each component within the equipment. PdM greatly relies on engineering
techniques and statistical tools to process the data and analyze the health condition
in order to predict possible equipment failure (Lee, Ardakani, Yang, & Bagheri,
2015). The prediction of the equipment condition is based on the finding that most
types of failures, which occur after a certain degradation process from a normal
state to abnormalities, do not happen instantaneously (Fu et al., 2004). Through
degradation monitoring and failure prediction, PdM reduces the uncertainty of
maintenance activities and enables identifying and solving problems before potential
damage. Condition-based maintenance, the alternative term for PdM, imposes more
emphasis on real-time inspections using RFID devices and wireless sense networks
(WSNs). The three key steps of CBM are monitoring and processing, diagnosis and
prognosis, and maintenance decision making.
BIG DATA ACHITECTURE
Big Data is not only a matter of volume increased of the collected data, but also
includes the evolution of data peculiarities, namely data variety and data velocity.
The intrinsic pattern of comprehensive data becomes the major driver for compa-
Big Data Analytics for Predictive Maintenance Strategies
53
nies to investigate the use of big data analytics. The features of big data are broadly
recognized as “4Vs” – i.e. volume, velocity, variety and value (Xin & Ling, 2013).
• Data Volume: Data volume is the primary attribute of data. Due to the dra-
matically boost in data volume and the rising demand of data storage, data-
base management has been modified as petabyte-scale data management. The
storage needs can be easily satisfied by increasing data warehouse capacities
but at the same time, the data transfer and processing will become too slow to
be operated and handled by a single computer. Therefore, a super-computer
or high capacity server is required for operation. On the other hand, the cost is
also a major concern for implementing with advanced computing equipment.
The data growth in the supply chain and logistics system becomes too large
and too complex in the traditional data warehouse and system architecture.
In addition, big data analytics makes good use of multiple modes during the
computation. This Big Data Analytics method can handle large volume data.
• Data Velocity: With the use of automatic data retrieval systems (e.g. sen-
sor network, wireless network and electronic data interchange through the
Intranet and Internet), the speed of data transaction is rapidly increased.
Moreover, this allows companies to collect instant information and accelerate
the speed of data production and streaming. The data transfer and transfor-
mation are more intensive nowadays. If the company chooses to extract real
time information from several automatic data retrieval systems, the speed of
data processing from big data must be as expeditious as possible to meet the
requirement of quick response.
• Data Variety: The analytics process of data mining has been expanded from
structured data to unstructured data, typically the images, videos, audio and
text, etc. The systematic relationship is longer limited by numerical results,
but also prohibits pattern finding in unstructured data. This motivates com-
panies to investigate semi-structured and unstructured data for their decision
making process. The major purpose of big data analytics is to resolve the
problem of incompatible data formats and non-aligned data structures of un-
structured data with data mining techniques.
• Data Value: Big data analytics creates value for data mining in order to find
the intrinsic and multidimensional attributes from the enormous amounts of
data. To a certain extent, a big data-driving model can perform as a support
vector machine to establish supervised learning, which allows the model to
adapt and evolve from time to time. Value creation is a significant process
contributing to organizations’ continuous improvement and demand predic-
tion. We need to understand that big data analytics is not only in statistical
analytics, but also in more complicated and tailor-made analytics. In general,
Big Data Analytics for Predictive Maintenance Strategies
54
big data analytics come into existence for resolving data storage problems as
well as providing valuable insights for organizations.
The acquisition and processing of big data largely improves the transparency
along the supply chains providing accurate and timely information for managerial
decision making. Companies and organizations operate on the huge amounts of
data by classifying trends and identifying patterns to produce invaluable knowledge.
Meanwhile the flood of big data with high speed and many variations has chal-
lenged the limited storage and conventional data mining methods. Challenges are
also from processing and analyzing the large amount of unstructured data which are
the major components in the big data acquired. Technologies have been advancing
towards better performance in the big data context regarding integrated platforms,
predictive analytics, and visualization (Lee, Kao, & Yang, 2014). Big data and
predictive analysis are strongly interconnected. Without proper analytics, big data
is just a deluge of data, while without big data, predictive analytics, the strength
of statistics, modeling, and data mining tools for analyzing current and historical
conditions will be undermined.
BIG DATA ANALYTICS
The descriptive tasks of big data analytics identify the common characteristics of
data with the purpose of deriving patterns and relationships existed in the data. The
descriptive functions of big data mining include classification analysis, clustering
analysis, association analysis, and logistic regression.
