Perceptual Linear Prediction for Machine Learning-Based Fault Classification in Electric Motors

This paper proposes a novel acoustic-based fault detection system for electric motors, incorporating Perceptual Linear Prediction (PLP) for feature extraction and an Artificial Neural Network (ANN) for fault classification. By leveraging PLP's capability to mimic human auditory perception, the system enhances robustness under varying noise conditions. Experimental results show perfect accuracy on training data across four fault categories: normal, rotor unbalance, bearing fault, and combination fault. However, cross-validation on unseen data revealed several misclassifications, particularly between rotor unbalance and normal conditions, indicating a need for further refinement to improve generalization. Despite these limitations, the results underscore the importance of cross-validation in evaluating real-world performance and highlight the potential of PLP-based features for non-invasive, real-time motor fault diagnosis. A comparative analysis with Linear Frequency Cepstral Coefficients (LFCC) further demonstrates that PLP achieves faster convergence, greater training stability, and better generalization. These findings suggest that PLP is a more efficient and reliable feature extraction method for motor fault classification, supporting the development of advanced diagnostic tools in industrial environments.