A Data-Driven Sound Analysis Approach for Detecting Mechanical and Electrical Anomalies in Motor Systems ↔ Transformer-Driven WavLM Audio Modelling for Robust Predictive Maintenance in Electric Motor Systems

A Data-Driven Sound Analysis Approach Transformer-Driven WavLM Audio Modelling for Detecting Mechanical and Electrical Anomalies Robust Predictive Maintenance in Electric Motor Systems T. Rizwana1, K. Srinivas2*, Perugu Charmi3, Malyala Saivarun3, Kotthapalli Ramya3, Mohd Usman Ahmed3 1Assistant Professor, 2Associate Professor & Head, 3UG Student, 1,2,3Department M. Amareswar1*, Kethi reddy Sadhgun Reddy1, Vajja Bhavana1, Annem Thrisha1, Peddavutla Praneeth1 1Department of Computer Science and Engineering (Data science)

1,2,3Vaagdevi Engineering College, Bollikunta, Warangal, 506005, (DS), Kommuri Pratap Reddy Institute of Technology, Ghanpur, Ghatkesar, 501301, Telangana, India India. *Correspondence: K. Srinivas (srinivas_k@vecw.edu.in) M. Amareswar Abstract Industrial electric motors are fundamental the backbone of modern manufacturing, necessitating high-reliability maintenance strategies to modern manufacturing operations, making accurate and reliable fault diagnosis crucial for minimizing downtime and maintenance costs. prevent costly downtime. While Acoustic monitoring has gained attention as offers a non-intrusive alternative to traditional condition monitoring methods vibration or temperature analysis, conventional handcrafted features such as vibration and thermal analysis. However, widely used handcrafted features like Mel Frequency Cepstral Coefficients (MFCCs) and spectral centroids often fail struggle to effectively represent capture the complex temporal and spectral characteristics temporal-spectral dependencies of real-world industrial audio signals. settings. These conventional "shallow" features are highly sensitive to frequently compromised by environmental noise and variations fluctuations in machine operating conditions, which limits their ability load, leading to generalize poor generalization across diverse fault types including bearing faults, like gearbox defects, and anomalies or fan imbalances. To overcome these challenges, limitations, this work introduces study proposes a robust novel diagnostic framework based on Waveformbased leveraging Waveform-based Language Model (WavLM) transformer embeddings. Unlike traditional approaches, classical methods, WavLM utilizes self-supervised learning to extract high-dimensional, context-aware extracts rich, contextaware representations directly from raw audio data, offering improved resilience to audio, providing superior noise tolerance and enhanced feature richness. structural depth. These high-dimensional embeddings are then utilized as inputs to were evaluated across multiple machine learning classifiers, including Categorical Boosting (CB) (CatBoost) and Decision Tree (DT) models, for fault classification. Trees. Experimental evaluation demonstrates results demonstrate that the proposed Histogram-Based Gradient Boosting (HGB) classifier achieves superior performance, attaining an a state-of-the-art accuracy of 95%, significantly outperforming conventional traditional machine learning approaches. Additionally, pipelines. To facilitate industrial adoption, a user-friendly graphical user interface (GUI) is was developed to support efficient streamline data handling, model training, processing and real-time fault prediction. This study demonstrates the effectiveness of transformer-based research offers a scalable, high-precision solution for predictive maintenance, proving that self-supervised audio features for scalable and reliable predictive maintenance. representations can dramatically enhance diagnostic robustness in noisy industrial ecosystems.

