Explainable CNN–GRU learning on Mel-spectrogram acoustic signals for bearing fault diagnosis under small-sample experimental conditions

Abstract

Reliable bearing fault diagnosis is essential for predictive maintenance of rotating machinery, particularly in applications where contact-based vibration sensors are difficult to install or maintain. This study proposes an explainable acoustic fault diagnosis framework based on Mel-spectrogram representations and a hybrid Convolutional Neural Network–Gated Recurrent Unit (CNN–GRU) architecture. Acoustic emission signals collected from a controlled bearing fault simulator are first segmented and transformed into Mel-spectrograms to represent the time–frequency structure of normal and faulty bearing conditions. The CNN module extracts localized spectral patterns from the Mel-spectrogram images, while the GRU module models temporal dependencies along the spectrogram time axis with fewer recurrent parameters than conventional LSTM-based designs. To address the interpretability limitations of deep learning models, Grad-CAM, Integrated Gradients, and SHAP are employed to analyze the frequency–time regions contributing to the model decisions. In addition, the explanation maps are compared with fault-relevant spectral regions derived from fault characteristic frequency analysis to evaluate whether the model focuses on physically meaningful patterns rather than arbitrary image regions. The proposed framework achieved an high accuracy of in the initial experiment and was further evaluated through repeated validation to assess performance stability under small-sample conditions. The results demonstrate that acoustic sensing combined with explainable CNN–GRU learning can provide a non-contact and interpretable alternative for bearing fault diagnosis. However, the limited dataset size remains an important constraint, and future studies should validate the framework on larger record-level and cross-domain datasets.