To address the limitations of single deep learning methods in comprehensively characterizing acoustic emission signals for bearing fault diagnosis

particularly the low accuracy and poor generalization capability. This paper proposes an innovative bearing fault acoustic emission signals diagnosis method based on multi-scale convolutional long short-term memory (MSCNN_LSTM) with an improved attention mechanism. The method employs a multi-scale convolutional neural network to extract both global and detailed features from bearing fault acoustic emission signals

enriching the fault related information. These features are then fused and fed into an LSTM network to capture temporal characteristics. An improved attention mechanism is proposed to assign adaptive weights to fault features

eliminating redundant information. A Softmax layer is employed to classify different bearing fault types. In order to verify the diagnosis effect of this method on bearing faults

a bearing fault experimental bench was built

and an experimental study on acoustic emission detection of bearing faults in rotating machinery was carried out

and the average accuracy of the diagnosis of bearing faults under different rotational speeds reached 99.47%. It shows that the proposed method has good generalization ability and robustness under variable working conditions and strong noise background. The study provides an effective method for bearing fault acoustic emission signal diagnosis.