See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/340443898

IoT based Predictive Maintenance in Manufacturing Sector

Conference Paper · April 2020

CITATIONS READS

2 418

3 authors:

Shikhil Nangia Sandhya Makkar

Lal Bahadur Shastri Institute of Management Lal Bahadur Shastri Institute of Management
2 PUBLICATIONS 13 CITATIONS 16 PUBLICATIONS 57 CITATIONS

SEE PROFILE SEE PROFILE

Rohail Hassan
Universiti Utara Malaysia
82 PUBLICATIONS 2,345 CITATIONS

SEE PROFILE

All content following this page was uploaded by Rohail Hassan on 07 April 2020.

The user has requested enhancement of the downloaded file.

 International Conference on Innovative Computing and Communication (ICICC 2020)

IoT based Predictive Maintenance in Manufacturing Sector

Shikhil Nangiaa,* ,Sandhya Makkarb, Rohail Hassanc

a,b
Lal Bahadur Shastri Institute of Management, 11/07, Dwarka Sector – 11,New Delhi, 110075, India
c
Othman Yeop Abdullah Graduate School of Business (OYAGSB), Universiti Utara Malaysia, 06010 UUM Sintok, Kedah Darul Aman, Malaysia.

*Corresponding author. E-mail address: [email protected]

Abstract: Industry 4.0 enables technological trends like Big Data Analytics and Machine Learning techniques to converge into and merge with traditional
manufacturing processes, resulting in smart manufacturing. Smart manufacturing techniques leverage the use of Industrial Internet of things (IIoT)
technology using IoT sensors that are fitted on physical assets to enhance manufacturing processes. IoT Sensors enable smart manufacturing facilities
capable of autonomously exchanging information, which can be used to drive business decisions more accurately. Businesses that adopt Smart
manufacturing techniques lead to a competitive advantage for these firms as they can bring in higher profit margins, reduced maintenance costs, energy
savings, and better-quality products. This study proposes an architecture for IIoT based predictive maintenance. A case study from ancillary automobile
industry is presented to demonstrate a predictive model for predicting sudden breakdown in industrial machines, thereby enabling the production and
maintenance cycle to be ‘smart.’

Keywords: Industry 4.0, Analytic Models, Machine learning, Smart Manufacturing, Sensors

methodology, and machine learning models are presented in Section 4 of

1. Introduction the case of manufacturing firm. Section 5 describes experiments, results,
and discussion. Finally, conclusion and future scope are presented in
In today’s highly competitive business environment, organizations want to Section 6.
increase production and operational efficiency, employee safety and drive
financial results. Many manufacturing organizations are yet to leverage 2. Literature Review
the use of big data and machine learning analytics due to lack of smart
analytic tools and capabilities (Jay Lee, Kao, & Yang, 2014). The IoT (Internet of Things) has spread its applications into various verticals
transition and advancements in manufacturing technologies have defined such as healthcare (Farahani & Firouzi et.al, 2018), manufacturing
the industrial revolution. Industry 1.0 began in the 1760s, and it led to (Selcuk, 2016), oil and gas (Chiang & Zhang, 2016) and many more. With
advent of mechanization, steam power, and weaving loom. In the year the use of Machine learning analytics, huge amounts of data generated
1870, mass production driven by electricity and assembly lines provided from IoT sensors fitted on manufacturing assets can be leveraged to detect
momentum for Industry 2.0. The third industrial revolution 3.0, beginning asset failure in advance. The objective is to predict failures in advance,
in the 1960s was the era of Automation and digital technologies – thereby increasing equipment uptime. The research on IoT based
electronic chips, mainframes, PCs and the Internet. All this paved the way Predictive maintenance is still in the preliminary stage. Earlier literature
for the fourth industrial revolution (4.0) in the 21st century by the (Lee, Ghaffari & Elmeligy, 2011) was mainly centered around Condition
confluence of emerging technological breakthroughs, from artificial Based Maintenance (CBM) Strategies (Shin, J.-H. & Jun, 2015), as
intelligence, 3D printing, nanotechnology, machine learning to quantum described in section 1, which was a part of Industry 3.0 revolution.
computing (Klaus Schwab, 2016). Literature by (Lee, Ghaffari & Elmeligy, 2011) discusses smart
maintenance systems that estimate the reliability of a system to detect any
The current study aims to deduce an approach for Predictive Maintenance failures beforehand. It is based on maintenance techniques such as
using IoT sensors that will aid manufacturing industries to predict Preventive Maintenance (PM) and CBM strategies. In the research by
machine failures before asset breakdown occurs. This paper discusses a (Selcuk, 2016), the latest trends, applications, and how PdM can help
case study from the manufacturing sector where the solution is developed improve reliability, efficiency, and protect the environment. Work done in
using Machine Learning’s (ML) binary classification Support Vector (Civerchia, Bocchino, Salvadori et.al, 2017) describes the system and
Machine, Classification and Regression Trees (C&RT), and Boosted component architecture of Industrial sensors in detail. Research by
Classification Trees. These algorithms are discussed in Section 5. (Mourtzis, Vlachou & Milas, 2016) discusses the industrial application of
IoT wireless sensors networks how its adoption generates volumes of big
The paper is organized as follows. Literature review on Predictive data that needs to be further analysed. The above work focuses on PdM
Maintenance (PdM) modeling and IoT sensors are presented in Section 2. and IoT applications, trends, CBM and Industrial IoT sensor architecture.
An Architecture for Maintenance modeling using IoT sensors is proposed Implementation of IoT sensor architecture using Machine learning
in Section 3. A brief overview of Problem definition, features, data set, algorithms has not yet been fully exploited.

