Physically interpretable Stockwell weight initialization and adaptive fusion average threshold for intelligent fault diagnosis of rolling bearing under noisy environment

Deep learning technology has significantly advanced the field of rolling bearing fault diagnosis, delivering impressive improvements in diagnostic accuracy in recent years. These breakthroughs have revolutionized intelligent fault diagnosis, allowing for the extraction of valuable information from large datasets without manual intervention. However, despite the progress, there remains limited research in tailored weight initialization methods and noise-reduction threshold algorithms, especially in noisy environments. To tackle these challenges, we propose an innovative fault diagnosis network termed as Stockwell adaptive fusion average threshold network (SAFATN), which leverages Stockwell weight initialization to capture fault-related features and provide interpretability. This approach integrates prior physical knowledge into the first convolutional layer, making it more suitable for bearing fault diagnosis in noisy conditions. Furthermore, an adaptive fusion average threshold algorithm is introduced, which designs to enhance interactions between spatial and channel dimensions, thereby reducing noise interference. Experimental results from two distinct bearing datasets underscore that SAFATN consistently outperforms other state-of-the-art methods, showcasing superior diagnostic accuracy and robustness in noisy environments.