Download as pdf or txt
Download as pdf or txt
You are on page 1of 36

The current issue and full text archive of this journal is available on Emerald Insight at:

https://www.emerald.com/insight/1355-2511.htm

LSTM based artificial intelligence Industry 4.0


transformation
predictive maintenance technique
for availability rate and OEE
improvement in a TPM
implementing plant through Received 5 July 2022
Revised 6 January 2023

Industry 4.0 transformation Accepted 10 February 2023

Roosefert Mohan, J. Preetha Roselyn and R. Annie Uthra


Department of EEE, SRM Institute of Science and Technology, Chennai, India

Abstract
Purpose – The artificial intelligence (AI) based total productive maintenance (TPM) condition based
maintenance (CBM) approach through Industry 4.0 transformation can well predict the breakdown in advance
to eliminate breakdown.
Design/methodology/approach – Meeting the customer requirement as per the delivery schedule with the
existing resources are always a big challenge in industries. Any catastrophic breakdown in the equipment
leads to increase in production loss, damage to machines, repair cost, time and affects delivery. If these
breakdowns are predicted in advance, the breakdown can be addressed before its occurrence and the demand
supply chain can be met. TPM is one of the essential operational excellence tool used in industries to utilize the
existing resources of a plant in a optimal way. The conventional time based maintenance (TBM) and CBM
approach of TPM in Industry 3.0 is time consuming and not accurate enough to achieve zero down time.
Findings – The proposed AI and IIoT based TPM is achieved in a digitalized data oriented platform to monitor
and control the health status of the machine which may reduce the catastrophic breakdown by 95% and also
improves the quality rate and machine performance rate. Based on the identified key signature parameters
related to major breakdown are measured using the sensors, digitalised by programmable logic controller
(PLC) and monitored by supervisory control and data acquisition (SCADA) and predicted in server or cloud.
Originality/value – Long short term memory based deep learning network was developed as a regression
forecasting model to predict the remaining useful life RUL of the part or assembly and based on the predictions,
corrective action has been implemented before the occurrence of breakdown. The reliability and consistency of
the proposed approach are validated and horizontally deployed in similar machines to achieve zero downtime.
Keywords Intelligent total productive maintenance, Industry 4.0 intelligent condition based maintenance,
Predictive maintenance, OEE, Long short term memory
Paper type Research paper

1. Introduction
With the rapid growth in automobile consumption, the demand in the automobile
manufacturing industry is growing rapidly (Gackowiec, 2019). Through Industry 3.0 to
Industry 4.0 transformation, the breakdown of a machine is eliminated which leads to
stoppage or reduction in speed of the machine or production, delay in delivery schedule,
decrease in production, increase in operational cost and affect the product quality (Roosefert
Mohan et al., 2021). Any unplanned breakdown or functional deterioration of a machine may
lead to stoppage or reduction in speed of machine or production, delay in delivery schedule,
decrease in production, increase in operational cost, affect the product quality and also reduce
earnings before interest, tax, depreciation and amortization EBITDA of an industry.
International Society of Automation claims that the global downtime cost per year is $ 647 Journal of Quality in Maintenance
Engineering
Billion. Studies claim that effective predictive maintenance reduces machine downtime by © Emerald Publishing Limited
1355-2511
30–50% and increases machine efficiency by 20–40%. DOI 10.1108/JQME-07-2022-0041
JQME Total productive maintenance (TPM) is one of the maintenance approaches used in
industries to reduce downtime (Mungani and Visser, 2013). Corrective maintenance (CM),
time based maintenance (TBM), breakdown maintenance (BM), condition based maintenance
(CBM), inspection and repair (IR) and maintenance prevention (MP) are the conventional
maintenance techniques which are used to achieve zero downtime (Roosefert Mohan et al.,
2021). Predictive maintenance can be achieved effectively using the Internet of things (IoT)
(Ran et al., 2019). Autonomous maintenance pillar activity, clean lubricate inspect and
tightening (CLIT) and PM pillar activity like TBM and condition based maintenance (CBM)
are carried out to replace the parts as per schedule to achieve zero downtime. The check
sheets are prepared by experts as per the recommendation of the equipment supplier, past
breakdown history and updated based on the result of the “five why analysis” of the
breakdown (Coccia, 2017). Figure 1 shows the conventional TPM approach where the
breakdown phenomenon are captured with duration and occurrences and restored as BM
technique and after restoration, Pareto and five why to analysis are carried out.
The human error in data capturing and the poor skill lead to breakdown. As the CBM
activity is time-based and not a continuous one, there is a chance of breakdown between
consecutive CBM or TBM. The proactive maintenance approaches are used to prevent the
failure of equipment before the occurrence of breakdown using the predefined breakdown-
related parameter threshold limit (Passlick et al., 2020). Conventional maintenance
approaches (Ran et al., 2019) have inaccurate mathematical degradation processes and also
manual intervention. There are many key performance indices (KPI) (Rimawan and Irawan,
2017) to measure the performance of the process or machines. One of the important KPI is
overall equipment efficiency (OEE) (Sonmez et al., 2018; Ng Corrales et al., 2020) which
measures how the manufacturing line is running compared to how well it could be running.
The causes of breakdown can be analyzed, addressed and eliminated through TPM based
approach in the existing Industry 3.0 (Rimawan and Irawan, 2017). The Predictive
Maintenance activities are based on the data collected to check the health status of the
equipment (Jimenez- Cortadi et al., 2020). The data may misguide the maintenance personnel
due to the following reasons: (1) In the traditional PdM, data is collected over a period of time
(1–3 months) and hence if a breakdown appears, after the data is collected it remains
undetected. (2) Due to the varying skill of the data-collecting persons, inaccurate data may be
collected for analysis. (3) At various load or speed conditions, the collected data may be
inaccurate (e.g. motor current may vary under the condition of semi-load and full load) and the
analysis may be misled. Though maintenance-related checklists are adhered to by 100% in

