Physics-guided deep learning for acoustic-based fault diagnosis

Abstract

Deep learning has been applied to classify car engine breakdowns based on acoustic features, but a major challenge remains: breakdown sounds are often very weak compared to engine and environmental noise, making accurate classification difficult. We handle this challenge in both feature extraction and model training phases. Firstly, we employed log Mel-spectrogram and Mel-Frequency Cepstral Coefficients (MFCC) for feature extraction process, effectively isolating breakdown-relevant patterns. These features were then used to train a bi-directional Gated Recurrent Unit (GRU) model with a novel enhanced hierarchical attention for Multi-Scale Feature aggregation, allowing the model to focus on informative audio segments while mitigating noise. This ensures better recognition of rare fault conditions. Also, this study integrates Physics-Informed Deep Learning (PIDL) with Finite Element Analysis (FEA) to generate synthetic failure sounds and enforce physics-based constraints, improving model robustness and interpretability for acoustic-based engine fault diagnosis. Experimental results demonstrated high performance, achieving classification accuracy, sensitivity, and specificity of 99.3%, 98.0%, and 99.1%, respectively. These findings confirm the effectiveness of our approach in detecting and classifying engine breakdowns, even in challenging acoustic environments.