CNN-based Rolling Bearing Fault Diagnosis Method with Quantifiable Interpretability

Convolutional neural network (CNN) has been widely used in bearing fault diagnosis. However, the diagnostic process of CNN is difficult for users to interpret and understand, and most visualization-based interpretability methods lack interpretable qualitative or quantitative evaluation criteria. Therefore, a novel fault diagnosis framework is proposed to solve the problems. First, this article proposes a processing method using singular value decomposition and multi-scale local binary pattern (HSN-MSLBP) to transform vibration signals into multiscale grayscale images with detailed local features. Then, guided gradient-weighted class activation mapping (Guided Grad-CAM) is applied to the visualization of the diagnostic model and combined with grayscale images to provide highly detailed and intricate pixel-level interpretability analysis. Based on this, an evaluation indicator called average score decrease (ASD) was proposed to quantitatively analyze the interpretability of the visualization methods, which described the degree of score reduction for the diagnostic model after feature removal across the entire grayscale image samples. The experiment was conducted on two datasets, the results indicate that the proposed method can offer superior interpretability of the CNN prediction process while maintaining high diagnostic accuracy.