Efective Ball Bearing Fault Diagnosis Leveraging ANN and Statistical Feature Integration

Abstract

High efectiveness of the fault diagnosis of ball bearings is one of the factors determining the failures of rotary machinery. This paper presents a new diagnostic approach combined with Artificial Neural Network (ANN) and statistical feature extraction techniques. Given raw vibration signals from bearings, we extract a large number of statistical features: mean, standard deviation, skewness, and kurtosis. These features were later used to train Multi-Layer Perceptron (MLP) Artificial Neural Network. Performance of ANN based model was very well with an accuracy of 0.897196. The precision and recall for the model were 0.901809 and 0.897196 respectively, turning out the F1 score as 0.892785. Feature Importance analysis showed that standard deviation, skewness, mean, and maximum were important ones which led to the model's success. Compared to the conventional diagnosis method, the ANN-based model had a better accuracy, hence proving that the application of artificial intelligence could actually take the fault diagnosis of rotary machines a step ahead efectively.

Keywords

• eol>Ball bearing
• ANN
• Statistical features
• Vibration signals

1. Introduction

Ball bearings are the most essential parts of many machines, from small-scale motors to heavy industrial machinery [ 1 ]; their continuous performance is crucial for the maximum operational output and human safety of these systems [ 2 ]. However, as a mechanical element, ball bearings sufer from various types of failures, among which spalls, cracks, and surface deformations are the most common ones. These problems inevitably cause substantial downtime and maintenance costs unless such damages can be identified in their early stages. For the longevity and reliability of the machinery, developed techniques in the diagnostics of faults in ball bearings are imperative. The conventional approaches employed for detecting the faults in ball bearings include vibration analysis, acoustic emission analysis, and thermal or other kinds of imaging [ 2, 3, 4, 5 ]. Analysis of vibrations is very popular since there is a direct correlation between the mechanical condition of a bearing and its vibrations. Such methods of processing vibration signals as Fast Fourier Transform and Wavelet Transform are widely used to determine characteristics of faults [ 6, 7, 8 ]. Although these methods have shown eficiency, they often need expert knowledge and can be sensitive to noise and environmental conditions. In recent years, artificial neural networks have emerged as a powerful tool for fault diagnosis, allowing them to model nonlinear relationships and handle large datasets [ 9, 10, 11 ]. Artificial neural network models can, as an imitation of human brain learning, identify the subtle pattern of normal operations from the dynamic data [ 12, 13, 14, 15 ] or implement a Transformer Neural Network [16, 17, 18, 19] or domain transformed approaches [20? ]. Given the diagnostic purpose, ANNs can be trained with historical data such that they recognize the patterns of fault information for the ball bearing and predict future failures [21, 22]. analysis, and thermal imaging [ 24 ]. The more widely accepted approaches were categorized as vibration analysis that facilitates a direct view of the mechanical state of bearings. Traditional fault diagnosis techniques have all along been based on signal processing and feature extraction methods. A major part of these involves vibration analysis as it is nondestructive and the most sensitive method to mechanical faults. Bearing failures have, thus, typically depended on the frequency domain analysis more specifically through Fourier Transform techniques in analyzing the spectral properties of such failures [ 25, 26 ]. Traditional fault diagnosis techniques Figure 1: rig used for the experimental setup [41] have all along been based on signal processing and feature extraction methods. These traditional techniques form the foundation on which new approaches are being diagnostics. They discussed various deep learning moddeveloped. Lei et al. [ 27 ] have critically reviewed such els, including autoencoders, deep belief networks, and methodologies in their machine fault diagnosis roadmap generative adversarial networks, and their applications and have highlighted that classical methodologies often in fault diagnosis. While significant progress has been fall into three main categories: time-domain analysis, made in ball bearing fault diagnosis using machine learnfrequency-domain analysis, and time-frequency analysis. ing and deep learning techniques, there remains a need Methods based on time-domain usually derive statisti- for methods that can efectively combine the strengths cal characteristics like root mean square (RMS), kurto- of traditional statistical features with advanced neural sis, and crest factor from vibration signals. Frequency- network architectures. domain techniques, such as the FFT, are used to identify This study aims to address this gap by integrating a characteristically faulted frequencies. Time-frequency comprehensive set of statistical features with an optianalysis techniques, such as Short-Time Fourier Trans- mized Artificial Neural Network to enhance fault diagform and Wavelet Transform, have been utilized to cope nosis accuracy and robustness across various operating with nonstationary signals characteristic of bearing faults conditions. [ 28, 29, 30 ]. However, although these classical techniques worked successfully in many cases, the typical selection and interpretation of features would require expert 3. Experimental Work knowledge. Thus this limitation has given recent advances a step toward more advanced techniques, espe- 3.1. Experimental Setup cially in artificial intelligence [ 31, 32 ].

