Multi Scale Denoising Neural Network Based on Position Self Attention and Bidirectional Long Short-Term Memory for Bearing Fault Diagnosis

Due to the multi-scale characteristics of vibration signals and their long-term dependence on temporal characteristics, traditional network models cannot accurately capture the multi-scale temporal features within them. This paper proposes a multi-scale denoising neural network (SAMCNN BiLSTM) based on position self attention and bidirectional long short-term memory for fault diagnosis in complex noisy environments, in response to the multi-scale temporal characteristics exhibited by vibration signals themselves and the random distribution of noise signals. Based on the characteristics of random distribution of noise, a position self attention structure with spatial feature filtering ability is constructed, embedded in a multi-scale convolutional network to extract multi-scale spatial features and adaptively allocate position weights for each feature. At the same time, in order to enhance the model's ability to extract temporal features while preventing redundant or missing temporal signals, a bidirectional long short-term memory network is introduced to extract long-term relevant features of vibration signals. Verified by a laboratory rolling bearing test bench, the experimental results have demonstrated that the proposed model can comprehensively extract bearing fault features while removing noise features. Compared with other models, it has higher diagnostic accuracy and stronger generalization ability in noisy environments.