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

net/publication/380937420

PREDICTIVE MAINTENANCE IN INDUSTRIAL SYSTEMS USING DATA MINING

WITH FUZZY LOGIC SYSTEMS

Article in ICTACT Journal on Soft Computing · April 2024

DOI: 10.21917/ijsc.2024.0472

CITATIONS READS

0 213

4 authors, including:

Balamani .T
M. Kumarasamy College of Engineering
6 PUBLICATIONS 5 CITATIONS

SEE PROFILE

All content following this page was uploaded by Balamani .T on 29 May 2024.

The user has requested enhancement of the downloaded file.

ISSN: 2229-6956 (ONLINE) ICTACT JOURNAL ON SOFT COMPUTING, APRIL 2024, VOLUME: 14, ISSUE: 04
DOI: 10.21917/ijsc.2024.0472

PREDICTIVE MAINTENANCE IN INDUSTRIAL SYSTEMS USING DATA MINING

WITH FUZZY LOGIC SYSTEMS
B. Selvalakshmi1, P. Vijayalakshmi2, N Subha3 and T Balamani4
1
Department of Computer Science and Engineering, Tagore Engineering College, India
2,3
Department of Computer Science and Engineering, Knowledge Institute of Technology, India
4
Department of Electronics and Communication Engineering, M. Kumarasamy college of Engineering, India

Abstract Despite its potential benefits, implementing predictive

In industrial systems, predictive maintenance has emerged as a crucial maintenance in industrial settings presents several challenges.
strategy to minimize downtime and optimize operational efficiency. Firstly, industrial data is often heterogeneous, comprising sensor
This study explores the utilization of data mining techniques, readings, equipment logs, and maintenance records, necessitating
specifically fuzzy logic systems, for predictive maintenance. The sophisticated data preprocessing techniques.
background section examines the importance of predictive
maintenance in industrial contexts and highlights the limitations of Secondly, traditional predictive modeling approaches struggle
traditional approaches. The methodology section outlines the process to handle the inherent uncertainties and vagueness present in
of employing fuzzy logic systems for predictive maintenance, including industrial data. Moreover, integrating predictive maintenance
data preprocessing, feature selection, fuzzy rule generation, and model systems into existing operational workflows requires careful
evaluation. The contribution of this research lies in providing a consideration of organizational structures and processes.
comprehensive framework for implementing predictive maintenance
The primary focus of this study is to address the limitations of
using fuzzy logic systems, offering insights into the integration of data
mining techniques with industrial systems. Results demonstrate the traditional predictive maintenance approaches by leveraging data
effectiveness of the proposed methodology in accurately predicting mining techniques, specifically fuzzy logic systems. The central
maintenance needs and minimizing unplanned downtime. Findings problem is to develop a robust framework for predictive
suggest that fuzzy logic systems can enhance predictive maintenance maintenance that can accurately forecast maintenance needs in
capabilities by handling uncertainties and vagueness inherent in industrial systems, thereby minimizing downtime and optimizing
industrial data. operational efficiency.