Figure 1.
Conventional
TPM-based approach
industry 3.0 using systematic TPM, it is a challenging task to achieve nil downtime since Industry 4.0
some of the breakdown-related key signature parameters are not monitored preciously unlike transformation
intelligent, machine learning-based CBM that may overcome these issues (Kateris et al., 2014).
Artificial Intelligence (AI) based TPM may reduce breakdown and improve the efficiency of
the machine (Angelo et al., 2021).
IoT in Industry 4.0 standard digitalizes the factory and monitor then control the operation
of the plant and process (Lu, 2017). For efficient analysing of the complex system in a plant
accurate and digitalized data are collected through intelligent sensors (Pech et al., 2021). In the
Machine learning and artificial intelligence-based predictive maintenance approach,
the signature parameter is filtered, analysed and predicted to trigger the alert to the
authenticated person through servers and clouds alert to address the breakdown before the
occurrence (Carvalhoa et al., 2019; Samanpour et al., 2018). The four pillars of IoT are Wireless
Sensor Network (WSN), Machine to Machine communication (M2M), Supervisory Control and
Data Acquisition (SCADA) and Radio Frequency Identifier (RFiD) to eliminate breakdown,
increase productivity and optimise resources. Predictive maintenance has a remarkable
benefit with the use of real-time detection and prediction algorithms (Çınar et al., 2020).
The vibration signal captured from a single point of an assembly can be used for predicting
the RUL of rotary equipment (Ruiz-Gonzalez et al., 2014). The spare parts can be shared
efficiently for loading in an optimised cost manner by using an accurate prediction method
(Uit Het Broek et al., 2021). The benefits of IoT-based predictive maintenance are as follows:
(1) sensory-based data collection reduces the maintenance cost of breakdown detection and
thereby decreases breakdown cost; (2) automatic predictive maintenance reduces
unnecessary maintenance and TBM by predicting the RUL well ahead of failure;
(3) automatic alert generation to the authenticated persons when parts need to be replaced
or maintenance activity to be carried out.
In smart manufacturing, the Internet of things, data mining and artificial intelligence
techniques are used in PdM for condition-based monitoring. By predicting the remaining
useful life (Qin et al., 2017) in advance, catastrophes can be eliminated to maximize machine
uptime. The collected data of signature parameters related to breakdown are processed using
Machine Learning (ML) models like Multi-Layer Perceptron (MLP), Autoregression
Integrated Moving Average (ARIMA) model and Support Vector Regression (SVR) model
(Santos Junior et al., 2019; Jeyasekar et al., 2017).
SVR is used to predict the cutting tool’s life (Liu et al., 2019). AI based predictive
maintenance technique is implemented to handle the health status of industrial equipment
(Hashemian and Bean, 2011). LSTM prediction model is used to predict the RUL of parts at
different working conditions (Çınar et al., 2020; Zhou et al., 2019). RUL prediction can be
achieved using model-based, data-driven or hybrid-based models (de Pater and Mitici, 2021).
Model based methods include filters whereas data-based prediction methods do not require
mathematical models or in-depth physical knowledge about equipment (Li and Zhao, 2020).
Support Vector Machine (SVM) performs well in classification models with a balanced data
set (Da Silva et al., 2020; Zhao et al., 2019). But in classification, there is a chance of ambiguity
situation (Luo et al., 2021). LSTM can be used as a regression model to forecast the RUL of
engines, parts or rotary equipment efficiently (Namuduri et al., 2020). The more human
dependencies are removed, the more accuracy of prediction can be achieved as mentioned in
Li et al. (2020). As in Jardine et al. (2006), data capturing and acquisition, processing and
decision-making are the three steps in CBM.
The prediction accuracy of conventional CBM is not accurate enough because of manual
error, poor skill and manual decision-making. Out of eight pillars of TPM, the four key pillars
of TPM namely Jishu Hozen (JH) pillar which is called as autonomous maintenance pillar, KK
Pillar, Quality Management (QM) pillar and Planned Maintenance (PM) pillar are sufficient to
improve OEE. The part of equipment which affects the quality of the product is called
JQME Quality-components (Q-components). In QM pillar activity, by maintaining the healthy status
of these Q-components which is related to quality, the quality rate is increased. Though KK
pillar and JH pillar activities like kaizens, putting tags, Hard To Access, Source of
Contamination, Visual Controls, One Point Lessons (OPL) and CLIT as discussed in Roosefert
Mohan et al. (2021) have adhered 100%, it is observed that it is difficult to achieve zero
downtime and improve OEE to an optimized level by implementing the conventional TPM.
Also, to run a machine with zero breakdowns above 20 years of life and improve the OEE to
an optimized level in industries are very essential and hence the PM pillar needs to be
addressed along with AI and IoT which enhances the system performance by eliminating the
human intervention, manual decision making and error in data capturing.
In the proposed work, Pareto analysis is used to identify the key signature parameters
related to the equipment which cause 95% of breakdown and by implementing LSTM based
prediction model the RUL is predicted to achieve zero downtime. Instead of using multiple
prediction models, LSTM models address all the signature parameters for prediction which
eliminates the overloading of the existing server with multiple prediction software models.
IoT based prediction is used to eliminate oil temperature breakdown which improves
availability rate and OEE further. The key signature parameter data captured by the analysis
of the sensor are fed into the Programmable Logic Controller (PLC) where digitization takes
place and forwarded to the SCADA and then to the server/cloud to carry out the LSTM based
prediction model to find the RUL. After predicting the RUL an alert or alarming message is
triggered to the authenticated persons in the remote area for preventive measure action
during a planned time to avoid catastrophic breakdown. By addressing these key signature
parameters, breakdown is eliminated to increase the availability rate, performance rate,
quality rate and OEE of the equipment since around 80% of the high-pressure moulding
machine’s operations are related to these key signature parameters.
The major contributions and novelty of the proposed work are as follows.
(1) Critical equipment is identified using ranking score calculation of the TPM approach
through which it is easy to identify the priority of equipment and based on that
signature parameters can be prioritized which optimizes the number of sensors to be
used which saves costs, time and overloading of systems.
(2) Development of a multivariant model with a single software prediction model instead
of multiple software to avoid overloading the existing server based on TPM
implemented Pareto analysis as a proactive approach to achieve zero breakdowns till
achieving world-class Availability Rate and OEE.
(3) Development of LSTM based prediction model to achieve zero downtime in a real-
time application based on Condition monitoring to improve the availability rate and
performance rate by eliminating minor stoppages and machine speed related to
equipment healthy status and quality rate by maintaining the Quality-components
(Q-Components) which affect the quality of the product as proactive approach.
(4) Real-time Taylor made a kaizen-based solution, based on five why analysis, suitable
to all heavy engineering industries to achieve zero downtime which increases
productivity with the available resources.

2. Proposed methodology
An IoT and AI-based prediction model in the industry predicts the breakdown of equipment
by collecting the data of related signature parameters from sensors. In the proposed work, Oil
Contamination Level (OCL), hydraulic pump pressure, mixer equipment motor-gearbox
assembly vibration and hydraulic oil temperature are the identified key signature parameters Industry 4.0
in the high-pressure mould-making line. The sensors deployed in the machinery, collect the transformation
data and transfer the information to the server. Figure 2 shows the layout of the artificial
intelligent-based TPM approach towards achieving zero breakdowns.
The sensors or Cyber-Physical System (CPS) captures the signals related to key
signature parameters and these signals are connected to PLC to get digitized data. SCADA
receives the captured data from PLC and sends these data to the server and then to the
cloud. The proposed LSTM model is implemented in the server for prediction and an alert is
triggered to the authorized person for proactive actions. If the breakdown related to a
particular phenomenon achieves zero downtime, it is standardized else based on the
brainstorming technique, additional sensors are deployed to capture the signature
parameter to achieve zero downtime for the same breakdown phenomenon. The
authenticated person can fetch any data related to predictive maintenance from the
cloud for further analysis at anytime from anywhere on his mobile or PC. The same
approach is horizontally deployed to all signature parameters and the validation is carried
out once in six months.

2.1 Ranking of machines to identify signature parameters and OEE calculation


In the proposed methodology, the critical equipment in a plant is identified based on the
ranking of the machine. Through ranking of machines, the criticality of the machine can be
identified which enables the maintenance team to ensure more priority to the critical
machines to identify key signature parameters for the RUL prediction model, spare parts
maintenance and planned maintenance to achieve zero downtime. Table 1 depicts the
scorecard for equipment ranking calculation based on the productivity, quality of parts
produced, maintenance, safety and age of the machine.
Figure 3 shows the equipment criticality and the type of maintenance required. Using this
ranking method, it is easy to identify the machine sensors which can be provided for key
signature parameter identification. Hence sensors are deployed for required part of the
equipment and overloading of PLC with more analog inputs and processing time is
eliminated and also overloading of SCADA with more number of tags and CPU of server can
be eliminated. Figure 4 depicts the method of performing ranking calculation and grouping
the machine rank/class as S, A, B and C based on the score. The minimum score obtained by
any equipment is 0 and the maximum is 25.

