Alexandria Engineering Journal 88 (2024) 298–309

Contents lists available at ScienceDirect

Alexandria Engineering Journal

journal homepage: www.elsevier.com/locate/aej

Original article

Machine learning and IoT – Based predictive maintenance approach for

industrial applications
Sherien Elkateb a, *, 1, Ahmed Métwalli b, Abdelrahman Shendy c, Ahmed E.B. Abu-Elanien d
a
Textile Engineering Department, Faculty of Engineering, Alexandria University, Egypt
b
Electrical (Electronics and Communications) Engineering Department, Arab Academy for Science, Technology and Maritime Transport, Alexandria, Egypt
c
Computer Engineering Department, Arab Academy for Science, Technology and Maritime Transport, Alexandria, Egypt
d
Electrical Engineering Department, Faculty of Engineering, Alexandria University, Egypt

A R T I C L E I N F O A B S T R A C T

Keywords: Unplanned outage in industry due to machine failures can lead to significant production losses and increased
Failure maintenance costs. Predictive maintenance methods use the data collected from IoT-enabled devices installed in
IoT working machines to detect incipient faults and prevent major failures. In this study, a predictive maintenance
Knitting machines
system based on machine learning algorithms, specifically AdaBoost, is presented to classify different types of
Machine learning
Predictive maintenance
machines stops in real-time with application in knitting machines. The data collected from the machines include
machine speeds and steps, which were pre-processed and fed into the machine learning model to classify six
types of machines stops: gate stop, feeder stop, needle stop, completed roll stop, idle stop, and lycra stop. The
model is trained and optimized using a combination of hyperparameter tuning and cross-validation techniques to
achieve an accuracy of 92% on the test set. The results demonstrate the potential of the proposed system to
accurately classify machine stops and enable timely maintenance actions; thereby, improving the overall effi­
ciency and productivity of the textile industry.

1. Introduction machines are used to produce knitted fabrics of various sizes, shapes,
and textures. The smooth operation of these machines is critical to the
The manufacturing industry is always looking for ways to improve quality of the final product, as well as to the efficiency of the production
efficiency, reduce downtime, and minimize costs. One area where im­ process. To achieve predictive maintenance for circular knitting ma­
provements can be made is in the maintenance of equipment, especially chines, an Internet of Things (IoT) system is developed to capture ma­
for machines used in the textile industry. Traditional maintenance chine speed and machine stops data. These data are then used to train a
practices have been based on reactive approaches, where equipment is machine learning model to predict the time that the machine is likely to
repaired after a failure occurrence [1]. However, predictive mainte­ stop or malfunction. The machine stops’ reasons are classified according
nance, which relies on the use of data analytics and machine learning to the proposed machine learning technique into maintenance, tool
techniques, has emerged as a more proactive and cost-effective solution change, or operator error [3,4].
[2]. The proposed system is expected to offer several benefits over
In this paper, we focus on the application of predictive maintenance traditional maintenance practices, including reduced downtime,
for circular knitting machines in the textile industry. Circular knitting improved machine availability, and increased productivity.

Abbreviations: AI,, Artificial Intelligence; CNN,, Convolutional Neural Network; DBN,, Deep Belief Network; FDP,, Fault diagnosis and prognosis; GP,, Gaussian
process; GraphQL,, Query Language; GridSearchCV,, Grid search cross-validations; IIoT,, Industrial IoT; IoT,, Internet of Things; JSON,, JavaScript Object No­
tation; KNN,, K-Nearest Neighbor; LR,, linear regression; LSTM,, Long Short-Term Memory; ML,, Machine Learning; NB,, Naïve bayes; MQTT,, Message Queuing
Telemetry Transport; PCB,, Printed circuit board; PWM,, Pulse width modulation; REST-API,, Representational State Transfer Application Programming Interface;
RF,, Random forest; RNN,, Recurrent neural network; RUL,, Remaining Useful Life; SCADA,, Supervisory control and data acquisition; SPC,, Statistical process
control; SVM,, Support vector machine; XGBoost,, Extreme gradient boosting.
* Corresponding author.
E-mail address: [email protected] (S. Elkateb).
1
ORCID: 0000-0002-5251-632X

https://doi.org/10.1016/j.aej.2023.12.065
Received 20 October 2023; Received in revised form 19 December 2023; Accepted 30 December 2023
Available online 20 January 2024
1110-0168/© 2024 The Author(s). Published by Elsevier BV on behalf of Faculty of Engineering, Alexandria University This is an open access article under the CC
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
S. Elkateb et al. Alexandria Engineering Journal 88 (2024) 298–309

