Neuro-Fuzzy Framework for Fault Prediction in Electrical Machines via Vibration Analysis

: The advent of Industry 4.0 has ushered in a new era of technological advancements, particularly in integrating information technology with physical devices. This convergence has given rise to smart devices and the Internet of Things (IoT), revolutionizing industrial processes. However, despite the push towards predictive maintenance, there still is a significant gap in fault prediction algorithms for electrical machines. This paper proposes a signal spectrum-based machine learning approach for fault prediction, specifically focusing on bearing faults. This study compares the effectiveness of traditional neural network algorithms with a novel approach integrating fuzzy logic. Through extensive experimentation and analysis of vibration spectra from various mechanical faults in bearings, it is demonstrated that the fuzzy-neuro network model outperforms traditional neural networks, achieving a validation accuracy of 99.40% compared to 94.34%. Incorporating fuzzy logic within the neural network framework offers advantages in handling complex fault combinations, showing promise for applications requiring higher accuracy in fault detection. While initial results are encouraging, further validation with more complex fault scenarios and additional fuzzy layers is recommended to fully explore the potential of fuzzy-neuro networks in fault prediction for electrical machines.


Introduction
Nowadays, electrical machines play a pivotal role across various applications and industries, significantly contributing to efficiency and productivity enhancements.Their extensive utilization underscores the importance of ensuring proper maintenance to uphold operational integrity.From small appliances to large-scale industrial machinery, electrical machines are integral components of daily life.Despite their benefits, various damages combined with environmental factors can adversely affect the performance of induction machines during operation, impacting their effectiveness, maintenance requirements, and lifespan.Given the indispensable nature of these motors across diverse industrial sectors, preventing such failures is crucial and necessitates vigilant attention [1].
In today's landscape, every energy system represents a complex mechanism vulnerable to damage.Monitoring numerous parameters is imperative to ensure operation and prevent unforeseen failures [2].Furthermore, certain fault patterns present in the signals can serve as indicators of imminent failure.Therefore, incorporating condition-based monitoring is essential to remain informed and to make informed decisions regarding machine maintenance.Rapid advancements in information technologies, particularly cloud computing and the Internet of Things (IoT), introduce enhanced diagnostic capabilities-predictive maintenance-leveraging big data and numerical models of systems for greater efficiency [3,4].Many studies focusing on intelligent health monitoring heavily rely on machine learning techniques [5].Artificial neural networks stand out as a widely used set of methods [6].Owing to their adaptability, nonlinearity, and capacity to approximate arbitrary functions, neural networks have demonstrated effectiveness in condition monitoring and prediction.Through training, they discern intricate dependencies between input and output data, enabling generalization [7].Consequently, neural networks are hailed as a versatile solution applicable to a myriad of problems.Authors in [8] propose a diagnostic method for bearing based on vibration signals and convolutional neural networks.In [9], a vibration-based bearing fault diagnosis approach with the usage of convolutional neural network is presented.The authors in [10] propose a diagnosis method that combines frequency-domain signal analysis with lightweight neural networks.In [11], the authors present a method that trains deep neural networks with convolutional autoencoder and kernel density estimation to define an optimal threshold for anomaly detection.The authors in [12] propose a method based on a multi-attention convolutional neural network for accurate fault diagnosis.In [13], the authors explore the effectiveness of image-based approaches for machine fault diagnosis via convolutional neural networks.
At the same time, fuzzy logic offers a flexible approach capable of handling variables within these parameters.This alternative decision-making method presents numerous advantages.The authors in [14] present an approach for wind turbine condition monitoring using a fuzzy inference system.In [15], partial discharge monitoring of medium-voltage switchgears is discussed.The authors in [16] propose a health monitoring method for lithium-ion batteries.In [17], the authors present techniques of fault diagnosis for a threephase voltage-source inverter.The authors in [18] introduce a multiclass adaptive neurofuzzy classifier combined with feature selection techniques for detecting and classifying faults in photovoltaic arrays.
As discussed, fuzzy logic approaches are predominantly utilized for fault classification rather than prediction.Hence, one of the distinctive aspects of this study was to explore the potential of fuzzy logic in prediction tasks, which represents a novel direction compared to prevailing methodologies in the field.This study examines two approaches to assess the accuracy of machine learning models and determine if results can be enhanced by integrating fuzzy logic with machine learning.The novelties and contributions of this manuscript are as follows: 1.This paper presents a novel integration of fuzzy logic into neural networks, creating a fuzzy-neuro network specifically for fault prediction in electrical machines.This innovative approach combines the decision-making strengths of fuzzy logic with the pattern recognition capabilities of neural networks, resulting in significantly higher accuracy for detecting complex fault patterns in bearings.2. This study introduces a unique methodology by applying machine learning to the signal spectrum of vibration data for predicting bearing faults.This focus on analyzing vibration spectra in the frequency domain provides a more precise and reliable means of fault detection, distinguishing it from traditional time-domain analysis techniques. 3.This research demonstrates that incorporating a fuzzy logic layer within a neural network framework achieves a validation accuracy of 99.40%, outperforming the traditional neural network's 94.34%.This improvement highlights the effectiveness of the fuzzy-neuro network in handling intricate fault combinations, showcasing its potential for applications requiring high accuracy in fault prediction.