Figure 2.
Artificial intelligence
based TPM approach
in PM pillar
JQME Ranking score of equipment
Evaluation items Contents Score

Production If the machine stops the entire plant will be stopped 5


If the machine stops the entire plant will produce less 3
Even if the machine stops production will not be affected 0
Quality By equipment problem, quality will be badly affected 5
Quality will be affected by the equipment problem 3
Quality will not be affected by the equipment problem 0
Maintenance Unexpected failure will cause large repair cost 5
Unexpected equipment failure will cause moderate repair cost 3
The cost of Repair is negligible 0
Safety Maintenance Death, injury or fire may occur in the event of machine failure 5
Failure of the machine can result in an accident 3
No damage and no secondary troubles are possible 0
Number of usage years More than 20 years 5
Table 1. More than 10 years and less than 20 years 3
Ranking score Less than 10 years 0
of equipment Source(s): Table created by Author

Figure 3.
Equipment criticality
and maintenance types

Figure 5 shows how many machines are coming under ranks or class S, A, B and C. Rank A
type machine is a more critical machine, B is less critical and C is a very less critical machine
which may be neglected and breakdown maintenance can be done. Special maintenance is to
be addressed in rank S equipment. Based on the ranking of machines, required key signature
parameters of the high-pressure moulding machines are identified which is sufficient for the
entire moulding line for digital transformation towards Industry 4.0 standard and to improve
the OEE. Few key signature parameters out of 468 numbers are shown in Figure 6.
To improve the OEE, the losses are analysed. Figure 7 displays the sixteen types of
losses defined by TPM related to equipment, manpower, material and energy. Each type of
loss is assigned a number for analysis purposes and addressed. Loss number 10 and Loss
number 13 are OEE affecting major losses related to manpower. In the proposed existing
moulding line setup, the loss numbers 10 and 13 can be ignored since measurement can be
done offline without disturbing the running line. The losses numbers 9 to 16, 2 and 4 can be
ignored for the OEE improvement which is not applicable to the existing foundry moulding
Industry 4.0
transformation

Figure 4.
Ranking score
calculation of
equipment
JQME

Figure 5.
Equipment rank

line setup. Loss number 3 which is tool change is automated in the equipment to achieve it
without affecting the cycle time. Hence the major loss is loss number 1 which is breakdown
loss. Figure 8 shows the procedure of calculating the availability rate, performance rate,
quality rate and OEE.

2.2 Pareto analysis to identify key signature parameters for zero breakdown
In the high-pressure moulding line of the foundry, the performance of the machine like
breakdown and defects is based on the quality of hydraulic oil since most of the operations
are controlled by hydraulic actuators or cylinders operated with an accuracy of ±1 mm.
The key signature parameters of oil namely, oil contamination, oil pressure variation,
temperature of the oil and motor-gearbox vibration are considered for analysis. Table 2
describes the data collected for the phenomenon of breakdown during the year 2017 as the
benchmark which restricts zero downtime. Table 3 depicts the relationship between the
number of phenomena related to key signature parameters. The proposed AI-based
prediction model predicts the potential damages causing the equipment by collecting the key
signature parameters data from sensors namely, motor-gearbox vibration, hydraulic oil
temperature, oil contamination level and hydraulic pressure.
In AI-based TPM approach, based on the Pareto chart analysis, more priority is given for
parameters related to 95% of the breakdown as shown in Figure 9. The identified key
parameters data are fed to the PLC, SCADA, local server with GPU and cloud. Using the key
signature parameters, AI-based time series prediction models are implemented in the local
server and alerts are sent to the authorized persons for the corrective measure before the
equipment failure.

2.3 Architecture of LSTM


In the proposed work, LSTM is used as the regression model using the signature parameters
captured by the sensors. The architecture of LSTM is shown in Figure 10. Once the data is
collected, it is analysed with LSTM based deep learning model that is used to identify the RUL
of the part. The collected historical data are trained and tested offline in the best-to-fit model.
From this data, the prediction is achieved using the proposed LSTM model where the online
signature parameter is collected through sensors and fed to LSTM. The predictive model is
regularly validated once in three months and tested for its accuracy.
The LSTM cell unit is built with an input gate, forget gate, cell and output gate through
which the information flows into and out of the cell. The cell has a linear activation function
before the forget gate as shown in Figure 11. The forget gate executes the linear function of its
Industry 4.0
transformation

Figure 6.
Details of signature
parameters with
corresponding sensors
JQME

Figure 7.
Sixteen losses in TPM

inputs and using the activation function gives its output either as 0 or 1. The cell also receives
inputs from other cells in the network and is added together and passed through the tanh
activation function which brings the values between 1 and 1. The memory cell status
related to Input Gate and Forget Gate status is shown in Table 4.
The output value of each cell in LSTM is mentioned in Equations (1) to (7).
Ft ¼ σ ðWf Xt þ Uf Ht−1 þ bF Þ (1)
It ¼ σ ðWI XI þ UI Ht−1 þ bI Þ (2)
Qt ¼ σ ðWQ Xt þ UQ Ht−1 þ bQ Þ (3)
 
^t ¼ tanh W ^ Xt þ U ^ Ht−1 þ b ^
C (4)
C C C

^t
Ct ¼ Ft Ct−1 þ It C (5)
Ht ¼ Qt ðtanhðCt ÞÞ (6)

The format value of x at a given time t is shown in Equation (7).


xðtÞ ¼ xðt  1Þ; xðt  2Þ; . . . :; xðt  nÞ (7)

where Xt is the input vector, Ht1 is the previous cell output, Ht is the current cell output, Ct1
is the previous cell memory, Ct is the current cell memory and U and W are Weight Vectors, b
is Bias for the gates. In the proposed model, Stochastic Gradient Descent with Momentum
(SGDM) and Adaptive Moment Estimation (ADAM) training optimizers are implemented and
compared for predicted value accuracy.

2.4 Performance metrics for predictive maintenance


The various Key Performance Indicators (KPIs) are Overall Equipment Efficiency (OEE),
Availability Rate (AR), Quality Rate (QR) and Performance Rate (PR).
Industry 4.0
transformation

Figure 8.
OEE calculation
JQME Year 2017 average Year 2017 average
S. No. Cause description mins/month occurrences/month

1 Oil Contamination 1,268 11


2 Vibration 732 6
3 Oil Pressure 668 4
4 Oil Temperature 438 4
5 Electrical Loose Connection 128 3
6 Drive faults 60 2
7 Lubrication System Failure 33 1
8 PLC and Control System Fault 9 0.5
Table 2. 9 Design Issues 6 0.3
Benchmark of major 10 Poor Skill 4 0.3
loss capturing 11 Others 2 0.7
signature parameters Source(s): Table created by Author

Signature parameter related to number of potential phenomenon


Related number of
Sl Signature breakdown Related number of Related number of minor
No parameter phenomenon defect phenomenon stoppage phenomenon

1 Hydraulic Oil 22 4 32
Pressure
Table 3.
Relation between 2 Oil Contamination 32 4 42
signature parameter 3 Motor Vibration 6 3 18
and breakdown 4 Oil Temperature 18 2 28
phenomenon Source(s): Table created by Author

Figure 9.
Pareto analysis
for captured loss
Industry 4.0
transformation

Figure 10.
Architecture
of proposed
prediction model

Figure 11.
LSTM cell structure

LSTM memory cell status


Forget gate Input gate Cell state

0 0 Delete the existing value


0 1 Over-write the value
1 0 The previous value is maintained Table 4.
1 1 Adding the existing vale to the previous value Memory cell status
Source(s): Table created by Author of LSTM network
JQME 
Overall Equipment Efficiency; OEE ð%Þ ¼ AR 3 PR 3 QR 10000