Furthermore, the system’s capability to identify the root cause of ma­ effectiveness in planning predictive maintenance through a sewing
chine stoppages would enable targeted maintenance interventions, machine needle case study. Elkateb et al. [10,11] introduced an IoT and
minimizing the risk of unplanned downtime. ML-based online monitoring system for knitting machines, contributing
In this paper, the main contributions are summarized as follows: significantly to predictive maintenance. This system facilitates real-time
tracking, statistical analysis, and issue resolution. Accordingly, it en­
• Novel system of data acquisition in textile industry is presented. In ables preventive maintenance and accurate productivity measurement.
which, a circular knitting machine sensor data and stops are real- Surucu et al. [12] extensively reviewed recent literature on the efficacy
time monitored. Machine profiles detailing production settings like of ML-based condition monitoring, emphasizing their significant con­
machine types and yarn configurations are introduced using tributions to predictive maintenance models. The study compared
improved database. The database allows easy management of mul­ models using deep learning and Bayesian optimization, employing a
tiple machine profiles and maintains detailed records of each shift Deep Belief Network (DBN) for feature extraction and a Gaussian pro­
activity. cess (GP) to optimize DBN hyper-parameters. Empirical results demon­
• The contribution involves the introduction of a novel dataset ob­ strated precise machine failure time prediction, surpassing conventional
tained from a textile factory operating under standard conditions. ML methods. Therefore, cross-case performance comparisons are insuf­
This dataset serves as a repository for preserving the machine’s ficient due to diverse complexities and unique contextual factors.
historical records, facilitating the application of predictive mainte­ Another investigation compared an intelligent Predictive Maintenance
nance methodologies through the categorization and classification of (PdM) system for industrial equipment, utilizing Industrial Internet of
instances of machine stoppages. Things (IIoT), Message Queuing Telemetry Transport (MQTT), and
• Specially designed Machine Learning (ML) - based model is used for machine learning (ML) algorithms [13]. Vibration, current, and tem­
classification of machine stops. Grid search cross validation is perature sensors collect real-time data from electrical motors to be
employed to optimize AdaBoost classifier for high classification ac­ analyzed by five ML models: k-nearest neighbor (KNN), support vector
curacy for machine faults. machine (SVM), random forest (RF), linear regression (LR), and naïve
• A test rig is built in real circular knitting machine to realize the bayes (NB) for anomaly detection and failure prediction. The MQTT
proposed condition monitoring system and prove its functionality. protocol enables efficient communication between sensors, gateways,
and the cloud server. Random forest (RF) exhibits the highest accuracy
The remaining of this paper is organized as follows. Section 2 pre­ in operational motors, optimizing maintenance schedules to minimize
sents the related work on predictive maintenance for industrial equip­ downtime and costs [13].
ment. Section 3 introduces the IoT system architecture and the data In summary, recent research in predictive maintenance for knitting
collection process. Section 4 details the used machine learning meth­ machines has shown promising results in improving maintenance effi­
odology, including data preprocessing, feature selection, and model ciency and reducing costs. These studies have utilized a variety of al­
training. Section 5 presents the results of experimental work, including gorithms and different attributes to predict Remaining Useful Life (RUL)
accuracy metrics and performance comparison with other maintenance and to diagnose various faults in the machine. However, there is still a
strategies. Finally, Conclusion section highlights the significance of the need for further research to develop more accurate, comprehensive, and
proposed system and gives suggestions for future research. efficient predictive maintenance systems for circular knitting machines.
Diagnosis of different machine faults to achieve comprehensive predic­
2. Related work tive maintenance systems were not well covered in the literature.
Moreover, applications of the developed methods on real working ma­
In recent years, researchers have shown a growing interest in the chines outside the laboratory environment were not well covered to
development of predictive maintenance systems for circular knitting prove the applicability in real conditions. The proposed work presents a
machines. Gao et al. [5] proposed a deep learning-based fault diagnosis predictive maintenance strategy that predicts machine stoppage and
method for circular knitting machines. Their system uses a Convolu­ reason of stoppage (failure), so that long machine failures can be avoi­
tional Neural Network (CNN) to automatically extract features from ded. Multiple failure reasons are taken into consideration by using
vibration signals, followed by a SoftMax classifier to classify the fault multiple sensing devices to have comprehensive maintenance strategy.
types. The experimental results demonstrated that their method ach­ ML-based classifier can access readings of these devices through IoT
ieved a promising accuracy in fault diagnosis for circular knitting ma­ system. Its unique features, including real-time monitoring, a sophisti­
chines. However, CNN needs a large amount of training data to achieve cated ML model, and a comprehensive database, indicate a departure
acceptable performance. Udo and Muhammad [6] introduced a pre­ from conventional approaches to modern applicable approaches. The
dictive maintenance system for wind turbines using SCADA data, proposed predictive maintenance system is implemented on real circular
employing XGBoost and Long Short-Term Memory (LSTM) models for knitting machine to prove its ability for practical applications. The
gearbox and generator monitoring. Statistical Process Control (SPC) system shows outstanding performance and high accuracy. Moreover, it
assesses anomalies, demonstrating effectiveness in fault detection for six holds substantial promise for minimizing downtime, enhancing machine
wind turbines, aiding in early intervention and cost-effective dynamic availability, and optimizing productivity in the textile industry.
maintenance strategies. However, this system is not tested on knitting
machines. 3. System setup
Lee et al. [7] explored recent advances in maintenance methods for
manufacturing industries, emphasizing the shift from reliability In this paper, we present an IoT-based system for predictive main­
improvement to flexible and customizable maintenance scheduling in tenance of circular knitting machines. The system is comprised of a
the era of smart manufacturing. Singha et al. [8] delved into the inte­ custom-built Printed Circuit Board (PCB) with an ESP32 microcontroller
gration of Artificial Intelligence (AI) and ML in the knitting industry, and a variety of sensors, including speed sensors, and machine stop
highlighting their transformative impact. The study emphasized the detection sensors. The machine stop detection sensors are designed to
comprehensive application of these technologies across various stages: detect different types of machines stops, including lycra stop, gate stop,
product sourcing, design, production, distribution, and sales. The idle stop (which is triggered manually through human interaction),
incorporation of AI and ML facilitates advancements in fiber classifica­ feeder stop, needle stop, and completed roll. The system is deployed on
tion, thread prediction, fault identification, and dye recipe prediction; two different PAILONG circular knitting machines, with an ambient
thereby, aiding predictive maintenance in knitting industry. A devel­ temperature of 26 degrees Celsius. The collected data is stored in
oped fuzzy decision-making system is developed [9]. It demonstrates its mongoDB which is an open-source NoSQL database management

299
S. Elkateb et al. Alexandria Engineering Journal 88 (2024) 298–309

Fig. 1. System setup.