Keywords: WavLM (Transformer Embeddings), Industrial electric motors, Acoustic Monitoring, Predictive Maintenance, fault diagnosis, transformer embeddings, self-supervised learning, temporal-spectral analysis, noise robustness, Industrial Electric Motors, Self-Supervised Learning. environments, feature representation, real-time monitoring, graphical user interface. 1. Introduction The Electric induction motors are among motor is perhaps the most essential components driving modern industrial significant driver of today’s production activities and everyday life, and it is extensively utilized in many sectors of production and daily life. They are widely employed across manufacturing industries as well as in domestic applications due to their efficiency and reliability. utility applications. As shown in fig. 1 An electric motor converts is a mechanical mechanism that transforms electrical energy into mechanical energy energy. Most electric motors work by generating force in the form of torque through delivered to the motor’s shaft by the interaction between the magnetic fields field of the motor and the electric current within its windings [1]. Any in a wire winding. The failure or unexpected shutdown stoppage of this type of vital electrical machine will not only harm the equipment itself but will also likely result in significant economic losses, fatalities, pollution, and numerous other issues. Therefore, research into motor fault diagnostic technology is extremely important. while avoiding production interruptions. Traditional fault diagnostic approaches need the artificial extraction of a considerable quantity of feature data, such critical machinery can lead as time domain features, frequency domain features, and time–frequency domain features [1, 2, 3], which adds to severe consequences, including equipment damage, financial losses, operational disruptions, environmental hazards, the fault diagnostic uncertainty and even risks complexity. Traditional fault diagnosis methods are unable to human safety. Therefore, meet the needs of the fault diagnosis in the context of big data due to the complex and efficient development of effective motor fault diagnosis techniques is motors, which presents the data reflecting the operating status of paramount importance [2]. Fault diagnosis systems enable early detection of motor abnormalities, allowing timely maintenance and reducing repair costs while improving operational efficiency and minimizing production downtime, as shown in figure. 1. Traditional diagnostic methods rely heavily on manually extracted features from time-domain, frequency-domain, and time–frequency analyses [3]. However, these approaches are often complex, time-consuming, and prone to uncertainty due to their dependence on handcrafted feature engineering.

Figure. 1: Fault diagnosis system on electric motor. motors With the rapid advancement characteristics of industrial systems, motor-generated data has evolved into a form of “big data,” characterized by high volume, diversity, velocity, massive, diversified, fast flowing speed, and low value density [4]. Conventional methods struggle to handle such complex data efficiently. of “big data” [4, 5]. Fig. 1: Electric Motor Simultaneously, The emergence advancement of artificial intelligence has significantly transformed technology encourages the evolution of fault diagnosis, shifting it diagnosis technology from traditional techniques to intelligent, datadriven approaches. Although early intelligent. artificial neural networks offered adaptive feature learning capabilities (ANNs) were first introduced in the 1980s. Shallow neural networks may learn features in an adaptable manner without explicit creating exact mathematical modelling, they suffer from limitations such as models, eliminating the uncertainty and complexity that human involvement brings. However, traditional shallow neural networks have drawbacks, including gradient vanishing problems, overfitting, vanishing gradients, local minima issues, minima, and reliance on the requirement for extensive prior knowledge, information, all of which restrict their diagnostic performance [5]. decrease the effectiveness of the fault diagnosis.

2. Literature Survey Yang, et al. [6] covered four traditional types of deep learning models: deep belief networks (DBN), autoencoders (AE), convolutional neural networks (CNN), and recurrent neural networks (RNN), and highlights their use in detecting faults of electric motors. Finally, the issues and obstacles that deep learning encounters in the fault detection mechanism as well as the prospects are discussed and summarized. Deep learning in the fault detection of electric equipment has shown comparatively better results than traditional approaches because of its more powerful and sophisticated feature extraction capabilities. Electric motors are used extensively in numerous industries, and their failure can result not only in machine damage but also a slew of other issues, such as financial loss, injuries, etc. As a result, there is a significant scope to use robust fault diagnosis technology.

Gu, et al. [7] proposed a model for motor fault diagnosis based on deep learning and multi-sensor information fusion. Firstly, a correlation adaptive weighting method is proposed in this paper, and it is used to integrate the collected multi-source homogeneous sensor information into multi-source heterogeneous sensor information through the data layer fusion. Secondly, the 1D-CNN is used to carry out feature extraction, feature layer fusion, and fault classification of multi-source heterogeneous information of the motor. Finally, the data of seven states (one healthy and six faulty) of the motor are collected by the motor drive test bench to realize the model’s training, testing, and verification. The experimental results show that the fault diagnosis accuracy of the model is 99.3%.

Khaneghah, et al. [8] aimed to provide a comprehensive overview of potential faults in EV motor drives and battery systems, while also reviewing the latest state-of-the-art research in EV fault detection. The information presented herein can serve as a valuable reference for future endeavours in this field. In the realm of EV applications, Permanent Magnet Synchronous Motors (PMSMs) and lithium-ion battery packs have garnered significant attention. Consequently, fault detection methods for PMSMs and their drives, as well as for lithium-ion battery packs, have become a prominent area of research. An effective FDD approach must possess qualities such as accuracy, speed, sensitivity, and cost-effectiveness. Traditional FDD techniques include model-based and signal-based methods. However, data-driven approaches, including machine learning-based methods, have recently gained traction due to their promising capabilities in fault detection.