Bearing Faults and Experimental Setup
Bearings are pivotal components within rotating machinery.Various internal and external factors can impact bearing performance during motor operation, including material fatigue, improper lubrication, contamination, mechanical damage, and shaft currents.
Any unforeseen failure can have fatal consequences in production.To avoid such risks, consistent condition monitoring remains imperative.
A substantial amount of qualitative data is essential to effectively train the system to predict potential machine failures.The main challenge revolves around the quantity and quality of training data available.For this reason, an experimental test bench was devised for data gathering, as illustrated in Figure 1.As shown, the setup comprises a testing machine, loading machine, and acquisition system (Dewetron, Grambach, Austria).The tests were conducted at the rated speed in steady state.The parameters of the testing and loading motor are presented in Table 1.To ensure the accuracy of fault pattern training, we considered various operational conditions of the rotating machine.Signals were extracted from parameters such as current, voltage, torque, speed, and vibration to analyze fault impacts.However, the vibration spectrum plays a pivotal role in analyzing damaged bearings.For this reason, the vibration signals were prioritized.During the experiments, vibration measurements were carried out utilizing a triaxial accelerometer with a range of ±100 g, strategically positioned over the shaft.Testing encompassed different motor loads, spanning from 0% to 100%.Furthermore, data collection occurred in diverse control environments, including grid-fed, scalar control, and direct torque control systems.
As shown in Figure 2, this study focused on the most prevalent mechanical bearing faults: faulty outer ring, faulty inner ring, and damaged cage.The faults were manually implemented to the healthy bearings.Both healthy and faulty cases were placed in the test motor for examination.Data were meticulously collected and recorded through the acquisition system, resulting in an extensive array of datasets detailing the bearings' condition.

