Sensor-based bearing fault diagnosis using a hybrid CNN–LSTM framework with explainable AI

This study addresses the problem of non-contact bearing fault diagnosis using time-series sensor data. While vibration-based monitoring is widely used, its dependency on physical contact limits practical deployment in certain industrial environments. To overcome this limitation, this work proposes a hybrid CNN – LSTM framework utilizing Mel-spectrogram representations combined with explainable AI techniques (Grad-CAM, Integrated Gradients, and SHAP). The dataset consists of 48 segmented samples derived from four operating conditions. A 5-fold cross-validation strategy is employed. Experimental results demonstrate an accuracy of 97.92%. Ablation studies confirm that both the Mel-spectrogram representation and hybrid architecture contribute significantly to performance. The proposed framework provides not only high accuracy but also interpretable diagnostic insights, making it suitable for predictive maintenance applications. However, due to dataset limitations and sensing modality ambiguity, further validation on real-world datasets is required.