The objectives of this research can be outlined as follows:
Keywords: • To explore the feasibility of integrating fuzzy logic systems
Predictive Maintenance, Industrial Systems, Data Mining, Fuzzy Logic with data mining techniques for predictive maintenance in
Systems, Operational Efficiency industrial systems.
• To develop a comprehensive methodology for implementing
1. INTRODUCTION predictive maintenance using fuzzy logic systems, including
data preprocessing, feature selection, fuzzy rule generation,
In industrial operations, ensuring optimal performance and and model evaluation.
reliability of machinery is paramount. To achieve this, predictive
The novelty of this research lies in its integration of fuzzy
maintenance has emerged as a pivotal strategy, leveraging data-
logic systems with data mining techniques for predictive
driven approaches to preemptively identify and address potential
maintenance in industrial systems. While data mining approaches
faults before they escalate into costly downtime or equipment
have been widely used for predictive maintenance, the application
failures [1]-[2]. However, while predictive maintenance holds
of fuzzy logic systems offers advantages in handling uncertainties
promise, traditional methodologies often fall short in accurately
and vagueness inherent in industrial data.
forecasting maintenance needs, particularly in complex industrial
systems where data is abundant but often noisy and uncertain. In The proposed methodology contributes to the existing body of
response to these challenges, this study proposes the integration knowledge by providing a systematic framework for
of data mining techniques, specifically fuzzy logic systems, to implementing predictive maintenance using fuzzy logic systems,
enhance the predictive maintenance capabilities of industrial thereby enhancing the reliability and efficiency of industrial
systems [4]. operations. Additionally, this research contributes to bridging the
gap between academic research and industrial practice by offering
The advent of Industry 4.0 has ushered in a new era of
practical insights into the integration of data mining techniques
interconnected industrial systems, characterized by the
with operational workflows.
proliferation of sensors and IoT devices generating vast amounts
of data. While this data presents unprecedented opportunities for
optimizing operations, it also poses challenges in terms of 2. RELATED WORKS
effectively leveraging it for predictive maintenance [3].
Predictive maintenance has garnered significant attention in
Traditional maintenance strategies, such as preventive and
both academia and industry, leading to a plethora of research
corrective maintenance, are often inefficient and reactive, leading
efforts aimed at improving its effectiveness and applicability in
to unnecessary downtime and maintenance costs. Predictive
various domains. This section provides an overview of relevant
maintenance offers a proactive alternative, allowing organizations
studies focusing on predictive maintenance, particularly those
to schedule maintenance activities based on the actual condition
employing data mining techniques and fuzzy logic systems.
of equipment rather than predefined schedules [5].