Good Produced Count 3 Cycle Time


¼ 3 100
Total Machine Planned Running Time
(8)
Machine Operating Time
Availability Rate AR ð%Þ ¼ 3 100 (9)
Planned Production Time
where,

Machine Operating Time ¼ Panned Available Time for Production – Machine Stopped Time (10)
Total Count Produced 3 Cycle Time
Performance Rate; PR ð%Þ ¼ 3 100 (11)
Machine Run Time
Good Produced Count
Quality Rate; QR ð%Þ ¼ 3 100 (12)
Total Produced Count
To evaluate the performance of the prediction model, two metrics namely Root Mean
Squared Error (RMSE) and Root Mean Squared Percent Error (RMSPE) are calculated.
RMSE measures the difference between the observed and predicted values as in
Equation (13)
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n  2
1 X ^t
RMSE ¼ Yt Y (13)
n i¼1

Root Mean Squared Percent Error (RMSPE) measures the percentage difference between the
observed and predicted values as in Equation (14).
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
!2
u
u1 X n
Yt  Y ^t
RMSPE ¼ t (14)
n i¼1 Yt

^ t denotes the predicted value at time t.


where Yt denotes the observed value at time t and Y

3. Description of industrial machinery setup


The layout of the foundry is shown in Figure 12 where the moulding machine is making the
mould using a green sand mould. In the green sand mould, liquid metal is poured to make
Gray or SG iron castings using a pouring unit from where the liquid metal is received from
induction furnaces after iron scrap melting and a chemical addition process. In the punchout
unit castings and sand are separated. The return sand is sent to the cooler through conveyor
belts which use air and water for cooling the hot sand and sent to the hopper through a bucket
elevator for storage hopper. Based on the requirement the mixer takes this returned sand
from the storage hopper and adds new sand, water, bentonite, coal dust etc and sends it to the
moulding machine for making green sand mould.
Figure 13 depicts the installed location of oil contamination, oil pressure and oil
temperature RTD sensors in the hydraulic power pack.
Industry 4.0
transformation

Figure 12.
Layout of foundry
machinery

Figure 13.
Layout of hydraulic
power pack

Plate 1 shows the OCL sensor make by Hydac and type CS1000 with the output of 4 to 20 mA
provided to measure the OCL in NAS 1638 standard. The cleanliness level of hydraulic oil
from NAS 0 to NAS 14 is proportional to the sensor output of 4.8 to 19.2 mA. Plate 2 shows
JQME that Hydac make transducer for pressure type HDA3744-A-400, 0–400 bar, the accuracy of
±5%, two wire type is mounted in hydraulic pump delivery line to measure 0–400 bar
pressure and gives the proportional output of 4 to 20 mA The standard set pressure is
230 bar ± 5% and hence the signal generated by the sensor is equal to 13.2 mA ± 5%.
Plate 3 shows the vibration sensor mounted on the shaft side of the 160 kW motor where
fluid couplings with pullies and V-belt are used to drive the gearbox of the mixer. The Siemens
make vibration sensor type SIPLUS CMS2000, with a maximum measurement range of 10 gm,
sensitivity is 500 mV/gm ± 10% is fixed at the drive end of the motor. The threshold value of
vertically foot-mounted 160 kW, 415 V, 3 phase, 1,485 rpm, 50 Hz motor with frame size 315 M,
CLASS III with IM2011(IMV15) is 3.6 M/Sec2 as per the standard ISO 10816-1 is 183.542 mV.

Plate 1.
Oil contamination
sensor

Plate 2.
Hydac pressure
transducer HDA3744-
A-400 in the hydraulic
system

Plate 3.
Vibration sensor
4. Results and discussion Industry 4.0
In the proposed work oil contamination level, oil temperature and motor-gear box assembly transformation
vibrations are addressed using LSTM-based prediction model and oil temperature is controlled
by IoT based prediction model. The key signature parameters identified namely, oil
contamination, motor vibration and oil pressure are collected over a time interval of 3 min
through PLC. The number of data fed to LSTM for prediction is 64 over a time interval of
192 min. The oil contamination level is captured by the CS 1000 Hydac sensor as shown in
Plate 1. A Hydac make pressure transducer type HDA3744-A-400 is used to measure the
hydraulic pressure of 0 to 400 bar as shown in Plate 2 to capture the hydraulic oil pressure data.
Siemens make SIPLUS CMS2000 vibration sensor with a maximum measurement range of 10
gm is provided to measure the vibration data as shown in Plate 3. Based on the
recommendation of the equipment or parts supplier and based on the history of a machine
breakdown, the threshold values are fixed. The proposed LSTM based prediction is developed
in MATLAB 2018B version which is tested with two training options namely, ADAM and
SGDM optimizers. The training and testing data are assigned as 80% and 20% respectively.

4.1 Hydraulic oil contamination


In the proposed work, the recommended OCL should be less than NAS 7, since 80% of the
high-pressure moulding line operations are controlled by hydraulics using high precision
Rexroth make servo and proportional valves with a positioning accuracy of ±1 mm for a
linear load of up to 15 Tons and with a velocity of 1800 mm/sec. The training is carried out for
200 epochs in both optimizer cases. ADAM and SGDM training with 80% of the data set and
testing with 20% data set is performed for the oil contamination LSTM prediction model.
It uses 200 epochs to reduce RMSE. Each epoch uses an iteration of 64 data similar to all
proposed LSTM prediction model and the remaining 20% of data is used for testing.
Figures 14 and 15 shows the ADAM and SGDM training respectively with 80% of the data
set and 20% for testing. It uses 200 epochs to reduce RMSE where each epoch uses an
iteration of 64 data. Each data is captured at a time interval of 3 min. The prediction will be
performed at a time interval of every 192 min and the RUL will be updated to avoid
catastrophic breakdown.

Figure 14.
Adam training for OCL
JQME

Figure 15.
Sgdm training for OCL

The forecasted result of oil contamination using ADAM and SGDM optimizer are provided in
Figure 16(a) and (b) respectively. The horizontal axis represents the forecasted period in
minutes.
Using the ADAM training function, it is predicted that the oil contamination is reaching the
NAS 7 equivalent threshold value of 12 mA in 20,738 min which is equivalent to 14 days, 9 h and
38 min. Similarly, using the SGDM training function the oil contamination is reaching the
threshold value of 12 mA in 20,564 min which is equivalent to 14 days, 6 h and 44 min. From
these observed values necessary maintenance-related activities can be done before breakdown.
Figure 17 shows the ground truth value of the Hydraulic oil contamination prediction model
and compares the predicted value obtained from ADAM and SGDM optimizers.
The actual and predicted values of oil contamination value using both optimizers are
provided in Table 5 which clearly shows that in SGDM, the forecasted values are closer to the
actual values when compared to the conventional and ADAM-based LSTM predicted model.
The RMSE and RMPSE values obtained using ADAM and SGDM models for oil
contamination are provided in Table 6. Based on the analysis, it is understood that SGDM
training performs better than the ADAM function in the LSTM model for oil contamination
prediction.