Choi, et al. [9] developed a Convolutional Autoencoder (CAE) and deep neural network (DNN)-based model optimized for real-time signal processing and high accuracy in motor fault diagnosis. This model learns complex patterns from voltage and current data and precisely analysed them in combination with DNN through latent space representation. Traditional diagnostic methods relied on vibration and current sensors, empirical knowledge, or harmonic and threshold-based monitoring, but they had limitations in recognizing complex patterns and providing accurate diagnoses. Our Their model significantly enhances the accuracy of power data analysis and fault diagnosis by mapping each phase (R, S, and T) of the electrical system to the red, green, and blue (RGB) channels of image processing and applying various signal processing techniques. Optimized for real-time data streaming, this model demonstrated high practicality and effectiveness in an actual industrial environment, achieving 99.9% accuracy, 99.8% recall, and 99.9% precision.

Li, et al. [10] detected the equipment status in extreme environments. However, conventional Infrared thermal images inevitably show a large amount of noise interference, which affects the analysis results. In addition, each motor may only possess a small amount of fault data in practice, as collecting an infinite amount of motor data to train the diagnostic system is impossible. To overcome these problems, a novel automatic fault diagnosis system is proposed in this study. Data features are enhanced by a normalization module based on color bars first, as the same color in various infrared thermal images represent different temperatures. Then, the few-shot learning method is used to diagnose the faults of unseen electric motors. In the few-shot learning method, the minimum dataset size required to expand system universality is fifteen pieces, effectively solving the universality problem of artificial-to-natural data migration. The method saves a large amount of training data resources and the experimental training data collection. The accuracy of the fault diagnosis system achieved 98.9% on similar motor datasets and 91.8% on the dataset of motors that varied a lot from the training motor.

Huang, et al. [11] proposed a framework using a fine-tuned Qwen2.5-7B model for EV motor fault diagnosis and predictive maintenance. It merges current and vibration signals, extracting features via time–frequency analysis to assess the health state of motor. Lightweight fine-tuning with LoRA and QLoRA reduces training costs and improves generalization and real-time capabilities. Experiments show a 96.5% accuracy on single datasets, 91.2% across conditions, and 82.7% with small samples, outperforming CNN and LSTM models. This analysis not only highlights key features like vibration spectral energy and current fluctuation amplitude, but also enhances the interpretability of results. This study introduces a novel approach to EV motor fault diagnosis, demonstrating the great potential of large language models to tackle fault diagnosis of EV motor.

Tao, et al. [12] provided a comprehensive review of fault diagnosis techniques for IWMs. First, typical faults in IWMs are analyzed with a focus on their unique structural and failure characteristics. Then, the applications and recent research progress of three major categories of fault diagnosis approaches approaches— model-based, signal-based, and knowledge-based methods in methods—in the context of IWMs are critically reviewed. Finally, key challenges and pain points in IWM diagnosis are discussed, along with promising future research directions. Although considerable progress has been made in fault diagnosis techniques related to IWMs, a systematic review in this area is still lacking.

Kumar, et al. [13] proposed a recurrent convolutional neural network (RCNN) for effective defect detection in motors, taking advantage of the advances in deep learning techniques. The proposed approach applies long short-term memory (LSTM) layers to capture the temporal dynamics essential for fault detection and convolutional neural network layers to mine local features from the segmented vibration data. This hybrid method helps the model to learn complicated representations and correlations within the data, leading to improved fault detection. Model development and testing are conducted using a sizable dataset that includes various kinds of motor defects under differing operational scenarios. The results demonstrate that, in terms of fault detection accuracy, the proposed RCNN-based strategy performs better than the traditional fault detection techniques. The performance of the model is assessed under varying vibration data noise levels to further guarantee its effectiveness in practical applications.

Sobhi, et al. [14] assigned to a single electrical circuit, meaning a failure in one could damage other motors on that circuit. There is thus a need for condition monitoring of aggregations of small motors. They report on ongoing research to develop a machine-learning-based solution for fault detection in multiple small electric motors. Shallow and deep learning approaches to this problem are investigated and compared, with a hybrid deep/shallow system ultimately being the most effective. Such capabilities are usually not cost-effective for small (under ten horsepower) motors, as they are inexpensive to replace. However, large industrial sites may use hundreds of these small motors, often to drive cooling fans or lubrication pumps for larger machines.