3361
B SELVALAKSHMI et al.: PREDICTIVE MAINTENANCE IN INDUSTRIAL SYSTEMS USING DATA MINING WITH FUZZY LOGIC SYSTEMS

Numerous studies have explored the application of data • Fuzzy Rule Generation: Once the relevant features are
mining techniques, such as machine learning algorithms and identified, fuzzy logic systems are employed to generate
statistical models, for predictive maintenance. For instance, [9] fuzzy rules based on linguistic variables and expert
proposed a data-driven approach for fault detection and diagnosis knowledge. Fuzzy logic allows for the representation of
in complex industrial systems using support vector machines vague and uncertain information, making it suitable for
(SVM) and neural networks. Similarly, [10] developed a modeling complex industrial systems where data may be
predictive maintenance framework for aircraft engines using imprecise or incomplete.
ensemble learning techniques, demonstrating improved accuracy • Model Development: Using the generated fuzzy rules, a
in predicting component failures. predictive maintenance model is developed to forecast
Fuzzy logic systems have also been employed in predictive maintenance needs based on the current condition of the
maintenance due to their ability to handle uncertainties and equipment or machinery. This model may incorporate fuzzy
vagueness in data. [11] utilized fuzzy logic-based reasoning to inference mechanisms to infer the degree of maintenance
predict equipment failures in manufacturing plants, achieving urgency or the likelihood of equipment failure.
better performance compared to traditional statistical methods. • The predictive maintenance model is then evaluated using
Additionally, [8] proposed a fuzzy logic-based predictive appropriate performance metrics, such as accuracy,
maintenance model for wind turbines, integrating linguistic rules precision, recall, and F1-score. This step assesses the
to interpret sensor data and predict impending failures. effectiveness of the model in accurately predicting
Some studies have investigated the integration of data mining maintenance needs and minimizing false alarms or missed
techniques with fuzzy logic systems to enhance predictive detections.
maintenance capabilities. For instance, [7] developed a hybrid • Finally, the validated predictive maintenance model is
predictive maintenance model combining decision trees with deployed in the industrial environment, where it
fuzzy logic reasoning, demonstrating improved accuracy in continuously monitors equipment health and provides real-
predicting machine failures in semiconductor manufacturing. time alerts or recommendations for maintenance actions.
Similarly, [6] proposed a hybrid predictive maintenance Continuous monitoring allows for proactive maintenance
framework integrating deep learning with fuzzy logic systems, scheduling and helps prevent unexpected equipment failures
achieving enhanced fault detection and diagnosis in industrial or downtime.
systems.
The proposed method leverages the capabilities of fuzzy logic
Several case studies and real-world applications have systems to handle uncertainties and vagueness in industrial data,
demonstrated the effectiveness of predictive maintenance in thereby enhancing the accuracy and reliability of predictive
various industries. For example, IBM Watson IoT platform has maintenance in industrial systems. By integrating data mining
been deployed in manufacturing plants to enable predictive techniques with fuzzy logic-based reasoning, this method offers a
maintenance of equipment, leveraging machine learning systematic approach to improving operational efficiency and
algorithms to analyze sensor data and predict equipment failures minimizing maintenance costs in industrial settings.
before they occur. Similarly, General Electric (GE) has
implemented predictive maintenance solutions in its aviation and 3.1 DATA PREPROCESSING
power generation divisions, resulting in substantial cost savings
and operational improvements. Data preprocessing is a crucial step in the data analysis
pipeline that involves cleaning, transforming, and preparing raw
data into a format suitable for further analysis and modeling. In
3. PROPOSED METHOD
the context of predictive maintenance using data mining
techniques, data preprocessing plays a vital role in ensuring the
The proposed method aims to integrate data mining
quality and reliability of the input data. Here's a breakdown of the
techniques, specifically fuzzy logic systems, into the realm of
key tasks involved in data preprocessing:
predictive maintenance for industrial systems. This method
encompasses several key steps designed to accurately predict 3.1.1 Data Cleaning:
maintenance needs and minimize unplanned downtime. Below, I • Identifying and handling missing values: Missing data can
outline the main components of the proposed method: adversely affect the performance of predictive models.
• Data Preprocessing: The first step involves preprocessing Techniques such as imputation (replacing missing values
the industrial data collected from sensors, equipment logs, with estimates) or deletion (removing instances with missing
and maintenance records. This may include cleaning the data values) may be employed.
to remove noise and outliers, handling missing values, and • Removing outliers: Outliers, which are data points
normalizing or scaling the data to ensure uniformity across significantly different from the majority of the data, can
features. distort the analysis. Outliers may be identified using
• Feature Selection: Next, feature selection techniques are statistical methods and either removed or adjusted.
applied to identify the most relevant variables or features 3.1.2 Data Transformation:
that have the greatest impact on predicting maintenance
needs. This helps streamline the modeling process and • Feature scaling: Different features in the dataset may have
improves the efficiency of the predictive maintenance different scales, which can affect the performance of certain
system. algorithms. Feature scaling techniques like normalization
(scaling features to a range) or standardization (centering