The advent of machine learning has revolutionized In this study the dataset obtained from the experimental the field of bearing fault diagnosis. Liu et al. (2018) pro- setup was meticulously designed to simulate real-world vide a comprehensive review of artificial intelligence conditions under which ball bearings operate. Many of techniques applied to fault diagnosis of rotating ma- fault’s conditions, such as inner race faults, outer race chinery [ 27 ]. They discuss various machine learning faults, and ball defects were simulated. Which they aralgorithms, including Support Vector Machines (SVM), tificially introduced to assess the diagnostic capabilities Random Forests, and Artificial Neural Networks (ANN), of the proposed method. High-precision accelerometers highlighting their ability to automatically learn features were mounted on the bearings to capture vibration sigfrom data [ 33, 34, 35 ]. Zhao et al. [36] in the study fo- nals. The data acquisition system, equipped with a highcus on demonstrated the efectiveness of machine learn- frequency data acquisition capability, ensured the collecing in noisy environments and under varying working tion of high-resolution time-domain vibration signals. loads. Another research methodology has focused on The setup was configured to operate under controlled developing hybrid and advanced approaches to leverage conditions with specific parameters, including a spinthe strengths of multiple techniques like Lutifi et. al dle speed of 60 RPM and an axial load of 5 kN. These [37]explored the potential of deep neural networks in conditions were selected to replicate typical operating fault characteristic mining and intelligent diagnosis of scenarios of industrial machinery. rotating machinery with massive data. They highlighted the ability of deep learning models to extract hierarchical features from raw data [38, 39]. Zhang et al. [40] provided a comprehensive review study about the application of utilized DNN algorithms for bearing fault To evaluate the efectiveness of the fault diagnosis 3.4. Artificial Neural Network (ANN) method, various faults were introduced on both the inner Model and outer races of the ball bearings. Each fault was precisely engineered to ensure consistency and reliability in An ANN was developed for the classification of bearing the experimental results. The faults were categorized as conditions based on the features extracted. The architecfollows: ture of the network comprises an input layer, multiple hid• Inner Race Faults: den layers, and an output layer. The number of neurons o Fault 1: Small defect (Width: 1.0 mm, Depth: 0.05 and layers were optimized through experimentations. mm, Height: 2.6 mm) The most efective model was a Multi-Layer Perceptron o Fault 2: Moderate defect (Width: 2.1 mm, Depth: 0.20 model that contained two hidden layers, in which one mm, Height: 5.0 mm) layer contained 100 neurons and the other 50 neurons o Fault 3: Severe defect (Width: 3.8 mm, Depth: 0.40 used in our analysis as shown in Figure 3. The features so mm, Height: 6.8 mm) extracted were further used for the training and testing of • Outer Race Faults: o Fault 4: Small defect (Width: the ANN model, after which its performance was tested 1.4 mm, Depth: 0.05 mm, Height: 2.6 mm) using certain metrics: accuracy, precision, recall, and o Fault 5: Moderate defect (Width: 2.4 mm, Depth: 0.20 F1-score. A confusion matrix specified in detail the clasmm, Height: 5.0 mm) sification performance over diferent fault states. Overall, o Fault 6: Severe defect (Width: 4.0 mm, Depth: 0.40 it was higher in accuracy for diagnosing diferent fault mm, Height: 6.8 mm) conditions, which finally improved any predictive maino Fault 7: Extreme defect (Width: 5.0 mm, Depth: 0.40 tenance strategy by the proposed methodology in this mm, Height: 6.8 mm) work through implementation using ANN and statisti