• Classification Analysis: Classification is a typical learning models used in
big data analytics, which aims to build a model for making prediction on data
feature from the predefined set of classes according to certain criteria. A rule-
base classification is used to extract IF-THEN rules to classify as different
categories. The examples include neural network, decision trees and support
vector machine.
• Clustering Analysis: Clustering analysis is defined as the process of group-
ing data into separate cluster of similar objects, which helps to segment and
acquire the data features. Data can be divided into different subgroups ac-
cording to the characteristics. The practitioners may formulate appropriate
strategies for different clusters. The common example of clustering technique
are K-means algorithm, self-organizing map, hill climbing algorithm and
density-based spatial clustering.
Big Data Analytics for Predictive Maintenance Strategies
55
• Association analysis: Association model helps the practitioners to recognize
groups of items that occur synchronously. Association algorithm is devel-
oped for searching frequent sets of items with minimum specified confidence
level. The criteria support and confidence level helps to identify the most
important relationships among the related items.
• Regression Analysis: Regression represents the logical relationship of the
historical data. The focus in regression analysis is to measure the dependent
variable given one or several independent variables, which support the con-
ditional estimation of expected outcome using the regression function. Linear
regression, non-linear regression and exponential regression are the common
statistical method to measure the best fit for a set of data.
MAINTENANCE STRATEGIES
The Big Data platform has the ability to handle huge amounts of data in manufactur-
ing or production logistics databases along with the development of computerized
maintenance management systems (CMMS), which assist decisions making so as
to formulate maintenance strategies.
Maintenance procedures will be undertaken when a machine failure has occurred
in the CrM strategy. Manufacturers are required to keep components inventories
for maintenance, repair and operations (MRO) in order to prevent disruption of the
overall production by failure of machine parts or equipment. Compared to the CrM
strategy, maintenance performance in the PvM strategy follows a fixed time, interval
basis or condition based schedules to avoid fatal machine failure. The design of PvM
is a protective, process-oriented approach in which machine failure and downtime
cost could be reduced by taking proper prevention and prevention to smoothen the
production. Decisions on maintenance schedules is based on a machine’s physical
properties or asset condition. Extra resources are spent on non-value-added activi-
ties to estimate and measure the condition rules for PvM (Exton & Labib, 2002).
However, PvM attempts to provide an empirical basis for the development of a
framework design of manufacturing flexibility at machine idle periods and during
maintenance activities. The assumption behind a PvM policy is that the machine
failure follows the bathtub curve in Figure 1. Scheduled maintenance happen in the
wear-out phase in order to reduce the failure rate (Sikorska, Hodkiewicz, & Ma,
2011). However, the most conspicuous deficiency in PvM is still the apparent ran-
dom failure within the useful life period. The impact of failure in a critical machine
is a tremendous risk to the downtime costs and, it in turn becomes bottleneck in
production logistics operations.
Big Data Analytics for Predictive Maintenance Strategies
56
To remedy random machine failure in maintenance management, PdM has been
well developed carrying out observation of the machine degradation process and
symptoms from normal to flawed situations (Wu, Gebraeel, Lawley, & Yih, 2007).
PdM is a sensor-based content-awareness philosophy based on the foundation of
“Internet of Things”. The intelligent maintenance prediction support system moni-
tors the machine status by utilizing real time sensory data (Kaiser & Gebraeel, 2009).
Advance maintenance in PdM policy is able to provide insights for maintenance
scheduling in advance in order to eliminate unanticipated machine breakdowns, and
minimize maintenance costs as well as downtime, before the occurrence of random
machine failure (Garcia, Sanz-Bobi, & del Pico, 2006).
The important factors associated with PdM in Maintenance Policy Management
(MPM) emphases criticality, availability of sensory data, reliability, timeliness,
relevance and knowledge-oriented strategy.
• Criticality in Failures: PdM strategy has been heavily concentrated in real
time machine condition monitoring through diagnostics and prognostics for
reimbursement of foreseeable machine downtime cost. In reliability study,
the critical assets must have a higher rank in priority formulate PdM strategy
to predict the most likely time for the next machine breakdown and random
error, as this will have the greatest impact on the production operations This
Figure 1. Classical Bathtub Curve
(Klutke, Kiessler, & Wortman, 2003)
Big Data Analytics for Predictive Maintenance Strategies
57
changes the maintenance objective from avoiding breakdown to accepting
downtime and taking maintenance action ahead of the schedule.
• Availability of Sensory Data: PdM policy is highly dependent on extract-
transform-load (ETL) operational data in close condition-based monitoring.
The current operational status and abnormal performance could be assessed
by equipping sensors to identify failure modules or machines. Lack of sen-
sory data may result in unpromising maintenance prediction.