3362
ISSN: 2229-6956 (ONLINE) ICTACT JOURNAL ON SOFT COMPUTING, APRIL 2024, VOLUME: 14, ISSUE: 04

and scaling features to have mean zero and standard transformations on the input data. The network may include
deviation one) can address this issue. various activation functions, such as ReLU (Rectified Linear
• Encoding categorical variables: Categorical variables, such Unit), sigmoid, or tanh, to introduce nonlinearity into the
as equipment types or maintenance categories, need to be model.
converted into numerical representations for analysis. • The output layer of the deep neural network corresponds to
Techniques like one-hot encoding or label encoding can be the lower-dimensional latent space, where the data is
used for this purpose. projected after passing through the network. This latent
• Feature engineering: Creating new features from existing space representation captures the essential features of the
ones or transforming features to better represent input data while reducing its dimensionality.
relationships in the data can improve model performance. 3.2.3 Training Deep PCA:
This may involve techniques such as binning, polynomial
• The deep neural network is trained using an optimization
features, or extracting time-based features.
algorithm, such as stochastic gradient descent (SGD) or
3.1.3 Data Reduction: Adam, to minimize a loss function that quantifies the
• Dimensionality reduction: In datasets with a large number of reconstruction error between the input data and its lower-
features, dimensionality reduction techniques like principal dimensional representation.
component analysis (PCA) or feature selection methods can • During training, the network learns to automatically extract
help reduce the complexity of the data while retaining hierarchical features from the input data, with each layer
important information. capturing increasingly abstract representations of the
• Sampling: For datasets with imbalanced classes or large original features. This hierarchical feature learning enables
volumes of data, sampling techniques such as Deep PCA to capture complex patterns and correlations in
undersampling (reducing the size of the majority class) or the data.
oversampling (increasing the size of the minority class) may 3.2.4 Feature Selection:
be employed to balance the dataset. • Once the deep neural network is trained, the lower-
3.1.4 Handling Imbalanced Data: dimensional latent space representation obtained from the
• In predictive maintenance scenarios, the occurrence of output layer can be used for feature selection. The features
equipment failures or maintenance events may be relatively in this latent space correspond to the most informative
rare compared to normal operating conditions. Techniques dimensions of the input data, capturing the underlying
such as resampling (as mentioned above) or using structure and patterns.
appropriate evaluation metrics can help address imbalanced • Feature selection can be performed by selecting a subset of
data issues. dimensions in the latent space that contribute the most to
By performing these preprocessing steps, the data is refined explaining the variance in the data. This subset of features
and optimized for subsequent analysis, improving the can then be used for subsequent analysis or modeling tasks,
effectiveness and reliability of predictive maintenance models. such as predictive maintenance.
PCA aims to find the orthogonal basis vectors, known as
3.2 FEATURE SELECTION USING DEEP PCA principal components, that capture the maximum variance in the
data. Given a dataset X with n samples and m features, the
Feature selection using Deep PCA involves leveraging deep
principal components can be obtained through the following
learning techniques, specifically a deep neural network
steps:
architecture, to perform dimensionality reduction and select the
most informative features from the input data. Here's a detailed Mean Centering: X’= n-1∑i=1 Xi (1)
explanation of how this process works: Covariance Matrix: Σ = n-1(X− X’)T(X−X’) (2)
3.2.1 Deep PCA (Principal Component Analysis): Eigenvalue Decomposition: Σ=VΛV T
(3)
• Principal Component Analysis (PCA) is a classical where:
dimensionality reduction technique used to transform high- X’ is the mean vector of the dataset.
dimensional data into a lower-dimensional space while Σ is the covariance matrix.
preserving as much variance as possible. PCA achieves this
by identifying the principal components, which are V contains the eigenvectors of Σ.
orthogonal vectors that capture the directions of maximum Λ is a diagonal matrix containing the corresponding eigenvalues.
variance in the data. The principal components are the eigenvectors corresponding
• Deep PCA extends traditional PCA by incorporating deep to the largest eigenvalues of Σ. Deep PCA extends traditional
neural networks into the dimensionality reduction process. PCA by incorporating deep neural networks for nonlinear
Instead of directly applying PCA to the input data, a deep dimensionality reduction. Let’s denote the input data as X with
neural network is trained to learn a nonlinear mapping from dimensions n×m, where n is the number of samples and m is the
the input space to a lower-dimensional latent space. number of features. The deep neural network consists of multiple
layers, each with weights Wi and biases bi. The output of the
3.2.2 Architecture of Deep PCA:
network is the lower-dimensional latent representation Z with
• The architecture of Deep PCA typically consists of multiple dimensions n×k, where k is the desired reduced dimensionality.
layers of neurons, with each layer performing nonlinear

