Acoustic Fault Diagnosis of Industrial Pumps Using Interpretable Deep Learning and SHAP Analysis

Abstract

Recently, acoustic diagnostics in industrial machine systems have demonstrated advantages over traditional vibration analysis methods, providing more efficient fault detection under industrial noise interference conditions. Here, w present a comparative analysis of eight neural network architectures for acoustic fault diagnosis in industrial pumping systems utilizing the MIMII acoustic dataset. Eight architectures, including SimpleDense, Conv1D, ResidualDense, ResNet1D, LSTM, InceptionTime, Transformer and Temporal Convolutional Network, were evaluated using extracted MFCC coefficients and spectral features. Temporal Convolutional Network showed the best performance with an AUC of 0.997 ± 0.003 and an accuracy of 97.8 % ± 1.8 %, outperforming the other architectures by 1.6-3.7 % in terms of AUC. Residual Dense performed competitively with an accuracy of 97.8 % and AUC of 0.987, while InceptionTime showed the lowest performance with an accuracy of 96.2 %. SHAP-based interpretability analysis revealed different architectural dependencies in feature usage: sequential models show sensitivity to autocorrelation features, while convolutional architectures show sensitivity to spectral centroid and MFCC mean order coefficients. The results obtained are in line with the Industry 4.0 concept, offering new approaches to the implementation of predictive maintenance systems for pumping equipment based on the analysis of acoustic signals.