An improved attention mechanism rolling element bearing fault diagnosis based on convolutional neural network and bi-directional long short-term memory

In this work, an improved attention mechanism rolling element bearing (REB) fault diagnosis method based on a convolutional neural network (CNN) and a Bi-directional long-short term memory (BiLSTM) is proposed. The original REB fault signals of 10 different fault types are selected from the bearing standard database of Case Western Reserve University (CWRU). The variational mode decomposition optimized by crested porcupine optimizer algorithm (CPO-VMD) is used for pre-processing REB fault signals into multiple intrinsic mode functions (IMF). The Squeeze-and-Excitation (SE) attention is inserted into the convolutional module to adjust the weight relationship between channels, thereby extracting more feature information. After the BiLSTM network structure, self-attention is introduced to focus on important fault characteristics. The experimental results show that the training time of the CNN-SENet-BiLSTM-Self-Attention REB fault diagnosis model proposed in this paper is 39 s, with an accuracy of 99.67%. Compared with the traditional CNN-LSTM, CNN-BiLSTM, CNN-SENet-BiLSTM, CNN-BiLSTM-Self-Attention, and CNN-SENet-LSTM-Self-Attention models, the accuracy rates increase by 9%, 8%, 1.7%, 0.7%, and 0.3%, respectively.