• Reliability: Maintaining critical machine performance and leveraging the
overall cost to sustain production are the major targets of PdM policy. The
system must provide the correct measures and reliable performance in pre-
diction to address feasible and foreseeable machine failure, and build confi-
dence in operation.
• Timeliness: The prediction for maintenance modules must have a high level
of confidence level before the undesired event occurs, and the data size and
data transmission speed administered in a timely manner. The time series of
the maintenance schedule and delivery of MRO should be taken into con-
sideration the maintenance management in order to facilitate the production,
with zero tolerance of equipment failure.
• Relevance: The MPM system needs to be developed based on the opinions of
experts. The collected sensory information must be recorded and analyzed on
a real time basis. In order to improve data quality, extraction of relevant data
for maintenance decision making is crucial in regard to engineering aspects.
Inappropriate integration of a sensor and machine may cause poor estimation
and inaccuracy prediction of the current machine performance.
• Knowledge-Objective Oriented Strategy: The concept of the PdM strategy
involves a belief that the implicit knowledge from collaboration of sensory
information did contribute to the maintenance in advanced. The knowledge
transfer system facilitates the disclosure of implicit information to maximize
production efficiency and minimizes the adverse impact of idling time un-
der maintenance and unawareness of potential failure. The decision of PdM
policy could be assessed by the involvement of Big Data Mining Techniques
to detect and defeat anomalies at an early stage.
PREDICTIVE MAINTENENACE IN BIG DATA FRAMEWORK
The ideology of a PdM is to create transparency of the machine condition and in
the utilization of available information for maintenance decision making. In Figure
2, the framework of a big data platform in PdM is designed for closer integration
of data acquisition and the maintenance decision support system (MDSS), which
Big Data Analytics for Predictive Maintenance Strategies
58
highlights the dataflow process in diagnostics and prognostics modeling for PdM.
Case examples are provided in the following section to describe the framework of
MPM for a semiconductor machine under Big Data Platform, and operations of
MDSS. The operating data from vibration, heat and pressure sensors, which provide
sensory information stored in the Big Database, are embedded on the semiconductor
machine to evaluate the machine condition. The diagnostics and prognostics pro-
cess may involve real time data and historical data for data mining procedures. The
advantage of Big Data Architecture are capable to manage huge units of data and
perform ETL in a timely manner by using appropriate data processing algorithms,
such as Map-Reduce technique. The Big Data platform can build an intelligent agent
Figure 2. Predictive maintenance model in big data platform
Big Data Analytics for Predictive Maintenance Strategies
59
to connect the PdM module and the sensor knowledge from the actual machine.
The PdM module consists of two major systems: a diagnostics system and a prog-
nostics system. PdM analysis is a powerful tool to identify machine/components
failure, and provide surprisingly accurate future breakdown using time-series data
by well-developed analytic processes. The predictive maintenance module closely
supervises the machine condition and constantly aids the analytics process of di-
agnosis and prognosis. The results from Artificial Intelligence (AI) will then be
further processed and evaluated by the MDSS. Certain operational guidance and
estimation of failure events, such as foreseeable situations, time to breakdown and
estimated downtime, are recommended for maintenance decision making. With the
implementation of the big data platform, more precise sensory data acquisition and
accurate maintenance decisions could be made from the suggested algorithms in
order to manage critical machines, with the objectives of maintenance in advanced
or just-in-time (JIT) maintenance for supply chain management.
It is practical for a decision makers to select appropriate analytic processes and
recognize the functionalities of algorithms for maintenance planning. Therefore, a
comprehensive discussion of algorithm design is presented herein. The common
practices of AI techniques can be classified as Knowledge Based System (KBS),
Data Mining (DM) and Machine Learning (ML) (Faiz & Edirisinghe, 2009).
• Knowledge Based System: These types of analytics process require logical
deduction and cognitive reasoning to resolve complex problems and support
decision making. KBS attempts to extract rules for algorithm contexts by
human intelligence and expert opinion, which are of practical significance.
The practical necessity of KBS is increased due to the advancement of sen-
sor-based PdM. KBS encourages a more flexible way to increase quality in
problem solving and in extracting relevant data into knowledge for decision
making, i.e., machine failure identification, classification in maintenance
policies. A variation of sensory information causes data-booming in the ana-
lytics process. The rule-based and inference engine expert system have the
ability to simulate a human expert in reducing the complexity in MPM and in
discovering hidden machine failures.
• Data Mining: The goal of the DM process is to create a constructive model
of patterns recognition and feature analysis, which is able to classify data
into groups, detect irregular features and measure the dependencies of data.