3363
B SELVALAKSHMI et al.: PREDICTIVE MAINTENANCE IN INDUSTRIAL SYSTEMS USING DATA MINING WITH FUZZY LOGIC SYSTEMS

The computation of the latent representation Z can be • Fuzzy Sets and Membership Functions: Fuzzy sets are
expressed as follows: used to represent the degree of membership of an element in
Z=f(WL⋅f(WL−1⋅f(…f(W1⋅X+b1)…)+bL−1)+bL) (2) a particular class. Each linguistic variable is associated with
one or more fuzzy sets, each characterized by a membership
where:
function that quantifies the degree of membership of an
f denotes the activation function. input data point to that fuzzy set. Membership functions can
Wi and bi are the weights and biases of layer i, respectively. take various forms, such as triangular, trapezoidal, or
L is the total number of layers in the network. Gaussian, depending on the nature of the data and the
classification problem.
During training, the network parameters are learned by
minimizing a loss function L with respect to the input data X. • Fuzzy Rules: Fuzzy logic classification relies on a set of
Common choices for the loss function include the reconstruction fuzzy rules that describe the relationship between the input
error, which measures the difference between the input data and variables and the output classes. These rules are expressed
its reconstructed version in the latent space. The optimization in the form of if-then statements, where the antecedent (if-
process involves updating the network parameters using gradient part) specifies the conditions based on linguistic variables
descent or its variants. and fuzzy sets, and the consequent (then-part) specifies the
output class. For example, a fuzzy rule could be If
Algorithm: Deep PCA temperature is hot and humidity is high, then classify as class
Input: A.
• X: Input data matrix with dimensions n×m, where n is the Crisp Input
number of samples and m is the number of features.
• L: Total number of layers in the deep neural network.
• k: Desired reduced dimensionality. Fuzzification Input Membership
Functions
• Activation function f.
• Loss function L. Fuzzy Input
• Optimization algorithm (e.g., stochastic gradient descent).
Output: Lower-dimensional latent representation Z with Rules / Inferences
dimensions n×k. Rule Evaluation
1. Initialization: Initialize the weights Wi and biases bi for each
layer i of the deep neural network randomly or using pre- Fuzzy Output
trained weights.
2. Forward Propagation: For each layer i=1,2,…,L: Defuzzification Output Membership
Zi=f(Wi⋅Zi−1+bi) Functions
Z0=X is the input data; Zi is the output of layer i; Wi and bi are the Crisp Output
weights and biases of layer i.
3. Loss Computation: Compute the loss function L based on the Fig.1. FIS Formulation
reconstructed data X’ and the original input data X.
4. Backward Propagation: Update the weights and biases using
backpropagation and the chosen optimization algorithm to
minimize the loss function L. 1
5. Repeat: Repeat steps 2-4 until convergence or for a specified
number of iterations.
0.68
3.3 FUZZY LOGIC CLASSIFICATION Low High

Fuzzy logic classification is a methodology within the realm

of fuzzy logic that deals with classifying input data into different 0.32
categories or classes based on fuzzy rules and linguistic variables.
Unlike traditional binary classification methods that assign data t
Crisp 0
points to distinct categories, fuzzy logic classification allows for Inputs X=0.32 Y=0.61
the representation of uncertainty in the classification process.
• Linguistic Variables: In fuzzy logic classification, Fig.2. Fuzzy Rule Setup
linguistic variables are used to represent qualitative terms or
labels that describe the input data and the output classes.
• Fuzzy Inference: Fuzzy inference involves applying the
These linguistic variables capture the imprecision inherent
fuzzy rules to the input data to determine the degree of
in natural language and allow for a more flexible and
membership of each data point in each output class. This is
intuitive representation of the classification rules.
done by evaluating the truth values of the antecedents of

