A Data-Driven Sound Analysis Approach for Detecting Mechanical and Electrical Anomalies in Motor Systems

Abstract

Industrial electric motors are fundamental to modern manufacturing operations, making accurate and reliable fault diagnosis crucial for minimizing downtime and maintenance costs. Acoustic monitoring has gained attention as a non-intrusive alternative to traditional condition monitoring methods such as vibration and thermal analysis. However, widely used handcrafted features like Mel Frequency Cepstral Coefficients (MFCCs) and spectral centroids often fail to effectively represent the complex temporal and spectral characteristics of industrial audio signals. These conventional features are highly sensitive to environmental noise and variations in machine operating conditions, which limits their ability to generalize across diverse fault types including bearing faults, gearbox defects, and fan imbalances. To overcome these challenges, this work introduces a robust diagnostic framework based on Waveformbased Language Model (WavLM) transformer embeddings. Unlike traditional approaches, WavLM utilizes self-supervised learning to extract high-dimensional, context-aware representations directly from raw audio data, offering improved resilience to noise and enhanced feature richness. These embeddings are then utilized as inputs to multiple machine learning classifiers, including Categorical Boosting (CB) and Decision Tree (DT) models, for fault classification. Experimental evaluation demonstrates that the proposed Histogram-Based Gradient Boosting (HGB) classifier achieves superior performance, attaining an accuracy of 95%, significantly outperforming conventional machine learning approaches. Additionally, a user-friendly graphical user interface (GUI) is developed to support efficient data handling, model training, and real-time prediction. This study demonstrates the effectiveness of transformer-based audio features for scalable and reliable predictive maintenance.