Fig. 2. System deployment.

system. The data are then used for further analysis and machine maintenance system for circular knitting machines using IoT. We built a
learning-based predictive maintenance. The developed system provides specially designed PCB featuring ESP32 and interface with sensors to
real-time monitoring for the circular knitting machines, enabling timely detect machine stops, records data in MongoDB through a logger system,
detection of potential faults and efficient maintenance scheduling. The and enables real-time monitoring for proactive maintenance. The PDB
machine built in stop detection sensors help to identify the root cause of function and connection with sensors are tested and verified. Integration
machine stops and enable operators to take corrective actions for of GraphQL (Query Language) enhances REST-API efficiency. Moreover,
reducing unplanned downtime. Fig. 1 shows a block diagram of the the adaptable database framework improves operational efficiency
system deployed on the knitting machine. across different machine types. The system ensures a secure and effec­
The machine speed and yarn feeder sensors are deployed to capture tive environment for utilizing machine data. The data connection to and
the speeds. The machine stops are used to detect any triggered event from the server was tested for secure data traffic. The data is retrieved
fired by feeder, gate, ideal, lycra, or needle. The pulse width modulation from the server to a personal computer via web portal or a mobile
(PWM) signals that represent machine operation/stop are captured and application. The retrieved data is compared with the original sensor data
delivered through the ESP32 internal storage. The ESP32 is used as a for assuring high data quality. Fig. 2 shows the system deployment in the
wireless module to deliver the data through the access point to the server main operator gate of the knitting machine. The ESP32 is mounted on a
of which the MongoDB database is used. particularly designed PCB to facilitate collection of data. The output
data of the different sensors are transferred to the ESP32 through the
PCB available connectors. Once the system is deployed, it starts col­
3.1. IoT infrastructure lecting the speeds and stops data.
The produced PCB has the ESP32 stack on it where the connector
The primary objective of this stage is to establish a resilient and pins are connected through cables to the machine sensors.
secure environment for the monitoring system using IoT within a swiftly In pursuit of this objective, we have judiciously decided to transfer
changing environment. This undertaking is imperative to facilitate the our system architecture from its rudimentary state, as depicted in Fig. 3-
seamless acquisition and analysis of machine-generated logs. The a, to a more sophisticated form in order to reflect the current
experimental procedures involve the implementation of a predictive

300
S. Elkateb et al. Alexandria Engineering Journal 88 (2024) 298–309

Fig. 3. The IoT system: a. IoT basic architecture, b. State of art IoT deployed system, c. system synchronization.

301
S. Elkateb et al. Alexandria Engineering Journal 88 (2024) 298–309

Fig. 4. Database Framework.

advancement in the field, which is reflected on quality of data collection By segregating the Representational State Transfer Application Pro­
and accuracy of incipient fault detection as shown in Fig. 3-b. In the gramming Interface (REST-API) from the machine logs and seamlessly
refined Fig. 3-b architecture, we have introduced a state-of-the-art integrating it with our user-facing platforms such as web portals or
component known as the "logger." This advanced entity meticulously mobile applications; we enhance the speed and responsiveness of our
captures and registers every event transmitted from the machine via the REST-API services. This separation optimizes system resources by alle­
Message Queuing Telemetry Transport (MQTT) broker. These critical viating the burden on the machine logs, consequently laying the
events are promptly stored in a high-performing NoSQL database, groundwork for real-time data transmission instead of transmission at
thoughtfully chosen to optimize the speed and efficiency of our logger fixed intervals.
system; thereby, ensuring swift and accurate log processing. The logger In order to optimize REST-API performance, we have embraced
also diligently records each event’s creation date and most recent update cutting-edge GraphQL technology, surpassing the limitations of con­
date, enabling the construction of a precise chronological graph of ventional REST-APIs. This groundbreaking transition provides us with
events. Fig. 3-c shows the system integration and synchronization. This the capabilities to overcome the inherent inefficiencies associated with
innovative capability empowers us to precisely reconstruct the ma­ server requests, particularly when faced with complex data retrieval
chine’s operational state during specific time intervals, facilitating scenarios. Traditional REST-APIs present suboptimal options when
comprehensive analysis and precise identification of stoppages and their accessing database tables with numerous records and columns. One
underlying causes. Moreover, the logger adeptly eliminates redundant possibility requires separating the requests for each platform, resulting
stoppages and accurately calculates and manages stoppage durations. in redundant efforts. The alternative mandates retrieving all columns
For instance, in cases where an ongoing needle break stoppage coincides within a single request, overburdening the computer/mobile processor
with another concurrent stoppage, the stop duration manager will dili­ and causing unwarranted delays in data display. After meticulous
gently calculate the duration of the needle stoppage while disregarding research and evaluation, GraphQL has supplied us with a solution to this
the concurrent stoppage until the needle break is resolved. Additionally, challenge. By leveraging GraphQL, each platform can precisely specify
the stop duration manager proficiently handles manual stoppages the required columns through a solitary request; accordingly, efficiency
initiated by workers. It accurately computes the duration of these in­ is significantly enhanced, and data retrieval processes are notably
terruptions and integrates them into the optimal machine runtime. expedited.
Furthermore, when a reason is assigned to a designated stop duration, Within the Database framework of Fig. 4, our primary focus centers
the stoppage duration supervisor adeptly oversees and allocates it on augmenting the dynamism and adaptability of the database infra­
correspondingly. The logger, then, can be configured to trigger various structure to accommodate all machine types without exceptions. To
actions, including real-time notifications or emails to managers, achieve this goal, we have introduced machine profiles that encapsulate
ensuring immediate awareness of machine stoppages and their root the indispensable production settings, encompassing machine types,
causes. feeder configurations, yarn counts, and the allocation of feeders for each