4.2 Hydraulic oil pressure


In the proposed oil pressure prediction model, 80% of the collected offline hydraulic oil
pressure data is used for training and the remaining 20% of data is used for testing.
The model is trained with both ADAM and SGDM optimizers and the results are compared
using the metrics RMSE and RMPSE. In the proposed work, the required oil pressure value is
230 bars in ±5%. If the pressure exceeds or reduces by 10%, then the system will start
malfunctioning which may affect the speed of the machine and the quality of the product as it
depends on the positioning velocity and squeezing pressure which may affect the positioning
accuracy controlled by proportional and directional control valves.
Figures 18 and 19 shows the training of ADAM and SGDM optimizer respectively for oil
pressure. Similarly, LSTM-based testing and training are performed for the mixer gearbox
motor vibration prediction model also. Figure 20(a) and (b) show the forecasted hydraulic oil
Industry 4.0
transformation

Figure 16.
LSTM-based predicted
value of oil
contamination value
JQME

Figure 17.
Ground truth value vs
LSTM predicted value
of oil contamination

RUL of oil prediction accuracy conventional versus LSTM (based on oil contamination)
Conventional LSTM LSTM Predicted error Predicted Predicted
TBM-based Actual RUL RUL % error % in error % in
Sl CBM model RUL in (ADAM) (SGDM) conventional LSTM LSTM
no. RUL in mins mins in min in min CBM model (ADAM) (SGDM)

1 150,000 231,281 231,988 231,716 54.187 0.306 0.188


2 150,000 151,438 150,962 151,754 0.959 0.314 0.209
3 150,000 146,765 146,316 146,478 2.157 0.306 0.196
4 150,000 113,743 114,088 113,967 24.171 0.303 0.197
5 150,000 99,384 99,678 99,579 33.744 0.296 0.196
Mean Square Error 932.950 0.093 0.039
Table 5.
Comparison MSE
of predicted oil RMSE 30.544 0.305 0.197
contamination with Note(s): Negative value in Error% represents Premature Failure than the actual RUL
actual value Source(s): Table created by Author

Table 6. Training option RMSE RMPSE %


Comparison
of accuracy between ADAM 0.3057 14.2917
ADAM and SGDM SGDM 0.1972 13.2911
functions Source(s): Table created by Author

pressure value using both optimizers. The oil pressure is reaching the threshold value of
12.28 mA which is less than 10% of the standard hydraulic pressure of 230 Bar in 51,683 min
which is equivalent to 35 days, 21 h and 23 min predicted using the ADAM function. It means
the RUL of the hydraulic pump is 35 days, 21 h and 23 min. Similarly, the threshold value of
12.28 mA is reached using the SGDM function in 51,446 min which is equivalent to 35 days,
17 h and 26 min. From the predicted RUL, the breakdown related to hydraulic pressure can be
eliminated before its occurrence by servicing the pump or can be replaced during the planned
shutdown time .
Industry 4.0
transformation

Figure 18.
Adam training of oil
pressure

Figure 19.
SGDM training of oil
pressure
JQME

Figure 20.
(a)ADAM trained
prediction (b)SGDM
trained prediction
Figure 21 shows the ground truth value of the hydraulic oil pressure prediction model is Industry 4.0
compared with the predicted value obtained from ADAM and SGDM training optimizers. The transformation
predicted value with the SGDM training optimizer is almost equal to the ground truth value.
The accuracy of LSTM based oil pressure prediction model is compared with the
conventional system. The observed and predicted values of hydraulic oil pressure using the
conventional system, ADAM and SGDM-trained optimizers are provided in Table 7 which
depicts that in SGDM, the forecasted values are closer to the actual values than the
conventional and ADAM-based LSTM predicted model.
The RMSE and RMPSE obtained using ADAM and SGDM optimizers for oil pressure are
provided in Table 8. Based on the prediction analysis, it is clear that SGDM training performs
better than the ADAM function in the LSTM model for hydraulic oil pressure forecasting.

4.3 LSTM-based prediction result of motor vibration


In the proposed motor vibration prediction model, 80% and 20% of the motor vibration data
are used for training and testing the model respectively. The proposed model is trained with

Figure 21.
Oil pressure ground
truth vs LSTM
predicted value

RUL of hydraulic system prediction conventional and lstm-based oil pressure developed
Conventional LSTM LSTM Predicted error Predicted Predicted
TBM-based Actual RUL RUL % error % in error % in
Sl CBM model RUL in (SGDM) (ADAM) conventional LSTM LSTM
no. RUL in mins mins in min in min model (SGDM) (ADAM)

1 150,000 182,423 184,228 179,602 21.615 0.989 1.546


2 150,000 99,526 100,511 101,086 33.649 0.990 1.567
3 150,000 176,394 174,687 179,121 17.596 0.968 1.546
4 150,000 167,832 169,501 165,248 11.888 0.994 1.540
5 150,000 78,676 77,901 79,891 47.549 0.985 1.544
Mean Square Error 862.277 0.971 2.399
MSE Table 7.
RMSE 29.365 0.985 1.549 Comparison of oil
Note(s): Negative value in Error% represents Premature Failure than the actual RUL pressure actual and
Source(s): Table created by Author predicted value
JQME both ADAM and SGDM optimizers separately and the predicted RUL are compared. In this
proposed work, the motor vibration threshold value should be below 2.3 mm/sec or 3.6 mm/
sec2 which is equivalent to 183.542 mV. The predicted motor vibration value using the
optimizers ADAM and SGDM are provided in Figure 22(a) and (b) respectively. The motor
vibration is reaching the threshold value of 183.542 mV in 27,478 min which is equivalent to
19 days, 1 h and 58 min as predicted using the ADAM function for necessary maintenance
activities. From the SGDM optimizer, it is observed that the motor vibration is reaching the
threshold value of 183.542 mV in 27,617 min which is equivalent to 19 days, 4 h and 17 min.
From this predicted value, the cause of motor vibration can be addressed before the
breakdown.
Figure 23 shows the ground truth value of the mixer gearbox motor vibration prediction
model compared with the predicted value obtained from ADAM and SGDM optimizers. The
predicted value and the ground truth value with ADAM and SGDM optimizer are nearly
matching with the ground truth value whereas the predicted value using SGDM optimizer is
much closer to the ground truth value.
The observed and predicted values of motor vibration value using ADAM and SGDM are
provided in Table 9 which depicts that in SGDM, the forecasted values are closer to the
observed values than ADAM based LSTM predicted model. The RMSE and RMPSE obtained
using ADAM and SGDM model for oil motor vibration is provided in Table 10. Based on the
predicted value, it is clear that SGDM training performs better than the ADAM function in the
LSTM model for motor vibration prediction.

4.4 Comparison of LSTM, ARIMA and conventional prediction models


The model SGDM, ADAM optimizer-based LSTM is compared with the ARIMA model and
conventional model with 50 numbers of trials. It is observed that SGDM-based LSTM is
having better accuracy as shown in Table 11.