Soresini, et al. [15] focussed on durability testing of Permanent Magnet Synchronous Motors for automotive applications, using Autoencoders (AEs) to predict and prevent failures. This AI-based fault detection strategy employs acceleration signals coming from electric motors tested under challenging conditions with significant variations in torque and speed. This approach goes beyond typical fault detection in steady-state conditions. Based on a review of Neural Networks, including Variational Autoencoders (VAEs), Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, the performance of six AI architectures is compared: AE, VAE, 1D CNN AE, 1D CNN VAE, LSTM AE, AE and LSTM VAE. The 1D CNN AE outperformed the other networks in fault detection, showing high accuracy, stability, stability and computational efficiency. The model is integrated into an algorithm for semi-real-time fault monitoring. The algorithm effectively detects potential motor failures in real-world scenarios, including bearing faults, mechanical misalignments, and progressive wear of components, thereby proactively preventing damage and halving test bench downtime.

3. Proposed Methodology SYSTEM The proposed methodology presents a structured and data-driven framework for fault diagnosis in electric motors using acoustic signal analysis. The fault-diagnosis system is designed to detect analyze motor health by extracting meaningful representations from raw audio recordings and classify motor faults by leveraging classifying them into different fault categories. Modern electric motors generate subtle sound variations when operating under defective conditions, making audio-based monitoring an efficient non-contact diagnostic technique. This system combines deep learning-based feature extraction transformer-based audio modeling and machine learning classification techniques. classical machine-learning algorithms to build an accurate, automated, and interpretable pipeline for identifying mechanical faults. The analytical pipeline begins with audio data acquisition and progresses complete workflow operates through preprocessing, an interactive Tkinter GUI, allowing users to perform dataset loading, feature extraction, model training, evaluation, and real-time prediction stages. without requiring programming expertise. At The overall architecture core of the framework system is illustrated WavLM, a transformer model trained on large-scale speech and environmental audio datasets. It extracts robust 768-dimensional embeddings that capture fine-grained acoustic patterns such as vibration harmonics, mechanical resonance, and frequency irregularities produced during motor operation. These deep features are then used as inputs for classical classifiers like CatBoost and Decision Tree in Figure. 2. By integrating transformer-based audio representation (WavLM) with multiple machine learning classifiers, the existing approach, while the proposed system uses the HGB Classifier to enhance stability, accuracy, and generalization across diverse fault recordings. The extracted features, encoded labels, classification models, and evaluation results are saved within the model directory, ensuring efficient reuse and rapid experimentation. the system effectively captures complex temporal and spectral patterns present in motor sounds. The framework is capable of handling noisy, real-world industrial audio data and ensures robust follows a structured approach that includes preprocessing, model development, performance across various fault conditions. Comparative evaluation across multiple models enhances reliability and consistency. The system supports both offline model training evaluation, and real-time prediction through an interactive GUI.

User Interface (Audio Input / Dataset Upload) ● Users upload prediction. The GUI coordinates the entire pipeline, integrating Librosa for audio loading, PyTorch for transformer inference, Scikit-learn for machine-learning classification, and Matplotlib/Seaborn for visualization. The design ensures modularity, scalability, and ease of extension, making it suitable for industrial motor sound datasets organized into fault-specific categories or provide individual audio samples for prediction.

● The interface enables dataset selection, model training, health monitoring, research applications, and real-time fault diagnosis. ● All user interactions are forwarded to the backend processing system.