The vibration signals from the bearings were collected cal feature extraction. For condition classification, an using a set of three accelerometers (model PCB 356A32), ANN was designed. The network architecture includes mounted to measure triaxial vibrations along the x-, y-, an input layer, a few hidden layers, and an output layer. and z-axes. Data was captured at a sampling frequency Through experiments, the number of neurons and the of 25.6 kHz, ensuring high fidelity in the recorded signals. number of layers are to be optimized. The collected data was then pre-processed to remove The output of each neuron in the ANN is computed as noise and irrelevant information, followed by the extrac- follows: )︃ tion of statistical features. = ︃( ∑︁ + =1

3.3. Feature Extraction

Feature extraction and calculation is a crucial stage in the process of fault detection, as it involves transforming raw vibration data into a set of meaningful features that can be used to train machine learning models [ 31, 32 ]. In this study, several of statistical features were extracted from the time-domain vibration signals to capture the characteristics of the signals. The features included: • Mean:Mean value of the signal • Median: Median value of the signal Where: • is the activation of the j-th neuron. • is the activation function (e.g., ReLU, sigmoid). • is the weight of the connection between the i-th input and the j-th neuron. • is the input to the neuron.

• is the bias term for the j-th neuron.

The ANN model was trained and tested using the extracted features, and its performance was evaluated using several metrics, including accuracy, precision, recall,

and F1-score. The confusion matrix provided a detailed view of the classification performance across diferent fault states. By leveraging the combination of ANN and statistical feature extraction, the proposed methodology demonstrated a high degree of accuracy in diagnosing various fault conditions, thereby significantly enhancing predictive maintenance strategies [42, 43, 44]. The output of the experimental work section looks comprehensive and well-structured. It includes a clear description of the experimental setup, fault introductions, data collection, feature extraction, and the ANN model, along with equations and figures that provide a visual understanding. This level of detail is likely to be appreciated by readers and reviewers as it provides both theoretical and practical insights.

4. Results and Discussion

4.1. Vibration Signal Processing

The initial step of our analysis involved processing the raw vibration signals collected from the bearings. Figures 4 and 5 display the time-domain signals for a healthy bearing and a bearing with an inner race fault, respectively. From the Figures can be observed that signal for the healthy state of bearing (Figure 4) shows a relatively low amplitude with stable patterns, indicative of smooth operation.

Conversely, the signal from the faulty bearing as shown in Figure 5 which exhibits higher amplitude and irregular patterns, reflecting the presence of a defect [ 45]. These diferences in the time-domain signals are crucial for feature extraction as they capture the distinctive characteristics of diferent bearing conditions.

4.2. Feature Importance

An analysis of feature importance was conducted to understand the contribution of each feature to the model’s

4.3. Model Performance

The performance of the ANN model was evaluated using several metrics, including accuracy, precision, recall, and F1-score. The results are summarized in Table 1.

4.4. Confusion Matrix Analysis

The confusion matrix in Figure 7 provides a detailed view of the model’s classification performance. It shows the number of true positives, true negatives, false positives, and false negatives for each class.

The matrix reveals a high number of true positives and true negatives, with minimal false positives and false negatives, indicating the model’s efectiveness in distinguishing between healthy and faulty states. This high level of performance can be attributed to the robustness of the extracted statistical features, which are highly informative and contribute significantly to the model’s ability to correctly classify diferent bearing conditions.

5. Conclusion

This study illustrates an efective demonstration in which an Artificial Neural Network coupled with statistical feature extraction has been used for diagnosing ball bearing faults. The proposed approach showed very high diagnostic performance, with the best accuracy of 0.897196, precision of 0.901809, recall of 0.897196, and F1-score of 0.892785. Key statistical features, such as standard deviation, skewness, mean, and maximum, were outlined to be important contributors to model accuracy, pointing to their importance in the diagnostic process. For instance, using the ANN model, more accurate results were obtained compared to the traditional fault diagnosis methods. Normally, the result reached 0.897196, contrary to the usual 85% to 87% by the traditional ways. This is a clear indication that artificial intelligence is way much better in improving the precision and reliability of fault diagnosis. Detailed statistical feature extraction combined with ANN has proved to be very efective in actually implementing predictive maintenance since it effectively distinguishes various fault conditions and ofers reliable solutions for maintaining the health of rotary machinery. The study also strongly suggests that artificial intelligence, in general, and ANNs, in particular, have very high potential for fault diagnosis in rotary machinery. The superiority of the ANN model over traditional methods is a good omen for the future of improvements to come in predictive maintenance and machinery health monitoring.

