Transformer-Driven WavLM Audio Modelling for Robust Predictive Maintenance in Electric Motor Systems

Abstract

Industrial electric motors are the backbone of modern manufacturing, necessitating high-reliability maintenance strategies to prevent costly downtime. While acoustic monitoring offers a non-intrusive alternative to traditional vibration or temperature analysis, conventional handcrafted features such as Mel Frequency Cepstral Coefficients (MFCCs) and Spectral Centroids often struggle to capture the complex temporal-spectral dependencies of real-world industrial settings. These "shallow" features are frequently compromised by environmental noise and fluctuations in machine load, leading to poor generalization across diverse fault types like gearbox anomalies or fan imbalances. To overcome these limitations, this study proposes a novel diagnostic framework leveraging Waveform-based Language Model (WavLM) transformer embeddings. Unlike classical methods, WavLM extracts rich, contextaware representations from raw audio, providing superior noise tolerance and structural depth. These high-dimensional embeddings were evaluated across multiple classifiers, including Categorical Boosting (CatBoost) and Decision Trees. Experimental results demonstrate that the proposed Histogram-Based Gradient Boosting (HGB) classifier achieves a state-of-the-art accuracy of 95%, significantly outperforming traditional machine learning pipelines. To facilitate industrial adoption, a graphical user interface (GUI) was developed to streamline data processing and real-time fault prediction. The research offers a scalable, high-precision solution for predictive maintenance, proving that self-supervised audio representations can dramatically enhance diagnostic robustness in noisy industrial ecosystems.