302
S. Elkateb et al. Alexandria Engineering Journal 88 (2024) 298–309

yarn type. Moreover, our optimized database structure enables seamless Table 1
management of multiple profiles for each machine, facilitating effortless The range of environmental parameters.
transitions between different operational configurations. Furthermore, Parameter Range
our refined database structure empowers us to meticulously track and
Temperature 20 ◦ C to 27 ◦ C (68◦ F to 81◦ F)
maintain a comprehensive historical record of each shift’s activities. Humidity 45% to 55%
This pivotal functionality facilitates the seamless distribution of orders Airflow and Ventilation 6 to 10 air changes per hour
across multiple shifts or machines, facilitating automated handling and Lighting 500 to 750 lux
improving overall operational efficiency. By undertaking these trans­ Noise Level Below 85 decibels (dB)
Particulate Matter (PM) Below 10 mg/m3
formative steps in our monitoring system using IoT, we are poised to
establish a secure and highly efficient environment, enabling us to
capture and leverage machine data effectively while enhancing opera­ smart textiles, landfill effects based on fibers, and overall textile con­
tional performance and decision-making processes. sumption for sustainable practices and informed policy development.
Kohli and Dua [17] explored essential elements of Indoor Environmental
3.2. Environmental conditions Quality (IEQ), covering building parameters, lighting, noise, humidity,
and acoustic quality. Emphasizing their impact on comfort, health, and
From the perspective of a knitting machine, the environmental productivity, the discussion explores thermal comfort, Sick Building
conditions may have a significant impact on the quality of the knitted Syndrome, and air temperature. Tailored recommendations for special
products and the performance of the machine. Some of the key envi­ needs individuals enhance health, wellbeing, comfort, and safety. IEQ
ronmental conditions that are important to consider for a knitting ma­ significantly influences occupants’ health, underscoring its importance
chine include: for exposed groups. Overall goals encompass minimizing
Temperature: The temperature of the environment can affect the building-related health issues and ensuring high-quality indoor envi­
viscosity of the lubricants used in the machine and also impact the ronments. Another study aimed to assess occupational noise exposure in
thermal expansion and contraction of the machine components. the textile industry. An initial anonymous survey was conducted, fol­
Humidity: The humidity of the environment can affect the static lowed by evaluating factory workers’ noise exposure levels and selecting
electricity generated during the knitting process, which can cause yarn ear protectors based on legal criteria. Results showed that most jobs
to stick together and impact the quality of the final product. exceeded legal noise limits, with no workers using hearing protection.
Air quality: The air quality of the environment can impact the Symptomatic issues, like tinnitus, and reduced hearing equity correlated
performance of the machine and the quality of the knitted product. Poor with years of exposure [18]. Hameed et al. [19] studied the effect of high
air quality can cause dust and other contaminants to accumulate on the noise levels in the knitting department pose a risk of Noise-Induced
machine, which can impact its performance. Good ventilation and air Hearing Loss (NIHL). Workers, despite experiencing symptoms like
filtration are important to maintain a clean and healthy environment for tinnitus, often neglect hearing loss complaints. Factors include age,
the knitting machine. duration of noise exposure, health conditions, and limited worker
Lighting: Adequate lighting is important for the operator to be able awareness. Table 1 outlines crucial environmental factors essential for
to monitor the performance of the knitting machine and to detect any the optimal operation of a knitting machine.
defects in the knitted products.
Noise level: Knitting machines can generate noise during operation,
which can be harmful to the operator’s hearing and impact their ability 3.3. Data collection
to detect any issues with the machine.
Vibrations: The vibrations generated by the knitting machine can The data collected through MongoDB as shown in Fig. 1, are stored in
impact its performance and affect the quality of the knitted products. JavaScript Object Notation (JSON) format. Then, the data are converted
Adequate vibration isolation and damping measures are important to to tabular form where the preliminary data shape acquired is made of
reduce the impact of vibrations on the machine and its environment. events and logs with a shape of 81000 records from 9 attributes. The
Overall, maintaining a controlled and stable environment with events represent the time of which any machine stop is fired. On the
suitable environmental conditions is essential to ensure the optimal other hand, the logs represent each record of the working machine. The
performance of a knitting machine and to produce high-quality knitted records were acquired during a period of 48 days. It is formed from the
products. machine speeds and machine stops where the attributes are the
Several studies have investigated the impact of working conditions, following:
including temperature, humidity, lighting, and noise, on the perfor­ Created At: A timeseries attribute which stores a record’s starting
mance and quality of knitting machines. For instance, Seitablaiev et al. time information.
[14] explored factors affecting thermal comfort and indoor air quality, Stitch Length: Stitch length is determined by the needle pitch, which
providing a comprehensive understanding of environmental pollutants is the distance between two adjacent needles in a circular knitting
and their health implications. This study aligns with those findings, machine.
aiming to optimize indoor conditions for user comfort by analyzing Yarn Feeder Speed: Yarn feeder sensors are capable of measuring
bioclimatic parameters, thermal comfort, and air quality, with a focus on yarn belt at high speeds, typically ranging from several hundred to
the consequential impact of pollutants on health and productivity. several thousand measurements per minute. The speed is measured in
Rassel and Hoque [15] emphasized the significant influence of the revolution per minute.
working environment on yarn breakage, affecting both production ef­ Production Rate: The amount of fabric produced within a given
ficiency and fabric quality. Addressing these factors is essential for time period, usually expressed in terms of weight or length. It is a
enhancing competitiveness, reducing costs, and minimizing the envi­ measure of the efficiency and productivity of the machine and is influ­
ronmental impact associated with increased power consumption. enced by various factors such as machine speed, stitch density, yarn
Moazzem et al. [16] conducted a comprehensive review of environ­ type, and machine settings. It is measured in kilograms per hour.
mental impacts throughout the textile supply chain, emphasizing the Machine Progress: It refers to the movement of the knitting ele­
substantial contributions of textile production and usage stages. Future ments, such as needles and sinkers, as well as the fabric, around the
research recommendations focus on evaluating fiber mixing, recycling cylinder or the needle bed. It indicates the position of the knitting ele­
options, and technical textiles. This study explores gaps in the envi­ ments and the fabric relative to the machine cycle and is an important
ronmental impact of the textile industry, emphasizing the evaluation of parameter for controlling the knitting process and ensuring consistent