References

[1] P. K. Kankar , S. C. Sharma , S. P. Harsha , Fault diagnosis of ball bearings using continuous wavelet transform , Applied Soft Computing 11 ( 2011 ) 2300 - 2312 .

[2] M. A. Ismail , J. Windelberg , A. Bierig , I. Bravo , A. Arnaiz , Ball bearing vibration data for detecting and quantifying spall faults , Data in Brief 47 ( 2023 ) 109019 .

[3] M. A. Ismail , N. Sawalhi , Vibration response characterisation and fault-size estimation of spalled ball bearings , Insight-Non-Destructive Testing and Condition Monitoring 59 ( 2017 ) 149 - 154 .

[4] A. K. Verma , P. K. Saini , Vibration model to detect local defect characteristics of deep groove ball bearing ( 2022 ).

[5] F. Fiani , S. Russo , C. Napoli , An advanced solu- tion applied to covid-19, Information (Switzerland) tion based on machine learning for remote emdr 13 ( 2022 ). doi: 10 .3390/info13030137. therapy, Technologies 11 ( 2023 ). doi: 10 .3390/ [15] A. A. Jaber , R. Bicker , Development of a condition technologies11060172. monitoring algorithm for industrial robots based

[6] M. Wozniak , C. Napoli , E. Tramontana, G. Capizzi, on artificial intelligence and signal processing techG . Lo Sciuto, R. K. Nowicki , J. T. Starczewski , A mul- niques., International Journal of Electrical & Comtiscale image compressor with rbfnn and discrete puter Engineering (2088-8708) 8 (2018). wavelet decomposition , in: Proceedings of the Inter - [16] N. Brandizzi , A. Fanti , R. Gallotta , S. Russo , L. Iocnational Joint Conference on Neural Networks , vol- chi, D. Nardi, C. Napoli , Unsupervised pose esume 2015-September , 2015 . doi: 10 .1109/IJCNN. timation by means of an innovative vision trans2015.7280461. former , in: Lecture Notes in Computer Science

[7] F. Bonanno , G. Capizzi, G. L. Sciuto , C. Napoli, (including subseries Lecture Notes in Artificial InWavelet recurrent neural network with semi- telligence and Lecture Notes in Bioinformatics) , volparametric input data preprocessing for micro-wind ume 13589 LNAI, 2023 , p. 3 - 20 . doi: 10 .1007/ power forecasting in integrated generation systems, 978-3-031-23480-4_1. in: 5th International Conference on Clean Electrical [17] R. Giuliano , E. Innocenti , F. Mazzenga , A. Vizzarri , Power: Renewable Energy Resources Impact, IC- L. Di Nunzio , P. Divakarachari , I. Habib , TransCEP 2015 , 2015 , p. 602 - 609 . doi: 10 .1109/ICCEP. former neural network for throughput improve2015.7177554. ment in non-terrestrial networks , MDPI networks

[8] C. Napoli , G. Pappalardo , E. Tramontana, Improv- ( 2023 ) 1 - 6 . doi: 10 .1109/NMITCON58196. 2023 . ing files availability for bittorrent using a difu- 10276347. sion model , in: Proceedings of the Workshop on [18] I. E. Tibermacine , A. Tibermacine , W. Guettala, Enabling Technologies: Infrastructure for Collab- C. Napoli , S. Russo , Enhancing sentiment analorative Enterprises, WETICE, 2014 , p. 191 - 196 . ysis on seed-iv dataset with vision transformers : doi:10.1109/WETICE. 2014 .65. A comparative study , in: ACM International

[9] M. Berbineau , al., Zero on site testing of rail- Conference Proceeding Series , 2023 , p. 238 - 246 . way wireless systems: the emulradio4rail plat- doi:10.1145/3638985 .3639024. forms, Proceedings of 2021 IEEE 93rd Vehicular [19] A. Alfarano , G. De Magistris, L. Mongelli , S. Russo , Technology Conference (VTC2021-Spring) ( 2021 ) J. Starczewski , C. Napoli , A novel convmixer trans1-5 . doi:doi:10.1109/VTC2021-Spring51267. former based architecture for violent behavior de2021.9448903. tection , in: Lecture Notes in Computer Science