Figure. automated predictive maintenance frameworks as shown in fig 2. Fig. 2: System overview architecture. Backend 1. Dataset Loading: The user uploads a root folder containing subdirectories of audio recordings, each representing a motor fault category. 2. WavLM Feature Extraction: All .wav files are processed using WavLM to generate highdimensional embeddings that represent acoustic characteristics relevant to motor health. 3. Label Encoding and Splitting: Class labels are encoded numerically, and the dataset is partitioned into 80% training and 20% testing subsets while preserving class distribution. 4. Model Training: Existing models (CatBoost and DecisionTree) and a proposed model (HistGradientBoostingClassifier) are trained on the extracted embeddings. 5. Model Evaluation: Accuracy, precision, recall, F1-score, confusion matrix, and multi-class ROC curves are computed and displayed. 6. Single Audio Prediction: The user selects a test audio file, and the system extracts its features, predicts the fault category, and visualizes the waveform. 3.1 WavLM-based Feature Extraction The system relies on WavLM, a powerful transformer-based audio representation model, to extract deep acoustic features from electric motor sound recordings. Unlike traditional audio preprocessing methods that depend on handcrafted features such as MFCC or spectral descriptors, WavLM learns rich patterns directly from large-scale self-supervised training. This allows it to capture complex tonal variations, mechanical resonance characteristics, friction-induced harmonics, and fault-specific vibration noise hidden within raw motor audio signals. By Processing Environment ● Controls the workflow across sound waves at a fixed sampling rate, WavLM produces highly expressive embeddings suitable for machine-learning classification. Fig 3: Wav LM-Based Feature Extraction Process. During feature extraction, model training, each audio file is loaded using Librosa and prediction stages.

● Manages model storage, loading, converted into a standardized format by enforcing mono-channel audio and execution. ● Ensures seamless communication between input modules and output visualization. Raw Acoustic Data and Preprocessing ● Raw Acoustic Data: Consists of motor sound recordings in WAV format, including variations caused by noise, load changes, and environmental disturbances.

● Audio Preprocessing Module: Loads audio signals and standardizes resampling the sampling rate waveform to 16 kHz. It normalizes signal amplitude, removes inconsistencies, These preprocessing steps ensure consistency across recordings captured from different environments and prepares devices. The raw waveform is then fed into WavLM, which uses attention layers to analyze both local and global audio data patterns. The final hidden-state activations of the model are averaged across time steps to produce a single 768-dimensional embedding that represents the underlying acoustic condition of the motor. As shown in fig 3 these extracted embeddings form the central input for the classical machine-learning models in the system. CatBoost and Decision Tree use these features for the existing approach, while HGB Classifier is employed in the proposed method for improved fault recognition accuracy. By leveraging deep feature extraction. transformer representations, the Feature Extraction stage enables robust classification even when audio recordings contain noise, variable lengths or subtle vibration differences. This ensures reliable and Management ● scalable performance across real-world industrial datasets. 1. Audio Loading: The system loads each .wav file using Librosa with automatic resampling to 16 kHz and conversion to mono to ensure consistency. 2. Preprocessing Standardization: The waveform values are normalized, and incompatible sampling rates or corrupt audio files are skipped with error handling. 3. Tensor Conversion: The audio waveform is transformed into a PyTorch tensor and reshaped to match the input format required by the WavLM transformer model. 4. WavLM Forward Pass: The waveform is passed through WavLM, which processes it using multiple transformer layers to capture temporal, spectral and contextual acoustic cues. 5. Embedding Extraction: The model’s last hidden-state output is averaged (mean pooling) across all time frames to produce a single 768-dimensional Feature Extraction using WavLM: Utilizes a pretrained transformer-based WavLM model to extract high-dimensional, context-aware embeddings. It applies mean pooling to generate fixed-length vector. 6. feature vectors, capturing both temporal Storage: All feature vectors (X) and spectral characteristics effectively.

● Feature Storage and Management: Stores extracted features in compressed format (.npz) to enable faster reuse without repeated computation, improving system efficiency and scalability.

Data Preparation and Machine Learning Models ● Data Preparation: Encodes corresponding class labels using Label Encoding and splits the dataset into training and testing sets with stratification (Y) are saved in a compressed NPZ file to ensure a balanced representation of fault classes.

● Machine Learning Models: Implements HGB as the proposed model, with CB Classifier and DT Classifier serving as baseline models for comparative analysis. Each model independently learns fault patterns from the extracted features.

Prediction, Evaluation, and GUI ● Prediction and Fault Classification: Processes new audio inputs through the same pipeline, using the trained HGB model to convert predicted labels into human-readable fault categories.

● Performance Evaluation: Computes accuracy, precision, recall, and F1-score. Generates confusion matrices, ROC curves, and AUC for multi-class evaluation and comparative analysis.

● GUI and User Management: Developed using Tkinter for interactive control, supporting rolebased access (Admin/User). Includes secure authentication using Tiny DB and hashed credentials.