Table 8.
Comparison of Training option RMSE RMPSE %
accuracy between
ADAM and SGDM ADAM 1.5491 8.3733
functions for oil SGDM 0.9849 7.8460
pressure Source(s): Table created by Author

Figure 22.
LSTM-based predicted
value of motor gearbox
Vibration
Industry 4.0
transformation

Figure 23.
Motor-gearbox
vibration ground truth
versus LSTM

RUL of motor and gearbox assembly prediction -conventional and LSTM-based on vibration
Conventional LSTM LSTM Predicted error Predicted Predicted
TBM-based Actual RUL RUL % error % in error % in
Sl CBM model RUL in (SGDM) (ADAM) conventional LSTM LSTM
no. RUL in mins mins in min in min model (SGDM) (ADAM)

1 86,400 92,345 93,755 94,045 6.881 1.527 1.841


2 86,400 90,876 92,236 89,316 5.181 1.497 1.717
3 86,400 98,584 97,144 100,314 14.102 1.461 1.755
4 86,400 66,481 67,511 65,301 23.054 1.549 1.775
5 86,400 124,906 122,956 127,406 44.567 1.561 2.002
Mean Square Error 558.156 2.308 3.314
MSE Table 9.
RMSE 23.625 1.519 1.821 Comparison of motor
Note(s): Negative value in Error% represents Premature Failure than the actual RUL vibration actual and
Source(s): Table created by Author predicted value

Training option RMSE RMPSE %


Table 10.
ADAM 1.82 11.20 Comparison of ADAM
SGDM 1.52 8.965 and SGDM functions
Source(s): Table created by Author for motor vibration

The LSTM-based prediction model is validated from 50 epochs to 200 epochs with LSTM
layer 3, Layer 4 and Layer 5 for all three OCL, Oil Pressure and gearbox motor vibrations as
shown in Figure 24. At 200 epochs with LSTM 3 layers, the Mean Square Error shows a less
value for the SGDM optimizer. Similarly, it is validated for ADAM optimizer also where the
MSE value is less at 3 layers with 200 epochs.
JQME

Table 11.
Comparison of
prediction accuracy
Oil pressure prediction Oil contamination prediction Vibration prediction
Parameter Conventional ARIMA ADAM SGDM Conventional ARIMA ADAM SGDM Conventional ARIMA ADAM SGDM

RMSE 44046.821 276.197 176.27 69.116 33236.453 228.784 166.428 61.231 39891.917 312.462 201.218 104.464
RMSPE 47.839 0.229 0.124 0.041 34.234 0.168 0.101 0.039 41.453 0.273 0.146 0.071
Source(s): Table created by Author
4.5 IoT-based hydraulic oil temperature prediction Industry 4.0
The four identified key signature parameters are Oil temperature, Oil contamination level, transformation
Hydraulic oil pressure and Motor Vibration. But the oil temperature can be predicted using
various threshold values using IoT and an alert can be triggered to the authenticated to
address the breakdown before the occurrence.
4.5.1 Experimental industrial setup for hydraulic oil temperature. Figure 25(a) shows the
PT100 RTD sensor whose resistance can vary from 100 to 138.505 Ohm for 0 to 100 0C
mounted on the hydraulic oil tank with an oil capacity of 9,000 litres. SCADA screen as shown
in Figure 25(b) is arranged to display the four levels of oil temperature which is connected to
the PLC. The maximum allowed oil temperature in the oil is 55 0C and it cut off all machine
operations when the oil temperature is exceeded. This RTD sensor is connected to PLC for
digitalizing and processing purposes, the measured value is stored and monitored by SCADA
and the temperature is controlled by PLC.
There are four stages in SCADA namely (1) oil temperature very high status when the
temperature exceeds 558C which will suddenly trip the machine (2) oil temperature high
status, when the temperature is in between 458C to 54.98C where the alarm message is sent to
the authenticated person through local server and cloud for necessary corrective action to
service the cooling system before breakdown (3) oil temperature normal status in which, the
oil temperature is 208C to 44.98C which is normal to run the machine (4) oil temperature too
cool status in which the oil temperature is less than 208C hence, an alert is sent to the

Figure 24.
Comparison of LSTM-
SGDM model with
different layers and
epochs

Figure 25.
(a) RTD installed in
hydraulic tank (b)
hydraulic temperature
in SCADA
JQME authorized to service the failed heating system or to switch on the spare cooling system to
maintain oil flow and viscosity before the occurrence of breakdown. Figure 26 shows the
downtime reduction trend due to Hydraulic pressure from Q1 of the year 2015 to Q4 of 2020.
The breakdown due to hydraulic pressure approached zero from Q1 of 2020 after the
implementation of the LSTM prediction model for Hydraulic pressure.
Figure 27 displays the downtime due to the vibration of the motors over a period of time
from Q1 of the year 2015 to Q4 of 2020. The breakdown due to motor vibration approached

Figure 26.
Breakdown
trend-hydraulic
pressure

Figure 27.
Breakdown
trend-vibration
zero from Q3 of 2019 after the implementation of the LSTM prediction model for motor Industry 4.0
vibration. transformation
Similarly, in Q4 of 2018, LSTM based oil contamination level prediction model is
implemented in the high-pressure moulding line. The breakdown due to oil contamination
attained zero downtime from Q4 of 2018. Figure 28 displays the downtime trend due to
hydraulic oil temperature over a period of time from Q1 of the year 2015 to Q4 of 2020.
The breakdown due to hydraulic oil temperature reached zero in Q3 of 2018 after the
implementation of IoT-based oil temperature prediction.
Figure 29 displays the moulding plant downtime reduction trend from the period Q1 of the
year 2015 to Q4 of the year 2020. Here the downtime is reduced from 5,832 min to 471 min
which is a reduction in downtime of 91.92% after the implementation of the proposed

Figure 28.
Breakdown
trend-hydraulic oil
temperature

Figure 29.
Plant breakdown trend
JQME LSTM-based TPM approach in the high-pressure moulding plant. From the period Q3 of 2018
IoT based prediction model kaizen K1 is implemented. From Q4 of 2018, LSTM-based oil
contamination level prediction model K2 is implemented in the foundry line. From the period
Q3 of 2019, LSTM-based prediction model K3 for oil pressure variation is implemented and
from Q1 of 2020, prediction model K4 due to motor vibration is introduced. From Figure 30, it
is clear that by implementing the conventional TPM, the availability rate is raised from 83.1%
to 87.2% and after implementing the LSTM approach, it is raised from 87.2% to 98.1%. Hence
the proposed systematic TPM approach with LSTM yields a 12.5% increase in
availability rate.
The performance rate is a factor of operator motion, speed loss and minor stoppage loss.
The increase in hydraulic oil temperature and the variation in pressure may lead to
positioning issues and causes minor stoppages. In the proposed operator motion, the speed of
the machine is getting increased and the minor stoppages are becoming less when the
hydraulic oil pressure, temperature and contamination is controlled which eliminates the
malfunction of servo valves and proportional valves. Table 12 depicts the minor stoppages
eliminated after the implementation of LSTM. Figure 31 depicts that, conventional TPM
approach the PR increased from 78.2% to 85.3% and after the implementation of IoT and
LSTM-based prediction model it increased from 85.3% to 90.8%.
Though the quality rate is a factor of the quality of the raw material, healthy status of the
machine and the process parameter. The quality of raw material, process and process
parameters are maintained over a period of time of TPM implementation through QM pillar.
The Identified Q-Components which control hydraulic pressure for squeezing the sand
mould, and the hydraulic temperature related to change in viscosity related to speed variation
of the equipment, leakages, cylinders seal failures etc causes soft, crack and broken sand
moulds which are the major rejections. Hence after QM pillar saturated activity, the quality is
proportional to the health status of the machine. Though different approaches are made to
control mould rejections, Table 13 shows the eliminated rejections due to proposed LSTM
approach on Q-Components. Figure 32 depicts that by implementing conventional TPM, the
QR is raised from 88.2% to 90.3% and after implementing IoT and AI-based prediction model
it is raised from 90.3% to 94.6%. Hence the systematic TPM approach with the proposed
work yields a better result in QR.