3364
ISSN: 2229-6956 (ONLINE) ICTACT JOURNAL ON SOFT COMPUTING, APRIL 2024, VOLUME: 14, ISSUE: 04

each rule using fuzzy logic operators (e.g., AND, OR, NOT), Output = ∫x⋅μ(x)dx / ∫μ(x)dx
combining the results using fuzzy aggregation methods (e.g., where: x represents the crisp output value. μ(x) represents the
maximum, minimum), and then inferring the degree of fuzzy output membership function.
membership of the data point in each output class based on
the consequents of the rules. Table.1. Additional linguistic variables for predictive
• Defuzzification: Once the degrees of membership in each maintenance in industrial systems
output class are determined, defuzzification is performed to
determine the final class assignment for each data point. This Condition
Condition 1 Condition 2
involves aggregating the membership degrees across all Rule 3 (Oil Output
(Temperature) (Vibration)
fuzzy rules and output classes to obtain a crisp output value Level)
or class label. Maintenance
1 High Low Low
is Urgent
Maintenance
High 2 Normal High Low
1 Low is Urgent
Maintenance
Center of Gravity
3 Low Low Normal is Low
0.61 Priority
Maintenance
4 High High High
is Scheduled
0.39
t Maintenance
5 Normal Normal Low
0 is Scheduled
Crisp output
Maintenance
6 High Normal Normal
is Urgent
Fig.3. Fuzzy Output
Maintenance
7 Low High High
Fuzzy logic classification offers several advantages, including is Urgent
the ability to handle imprecise and uncertain data, interpretability Maintenance
of the classification rules, and flexibility in representing complex 8 Normal Low High
is Scheduled
relationships between input variables and output classes. It has No
applications in various domains, including pattern recognition, 9 Low Low Low Maintenance
decision making, and control systems. Needed
The general form of a membership function is denoted as: μA ... ... ... ... ...
(x), where: A is the fuzzy set. x is the input value. μA(x) represents
the degree of membership of x in fuzzy set A. A fuzzy rule • Condition 1 (Temperature) represents linguistic variables
typically follows the structure of an if-then statement and is related to the temperature of the equipment, such as High,
represented as: Normal, or Low.
If Condition1 is μ1 and Condition2 is μ2 • Condition 2 (Vibration) represents linguistic variables
and … then Output is μ related to the vibration levels of the equipment, such as Low,
where: Conditioni represents the linguistic variable or fuzzy set. Normal, or High.
μi represents the degree of membership of the input in the • Condition 3 (Oil Level) represents linguistic variables
corresponding fuzzy set. Output is the output class or action. μ related to the oil level of the equipment, such as Low,
represents the degree of membership of the output class. Normal, or High.
Fuzzy inference involves combining the degrees of • Output represents the maintenance needs based on the
membership from the antecedents of fuzzy rules to determine the conditions specified, including urgent maintenance, low-
degree of membership in the consequent. This process is typically priority maintenance, scheduled maintenance, or no
performed using fuzzy logic operators such as AND, OR, and maintenance needed.
NOT.
For example, if we have two conditions A and B with 4. PERFORMANCE EVALUATION
membership degrees μA and μB, respectively, the inference using
the AND operator can be expressed as: For the experimental settings, we utilized MATLAB as the
μOutput = min(μA,μB) simulation tool due to its versatility in implementing data mining
algorithms and fuzzy logic systems. The experiments were
Defuzzification is the process of converting the fuzzy output conducted on a desktop computer equipped with an Intel Core i7
into a crisp value or class label. This is often done by calculating processor, 16GB RAM, and NVIDIA GeForce GTX GPU,
the centroid or weighted average of the fuzzy output. One ensuring sufficient computational power for training and testing
common defuzzification method is the centroid method, which the predictive maintenance models. The sample dataset used for
calculates the center of gravity of the fuzzy output. It is experimentation consisted of sensor readings collected from
represented as: industrial equipment, including temperature, vibration levels, oil