Methodology
The suggested methodology encompasses several key steps aimed at effectively predicting faults in electrical machines.Initially, the process begins with data gathering directly from the machine utilizing sensors, ensuring comprehensive coverage of relevant information.Subsequently, the collected data undergo thorough pre-processing, which includes denoising techniques and conversion to the frequency domain, enhancing the quality and suitability of the data for analysis.Following pre-processing, the data are utilized for fault prediction using a combination of machine learning and fuzzy logic models.The methodology of the suggested approach is presented in Figure 3. Focusing on specific frequency components within the harmonic spectrum is crucial to ensure the timely detection of failures.These components are identifiable in the frequency domain through fast Fourier transforms of the machine's signal.Since bearing faults primarily manifest in the vibration spectrum, emphasis was placed on analyzing vibration patterns.Figure 4 presents an initial comparison of various mechanical faults on the vibration spectrum, potentially arising within the bearings.The fault notably impacts the frequency range of 0 to 1000 Hz the most prominently.However, for algorithm training purposes, the most significant range lies within 0 to 500 Hz.As seen, there are distinct regions where the fault impact is the most pronounced.However, it is essential to investigate the effect of faults on fundamental harmonics before delving into their influence on side harmonics.Then, understanding how damages alter side harmonics becomes a logical progression.Consequently, fault patterns can be extracted from each signal and compared with healthy signals, facilitating system training.The vibration signals captured by the system inherently contain some level of noise attributed to external factors.Although denoising techniques such as wavelet transformation or applying smoothing functions tailored to vibration signals are typically considered, they have yet to be factored into the preliminary analysis.For now, the signal remains untreated to assess the potential impact of denoising methods in subsequent stages.For denoising purposes, a combination of low-pass filtering and median filtering is employed.This method ensures effective noise reduction without compromising the integrity of the information.While there may be concerns about information loss, particularly in scenarios involving low power signals and the use of microcontroller boards, the data acquisition process utilizing Dewetron minimizes such risks.Therefore, the approach adopted here mitigates the potential for information loss to a significant extent.
Once the signals are collected, the samples are shifted to the frequency domain for initial analysis to see if the healthy and faulty cases have unique frequency components.For the initial study, the collected samples were classified as either healthy or faulty signals.The data samples are then further divided with a ratio of 70-30% into training and testing data sets.The total number of initial samples collected for this study is 4 million, with each data sample consisting of every unique frequency component.The overview of the algorithms is shown in Figure 5.Typically, conventional computer programs make inflexible "yes or no" decisions by employing logic operations based on two-valued logic, which functions with either "1/0" or "true/false".In contrast to this binary logic, fuzzy logic can handle variables within these parameters.The primary configuration of the fuzzy logic block is depicted in Figure 6.It comprises fuzzification, where linguistic variables undergo transformation into fuzzy sets via the utilization of membership functions.Fuzzy linguistic variables are employed to express qualities that extend across a specific spectrum.The central component of the fuzzy logic block comprises fuzzy rules or sentences constructed in the form of IF-THEN rules: The rule base manages linguistic sentences to perform logic operations within the fuzzy block.The inference engine executes fuzzy implications for arriving at solutions.Defuzzification then converts variables from fuzzy sets into tangible, real values.Fuzzy logic decisions offer distinct advantages, outlined as follows: (1) Rule-Based Sentences: Fuzzy logic employs rules presented in sentence form, facilitating readability for process operators.These rules can be constructed using everyday vocabulary, such as 'high', 'low', and 'increasing', enabling operators to integrate their practical experience directly.(2) Incorporation of Expert Knowledge: Fuzzy logic allows decision-makers to consider numerous inputs, leveraging the advantage of including expert knowledge in decision-making.(3) Natural Language Interface: Fuzzy logic algorithms interface with natural language, setting it apart from other methods.This characteristic enhances its accessibility and user-friendliness.( 4) Nonlinear System Capability: The fuzzy logic block generally operates as a nonlinear system and can handle multiple inputs and outputs.Variables can be combined in "IF-THEN" rules using the connectives AND/OR.Rules are executed in parallel and provide recommended actions, even if conflicts arise.The controller is responsible for resolving these conflicts.
However, there are also several disadvantages: (1) Tuning Parameters: Fuzzy logic decisions involve more tuning parameters than classical approaches, making them potentially more complex.Additionally, tracing data flow during execution is challenging, which complicates error correction.(2) Understanding and Implementation: While fuzzy logic has merits, other techniques are well understood, easily implemented, and widely adopted in various applications.(3) Lack of Simple Equation: Due to its structure, fuzzy logic lacks a straightforward equation and mathematical apparatus, making system analysis challenging.Consequently, ensuring system stability becomes a complex task.
In this study, the approach is to compare a normal neural network algorithm and a fuzzy neural network algorithm, where a fuzzy logic layer is defined between the neural networks to smooth out the results further and achieve higher accuracy.The general overview of the training algorithm is shown in Figure 7.The training and testing were performed in Python 3.7.The training for the machine learning algorithms is similar, with the only difference being the introduction of the fuzzy layer as one of the hidden layers of the algorithm.Different configurations were considered for the number of neurons and hidden layers to achieve optimized results for an artificial neural network, and the most optimized one was selected from within.To keep the comparison similar, the same configuration was also applied in the case of fuzzy logic integration.The other settings were also optimized based on the initial results of the training algorithm, whereas the most optimal activation function was selected for this scenario.In further studies, algorithms are also being trained using different activation functions to check on the results.There is also a check for early stoppage in case of overfitting in algorithm training.
For the initial study, the neural network training consists of 5 hidden layers, 1 input and 1 output layer.The number of neurons in each layer of the neural network is shown in Table 2 for each case; this is taken to have a better understanding of the comparison.Also, optimizing the algorithms further with changes in the number of neurons did not generate any different results.There is an early stop function to check for overtraining and to achieve the most optimized result for the trained algorithm.During the initial optimization phase, multiple tests were conducted to refine the algorithm, with an early stoppage implemented to prevent overtraining.Beyond a specific threshold, increasing the number of neurons resulted in a decline in algorithm performance.For future reference, a statistical model will be utilized to further optimize neuron selection, enabling comparisons of both models under identical conditions for valuable insights into future enhancements.Figure 8 shows the algorithm's pseudo code.It explains the algorithm's implementation step by step, including the hidden layers, where a fuzzy layer is used instead of the usual neural network layer for the second case.This is used to help give more options to the weight adjustment of neural network hidden layers and induce a control layer in between to achieve better results.