Moving toward a total productive manufacturing system, DM is an instru-
ment that is able to mine all kinds of manufacturing knowledge, such as job
shop scheduling, manufacturing process, quality control, yield improvement,
and even predictive maintenance strategy. In the data mining technique, the
accuracy of the information discovered increases along with the increase in
Big Data Analytics for Predictive Maintenance Strategies
60
gathered data from the sensors and the historical maintenance data, which is
able to foresee failure from pattern behavior of the operating machine data
and the increased reliability of MDSS.
• Machine Learning: ML is another dimension of the analytics process. The
mechanism of KBS and DM are either to discover knowledge and insight
beforehand for the working process of the algorithm, which is concerned in-
formation and knowledge extraction from massive data. However, ML deals
with automatic reasoning and artificial cognitive resolution by an intelligence
agent. ML works as an online measurement of a health detection system to re-
veal machine degradation and anomalies from the models. Self-learning and
reinforcement in ML, together with normal degradation allow the forecasting
of random machine failure effectively and efficiently in order to plan for the
best before failure occurs.
THE RELATIONSHIP BETWEEN
DIAGNOSTICS AND PROGNOSTICS
In most cases, there is a measureable process of degradation before a machine fails.
Figure 3 illustrates the degradation process of a system, sub-system, or component
into failure. Through functioning life, the system may continuously degrade to a
condition with an observable drop in its performance level and initial faults may
occur during the degradation progress. The incipient defects continuously prolifer-
ates and the severity gradually increases which causes the system to fail to perform
its required function and fails. In order to predict and prevent failures in advanced,
diagnostics and prognostics techniques are studied and employed to evaluate the
current health conditions and forecast future performance.
Figure 3. Fault to failure progression
Big Data Analytics for Predictive Maintenance Strategies
61
Figure 4 indicates the inputs and outputs of the machine diagnostics and prog-
nostics process. The diagnostics process determines how the system has degraded
and investigates the cause or nature of such a degraded condition. Faults are de-
tected, isolated and identified to create a diagnostic record, and the faults possi-
bilities are computed to find the potential failure pattern. The diagnostic operations
follow an analysing approach from effect to cause. On the other hand, prognostics
considers time as a vital factor and focuses on predicting the future condition of the
system or component and in calculating accurately the remaining useful life before
the failure. The prediction is conducted from cause to effect. The analysing and
computing process is based on current system conditions and future operational
requirements. Although similar information and knowledge bases are shared, the
difference between the two concepts now becomes obvious. Diagnostics is to in-
vestigate and determine a failure mode within a system; while prognostics is to
compute a rather accurate result of the remaining useful life before final failure.
The following sections present a diagnostics and prognostics system in Big Data
Analytics with a case example. The system flowchart is shown in Figure 5. Fault
identification is one of the diagnostics modules. The classification of machine
failure is critical to smoothen the production, as not all the failures require emer-
gency maintenance. The result from the diagnostic model is able to assist in the
development of prognostics model. The reliability of the prediction could be increas-
ingly improved from a known machine failure.
Figure 4. Inputs and outputs with the diagnostic–prognostic process structure
Big Data Analytics for Predictive Maintenance Strategies
62
The machine failure identification has been developed and incorporated with
the current decision making grid (DMG) in MDSS for the case company. The DMG
model works like a map which categorizes the machines according to a set of pre-
defined parameters/criterions. The criteria for this project are the downtime and the
frequency of failures from the sensor data and historical record (Exton & Labib,
Figure 5. Flowchart of the proposed MDSS for case company
Big Data Analytics for Predictive Maintenance Strategies
63
2002). Other criteria, such as cost, availability of MRO and the bottleneck problem,
can also be included by expanding it from merely a 2 dimensional analysis to a 3
or more dimensional analysis model. Higher dimensional models can produce a
more comprehensive and accurate analysis. The DMG then proposes the different
types of maintenance policies based on the state in the grid which then determine
the appropriate maintenance actions for the MDSS. These maintenance policies
subsequently lead to be the formation of the following strategies; Operate to failure
(OTF), fixed time maintenance (FTM), skill level upgrade (SLU), condition based
monitoring (CBM) and design out machine/component (DOM), as shown in Figure
6.
• Operate to Failure: There are too many low downtime and low downtime
frequency components which made it impossible or too expensive for apply-
ing scheduled PdM maintenance. These components/machines are allowed to
operate to failure as they are deemed as having minimal impact on the sys-
tem. Ideally, if operators of the machine are also maintenance technicians,
which agrees with the TPM policy, OTF ‘faults’ can be further reduced with-
out the need for reporting and waiting for the technician to bring the machine
back to normal operating state. A CrM approach is suggested for OTF
situation.