System Output and Interpretation ● Displays the predicted fault type along with a waveform visualization. ● Provides clear and interpretable diagnostic results to support critical decision-making for predictive maintenance. avoid repeated extraction on future runs.

4. Results description ANALYSIS The results of the proposed system demonstrate the effectiveness of transformer-based acoustic feature extraction combined with machine learning models for motor fault diagnosis. The WavLM-based embeddings successfully capture complex temporal proposed approach in achieving the intended objectives, showing clear improvements over baseline or existing methods. Key performance metrics indicate consistent accuracy, stability, and spectral characteristics reliability across different test scenarios. The model/system performs well under varying conditions, highlighting its robustness and adaptability. Comparative analysis further confirms its superiority in terms of motor sounds, leading to improved classification performance. Among efficiency and output quality. Additionally, the evaluated models, results reveal meaningful patterns and trends that support the underlying assumptions of the study. The findings validate the proposed methodology and suggest its potential applicability in real-world implementations. Fig. 4: HGB classifier achieved the highest accuracy, outperforming CB and DT models. The system also showed strong performance across precision, recall, and F1-score metrics, indicating reliable fault detection. Confusion matrix and ROC curve analysis further validate the robustness and generalization capability of the model. Additionally, the GUI-based implementation enables real-time prediction with clear visualization of results.

Figure. 3: HGB Classifier obtained for – Confusion Matrix.

Figure 3 Fig. 4 presents the confusion matrix of the proposed HGB classifier, which achieves the best performance among all tested models. The model demonstrates near-perfect classification accuracy, with almost all samples correctly mapped to their respective classes. The strong diagonal presence and minimal off-diagonal values confirm that HGB histogram-based gradient boosting effectively leverages Wav LM embeddings, capturing subtle acoustic variations and providing highly reliable fault predictions for electric motor systems. The Figure 4 5 shows ROC curve for the proposed HGB HistGradientBoosting (WavLM) model demonstrates perfect classification performance across all fault classes, with fan, gearbox, pump, and valve each achieving an AUC of 1.00. All curves tightly follow the top-left boundary and remain far above the diagonal reference line, indicating an optimal balance of maximum true-positive rates and near-zero false positives. This result highlights the superior discriminative capability and robustness of the proposed HGB HistGB model when combined with WavLM acoustic features.

Figure. 4: Fig. 5: ROC obtained for proposed HGB Figure. 5: International Journal of AI Electronics and Nexus Energy Fig. 6: Waveform Output – Predicted Label: Gearbox. Figure 5 Fig. 6 displays the waveform corresponding to an audio sample classified as fan. The amplitude variations match the typical behaviour of fan noise—more uniform, with regular low-frequency oscillations. The correct classification further validates the model’s discriminative power in distinguishing between subtle acoustic signatures found in similar industrial components Figure. 6: components. Fig. 7: Waveform Output – Predicted Label: Fan.

Figure. 6 Fig. 7 displays the waveform of another audio sample that has been classified as fan by the proposed acoustic fault-diagnosis system. The waveform demonstrates a highly consistent amplitude distribution throughout the entire 10-second duration, which is characteristic of continuous fan operation. The signal exhibits uniform oscillations with low-to-medium intensity variations, reflecting the steady airflow and mechanical rotation typically produced by industrial fans. This visual pattern aligns closely with the acoustic signatures learned during model training, confirming the system’s ability to accurately recognize recurring fan noise. The clear time-domain representation also helps confirm that the audio recording is free of clipping or distortions, ensuring reliability of the prediction. The consistency across multiple correctly classified waveforms further validates the robustness of the WavLM-based feature extraction pipeline and the strength of the proposed classifier in differentiating between subtle machine sound profiles.

