IMDRSN-BiLSTM for Rolling Bearing Fault Diagnosis

To tackle the problem of decreased accuracy of deep residual shrinkage networks (DRSN) in the presence of strong noise, this paper proposes an improved multi-scale deep residual shrinkage network (IMDRSN) combined with bidirectional long short-term memory (BiLSTM) for rolling bearing fault diagnosis. Firstly, raw fault data is transformed into time-frequency images, and the Xception module captures multi-scale information in the images. Secondly, multiple scales of residual shrinkage building units (RSBU) are used to denoise the captured image information. Thirdly, introduce a Xception module into each RSBU to enhance the model’s information retrieval capabilities. Incorporate a convolutional block attention module (CBAM) into each RSBU to strengthen the model’s focus on key features, and introduce an adaptive module to reduce the constant bias impact of soft thresholding between input and output. Finally, the BiLSTM module is employed to capture the dependencies within the time series data, and to perform the task of fault classification. The IMDRSN-BiLSTM model is applied to the rolling bearing fault diagnosis task on the case western reserve university (CWRU) dataset in noisy environments, and experimental outcomes demonstrate that the IMDRSN-BiLSTM model delivers higher precision and robustness in identifying bearing malfunctions.