• Fixed Time Maintenance: PvM is also called Fixed Time Maintenance. A
more flexible PvM can be chosen which is determined either by availability,
machine not in use, or by the severity of the faults. Faults with higher down-
time and frequency faults are allowed to have shorter fixed time maintenance,
Figure 6. Decision Making Grid by the case company
Big Data Analytics for Predictive Maintenance Strategies
64
whereas faults with lower downtime and frequency can extend the duration of
their fixed time maintenance.
• Skill Level Upgrade: This strategy falls on the high frequency and low
downtime grid. Faults that fall under this strategy are faults caused by human
errors, such as the accidental pressing of the wrong switch/button. For this
grid, the human factor oriented policy will be emphasized, either through the
human/operator or through the machine. From the human perspective, op-
erators can undergo skill level upgrading so that they can ratify the problem
personally without any assistance from a maintenance technician. Whereas
from the machine side, human errors can be reduced by designing more hu-
man oriented machine systems.
• Condition Based Monitoring: This strategy falls in the low frequency and
high downtime grid. This matches RCM policy where studies and measure-
ment need to be done to determine the underlying reliability condition of the
machines. Sensors are usually be applied to feed monitoring data for machine
learning in a PdM approach.
• Design Out Machine/Component: This strategy falls on the high frequency
and high downtime grid. Machines/components that are prescribed in this
grid region are usually unable to manage the current production level or are
subjected to substantial wear and tear after prolonged usage. Either the need
to upgrade to a better machine/components or a replacement of a new ma-
chine has to be undertaken.
FUZZY LOGIC FOR DIAGNOSTICS MACHINE FAILURE
Fuzzy logic is an approximate reasoning model to estimate the possible outcome
based on a set of rules. This method has been proven to be a prominent control
system that have been implemented in different engineering application, hardware
monitoring and conditional-based assessment. The process of Fuzzy Logic is a
simple, rule-based by using IF-THEN statements that imitate the decision making of
humans to classify the type of responsive performance by defining the intermediate
possibilities. Uncertainty in maintenance management, such as non-linear, impre-
cise and incomplete knowledge representation can be resolved by the adoption of
fuzzy logic. As a consequence, conclusions are derived with certainty factor from
a predefined fuzzy sets.
The research focus of predictive maintenance is to discover and recognize critical
random failure, involving low frequency and high downtime of maintenance. The
guidance from diagnostics in MDSS presents satisfactory knowledge rules to suggest
maintenance action afterwards. Fuzzy Logic in KBS is merely a suitable algorithm
Big Data Analytics for Predictive Maintenance Strategies
65
for machine faliure identification in an imprecise enironment. The linkage between
machine failure and classification of maintenance policies are frequently vague and
subject to various sensory information. In practice, however, a need to refine the
DGM model is required due to two scenarios. The first scenario is when data is
located close to each another but at different sides of the policy boundary, leading
to applying different strategies, despite being closely similar with one another. The
next scenario is when two data points are located at both ends of the boundary within
the same grid, leading to applying the same strategies, despite being far apart, as
shown in Figure 7. Therefore, the concept of Fuzzy Logic can be applied to reshape
the rigid boundaries,and make it more logical for the system.