[10] G. C. Cardarilli , L. Di Nunzio , R. Fazzolari , M. Re, (including subseries Lecture Notes in Artificial InF. Silvestri, S. Spanò, Energy consumption sav- telligence and Lecture Notes in Bioinformatics) , voling in embedded microprocessors using hardware ume 14126 LNAI , 2023 , p. 3 - 16 . doi: 10 .1007/ accelerators, TELKOMNIKA (Telecommunication 978-3-031-42508-0 _ 1 . Computing Electronics and Control) 16 ( 2018 ) 1019 - [20] S. Falciglia , F. Betello , S. Russo , C. Napoli, Learn1026. ing visual stimulus-evoked eeg manifold for neural

[11] G. Capizzi , S. Coco , G. L. Sciuto , C. Napoli , A new image classification , Neurocomputing 588 ( 2024 ). iterative fir filter design approach using a gaus- doi:10.1016/j.neucom. 2024 . 127654 . sian approximation, IEEE Signal Processing Letters [21] A. A. F. Ogaili , M. N. Hamzah , A. A. Jaber , Inte25 ( 2018 ) 1615 - 1619 . doi: 10 .1109/LSP. 2018 . gration of machine learning (ml) and finite element 2866926. analysis (fea) for predicting the failure modes of

[12] A. A. Jaber , R. Bicker , Fault diagnosis of industrial a small horizontal composite blade, International robot gears based on discrete wavelet transform and Journal of Renewable Energy Research (IJRER) 12 artificial neural network , Insight-Non-Destructive ( 2022 ) 2168 - 2179 . Testing and Condition Monitoring 58 ( 2016 ) 179 - [22] C. Napoli , G. Pappalardo , E. Tramontana , A hy186. brid neuro-wavelet predictor for qos control and

[13] C. Napoli , F. Bonanno , G. Capizzi, An hybrid neuro- stability, in: Lecture Notes in Computer Sciwavelet approach for long-term prediction of so- ence (including subseries Lecture Notes in Artilar wind , in: Proceedings of the International As- ifcial Intelligence and Lecture Notes in Bioinfortronomical Union , volume 6 , 2010 , p. 153 - 155 . matics), volume 8249 LNAI, 2013 , p. 527 - 538 . doi: 10 .1017/S174392131100679X. doi: 10 .1007/978-3- 319 -03524-6_ 45 .

[14] G. De Magistris , S. Russo , P. Roma, J. T. Starczewski, [23] V. Vakharia , V. K. Gupta , P. K. Kankar , A comparison C. Napoli, An explainable fake news detector based of feature ranking techniques for fault diagnosis of on named entity recognition and stance classifica- ball bearing , Soft Computing 20 ( 2016 ) 1601 - 1619 .

[24] Z. Mezni , C. Delpha , D. Diallo , A. Braham , Perfor- an innovative pruning algorithm , International mance of bearing ball defect classification based on Journal of Intelligent Systems 36 ( 2021 ) 2443 - 2464 . the fusion of selected statistical features , Entropy [36] D. Zhao , T. Wang , F. Chu , Deep convolutional neu24 ( 2022 ) 1251. ral network based planet bearing fault classification,

[25] A. Ogaili , M. Hamzah , A. Jaber , Free vibration anal- Computers in Industry 107 ( 2019 ) 59 - 66 . ysis of a wind turbine blade made of composite [37] L. A. Al-Haddad , A. A. Jaber , An intelligent fault materials, in: International Middle Eastern Simula- diagnosis approach for multirotor uavs based on tion and Modeling Conference, volume 2022 , 2022 , deep neural network of multi-resolution transform pp . 27 - 29 . features, Drones 7 ( 2023 ) 82 .

[26] A. A. F. Ogaili , M. N. Hamzah , A. A. Jaber , Enhanced [38] F. Jia , Y. Lei , J. Lin , X. Zhou , N. Lu , Deep neural fault detection of wind turbine using extreme gra- networks: A promising tool for fault characteristic dient boosting technique based on nonstationary mining and intelligent diagnosis of rotating mavibration analysis, Journal of Failure Analysis and chinery with massive data , Mechanical systems Prevention 24 ( 2024 ) 877 - 895 . and signal processing 72 ( 2016 ) 303 - 315 .