Results
Once the model is trained, it undergoes blind validation using test data samples that were separated at the start of the training process.The validation accuracy is calculated for both the traditional neural and fuzzy-neuro network models, as presented in Figure 9.In this study, the training was conducted over a maximum of 350 epochs, with an early stopping mechanism to prevent overtraining and ensure optimal machine learning model performance.The validation accuracy results for both models are summarized in Table 3 below: The fuzzy-neuro network significantly outperforms the traditional neural network, achieving a validation accuracy of 99.40% compared to 94.34%.An improvement of more than 5% demonstrates the good capability of the fuzzy-neuro network in accurately predicting faults.Moreover, the results suggest that the fuzzy-neuro network's advantage will be even more pronounced in more complex scenarios involving multiple types of faults.The model can further refine weight adjustments by incorporating additional fuzzy logic layers, leading to even higher accuracy.This scalability and adaptability make the fuzzy-neuro network a promising solution for sophisticated fault detection tasks in electrical machines.
The substantial increase in accuracy can be attributed to the integration of fuzzy logic within the neural network framework.Fuzzy logic allows the model to handle uncertainties and approximate reasoning more effectively than traditional binary logic, which is particularly beneficial in the context of fault prediction where data can be noisy and complex.
1.The fuzzy-neuro network's ability to manage complex fault patterns stems from its flexible rule-based approach.This leads to a more robust and nuanced fault detection system capable of identifying subtle differences between healthy and faulty states.2. The results indicate that the fuzzy-neuro network is more accurate and adaptable to varying conditions.As industrial machines operate under diverse and often unpredictable environments, the ability of the fuzzy-neuro network to adapt to different types of faults and operational scenarios is a significant advantage.3. The higher accuracy achieved by the fuzzy-neuro network also reduces false positives and negatives.This is crucial in industrial applications, where the cost of undetected faults (false negatives) or unnecessary maintenance (false positives) can be substantial.By minimizing these errors, the fuzzy-neuro network contributes to more efficient maintenance schedules and reduces the risk of unexpected machine downtime.
4. The fuzzy-neuro network model's high accuracy and robust performance shows potential for real-time fault detection and predictive maintenance systems.Implementing such a model in real time could provide continuous monitoring and immediate fault diagnosis, further enhancing operational efficiency and machine longevity.
The integration of fuzzy logic within neural networks presents a significant advancement in fault prediction for electrical machines.The fuzzy-neuro network's good accuracy, adaptability, and potential for real-time application make it a valuable tool for enhancing predictive maintenance and ensuring the reliability of critical industrial machinery.

Conclusions
Diagnostics of electrical machines has always been an essential aspect of research.With the revolution in industrial applications, more emphasis has been placed on predictive maintenance rather than scheduled maintenance.With more research moving towards condition monitoring and real-time fault detection, neural network-trained models have become essential to current fault detection methods.
This paper presents a conceptual short comparison between fault detection using neural network-trained models and fuzzy-neuro network-trained models.It introduces a technique to utilize fuzzy control logic within the neural network to train models that can be utilized for fault detection with better accuracy.It can be seen from the results that the fuzzy-neuro network-trained models show 5% more accuracy for validation cases as compared to the neural network-trained models for the same data set.
It can be interpreted that fuzzy-neuro models can be used to train more complex models especially where the number of faults integrated in a single model is more than five as the accuracy of neural network-trained models start going down with complex combinations.
The proposed method is still in its early stages of development and validation; hence, it needs to be applied to more complex cases to verify its accuracy.It might also need more optimization and testing to achieve the best configuration and combination of neural networks with fuzzy logic to achieve optimal results.Hence, more work is needed to find out the exact restrictions or limitations of the method and how they can be addressed.

Figure 3 .
Figure 3. Methodology of the suggested approach.

Figure 5 .
Figure 5. Overview of the flow of the algorithm.

Figure 6 .
Figure 6.The structure of the fuzzy logic block.

Figure 7 .
Figure 7. General overview of the training algorithm with fuzzy logic layer.

Figure 8 .
Figure 8. Pseudo code of the algorithm.

Figure 9 .
Figure 9. Validation results from (a) machine learning algorithm and (b) neuro-fuzzy machine learning algorithm.

Table 2 .
Number of neurons per layer for both neural network and fuzzy-neuro network training.

Table 3 .
Validation accuracy results for both algorithms.