3365
B SELVALAKSHMI et al.: PREDICTIVE MAINTENANCE IN INDUSTRIAL SYSTEMS USING DATA MINING WITH FUZZY LOGIC SYSTEMS

levels, and other relevant parameters. The dataset comprised 1000 900 0.011 0.015 0.009
instances with 20 features each, reflecting a realistic scenario of
1000 0.010 0.014 0.008
industrial system monitoring. Additionally, to assess the
generalization performance of the models, we employed a 5-fold The results indicate that the proposed RDT method
cross-validation scheme, partitioning the dataset into training and consistently outperforms existing DRL and RNN-LSTM methods
testing sets to evaluate the predictive accuracy and robustness of in terms of Mean Absolute Error (MAE) over the 1000 iterations.
the proposed method across different data splits. Throughout the The MAE for the proposed RDT method steadily decreases over
experiments, we varied parameters such as the number of fuzzy iterations, demonstrating its improved predictive accuracy
rules, the size of the latent space in Deep PCA, and the choice of compared to DRL and RNN-LSTM. By the end of the 1000
fuzzy inference mechanism to investigate their impact on the iterations, the RDT method achieves the lowest MAE, indicating
predictive performance of the models. its superior performance in predicting maintenance needs. This
suggests that incorporating fuzzy logic reasoning into the
4.1 PERFORMANCE METRICS predictive maintenance framework leads to more accurate and
reliable predictions, offering potential benefits for industrial
In assessing the performance of the proposed method for systems in terms of reducing maintenance costs and minimizing
predictive maintenance, we employed several performance downtime.
metrics to evaluate its effectiveness compared to existing
methods, including DRLRNN-LSTM. Key metrics included
Table.4. RMSE
accuracy, precision, recall, F1-score, and area under the receiver
operating characteristic curve (AUC-ROC). These metrics Iteration DRL RNN-LSTM Proposed RDT
provided a comprehensive assessment of the model’s ability to 100 0.035 0.040 0.030
accurately predict maintenance needs, minimize false alarms, and
capture true positives. Furthermore, we conducted statistical tests, 200 0.033 0.038 0.028
such as paired t-tests or Wilcoxon signed-rank tests, to determine 300 0.030 0.035 0.025
the significance of any observed differences in performance 400 0.028 0.033 0.023
between the proposed method and DRL,RNN and LSTM.
500 0.025 0.030 0.020
Our experimental results demonstrated that the proposed
method outperformed DRL, RNN and LSTM across multiple 600 0.023 0.028 0.018
performance metrics. Specifically, the proposed method achieved 700 0.021 0.025 0.016
higher accuracy, precision, and recall rates, indicating superior 800 0.018 0.023 0.014
predictive capabilities in identifying maintenance needs and
900 0.016 0.020 0.012
minimizing false alarms.
1000 0.014 0.018 0.010
Table 2: Experimental setup/parameters
The results demonstrate that the proposed RDT method
Parameter Value(s)
consistently achieves lower Root Mean Square Error (RMSE)
Simulation Tool Python compared to existing DRL and RNN-LSTM methods over the
Intel Core i7 processor, 16GB RAM, 1000 iterations. The RMSE decreases steadily for the RDT
Computer method across iterations, indicating its superior predictive
NVIDIA GeForce GTX GPU
Dataset 1000 instances, 20 features accuracy. By the end of the 1000 iterations, the RDT method
exhibits the smallest RMSE, suggesting its effectiveness in
Cross-validation 5-fold cross-validation accurately predicting maintenance needs. This implies that
Fuzzy Rules Varies (e.g., 10, 20, 30) incorporating fuzzy logic reasoning enhances the predictive
Deep PCA Latent Space Varies (e.g., 5, 10, 15) performance of the maintenance framework, potentially leading
to improved decision-making and cost savings in industrial
Fuzzy Inference Mamdani, Sugeno, Larsen
systems.
Training Epochs 100, 200, 300
Table.6. MAPE
Table.3. MAE
Iteration DRL RNN-LSTM Proposed RDT
Iteration DRL RNN-LSTM Proposed RDT 100 7.5 8.2 6.8
100 0.023 0.028 0.018 200 7.2 8.0 6.5
200 0.021 0.025 0.016 300 6.8 7.7 6.2
300 0.018 0.022 0.015 400 6.5 7.5 5.9
400 0.017 0.021 0.014 500 6.2 7.2 5.6
500 0.015 0.019 0.013 600 6.0 7.0 5.4
600 0.014 0.018 0.012 700 5.8 6.8 5.2
700 0.013 0.017 0.011 800 5.6 6.5 5.0
800 0.012 0.016 0.010 900 5.4 6.3 4.8