NANGIA, MAKKAR, HASSAN 1

Electronic copy available at: https://ssrn.com/abstract=3563559

 International Conference on Innovative Computing and Communication (ICICC 2020)

With the advent of Industry 4.0, the goal of Predictive Maintenance space where data is stored. Data from IoT sensors is transmitted to a
centralized cloud server or decentralized fog nodes via Wi-Fi, or in remote
systems is reaching Zero-Defect Manufacturing (Wang, 2016). The author
locations through Cellular networks or Bluetooth Low Energy (BLE)
defines PdM as a set of activities that can detect changes in a physical technology.
state of an asset (signs of possible equipment failure) so that the
appropriate maintenance work for maximizing the service life of an asset 3.4. Edge/Fog/Cloud Computing:
can be done optimally. The author (Wang, 2016) propose an Intelligent
Predictive Maintenance framework (IPdM) that gave an initial direction IoT sensors like Arduino, Raspberry PI, ESP8266, ESP32 have inbuilt
for the current research work. Overall, the goal is to create and deploy a computing power. Based on the criticality of the asset, processing of PdM
model for prediction of machinery failure using data collected from algorithms can be done at Edge, Fog, or Cloud. In Edge devices, the
processing is done locally. It is used to provide near to real-time insights
industrial sensors. The next section discusses the architecture of the PdM about Machine failure. In cloud computing, the server is located in a
model. centralized location, having big data centers (Chiang & Zhang, 2016).
Data generated from the IoT sensors in PdM applications are sent to
Cloud for further processing. Processing at cloud has high latency. Fog
3. IoT based Predictive Maintenance Architecture Computing has distributed architecture where fog nodes are distributed in
many locations, over large geographic areas (David Linthicum, 2016). It
The study proposes an architecture for Industrial IoT based predictive is the intermediate layer between Cloud and IoT Edge devices (Fig. 2) if
the assets are separated by the large distance it is recommended to
maintenance (Fig. 1). IoT sensors are used to monitor and collect real-
implement fog architecture over the cloud due to low latency and higher
time data. The data collected is then modelled into PdM model using computing power (Atlam, Walters & Wills, 2018).
Statistical tools and software. The following are the components of PdM
systems.