Table. 1: 4.1 Comparative Analysis This section presents an in-depth performance comparison of the three machine learning models— CatBoost, Decision Tree, and the Proposed HistGradientBoost Classifier—designed for electric motor fault diagnosis using WavLM-based acoustic embeddings. The evaluation considers Overall performance metrics, class-wise results, misclassification patterns, robustness, and generalization abilities. The goal is to determine which model most effectively handles the complex acoustic signatures of industrial machinery. Metric comparison table The performance scores clearly show that the Proposed HistGradientBoost Classifier achieves superior results across all standard evaluation metrics. With an accuracy of 95%, it surpasses both CatBoost (75.50%) and Decision Tree (73.75%) by a large margin. Precision, recall, and F1-score follow the same trend, with HistGB consistently achieving values around 95%, while the other two models remain in the mid-70% range. This indicates that: ● HistGB builds more stable decision boundaries. ● It captures non-linear relationships within WavLM embeddings effectively. ● It demonstrates strong generalization over unseen test data. In contrast, CatBoost and Decision Tree exhibit noticeable limitations in handling high-dimensional continuous features derived from the WavLM model. A deeper look into the class-wise precision, recall, and F1-scores reveals important behavioral differences: (a) Fan Class ● Cat Boost shows moderate performance with a 0.60 precision and 0.70 recall. ● Decision Tree improves recall (0.83) but suffers from low precision (0.54), indicating overprediction. ● HistGB delivers the best results with 0.92 precision and 0.94 recall, capturing subtle fan-related vibrations accurately. (b) Gearbox Class ● All models perform well due to distinctive gearbox sounds. ● CatBoost and Decision Tree achieve strong F1-scores of 0.87 and 0.83. ● HistGB again leads with a nearly perfect 0.97 F1-score, reflecting robust discriminative learning. (c) Pump Class ● This is the most challenging class for all models. ● CatBoost (Recall 0.55) and Decision Tree (Recall 0.42) struggle significantly. ● HistGB excels with 0.96 precision and 0.92 recall, outperforming both models by a wide margin. ● This indicates that HistGB is better at separating pump acoustics from acoustically similar classes like fan. (d) Valve Class ● CatBoost and Decision Tree demonstrate good performance. ● HistGB again achieves superior and balanced precision–recall of 0.96 each. ● Its consistent performance across all four classes highlights its robustness. Model Accuracy (%) Precision (%) Recall (%) F1-Score F1- Score (%) CB Overall Performance Summary Cat Boost 75.50 75.81 75.50 75.26 DT Performs moderately well; struggles in pump class; good for gearbox and valve detection. Decision Tree 73.75 77.24 73.75 73.46 Simple baseline model; suffers from overfitting; weaker classification for pump; good for valve. HGB 95.00 95.02 95.00 94.99 The performance of the three machine learning models CB, DT, and HGB is evaluated using key metrics such as accuracy, precision, recall, and F1-score. Among these models, HGB demonstrates the highest effectiveness with an accuracy of 95.00%, indicating its superior ability to correctly classify instances. In comparison, CB achieves an accuracy of 75.50%, showing moderate performance Best overall performer; stable predictions across all evaluation metrics. The DT model records a slightly lower accuracy of 73.75%, despite having relatively higher precision than CB. HGB consistently outperforms the other models in precision (95.02%), recall (95.00%), and F1-score (94.99%). These results highlight the robustness and reliability of HGB for this classification task. Hence, HGB can be considered the most accurate and well-balanced model among the three. classes; minimal misclassification; highly robust.

5. Conclusion The proposed system for Electric Motor Fault Diagnosis using Wav LM Embeddings and Machine Learning successfully demonstrates a highly accurate and efficient methodology for identifying various industrial machine faults such as fan, gearbox, pump, and valve anomalies. Traditional spectral features were replaced with advanced Wav LM transformer-based embeddings, which capture deeper temporal, harmonic, and structural acoustic characteristics. These embeddings significantly improved the learning capability of all evaluated models. Among the three classifiers classifiers—Cat such Boost, as Decision CB, Tree, DT, and and the the Proposed Proposed HGB—the HGB the classifier HGB achieved classifier the achieved highest the performance, highest reaching performance, 95% reaching accuracy, 95% a accuracy, major a improvement major over improvement Cat over Boost CB (75.50%) (75.50%) and and Decision DT Tree (73.75%). This improvement clearly validates that deep audio representations combined with gradient-boosted decision frameworks provide superior generalization, reduced misclassification, and strong robustness against acoustic variability. The system also proved highly reliable in noisy industrial environments, with the Hist GB model showing minimal confusion across all four classes. The developed GUI provides a user-friendly interface, enabling seamless dataset loading, feature extraction, model training, and real-time audio fault prediction. The This research demonstrates a substantial performance enhancement compared to traditional machine learning approaches and establishes a strong foundation for deploying scalable acoustic monitoring solutions in industrial settings.