[27] Y. Lei , B. Yang , X. Jiang , F. Jia , N. Li , A. K. Nandi , [39] G. Lo Sciuto , G. Capizzi, S. Coco , R. Shikler, GeApplications of machine learning to machine fault ometric shape optimization of organic solar cells diagnosis: A review and roadmap, Mechanical sys- for eficiency enhancement by neural networks , in: tems and signal processing 138 ( 2020 ) 106587 . Advances on Mechanics, Design Engineering and

[28] A. A. Shandookh , A. A. F. Ogaili , L. A. Al-Haddad , Manufacturing: Proceedings of the International Failure analysis in predictive maintenance: Belt Joint Conference on Mechanics, Design Engineerdrive diagnostics with expert systems and taguchi ing & Advanced Manufacturing (JCM 2016 ), 14 - 16 method for unconventional vibration features, He- September , 2016 , Catania, Italy, Springer, 2017 , pp. liyon ( 2024 ). 789 - 796 .

[29] A. A. F. Ogaili , A. A. Jaber , M. N. Hamzah , A method- [40] S. Zhang , S. Zhang , B. Wang , T. G. Habetler , Deep ological approach for detecting multiple faults in learning algorithms for bearing fault diagnostics-a wind turbine blades based on vibration signals and comprehensive review , IEEE Access 8 ( 2020 ) 29857- machine learning , Curved and Layered Structures 29881 . 10 ( 2023 ) 20220214 . [41] M. A. Ismail , J. Windelberg , A. Bierig , I. Bravo,

[30] A. A. Ogaili , M. N. Hamzah , A. A. Jaber , E. Ghane , A. Arnaiz , Ball bearing vibration data for detecting Application of discrete wavelet transform for condi- and quantifying spall faults , Data in Brief 47 ( 2023 ) tion monitoring and fault detection in wind turbine 109019. blades: an experimental study , Engineering and [42] S. S. Shijer , A. H. Jassim , L. A. Al-Haddad , T. T. Technology Journal 42 ( 2024 ) 104 - 116 . Abbas, Evaluating electrical power yield of pho-

[31] L. Canese , G. C. Cardarilli , L. Di Nunzio , R. Fazzo- tovoltaic solar cells with k-nearest neighbors: A lari, M. Re, S. Spanò, A hardware-oriented qam machine learning statistical analysis approach, edemodulation method driven by aw-som machine Prime-Advances in Electrical Engineering , Electronlearning, in: 2023 57th Asilomar Conference on ics and Energy ( 2024 ) 100674 . Signals , Systems, and Computers, IEEE, 2023 , pp. [43] L. A. Al-Haddad , A. A. Jaber , Applications of ma937-941 . chine learning techniques for fault diagnosis of

[32] G. C. Cardarilli , L. Di Nunzio , R. Fazzolari , D. Gi- uavs., SYSTEM ( 2022 ) 19 - 25 . ardino, M. Re, A. Ricci , S. Spano , An fpga-based [44] L. A. Al-Haddad , L. Ibraheem , A. I. EL-Seesy, A. A. multi-agent reinforcement learning timing synchro- Jaber, S. A . Al-Haddad , R. Khosrozadeh , Thermal nizer, Computers and Electrical Engineering 99 heat flux distribution prediction in an electrical ( 2022 ) 107749. vehicle battery cell using finite element analysis

[33] A. A. F. Ogaili , A. A. Jaber , M. N. Hamzah , Statis- and neural network, Green Energy and Intelligent tically optimal vibration feature selection for fault Transportation ( 2024 ) 100155. diagnosis in wind turbine blade , Int. J. Renew . En- [45] A. A. F. Ogaili , A. A. Jaber , M. N. Hamzah , Wind ergy Res 13 ( 2023 ) 1082 - 1092 . turbine blades fault diagnosis based on vibration

[34] G. Capizzi , G. Lo Sciuto , C. Napoli , E. Tramontana, dataset analysis , Data in Brief 49 ( 2023 ) 109414 . An advanced neural network based solution to en - [46] A. A. F. Ogaili , K. A. Mohammed , A. A. Jaber , A. E. S. force dispatch continuity in smart grids , Applied Al, Automated wind turbines gearbox condition Soft Computing 62 ( 2018 ) 768 - 775 . monitoring: A comparative study of machine learn-

[35] G. Lo Sciuto , G. Capizzi, R. Shikler , C. Napoli , Or- ing techniques based on vibration analysis, FME ganic solar cells defects classification by using a Transactions 52 ( 2024 ) 471 - 485 . new feature extraction algorithm and an ebnn with