Fig. 1 - Architecture Explained

3.1. IoT Sensors

Sensors or Cyber-physical systems can interact and communicate with Fig. 2 - Fog Computing, an extension of Cloud computing
each other. Sensors capture raw data. Few Popular Industrial IoT Enabled
Sensors measure Temperature, Humidity, Pressure, Current, Vibration, 3.5. Predictive Maintenance:
Air Quality, Footfall, Gas, and Weight. Another commonly used sensor is
the Photoionization Sensor, which is used to measure levels of volatile After the data has been transferred to Fog/Cloud/Edge nodes or a
inorganic compounds and gases. The ultrasonic sound sensor is used to centralized database, the next step is PdM algorithm design. The goal of
translate high-frequency sounds due to leaks. predictive maintenance is to predict asset failure. ML algorithms
(Supervised or unsupervised) are applied to model the data. The
3.2. Digital Signal: processing of the data from IoT sensors is done at fog/cloud or edge
nodes. The algorithms process data sent from IoT sensors embedded in
Data from sensors captured in the form of analog signals is converted into the asset and predict failures.
digital signals using Analog to digital converter. The digital signal is
machine-readable and is used for further analysis. The steps discussed above are generalized and applies to all PdM systems
in manufacturing. A case from the Auto Ancillary Company is discussed
3.3. Data Storage and Transfer: next.

Digital data stream generated is stored on organizations' local intranet or a

secure centralized or remote cloud location in the form of a database file.
IoT Sensors which are installed in remote locations without internet
connectivity have inbuilt Random Access Memory (RAM) and Storage

NANGIA, MAKKAR, HASSAN 2

Electronic copy available at: https://ssrn.com/abstract=3563559

 International Conference on Innovative Computing and Communication (ICICC 2020)

4. Case Problem Definition • RP: Outside room Pressure.

• IP: Inside Pressure of machine conduits for leak detection.
XYZ Pvt Ltd (name changed) is a manufacturer of automobile parts based • Temperature: actual temperature of machinery
in India. The manufacturing process at XYZ Company involves metallic • Output: The target variable predicting whether the equipment is
products, plastic, and polymer products manufacturing. The assembly line running or failed. It is a binary coded variable, where 0 depicts running,
of its metal manufacturing products consists of components like – motors, and 1 represents failure.
rotors, heat exchangers, robotic arms, etc. If any of these components get
faulty, the entire assembly line needs to be stopped resulting in downtime. 4.3. Methodology
The automobile sector incurs huge losses due to downtime (Advanced
Technology Services, 2006). XYZ Ltd. wants to predict machine failure in The approach is spread across six phases. (1) In the first phase, critical
advance to take preventive measures before actual machine downtime assets of the organization are identified. These include assets that are
occurs, thereby aiming to reduce its maintenance costs, increase critical to the organization, have frequent failure rates, and whose
productivity, improve product quality and save energy from smoother maintenance involves high costs. (2) Data is collected from IoT sensors
machines. installed on machines, as discussed in Section 3, is carried out in the
second phase. This data is stored in a database or cloud. (3) In the third
4.1. Data Collection from Critical Assets phase, the raw data is pre-processed. Much of raw data in databases may
contain missing values, outliers, obsolete or redundant data and
XYZ Ltd., as a pilot planned to implement IoT based predictive inconsistent values (Larose & Larose, 2015). The data need to be
maintenance on its heat exchanger unit whose function was to cool down prepared, so that clean data generates higher reliability and accuracy of
extremely high-temperature synthetic fluids being discharged from the prediction. (4) In the next phase, data were split into training and test data.
assembly line. The workstation had frequent issues of downtime on the Supervised machine learning algorithms were applied to create the PdM
clogging of its conduits due to chemical deposits and issues of heat model. The model is trained using training data and model performance is
leakage due to thermal cracks or perforations. The heat exchanger unit checked on the test data. (5) In the fifth phase, the output is analyzed and
was critical in operation, and there were employee safety risks associated is evaluated based on the accuracy of prediction of the ML algorithms.
with its frequent failure. XYZ Ltd. wanted a cost-effective solution to This is done using Receiver Operating Curve (ROC) metrics, Gains
monitor and predict failure. The workstation was connected with IoT charts, visualizations such as Key Performance Indicator charts,
sensors that are discussed next. dashboards, confusion matrix, misclassification error rates, and there are
several other performance metrics, which can be used as required. (6) In
4.2. Dataset and Features the final phase, the model is deployed to the enterprise. Deployment of the
model is done in the production stage. Machine learning tools like
The dataset used for demonstration is publicly available as original data Microsoft Azure ML Studio (Software), Tibco Statistica (Software), SAS
could not be shared. The sample size of the dataset is 944 observations Visual Data Mining and Machine learning (Software), Google AI
and ten features. The dataset has been cleaned and does not contain any platform (Software) and open-source software like R and Python can be
missing values or outliers. The dataset can be downloaded from used to model the data and trigger alarms on a prediction of failure.
https://bit.ly/2nwwwiE. Description of features in the dataset is as
follows: 5. Model Building
• footfall: The total number of workers in the shop floor at a particular
instance of time TIBCO Statistica version 13.5 (Software) is used to model the IoT sensor
• tempMode: temperature mode (1-7), where one is minimum, and data, using supervised learning algorithms. Statistica is a powerful
seven is maximum predictive analytics software that allows end-users to build models that
• AQ: Air Quality sensor, with modes between 1 to 7, where modes 1-5 can be packaged and published by businesses. TIBCO Statistica provides
are acceptable levels of air quality and mode 6 -7 are considered not a visual drag and drop mode of interaction for modelling known as a
acceptable workspace (Fig. 3).
• : Ultrasound Sensor, used to detect high or low-frequency sounds due
to leakage. 5.1. Steps for Model Building
• CS: Current Sensor to monitor the current flow. Its level is between 1-
7, where 1 in minimal level and 7 is maximum current flow The model (Fig. 3) that is deployed as a pilot for XYZ Ltd. A brief
• VOC: (PhotoIonization Detector) sensors used for detecting low description of each of the vital components of the workflow is as follows.
concentrations of Volatile Organic Compound to detect clogging in • Input dataset into the TIBCO Statistica workspace.
conduits of heat exchanger unit.