Figure 30.
Availability rate trend
Before LSTM After LSTM
Industry 4.0
Minor stoppages Occurrences/ Stoppages Occurrences/ Stoppages transformation
Sl no. phenomenon month minutes/month month minutes/month

1 TZ3 Pushing cylinder 98 103 4 5


position missing
2 TZ4 Pushing cylinder 92 96 3 4
position missing
3 ZL close missing 1 242 318 6 8
4 ZL close missing 2 198 211 4 6
5 Clamping device 1 64 69 3 3
position missing
6 Clamping device 2 57 62 2 2
position missing
7 ZE gripper 1 missing 117 139 5 5
8 ZE gripper 2 missing 123 144 4 4
9 Turn Table Cope 48 50 2 2
position missing
10 Turn Table Drag 53 55 2 2 Table 12.
position missing Mould rejections
Total 1,092 1,247 35 41 eliminated due to minor
Source(s): Table created by Author stoppages

Figure 31.
Trend of
performance rate

The healthy status and the performance of the machine are measured by the parameter, OEE.
Figure 33 depicts that by implementing conventional TPM, the OEE is increased by 13.2%
and after implementing LSTM approach it is further increased by 29.8%. Hence, a total
increase of 46.9% from the benchmark is achieved. In other words, this 46.9% increase can be
interpreted as 46.9% increase in productivity without any major structural change or without
any major investment in equipment. The systematic TPM approach with implementing IoT
and AI-based prediction model achieves a remarkable improvement in OEE.
JQME No of mould rejection
Sl no. Q-components Defect phenomenon eliminated/month

1 Squeezing Circuit Internal Mould Hardness soft inner side 284


proportional valve
2 Squeezing Circuit external Mould Hardness soft outer side 243
proportional valve
3 Hub Lifting Cylinder Servo valve Mould soft 148
4 Box Pusing Cylinder servo valve Mould crack due to mould 84
movement
5 Squeezing Circuit Internal Mould crack due to inside excess 223
proportional valve hardness
Table 13. 6 Squeezing Circuit external Mould crack due to outside 199
Mould rejections proportional valve excess hardness
eliminated due to Total 1,181
Q-components Source(s): Table created by Author

Figure 32.
Plant quality rate trend

4.5.2 Proactive approaches after prediction to achieve zero downtime. Since the process is
continuous in the proposed system, planned maintenance activities are carried out once in a
week and on public holidays. The four key identified phenomenon using Pareto analysis are
addressed as follows: Hydraulic Oil filtration is carried out using external body based on the
prediction of oil contamination and hydraulics tanks of 9,000 litres will be cleaned based on
the condition of contamination level of oil. Oil replacement will be done if the existing oil
cannot be filtered. In case of oil pressure, pumps been replaced and relief valve has been
checked or replaced and calibrated based on the required pressure. All loading valves and
cables will be checked and tightened related to the pump pressure circuit. Adaptor for relief
valve and relays have been replaced. Line leakages will be checked thoroughly. For Motor
vibrations, bearings will be replaced if necessary, spare motor or gear box will be replaced.
Mounting bolts, pullies will be tightened. Fluid coupling will be checked and replaced if
Industry 4.0
transformation

Figure 33.
Plant OEE trend

necessary, based on the condition. Gear box bearings and gears will be replaced if necessary.
Shaft dimensions will be checked and maintained. For hydraulic oil temperature, circulation
pump flow for oil will be checked and rectified. Heat exchanger will be cleaned and made
effectively for heat transfer. Cooling towers will be checked by checking fills, cooling tower
fan, bearings, motor windings, cooling pump flow will be checked. A spare will be kept for
replacement purpose.