A Fuzzy Logic design for semiconductor equipment is provided with two factors:
downtime records and frequency records as in Figure 8. All membership functions
(MF) are selected as trapezium shape, with two numerical inputs and one numerical
output involved, as shown in Figure 9. Defuzzification is a method of extracting a
crisp value from a fuzzy set as a representative value. In general there are five
methods, for defuzzifying a fuzzy set A in a universe of discourse Z. These methods
are the Centroid of Area (COA) or Center of Gravity, Mean of Maximum (MOM),
Bisector of Area (BOA), Smallest of Maximum (SOM) and Largest of Maximum
(LOM). COA is selected as the defuzzification strategy for the model. A fuzzy if-
then rule is applied on the interface, as shown in Figure 10. When all rules and
membership functions are settled, the Fuzzy Logic Rule Viewer and Fuzzy Logic
Surface Viewer can be examined in Figure 11. With the help of fuzzy logic, the
Figure 7. Decision making grid formaintenance policies
Big Data Analytics for Predictive Maintenance Strategies
66
known machine failure problem associated with breakdown time and frequency can
be reviewed for the current machine. Figure 12 shows the result of machine fault
identification from semiconductor A and B
Figure 8. Fuzzy logic: fuzzy inference system editor
Figure 9. Fuzzy logic membership input function editor: downtime and frequency
Big Data Analytics for Predictive Maintenance Strategies
67
ARTIFICIAL NEURAL NETWORK FOR PROGNOSTICS
MACHINE FAILURE
Prognostics is a model that is able to predict machine failure compared with normal
behavior on a real time basis. Maintenance in advanced is better adopted to the
necessity of repairs for an in-service machine in order to mitigate the disruption
of the overall production. A health condition assessment is required to check the
machine status frequently. Machine failure can be caused by normal degradation
and random failure. Basic KBS and DM are incapable of distinguishing random
failure and normal degradation, since machine degradation follows time progres-
sion rather than by rules. Therefore, an adoptive approach of machine learning to
prognosticate the next random failure is required. The Artificial Neural Network
Figure 10. Fuzzy logic MF rules and output: maintenance policies
Figure 11. Fuzzy logic surface viewer
Big Data Analytics for Predictive Maintenance Strategies
68
(ANN) model is a supervised learning method to estimate the Remaining Useful
Lifetime (RUL) of a machine from degraded failure, or to differentiate anomalies
from normal machine behavior. ANN can be used as a time series or forecasting
of unanticipated machine failure with online sensory data, which is able to learn
patterns from the training data set to distinguish the normal machine behavior and
any anomalies. The prediction in machine failure involves Big Data management
with the purpose of strengthening the data quality. Continuous data collection from
different sensors installed in the machine supports the quality of prediction.
Figure 13 illustrates the mechanism of the ANN model. The basic ANN algo-
rithm estimates the feasible outputs for prediction maintenance from observation
data or sensory information in order to understand the machine condition. ANN
provides assurance in resolving prediction and performance assessment with suffi-
Figure 12. Maintenance policies for seminconductor machine A and B
Big Data Analytics for Predictive Maintenance Strategies
69
cient continuous input data. Vibration, heat and temperature sensors are established
for the evaluation of the semiconductor machine. ANN is a simple artificial neuron
processing system, which included input nodes, number of hidden layers, weighted
factors of adjacent layers and output nodes. The hidden layers function as connectors
between the input nodes and output nodes to transform and extract features of the
input space. The prediction accuracy of ANN is dependent on the training process.
The training process allows adjustment the synaptic weighting of input and develops
the overall network by the construction of hidden layer. Before the actual running
of the prediction, correct data and wrong data must be available in the training of
supervised learning algorithm to ascertain its reliability in prediction. As a rule of
thumb, 60% of the dataset is needed for training and the remaining dataset is for
running test. Big Databases are able to collect huge amounts of historical and online
data to enhance the accuracy of an ANN model through trial and error.
The predictive tasks in MDSS are presented in the following three main func-
tions:
• Health Condition Assessment: The main goal is to monitor the machine
performance and machine components degradation. In MPM, considerable
maintenance actions could be resolved by component replacement. The sen-
sor network can effectively enhance MRO order processing and improve flex-
ibility in maintenance scheduling for the component substitution. The ma-
chine downtime caused by component damage could be reduced to meet the
aim of agile production recovery.
• Anomalies Detection Assessment: Anomalies detection is an approach in
root cause analysis. Machine anomalies may take place when the machine
are not working normally. The cause of consistency anomalies is not straight-
Figure 13. Artificial neural network
Big Data Analytics for Predictive Maintenance Strategies
70
forward to detect. ML algorithms are capable of measuring the machine per-
formance from the current and normal conditions for identifying anomalies.
The features of abnormality in machine performance are provided for further
inspection by the repair technicians.
• Remaining Useful Lifetime Assessment: The major challenge in asset man-
agement is to optimize the machine lifetime utilization. Different machine
usage may conclude variability of depreciation. In addition, longer lead time
in machine replacement causes an uncertainty in the production and supply
chain performance. The estimation of remaining useful lifetime by sensory
information can effectively leverage the machine lifetime utilization and pre-
diction for the machine replacement and performing a just-in-time machine
replacement policy.
• Scheduling Optimization for Maintenance in Advance: Prediction of fore-
seeable machine failure is used to assess the machine expected performance
in the future and optimize the maintenance schedules with less adverse ef-
fect on production disruption. Unanticipated machine failure may give a lon-
ger downtime and cease the manufacturing process. The prediction from the
ANN model emphasizes proactive maintenance and provides the right time
to conduct inspection. ANN captures the possible reasons, like the timeline
of the machine failure event, reasons, duration and relevance information of
the machines by triggered sensory information. With systematic maintenance
in advanced, unplanned machine stopping can be eliminated.