NANGIA, MAKKAR, HASSAN 3

Electronic copy available at: https://ssrn.com/abstract=3563559

 International Conference on Innovative Computing and Communication (ICICC 2020)

Figure 3: Model Building on Tibco Statistica

NANGIA, MAKKAR, HASSAN 4

Electronic copy available at: https://ssrn.com/abstract=3563559

 International Conference on Innovative Computing and Communication (ICICC 2020)

Figure 4: Gains Chart

NANGIA, MAKKAR, HASSAN 5

Electronic copy available at: https://ssrn.com/abstract=3563559

 International Conference on Innovative Computing and Communication (ICICC 2020)

• Training Testing Sample: Data was split 75:25 ratio as Training and Table 1 - Summary of Model Performance Metrics on Testing data
Test sample Metric C&RT Boosted Classification Trees SVM
• Predictor screening: An option to select dependent/independent
Precision 0.891 0.899 0.893
variables to be used for analysis. Variable ‘Output’ was set as the
Recall 0.914 0.908 0.894
categorical dependent variable. Rest of the variables in our dataset like
‘Footfall,’ ‘tempMode, ‘AQ’, ‘USS’, ‘CS’, ‘VOC’, ‘RP’, ‘IP’ and F1 Score 0.903 0.903 0.893
‘Temperature’ were set as continuous predictors Error Rate 0.099 0.097 0.106
• Algorithm: Users can select various Data mining and Machine
learning algorithms to model data. We have used binary classification
algorithms to model our data. A brief detail of classification algorithms
Fig. 4. shows the Gains chart. The chart interprets the utility that one can
used for our PdM model is
expect by using C&RT, SVM and Boosted tree algorithms, as compared
Support Vector Machine (SVM): SVM represents a powerful technique
to using baseline information only or in other words the gains chart
for the general classification, regression and outliers detection with an
represents the ratio of results obtained with and without application of any
intuitive model representation (Suykens & Vandewalle, 1999).
models. The curve with maximum area under curve represents optimal
Random Forest: An Ensemble method for classification and regression
model. As can be observed, Boosted Classification tree algorithm has
(Liaw & Wiener, 2002).
maximum area under curve.
Advanced Classification Trees (C&RT): Advanced decision trees for
PdM architectural framework and steps proposed by (Wang, 2016)
classification and regression (Lemon, Roy, et.al, 2003).
was implemented on our model using higher accuracy obtained by voted
Each algorithm generates a PMML (Predictive Model Markup Language)
average of three machine learning algorithms. The model was evaluated
script file. PMML is an open standard to represent predictive models.
based on Misclassification error rate, Precision, Recall, F1 score, and
• Rapid Deployment: This node applies PMML codes to testing data
Gains chart. It is worthwhile to note that incase of PdM where cost of a
and outputs predicted values and model comparison metrics. Rapid
false negative (when the production line failure occurs and the model is
deployment node evaluates methods that measure model performance.
unable to predict it) is higher than false positive and hence it is best to
Deployment of a model in Production is done by publishing PMML codes
focus on Recall metric to evaluate the model. The results show that the
to an enterprise by a web service.
model can predict failure with a maximum accuracy rate of 90.18% on the
• Reporting: Output model summary report and prediction accuracy.
IoT sensor dataset and using the C&RT algorithm model has the best
The algorithms are compared using the same training and testing samples.
recall.
The output data can be saved and exported to Tibco Spotfire (a tool for
The next step is to deploy the model to production via a web service
data modelling and visualization) or database export or simple export into
using PMML codes that were generated by Tibco Statistica. Further setup
static Excel/CSV files.
can be done to trigger alarm (through email, SMS notification) on
machine fault prediction and monitor system status and production line
5.2. Results and Discussion
data using real time dashboards.