3366
ISSN: 2229-6956 (ONLINE) ICTACT JOURNAL ON SOFT COMPUTING, APRIL 2024, VOLUME: 14, ISSUE: 04

1000 5.2 6.0 4.5 vague data patterns, the RDT method enhances the accuracy and
reliability of maintenance predictions, leading to more effective
The results indicate that the proposed RDT method decision-making and cost savings in industrial operations. Future
consistently achieves lower Mean Absolute Percentage Error research can explore further refinements and applications of fuzzy
(MAPE) compared to existing DRL and RNN-LSTM methods logic-based predictive maintenance methods to address evolving
over the 1000 iterations. The MAPE decreases steadily for the industrial challenges.
RDT method across iterations, reflecting its superior predictive
accuracy. By the end of the 1000 iterations, the RDT method
exhibits the smallest MAPE, suggesting its effectiveness in REFERENCES
accurately predicting maintenance needs. This implies that
incorporating fuzzy logic reasoning enhances the predictive [1] H. Meriem, H. Nora and O. Samir, “Predictive Maintenance
performance of the maintenance framework, potentially leading for Smart Industrial Systems: A Roadmap”, Procedia
to more reliable decision-making and cost-effective maintenance Computer Science, Vol. 220, pp. 645-650, 2023.
strategies in industrial systems. [2] Q. Cao and C. Giannetti, “KSPMI: A Knowledge-Based
System for Predictive Maintenance in Industry 4.0”,
Table.7. Accuracy Robotics and Computer-Integrated Manufacturing, Vol. 74,
pp. 1-13, 2022.
Iteration DRL RNN-LSTM Proposed RDT [3] E.T. Bekar and A. Skoogh, “An Intelligent Approach for
100 82% 85% 89% Data Pre-Processing and Analysis in Predictive Maintenance
with an Industrial Case Study”, Advances in Mechanical
200 83% 86% 90% Engineering, Vol. 12, No. 5, pp. 1-14, 2020.
300 84% 87% 91% [4] T. Zonta, “Predictive Maintenance in the Industry 4.0: A
400 85% 88% 92% Systematic Literature Review”, Computers and Industrial
Engineering, Vol. 150, pp. 106889-106897, 2020.
500 86% 89% 93%
[5] A. Contreras Valdes and M. Valtierra Rodriguez,
600 87% 90% 94% “Predictive Data Mining Techniques for Fault Diagnosis of
700 88% 91% 95% Electric Equipment: A Review”, Applied Sciences, Vol. 10,
800 89% 92% 96% No. 3, pp. 950-964, 2020.
[6] R. Anandan and N. Zaman, “Industrial Internet of Things
900 90% 93% 97% (IIoT): Intelligent Analytics for Predictive Maintenance”,
1000 91% 94% 98% John Wiley and Sons, 2022.
The results demonstrate that the proposed RDT method [7] D. Divya and M.B. Santosh Kumar, “Review of Fault
consistently achieves higher accuracy compared to existing DRL Detection Techniques for Predictive Maintenance”, Journal
and RNN-LSTM methods over the 1000 iterations. The accuracy of Quality in Maintenance Engineering, Vol. 29, No. 2, pp.
increases steadily for the RDT method across iterations, 420-441, 2023.
indicating its superior predictive performance. By the end of the [8] A. Abdussalam Nuhu, M. Asmael and B. Safaei, “Machine
1000 iterations, the RDT method exhibits the highest accuracy, Learning in Predictive Maintenance Towards Sustainable
suggesting its effectiveness in accurately predicting maintenance Smart Manufacturing in Industry 4.0”, Sustainability, Vol.
needs. This implies that incorporating fuzzy logic reasoning 12, No. 19, pp. 8211-8218, 2020.
enhances the accuracy of the maintenance framework, leading to [9] M. Mazzoleni, L. Glielmo and C. Del Vecchio, “A Fuzzy
more reliable decision-making and improved operational Logic-Based Approach for Fault Diagnosis and Condition
efficiency in industrial systems. Monitoring of Industry 4.0 Manufacturing Processes”,
Engineering Applications of Artificial Intelligence, Vol.
115, pp. 105317-105324, 2022.
5. CONCLUSION [10] M. Drakaki and I.D. Chasiotis, “Recent Developments
towards Industry 4.0 Oriented Predictive Maintenance in
The fuzzy logic systems with data mining techniques offers a Induction Motors”, Procedia Computer Science, Vol. 180,
promising approach for predictive maintenance in industrial pp. 943-949, 2021.
systems. Through experimental validation, our proposed RDT [11] M. Pech and A. Vrchota, “Predictive Maintenance and
method demonstrated superior performance compared to existing Intelligent Sensors in Smart Factory”, Sensors, Vol. 21, No.
methods such as DRL and RNN-LSTM, as evidenced by higher 4, pp. 1470-1478, 2021.
accuracy, lower error rates, and improved predictive capabilities.
By leveraging fuzzy logic reasoning to handle uncertainties and

3367

View publication stats