303
S. Elkateb et al. Alexandria Engineering Journal 88 (2024) 298–309

Table 2
Statistical distribution.
Stich Lengh Yarn Feeder Production Rate Progress Machine Speed Total Turn per Roll Event Duration

mean 1.993838 896.599572 4.833878 50.420655 14.005796 1613.460960 84.548607

std 0.426776 224.270651 1.209121 28.861632 3.488458 923.572234 3666.011328
min 0.000000 0.000000 0.000000 0.093750 0.000000 3.000000 0.000000
25% 2.001149 865.000000 4.663514 25.656250 13.000000 821.000000 0.000000
50% 2.145279 866.000000 4.668905 50.437500 13.000000 1614.000000 0.000000
75% 2.150113 906.000000 4.884559 75.562500 16.000000 2418.000000 0.000000
max 2.283511 1320.000000 7.116576 100.000000 21.000000 3200.000000 396659.205

Fig. 5. Count of stop’s types.

Fig. 6. Total seconds for each stop.

and uniform stitch formation. The machine progress can be measured in proximity sensor of the machine. It is measured in revolution per
terms of degrees or fractions of a revolution of the cylinder or the needle minute.
bed, or in terms of machine cycle time. For example, in a single-cylinder Total Turns per Roll: The total turn per roll is the accumulation of
circular knitting machine, one complete revolution of the cylinder cor­ the roll revolutions per minute starting from placing the roll until the
responds to 360 degrees of machine progress, while in a double-cylinder machine completes 3200 turns.
machine, one complete cycle of the cylinder and dial needles corre­ Updated At: A timeseries attribute which stores a record’s ending
sponds to 720 degrees of machine progress. time information.
Machine Speed: The speed of the roll turns which is captured by the Event Duration: Event duration is the time of which the event

304
S. Elkateb et al. Alexandria Engineering Journal 88 (2024) 298–309

Fig. 7. Average duration for each stop.

Fig. 8. Median event duration in seconds for each stop.

occurred. It is the subtraction of the ‘Updated At’ and the ‘Created At’ 4. Classification scheme
attributes.
Stoppage Type: The machine stops are categorized as follow: gate, To enable predictive maintenance of circular knitting machines, a
needle, feeder, Lycra, ideal and completed roll. classification scheme is developed to categorize machine stop events and
Table 2 shows the statistical distribution of the data. Figs. 5–8 show identify the root cause of machine failures. The classification scheme is
some insights about the data. Fig. 5 shows the count plot of each stop, based on the types of machines stop detected by the sensors in the IoT
while Fig. 6 depicts the sum of all event durations for each stop. Fig. 7 system. The machine stops are classified into different categories based
portrays the average duration for each stop, while the median of event on the component causing the stop and the type of stop (such as whether
duration for each stop type is presented in Fig. 8. According to Fig. 5, it is a planned or unplanned stop). The proposed classification scheme
during the 48 days, the most frequent stop occurred was idle stop as it allows for a more precise diagnosis of the problems encountered during
counts 1476 cases. Gate, feeder, complete roll, needle, and Lycra stops machine operation, which is crucial for efficient and timely maintenance
are counted as 852, 401, 373, 24 and 22 cases respectively. The median [1,2]. In the following subsections, we will describe a flow chart of the
is calculated as it is a robust statistical measure that is less sensitive to proposed classification scheme and how it is integrated into the overall
extreme values or outliers in the data set. The median values for feeder, IoT system for predictive maintenance of circular knitting machines. The
gate, ideal, Lycra and needle based on collected data are 12,2137,38 and classification scheme is shown in Fig. 9.
9, respectively. The data are retrieved from the database, subsequently, the data
preprocessing phase is carried out, where the data are normalized, then

305
S. Elkateb et al. Alexandria Engineering Journal 88 (2024) 298–309

Fig. 9. Classification scheme flowchart.

Fig. 10. Data normalization procedure.