In this section, we analyse the predictive maintenance model by

comparing the results of three machine learning models on the training-
6. Conclusion
testing dataset. A comparison and summary of error rates on testing data
[Table 1] show that Boosted Classification Trees output minimum error in
The proposed work discusses IoT based predictive maintenance and
predicting machine failure. The error rate is calculated based on a voted
implements it using a case from the manufacturing industry. The research
prediction of the output of all the three algorithms. If two out of three
paper initially discusses Industry 4.0 and an interconnected IoT based
algorithms predict Machine failure and the third one does not predict
Sensor ecosystem that can share and exchange information amongst each
machine failure, based on a maximum count of votes, the output would be
other. Data collected from sensors can be leveraged by manufacturing
that Machine failure would occur.
organizations to predict machine failure.
The next important metrics to analyse are Precision, recall and F1
Opportunities in IoT and implementation of Predictive maintenance in
score. Precision is the ratio of correctly classified observations. A recall is
industries will be huge. The ability to predict failure in advance will help
the ratio of correctly predicted positive events. A recall is also known as
organizations increase productivity, reduce equipment costs, improve
sensitivity (Ruiter, 2015). Recall metric signifies the ratio of correctly
working conditions and work environment safety, better product quality,
predicted positive events. F1 score is a harmonic average of Precision and
reduced waste in terms of consumables and energy savings. On average,
recall. F1 metric is used when class distribution is uneven. In the IoT
the industry has seen a 70%-75% reduction in machine downtime (Selcuk,
sensor dataset, the ratio of ‘Fail’ and ‘Pass’ classes are not equal. A
2016).
comparison of Precision, Recall and F1 metric is shown [Table 1].
The case discusses a pilot implication of PdM for a specific asset in

NANGIA, MAKKAR, HASSAN 6

Electronic copy available at: https://ssrn.com/abstract=3563559

 International Conference on Innovative Computing and Communication (ICICC 2020)

