Vibration spectrogram analysis for bearing fault diagnosis based on grad-cam for feature selection and statistical approach

This study aims to enhance bearing fault diagnosis by integrating eXplainable Artificial Intelligence technology, emphasizing improved accuracy and interpretability in vibration data analysis. In the initial phase, vibration data undergo transformation into spectrograms through short time Fourier transform analysis, followed by the application of the Grad-CAM algorithm to extract interpretable features, thus enhancing diagnostic precision. Subsequently, the second phase capitalizes on data that have undergone inverse transformation, highlighting features through reduced variance for outlier detection and bearing fault differentiation. The validity of this method is corroborated by analysis of datasets from the University of Ottawa and CWRU, affirming its versatility across diverse fault scenarios. The outcomes validate the methodology’s capability in accurately differentiating between healthy and fault conditons, linking outlier frequencies to fault severity, and enhancing model interpretability through Grad-CAM visualizations. Despite certain constraints, this methodology signifies a progressive stride in precise and interpretable bearing fault diagnosis, heralding advancements in its application to industrial contexts.