306
S. Elkateb et al. Alexandria Engineering Journal 88 (2024) 298–309

they are divided into training set and testing and validation set. The ▪ Shuffle column j of dataset D to generate a distorted
training set is 60% and the testing and validation set is 20% and 20% version of the data named D
̃ k,j
respectively. The system is trained by the training set and the testing and
▪ Calculate the score sk,j of model m on data D ̃ k,j
validation set is used to validate and test the proposed method. The
▪ Calculate importance ij for feature fj as follows.
results are validated readily since the machine stoppage data includes
the sensor reading and stoppage reason. The feature selection process is 1 ∑K

done through permutation importance technique in which the noisy ij = s − sk,j (2)
K k=1
features are excluded. In the end, the ML algorithm is applied then
optimized using grid search cross validation to determine the best Where K: the total number of points in a data set, and.
classification scores. k: the kth data point.

4.1. Data preprocessing

4.2. Machine Stops Classification
In this section, the data preprocessing workflow is illustrated, which
includes, the data normalization and the feature selection processes. Adaboost is a popular ensemble learning technique that aggregates
weak learners into a strong classifier. Moreover, it is short for Adaptive
4.1.1. Data normalization Boosting. Therefore, it is used in this study for classification purposes. In
The goal of normalization is to transform features to be on a similar the present work, multiple weak classifiers are trained on different
scale. This scale is centralized at the origin with peak of − 1 or 1 as subsets of the training data, and their predictions are combined to make
shown in Fig. 10. This process improves the training stability of the a final stronger prediction. During each iteration, Adaboost assigns
model. The ML models may suffer poor performance and high higher weights to misclassified data points to force the weak classifiers
complexity due to input data with outliers and high variance [18–21]. to focus on the most challenging examples. The final classifier is a
Hence, z-score technique is used to normalize each data point i in data weighted sum of the weak classifiers, where the weights are determined
j based on their individual performance. Adaboost has been widely used
category j ( Di ). It measures the data point distance to its mean value
in various purposes such as text classification, face recognition, and
divided by a standard deviation. z-score is a variation of scaling that
bioinformatics [24]. One of the main advantages of Adaboost is its
represents the number of standard deviations away from the mean. The
ability to handle complex data distributions and achieve high accuracy
main advantage of z-score is to ensure that the feature distributions have
with simple classifiers. Adaboost has also been shown to be robust to
mean of 0 and standard deviation of 1. The advantage of normalization
noise and outliers, making it suitable for real-world scenarios [25].
appears when there are a few outliers. The z-score normalization can be
Moreover, Adaboost is a flexible algorithm that can be combined with
expressed as
different weak learners, such as decision trees, SVMs, and neural net­
Dij − Dj works, to improve their performance [26]. Classification algorithm
Dij′ = (1) using Adaboost is given below.
σ Dj ( )
- Given: (x1 , y1 ), …, xm , ym where xi ∈ X , yi ∈ { − 1, + 1}.
where. - Initialize D1 (i) = 1/m for i = 1, …, m.
Dij′: normalized data point I in data category j, - For = 1, …, T:
Dij : data point I in data category j,
• Perform training for weak learner using distribution Dt .
Dj :mean of data category j, • Obtain weak hypothesis ht : X →{ − 1, + 1}.
σ Dj : standard deviation of data category j. • Select ht with low weighted error. The error is calculated as follows:

4.1.2. Feature selection εt = Pri∼Dt [ht (xi ) ∕

= yi ] (3)
Permutation importance is a technique that can be used to determine
the significance of features in a machine learning model. This method
involves randomly shuffling the values of a feature in the test dataset ( )
1− εt
and observing the resulting reduction in the model’s accuracy. A large • Choose αt = 12 ln εt
drop in accuracy indicates that the feature is important to the model’s • For i = 1, …, m, update
performance. The use of permutation importance in classification has
several benefits, such as aiding in feature selection and model evaluation Dt (i)exp( − αt yi ht (xi ) )
Dt+1 (i) = (4)
[15,16]. When conducting permutation importance using Adaboost, the Zt
feature importance is measured for the specific classifier used in the
where Zt is a normalization factor (It is selected such that Dt+1 is a
classification process. This approach can provide more accurate and
distribution).
relevant information since different classifiers may assign varying levels
- Find the final hypothesis:
of importance to features. Adaboost is a type of ensemble method that
( )
combines multiple weak classifiers to create a strong classifier. ∑ T

Measuring feature importance using Adaboost can provide insight into H(x) = sign αt ht (x) (5)
how the various weak classifiers in the ensemble contribute to the
t=1

model’s overall performance [21–23]. The steps of measuring feature The grid search cross-validations (GridSearchCV) are then applied.
importance using Adaboost are listed below: GridSearchCV’s a commonly used hyperparameter tuning technique in
machine learning. It is a systematic approach to search for the best
• Input: fitted predictive model m, tabular dataset (training or vali­ combination of hyperparameters by exhaustively testing all possible
dation) D. parameter combinations within a predefined range. The technique
• Calculate the reference score s of the model m on dataset D (for evaluates the performance of each parameter combination by cross
instance classifier accuracy). validating the model on different subsets of the training data. Grid­
• For each feature j: SearchCV is widely used in various applications such as image classifi­
o For each iteration k in 1, .., k, …, K: cation, natural language processing, and predictive modeling [27].

307
S. Elkateb et al. Alexandria Engineering Journal 88 (2024) 298–309

Table 3 Table 4
Permutation importance parameters. Permutation importance score for each feature.
Base estimator Number of Learning Number of Parameter Created Progress Total Turn Updated Event
estimators rate repeats At per Roll At Duration

Decision tree 50 1 10 Score 0.175 0.165 0.175 0.145 0.345

classifier
maximum depth
=8 Table 5
Grid search hyperparameter optimization.