5. Conclusion and future work


TBM and CBM used in the conventional TPM have certain demerits due to human error,
skill, poor manual decision making, intermittent data collection and the collected data
may not be available at the time of breakdown for further analysis and hence the duration
of downtime and the OEE improvement gets no further improvement. In the proposed
work, using ranking of equipment required signature parameters are identified and using
the Pareto analysis key signature parameters causing 95% of breakdown are identified
and addressed. LSTM-based AI and IoT-based TPM approaches are implemented to
predict the RUL of the equipment part by predicting the signature parameter data to
achieve zero downtime in the foundry industry. This leads towards zero downtime and
improves the machine availability rate. Since the predicted accuracy of the LSTM model is
much closer to the realistic value, the AR of the moulding line is improved and hence a
remarkable dimensional improvement of the OEE of the system is achieved. The proposed
model tracks the system degradation and keeps on improving the health status by
identifying multivariant signature parameters related to breakdown. In the proposed
work, AR is increased by 11.9%. The PR is raised by 5.5% and QR is raised by 4.3% that
means an increase of 28.1% in OEE and 28.1% increase in productivity without additional
manpower and equipment. A raise in OEE of 46.9% is achieved using the TPM
approaches. In future, drones will be used for plant maintenance and all the signature
parameters related to OEE improvement will be sensed and digitalised to achieve the OEE
value above 95%.
JQME References
Angelo, E., Gosavi, A. and Murray, S.L. (2021), “Total productive maintenance of make-to-stock
production-inventory systems via artificial-intelligence-based iSMART”, International Journal
of Systems Science: Operations and Logistics, Vol. 8 No. 2, pp. 154-166.
Carvalhoa, T.P., Soares, F.A.A.M.N., Vita, R., Francisco, R.d.P., Basto, J.P. and Alcala, S.G.S. (2019),
“A systematic literature review of machine learning methods applied to predictive
maintenance”, Computers and Industrial Engineering, Vol. 137, 106024.
Çınar, Z.M., Abdussalam Nuhu, A., Zeeshan, Q., Korhan, O., Asmael, M. and Safaei, B. (2020),
“Machine learning in predictive maintenance towards sustainable smart manufacturing in
industry 4.0”, Sustainability, Vol. 12, p. 8211.
Coccia, M. (2017), “The Fishbone diagram to identify, systematize and analyze the sources of general
purpose technologies”, Journal of Social and Administrative Sciences, Vol. 4 No. 4, pp. 291-303.
Da Silva, B.S., Inaba, F.K., Salles, E.O.T. and Ciarelli, P.M. (2020), “Outlier robust extreme machine
learning for multi-target regression”, Expert Systems with Applications, Vol. 140, 112877, doi: 10.
1016/j.eswa.2019.112877.
de Pater, I. and Mitici, M. (2021), “Predictive maintenance for multi-component systems of repairables
with remaining-useful-life prognostics and a limited stock of pare components”, Reliability
Engineering and System Safety, Vol. 214, 107761.
Gackowiec, P. (2019), “General overview of maintenance strategies – concepts and approaches”,
Multidisciplinary Aspects of Production Engineering, Vol. 2 No. 1, pp. 126-139.
Hashemian, H.M. and Bean, W.C. (2011), “State-of-the-art predictive maintenance techniques”,
IEEE Transactions on Instrumentation and Measurement, Vol. 60 No. 10, pp. 3480-3492.
Jardine, A.K.S., Lin, D. and Banjevic, D. (2006), “A review on machinery diagnostics and prognostics
implementing condition-based maintenance”, Mechanical Systems and Signal Processing, Vol. 20
No. 7, pp. 1483-1510.
Jeyasekar, A., Kasmir Raja, S.V. and Annie Uthra, R. (2017), “Congestion avoidance algorithm using
ARIMA (2,1,1) model-based RTT estimation and RSS in heterogeneous wired-wireless
networks”, Journal of Network and Computer Applications, Vol. 93, pp. 91-109.
Jimenez- Cortadi, A., Irigoien, I., Boto, F., Sierra, B. and Rodriguez, G. (2020), “Predictive maintenance
on the machining process and machine tool”, Applied Sciences, Vol. 10 No. 1, p. 224.
Kateris, D., Moshou, D., Pantazi, X.-E., Gravalos, I., Sawalhi, N. and Loutridis, S. (2014), “A machine
learning approach for the condition monitoring of rotating machinery”, Journal of Mechanical
Science and Technology, Vol. 28 No. 1, pp. 61-71.
Li, B. and Zhao, Y.-P. (2020), “Group reduced kernel extreme learning machine for fault diagnosis
of aircraft engine”, Engineering Applications of Artificial Intelligence, Vol. 96, 103968.
Li, Y., Zeng, Y., Qing, Y. and Huang, G.-B. (2020), “Learning local discriminative representations via
extreme learning machine for machine fault diagnosis”, Neurocomputing, Vol. 409 No. 7,
pp. 275-285.
Liu, Y.C., Hu, X.F. and Sun, S.X. (2019), “Remaining useful life prediction of cutting tools based on Support
vector regression”, IOP Conference Series: Materials Science and Engineering, Vol. 576, 012021.
Lu, Y. (2017), “Industry 4.0: a survey on technologies, applications and open research issues”, Journal
of Industrial Information Integration, Vol. 6, pp. 1-10.
Luo, J., Wong, C.-M. and Vong, C.-M. (2021), “Multinomial Bayesian extreme learning machine for
sparse and accurate classification model”, Neurocomputing, Vol. 423, pp. 24-33.
Mungani, D.S. and Visser, J.K. (2013), “Maintenance approaches for different production methods”,
South African Journal of Industrial Engineering, Vol. 24 No. 3, pp. 1-13.
Namuduri, S., Narayanan, B.N., Davuluru, V.S.P., Burton, L. and Bhansali, S. (2020), “Review—deep
learning methods for sensor based predictive maintenance and future perspectives for
electrochemical sensors”, Journal of The Electrochemical Society, Vol. 167, 037552.
Ng Corrales, L.d.C., Lamban, M.P., Hernandez Korner, M.E. and Royo, J. (2020), “Overall equipment Industry 4.0
effectiveness: systematic literature review and overview of different approaches”, Applied
Sciences, Vol. 10, p. 6469. transformation
utzner, L., Eilers, D. and Breitner, M.H. (2020), “Predictive
Passlick, J., Dreyer, S., Olivotti, D., Gr€
maintenance as an internet of things enabled business model: a taxonomy”, Electronic Markets,
Vol. 32 No. 3, s12525-021-00520-0.
Pech, M., Vrchota, J. and Bedna, J. (2021), “Predictive maintenance and intelligent sensors in smart
factory: review”, Sensors, Vol. 21, p. 1470.
Qin, A., Zhang, Q., Hu, Q., Sun, G., He, J. and Lin, S. (2017), “Remaining useful life prediction for
rotating machinery based on optimal degradation indicator”, Hindawi Shock and Vibration,
6754968.
Ran, Y., Zhou, X., Lin, P., Wen, Y. and Deng, R. (2019), “A survey of predictive maintenance: systems,
purposes and approaches”, IEEE Communications Surveys and Tutorials, Vol. 20, pp. 1-36.
Rimawan, E. and Irawan, A.P.B. (2017), “Analysis of calculation Overall equipment effectiveness
(OEE) in the implementation of total productive maintenance (TPM) PC 200-8 excavator grab
and magnet type case study in Cakratunggal steel mills company”, International Journal of
Scientific and Engineering Research, Vol. 8 No. 1, pp. 1363-1368.
Roosefert Mohan, T., Annie Uthra, R., Devaraj, D. and Umachandran, K. (2021), “Intelligent machine
learning based total productive maintenance approach for achieving zero downtime in
Industrial machinery”, Computers and Industrial Engineering, Vol. 157, 107267.
Ruiz-Gonzalez, R., Gomez-Gil, J., Gomez-Gil, F.J. and Martınez-Martınez, V. (2014), “An SVM-based
classifier for estimating the state of various rotating components in agro-industrial machinery
with a vibration signal acquired from a single point on the machine Chassis”, Sensors, Vol. 14,
pp. 20713-20735.
Samanpour, A.R., Ruegenberg, A. and Ahlers, R. (2018), “The future of machine learning and
predictive analytics”, Digital Marketplaces Unleashed, Springer-Verlag GmbH Germany.
unior, D.S.d.O., de Oliveira, J.F.L. and de Mattos Neto, P.S.G. (2019), “An intelligent
Santos J
hybridization of ARIMA with machine learning models for time series forecasting”, Knowledge-
Based Systems, Vol. 175, pp. 2-86.
Sonmez, V., Testik, M.C. and Testik, O.M. (2018), “Overall equipment effectiveness when production
speeds and stoppage durations are uncertain”, The International Journal of Advanced
Manufacturing Technology, Vol. 95, pp. 121-130.
Uit Het Broek, M.A.J., Teunter, R.H., de Jonge, B. and Veldman, J. (2021), “Joint condition-based
maintenance and load-sharing optimization for two-unit systems with economic dependency”,
European Journal of Operational Research, Vol. 295 No. 3, pp. 1119-1131.
Zhao, Y.P., Huang, G., Hu, Q.K., Tan, J.F., Wang, J.J. and Yang, Z. (2019), “Soft extreme learning
machine for fault detection of aircraft engine”, Aerospace Science and Technology, Vol. 91,
pp. 70-81.
Zhou, J.-T., Zhao, X. and Gao, J. (2019), “Tool remaining useful life prediction method based on LSTM
under variable working conditions”, The International Journal of Advanced Manufacturing
Technology, Vol. 104, pp. 4715-4726.

About the authors


Roosefert Mohan received B.E degree in Electronics and Communication from
Government College of Engineering, Tirunelveli, Madurai Kamaraj University, Tamil
Nadu, India in 1990 and received M.E (Power Electronics and Drives) degree from
St.Peter’s University, India. He is currently a PhD research scholar in SRM Institute of
Science and Technology, India. His area of research is Predictive maintenance, Zero
Down time, Industry 4.0, Total Productive Maintenance and IIoT.
JQME J. Preetha Roselyn received B.E (Electrical and Electronics Engineering) from Madras
University, India in 2022, received M.S in Power System by Research from Anna
University, Chennai and received PhD from SRM University, India. Currently working
as professor in the Department of Electrical and Electronics Engineering in SRM
Institute of Science and Technology, India. She has published 41 international
publications including 16 SCI indexed journals, 2 lecture notes, 25 conference
publications and a book and she has received many funded projects. She is appointed
as faculty mentor of Marine Technology Society, US which received the best student
chapter award for the year 2019. Her area of interest are power system stability, intelligent controllers for
microgrid, soft computing, power system security. J. Preetha Roselyn is the corresponding author and
can be contacted at: preetha.roselyn@gmail.com
R. Annie Uthra received B.E degree in Computer Science and Engineering from
Manonmaniam Sundaranar University, India. Obtained her M.S degree and PhD from
SRM University, Chennai, India. She is serving as professor and head of the
department of Computational Intelligence at SRM Institute of Technology.
Additionally, she was an Adjunct Associate Teaching Professor in the Institute for
Software Research in the School of Computer Science at Carnegie Mellon University,
Pittsburgh, USA from 2102 till 2018. She had served as visiting professor in Henan
University of Economics and Law, Zhengzhou, Henan Province, China, from year 2018
to 2020. The awards she received include: IET CLN Women Engineer 2016-17, outstanding reviewer
Award from JNCA (Elsevier), Best teacher award 2006 etc. Her scholarly and teaching interests include
Wireless Sensor Networks, Machine Learning, Positioning and Navigation, IoT, Energy Aware Routing
Techniques.

For instructions on how to order reprints of this article, please visit our website:
www.emeraldgrouppublishing.com/licensing/reprints.htm
Or contact us for further details: permissions@emeraldinsight.com

You might also like