the assembly line of the manufacturing industry. A model is created using Lemon, S. C., Roy, J., Clark, M. A., Friedmann, P. D., & Rakowski, W. (2003).
advanced analytics software, which tends to simplify the process. Classification and regression tree analysis in public health: methodological
review and comparison with logistic regression. Annals of behavioral
A successful implementation of Industrial PdM would involve any
medicine, 26(3), 172-181.
organisation to scale across multiple assets. A model having thousands of
Advanced Technology Services, Inc. (2006, March 29). Downtime Costs Auto
IoT sensors would generate huge volumes of data. Such huge volumes of Industry $22k/Minute - Survey. Retrieved from
data need to be stored and analysed using Big data analytics. Apart from http://news.thomasnet.com/
this future scope for our research would also involves implementing a Microsoft Azure Machine Learning Studio (Software). Retrieved July 29, 2019,
Prescriptive maintenance-based solution for predicting failure and what from https://studio.azureml.net/.
action need to be taken to resolve it. Tibco Statistica (Software). Retrieved July 29, 2019, from
https://tibco.com/products
SAS Visual Data Mining and Machine Learning (Software). Retrieved July 29,
REFERENCES 2019, from https://www.sas.com/en_us/software/visual-data-mining-
machine-learning.html
Google AI and Machine Learning (Software). Retrieved July 29, 2019, from
Atlam, H., Walters, R., & Wills, G. (2018). Fog computing and the internet of https://cloud.google.com/ai-platform/
things: a review. big data and cognitive computing, 2(2), 10. David Linthicum (2016). Edge Computing Vs. Fog Computing: Definitions
Farahani, B., Firouzi, F., Chang, V., Badaroglu, M., Constant, N., & and Enterprise Uses. Retrieved July 29,2019, from
Mankodiya, K. (2018). Towards fog-driven IoT eHealth: Promises and https://www.cisco.com/c/en/us/solutions/enterprise-networks/edge-
challenges of IoT in medicine and healthcare. Future Generation Computer computing.html
Systems, 78, 659-676. Andreas de Ruiter, Microsoft Inc. (2015) Performance measures in Azure ML:
Selcuk, S. (2017). Predictive maintenance, its implementation and latest Accuracy, Precision, Recall [Web log post]. Retrieved July 29, 2019, from
trends. Proceedings of the Institution of Mechanical Engineers, Part B: https://blogs.msdn.microsoft.com/andreasderuiter/2015/02/09/performance
Journal of Engineering Manufacture, 231(9), 1670-1679. -measures-in-azure-ml-accuracy-precision-recall-and-f1-score/.
Civerchia, F., Bocchino, S., Salvadori, C., Rossi, E., Maggiani, L., & Petracca,
M. (2017). Industrial Internet of Things monitoring solution for advanced Supplement Information: The data set used in the study can be downloaded
predictive maintenance applications. Journal of Industrial Information from https://bit.ly/2nwwwiE. The feature names have been modified for
Integration, 7, 4-12. better understanding of the case. Data Source: Sapiro, Virginia, Steven J.
Schwab, K. (2016). The Fourth Industrial Revolution by Klaus Schwab. Rosenstone, Donald R. Kinder, Warren E. Miller, and the National
In Geneva: World Economic Forum. Election Studies. American National Election Studies, 1992-1997:
Chiang, M., & Zhang, T. (2016). Fog and IoT: An overview of research Combined File [Computer file]. 2nd ICPSR version. Ann Arbor, MI:
opportunities. IEEE Internet of Things Journal, 3(6), 854-864. University of Michigan, Center for Political Studies [producer], 1999.
Mourtzis, D., Vlachou, E., & Milas, N. (2016). Industrial Big Data as a result
of IoT adoption in manufacturing. Procedia cirp, 55, 290-295.
Shin, J. H., & Jun, H. B. (2015). On condition based maintenance
policy. Journal of Computational Design and Engineering, 2(2), 119-127.
Wang, K. (2016). Intelligent predictive maintenance (IPdM) system–Industry
4.0 scenario. WIT Transactions on Engineering Sciences, 113, 259-268.
Larose, D. T. (2015). Data mining and predictive analytics. John Wiley &
Sons.
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.
Lee, J., Ghaffari, M., & Elmeligy, S. (2011). Self-maintenance and engineering
immune systems: Towards smarter machines and manufacturing
systems. Annual Reviews in Control, 35(1), 111-122.
Suykens, J. A., & Vandewalle, J. (1999). Least squares support vector machine
classifiers. Neural processing letters, 9(3), 293-300.
Liaw, A., & Wiener, M. (2002). Classification and regression by
randomForest. R news, 2(3), 18-22.

NANGIA, MAKKAR, HASSAN 7

Electronic copy available at: https://ssrn.com/abstract=3563559

View publication stats