One of the advantages of GridSearchCV is its ability to optimize the Parameter Train/test split ratio Number of estimators Learning rate

model hyperparameters in a systematic and efficient manner. The Range 0.1 to 1 10 to 150 0.1 to 1
technique provides a comprehensive search over a predefined parameter Optimum 0.4 100 0.1
grid, which enables the user to find the optimal hyperparameters
without relying on intuition or trial-and-error. Moreover, GridSearchCV
Table 6
can help prevent overfitting by using cross-validation to evaluate the Classification report.
performance of each parameter combination on independent subsets of
Precision Recall F1-Score Support (Samples)
the training data [28].
Complete 1 1 1 144
Feeder 0.91 0.66 0.76 163
5. Results
Gate 0.92 0.95 0.93 332
Idle 0.91 0.96 0.93 602
The AdaBoost classifier permutation importance parameters are lis­ Lycra 0.75 0.75 0.75 8
ted in Table 3. Permutation importance results in feature selection is Needle 0.44 0.36 0.4 11
shown in Fig. 11. accuracy 0.92 1260
macro avg 0.82 0.78 0.8 1260
The permutation importance test showed that the stich length, yarn weighted avg 0.92 0.92 0.91 1260
feeder speed and machine speed are considered as noisy data and does
not affect the output since there score is zero. The output here refers to
the stop type during the classification process. Table 4 shows the actual group with 602 samples, have a precision of 0.91, a recall of 0.96, and an
score for each feature that is not dropped. F1-score of 0.93. The "Lycra" stops, though limited in number (8 sam­
The event duration showed the best promising results in feature se­ ples), showcases a balanced precision, recall, and F1-score, all at 0.75.
lection process with score of 0.345. The parameters used by grid search The "Needle" stops, based on 11 samples, have a lower precision of 0.44,
to determine the optimum (best parameter), are presented in Table 5. a recall of 0.36, and an F1-score of 0.4. The overall accuracy of the
The AdaBoost model is applied after eliminating the unwanted fea­ model across all categories is 0.92 based on 1260 samples. The macro
tures and hyper-tuning the model. The model achieved 92% accuracy of average, which gives equal weight to each category, has a precision of
classifying the stop only using the event duration. The classification 0.82, recall of 0.78, and F1-score of 0.8. The weighted average,
results are reported in Table 6. considering the number of samples in each category, is 0.92 for preci­
In the classification report for knitting machine stops, the "Complete" sion, 0.92 for recall, and 0.91 for the F1-score.
category has a perfect score of unity with respect to precision, recall, and
F1-score, based on 144 samples. The "Feeder" stops have a precision of 6. Conclusion
0.91, recall of 0.66, and F1-score of 0.76, with 163 samples. The "Gate"
stops exhibit high precision and recall at 0.92 and 0.95, respectively, In conclusion, this study focuses on addressing the challenges posed
with an F1-score of 0.93 from 332 samples. The "Idle" stops, the largest by unplanned outages in the textile industry through the

Fig. 11. Permutation importance results.