MANAGERIAL IMPLICATIONS AND RECOMMENDATIONS
The entire Big Data framework requires human intelligence and expert opinion in
the design stage. It is difficult for the manufacturer to manage and, control big data
and select relevant information for MDSS. The availability of sensors and techno-
logical advancement enable explorative research of Big Data Analytics, and allows
organizations to expand their capability to enhance the data transparency of machine
status for manufacturers. The speedy data flow and collection of abundant data
through WSNs enhance the potential of analytical performance. The adoption of Big
Data in MDSS state a significant step in machine health condition diagnosis. The
proposed approach helps to mitigate machine failure during production and uncover
the hidden patterns through Big Data Analytics. However, management faces three
major challenges of transformation traditional maintenance to advanced MDSS.
Various industries noticed that the size of data has been exponentially increasing
and accelerating due to the comprehensive use of sensor network. The transition from
conventional database to non-relational database is not only upgrading the storage
Big Data Analytics for Predictive Maintenance Strategies
71
capacity, but also requiring an infrastructure and expertise to process, and handle
structured and unstructured data. Handling and understanding on the petabytes or
even exabyte of data has become a challenge for IT teams. More advanced capabil-
ity in data warehouses and network connection expedites real time data processing.
Due to the complexity of data type, the current processing techniques could not
meet the demand of Big Data Infrastructure. Due to the enormous data booming via
WSNs, a cohesive platform for processing structure and unstructured data becomes
an essential element for any enterprises. The major purpose of Big Data Infrastruc-
ture is to resolve the problem of incompatible data formats and non-aligned data
structures. The impact on inconsistency of unstructured data requires pre-processing
of data input to enhance the performance of Big Data Mining.
Investigating the unstructured data in manufacturing not only create the value
for the production engineer but also support MDSS with more vigorous and so-
phisticated Big Data Analytics. Several research paper mentioned that analyzing the
unstructured data is the first priority in decision-making and prediction (Li, Bagheri,
Goote, Hasan, & Hazard, 2013; Muhtaroglu, Demir, Obali, & Girgin, 2013; Wielki,
2013). However, not all the unstructured data can be beneficial to knowledge de-
velopment and decision-making process. The relevant machine data must be fit for
purpose of maintenance policy selection. Proper domain experts in place are critical
to interpret the sensory information for predictive maintenance during the design
stage of Big Data Analytics. Furthermore, enhancing data quality by the adoption
of suitable sensors in the machine is also an importance for the company. Big Data
offers tremendous insight to the diagnostics and prognostics of the machine status.
Nonetheless, the information reliability from predictive maintenance is only avail-
able with appropriate sensors selection and adoption. In today’s Big Data Analytics,
research focus has been shifted from Volume of data to quality data.
Regarding the complexity of Big Data Analytics in MDSS, collaboration of in-
dustrial expertise and scholars must be involved to have sufficient breadth and depth
of domain knowledge to design an appropriate Big Data Analytics for maintenance
strategies. The proactive approach to predict machine failure provides a high level
reliability for excellence in maintenance management. Further benefit can be sum-
marized as reducing the frequency of corrective maintenance, increasing machine
performance and enhancing overall production reliability.
FUTURE RESEARCH DIRECTIONS
Future research is oriented to the utilization of manufacturing information in the
Cyber Physical System (CPS). Big Data Analytics are able to achieve better trans-
parency of production, which provides knowledge insight to practitioners. With the
Big Data Analytics for Predictive Maintenance Strategies
72
technological advancement in Internet of Things and the utilization of sophisticated
prediction tools, specialized automotive networks in a manufacturing company could
be developed for real time monitoring and control at the strategical level. Incorporat-
ing the manufacturing computational intelligence into the machine health monitoring
allow the manufacturers to enhance the overall system reliability and production
efficiency, especially in reducing the machine downtime. Predictive maintenance
is not only about health assessment but also detect the abnormality of the machine
before it breaks down. To have a step forward from predictive maintenance, pro-
duction plant should realize the importance of just-in-time maintenance strategy
for the whole production process. Implementation of CPS synthesizes data from
WSNs to enable the remaining life prediction so as to improve the asset utilization.
Risk assessment and impact evaluation of machine failure will also allow produc-
tion engineers to estimate the production system reliability. This motive turns the
predictive manufacturing system into a “self-aware-and-self-adjustment” system,
with intelligent machines and sensors in Big Data era.