308
S. Elkateb et al. Alexandria Engineering Journal 88 (2024) 298–309

implementation of a predictive maintenance system. By leveraging IoT- [5] Y. Gao, C. Chai, H. Li, W. Fu, A deep learning framework for intelligent fault
diagnosis using automl-cnn and image-like data fusion, Machines 11 (10) (2023)
enabled devices and machine learning algorithms, specifically Ada­
932, https://doi.org/10.3390/machines11100932.
Boost, the study successfully classified different types of machines stops [6] W. Udo, Y. Muhammad, Data-driven predictive maintenance of wind turbine based
in real-time. The data collected from the machines, including machine on SCADA data, IEEE Access 9 (2021) 162370–162388, https://doi.org/10.1109/
speeds and steps, underwent pre-processing were used as inputs to the ACCESS.2021.3132684.
[7] J. Lee, J. Ni, J. Singh, B. Jiang, M. Azamfar, J. Feng, Intelligent maintenance
machine learning model. The model was trained and optimized using a systems and predictive manufacturing, J. Manuf. Sci. Eng. 142 (2020) 1–40,
combination of hyperparameter tuning and cross-validation techniques. https://doi.org/10.1115/1.4047856.
As a result, it achieved an impressive accuracy of 92% on the test set. [8] K. Singha, S. Maity, P. Pandit, Use of AI and machine learning techniques in
knitting, Ch.6, Text. Inst. Book Ser. (2022) 161–180, https://doi.org/10.1016/
The outcomes of this research demonstrate the potential of the proposed B978-0-323-85534-1.00021-0.
system to accurately classify various machine stops, encompassing gate [9] C. Baban, M. Baban, S. Darius, Using a fuzzy logic approach for the predictive
stops, feeder stops, needle stops, completed roll stops, idle stops, and maintenance of textile machines, J. Intell. Fuzzy Syst. 30 (2) (2016) 999–1006,
https://doi.org/10.3233/IFS-151822.
Lycra stops. By enabling timely maintenance actions, the system has the [10] S. Elkateb, A. Métwalli, A. Shendy, An Innovative Online Monitoring System in
capability to enhance overall efficiency and productivity within the Knitting Industry, The 16th Textile Bioengineering and Informatics Symposium;
textile industry. Implementing the proposed predictive maintenance Blended Conference. 412–419, August 22–25(2023). DOI: TBIS 10.3993/tbis
(2023).
system holds significant advantages for the textile industry such as [11] S. Elkateb, A. Métwalli, A. Shendy, K. Moussa, A. Abu-Elanien, Online monitoring-
mitigating machine failures, reducing unplanned downtime, minimizing based prediction model of knitting machine productivity, Fibres Text. East. Eur. 31
production losses, and reducing maintenance costs. Moreover, it em­ (4) (2023) 46–52, https://doi.org/10.2478/ftee-2023-0035.
[12] O. Surucu, S.Andrew Gadsden, J. Yawney, Condition monitoring using machine
powers businesses to adopt a proactive maintenance approach, resulting
learning: a review of theory, applications, and recent advances, Expert Syst. Appl.
in improved operational efficiency and resource optimization. Accord­ 221 (2023) 119738, https://doi.org/10.1016/j.eswa.2023.119738.
ingly, this study has significant potential impact in the textile industry. It [13] N. Mohammed, O. Abdulateef, A. Hamad, An IoT and machine learning-based
enhances machine lifetime, improves product quality, and improves predictive maintenance system for electrical motors, J. Eur. Des. Systèmes Autom.
56 (4) (2023) 651–656, https://doi.org/10.18280/jesa.560414.
manufacturer income. Future work includes expanding the data inputs [14] M. Özdamar Seitablaiev, F. Umaroğulları, Thermal comfort and indoor air quality,
by more stop types to cover all incipient faults. Moreover, the presented Int. J. Sci. Res. Innov. Technol. 5 (2018) 90–109.
technique will be applied on different types of machines. [15] M. Rassel, M. Hoque, Re-evaluation on causes of circular knitting machine
production efficiency and their impact on fabric quality, Eur. Sci. J. ESJ 15 (2019),
https://doi.org/10.19044/esj.2019.v15n21p448.
Funding [16] S. Moazzem, E. Crossin, F. Daver, L. Wang, Environmental impact of apparel supply
chain and textile products, Environ., Dev. Sustain. 24 (2022), https://doi.org/
10.1007/s10668-021-01873-4.
STDF grant number 43467. [17] N. Kohli, K. Dua, Indoor Environment: Light, Noise, Humidity and Temperature.
Ch.5:57-75, Integrated Publications, Rohini, Delhi-110085, India, 2022.
Declaration of Competing Interest [18] M. Alves, H. Simões, J. Pereira, J. de Figueiredo, A. Ferreira, Evaluation of
workers’ exposure to occupational noise in the textile industry (Case Study), SSDC
492 (2023).
I am the corresponding author and I declare that I have no known [19] H. Hameed, A. Eleue, F. Hussein, Noise induced hearing loss (NIHL) in wasit
competing financial interests or personal relationships that could have corporation textile industries, J. Otolaryngol. -ENT Res. 11 (2) (2019).
[20] A. Fisher, C. Rudin, All models are wrong but many are useful: learning a variable’s
appeared to influence the work reported in this paper.
importance by studying an entire class of prediction models simultaneously,
J. Mach. Learn. Res. 18 (1) (2018) 3754–3786.
Acknowledgements [21] A. Niculescu-Mizil, R. Caruana, Predicting good probabilities with supervised
learning, Proc. 22nd Int. Conf. Mach. Learn. (2005) 625–632.
[22] T. Hastie, R. Tibshirani, J. Friedman, The elements of statistical learning: data
This paper is based upon work supported by Science, Technology & mining. Inference, and Prediction, Springer,, 2009.
Innovation Funding Authority (STDF) under grant number 43467. [23] P. Braga, A. Oliveira, G. Ribeiro, S. Meira, Bagging predictors for estimation of
software project effort, Int. Jt. Conf. Neural Netw., Orlando, FL, USA (2007)
1595–1600, https://doi.org/10.1109/IJCNN.2007.4371196.
References [24] Y. Freund, R.E. Schapire, A decision-theoretic generalization of on-line learning
and an application to boosting, J. Comput. Syst. Sci. 55 (1) (1997) 119–139.
[1] W. Lee, H. Wu, H. Yun, H. Kim, M. Jun, J. Sutheralnd, Predictive maintenance of [25] C. Hsu, C. Lin, C. Chang, A practical guide to support vector classification.
machine tool systems using artificial intelligence techniques applied to machine Department of Computer Science, National Taiwan University, Taipei, Taiwan.
condition data, Procedia CIRP 80 (2019) 506–511, https://doi.org/10.1016/j. (2003).
procir.2018.12.019. [26] R.E. Schapire, A brief introduction to boosting, Proc. Sixt. Int. Jt. Conf. Artif. Intell.
[2] T. Zonta, C.André da Costa, R. da Rosa Righi, M. de Lima, E. da Trindade, G. Li, (1999) 1401–1406.
Predictive maintenance in the Industry 4.0: a systematic literature review, Comput. [27] J. Bergstra, Y. Bengio, Random search for hyper-parameter optimization, J. Mach.
Ind. Eng. 150 (2020) 106889, https://doi.org/10.1016/j.cie.2020.106889. Learn. Res. 13 (2012) 281–305.
[3] W. Lee, H. Wu, H. Yun, H. Kim, M. Jun, J. Sutherland, Predictive maintenance of [28] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel,
machine tool systems using artificial intelligence techniques applied to machine E. Duchesnay, Scikit-learn: machine learning in python, J. Mach. Learn. Res. 12
condition data, Procedia CIRP 80 (2019) 506–551, https://doi.org/10.1016/j. (2011) 2825–2830.
procir.2018.12.019.
[4] C. Mgbemena, F. Okeagu, Development of an IoT-based real-time remote
monitoring device for the maintenance of injection moulding machines in plastic
industries, UNIZIK J. Eng. Appl. Sci. 2 (1) (2023) 260–278.

309