CONCLUSION
In this book chapter, the predictive maintenance model and Big Data Analytics in
managerial aspects are presented. The feature extraction through Big Data Analyt-
ics can be beneficial to managing the machine condition and in predicting machine
failure. The MDSS in Big Data is able to suggest maintenance strategies and provide
insight for management to tackle maintenance issues. The proactive strategies in
maintenance can be achieved by embedded sensors and real time based machine
monitoring systems. Besides, the prediction from Big Data Analytics and suggested
analytics processes are well-designed to reduce the maintenance turnaround time
and substantially enhance the production system availability. MRO and maintenance
resources can be planned in advanced to facilitate the process during the machine
downtime. Moreover, it provides flexibility to design maintenance schedules to
mitigate the risk of unplanned stoppings. The overall maintenance efficiency can
be much improved by the implementation of predictive maintenance under a Big
Data platform.
ACKNOWLEDGMENT
The research is supported by The Hong Kong Polytechnic University. The authors
would like to thank the case company for providing the data. Our gratitude is also
extended to the Research Committee and the Department of Industrial and Systems
Big Data Analytics for Predictive Maintenance Strategies
73
Engineering of the Hong Kong Polytechnic University for support in this project
(RUE9) and (RU8H).
REFERENCES
Exton, T., & Labib, A. (2002). Spare parts decision analysis–The missing link in
CMMSs (Part II). Journal of Maintenance & Asset Management, 17(1), 14–21.
Faiz, R., & Edirisinghe, E. A. (2009). Decision making for predictive maintenance
in asset information management. Interdisciplinary Journal of Information, Knowl-
edge, and Management, 4, 23–36.
Fu, C., Ye, L., Liu, Y., Yu, R., Iung, B., Cheng, Y., & Zeng, Y. (2004). Predictive
maintenance in intelligent-control-maintenance-management system for hydroelec-
tric generating unit. Energy Conversion. IEEE Transactions on, 19(1), 179–186.
Garcia, M. C., Sanz-Bobi, M. A., & del Pico, J. (2006). SIMAP: Intelligent Sys-
tem for Predictive Maintenance: Application to the health condition monitoring
of a windturbine gearbox. Computers in Industry, 57(6), 552–568. doi:10.1016/j.
compind.2006.02.011
Jardine, A. K., & Tsang, A. H. (2013). Maintenance, replacement, and reliability:
theory and applications. CRC Press.
Kaiser, K. A., & Gebraeel, N. Z. (2009). Predictive maintenance management using
sensor-based degradation models. Systems, Man and Cybernetics, Part A: Systems
and Humans. IEEE Transactions on, 39(4), 840–849.
Klutke, G.-A., Kiessler, P. C., & Wortman, M. (2003). A critical look at the
bathtub curve. IEEE Transactions on Reliability, 52(1), 125–129. doi:10.1109/
TR.2002.804492
Lee, J., Ardakani, H. D., Yang, S., & Bagheri, B. (2015). Industrial big data analytics
and cyber-physical systems for future maintenance & service innovation. Procedia
CIRP, 38, 3–7. doi:10.1016/j.procir.2015.08.026
Lee, J., Kao, H.-A., & Yang, S. (2014). Service innovation and smart analytics for
industry 4.0 and big data environment. Procedia CIRP, 16, 3–8. doi:10.1016/j.
procir.2014.02.001
Li, L., Bagheri, S., Goote, H., Hasan, A., & Hazard, G. (2013). Risk adjustment
of patient expenditures: A big data analytics approach. Paper presented at the Big
Data, 2013 IEEE International Conference on. doi:10.1109/BigData.2013.6691790
Big Data Analytics for Predictive Maintenance Strategies
74
Muhtaroglu, F., Demir, S., Obali, M., & Girgin, C. (2013). Business model canvas
perspective on big data applications. Paper presented at the Big Data, 2013 IEEE
International Conference on. doi:10.1109/BigData.2013.6691684
Sikorska, J., Hodkiewicz, M., & Ma, L. (2011). Prognostic modelling options for
remaining useful life estimation by industry. Mechanical Systems and Signal Pro-
cessing, 25(5), 1803–1836. doi:10.1016/j.ymssp.2010.11.018
Wielki, J. (2013). Implementation of the big data concept in organizations-possi-
bilities, impediments and challenges. Paper presented at the Computer Science and
Information Systems (FedCSIS), 2013 Federated Conference on.
Wu, S.-j., Gebraeel, N., Lawley, M. A., & Yih, Y. (2007). A neural network integrated
decision support system for condition-based optimal predictive maintenance policy.
Systems, Man and Cybernetics, Part A: Systems and Humans. IEEE Transactions
on, 37(2), 226–236.
Xin, N. Y., & Ling, L. Y. (2013). How we could realize big data value. Paper pre-
sented at the Instrumentation and Measurement, Sensor Network and Automation
(IMSNA), 2013 2nd International Symposium on.
