A FAULT DIAGNOSIS METHOD BASED ON INTERACTIVE MULTIHEAD ATTENTION

Aimed at the complex vibration signals of rotating machinery and the difficulty in extracting effective fault features and the problem that the multi head attention fault diagnosis method has high computational complexity, an interactive multi-head attention fault diagnosis method is proposed. By using the method of dividing the input data feature maps into windows, the computational complexity of attention was reduced. The rolled feature maps scheme was used to establish the data interaction between windows and the data interaction between self-attention, which reduced the computational complexity and ensured the classification accuracy. The position information of the input data was integrated into the attention weight matrix to enhance the ability of neural network to distinguish the position of input data. In the experiment part, in order to test the performance of the algorithm, it was applied to the ZHS-2 multi-function motor flexible rotor test bed to verify. Compared with other datadriven fault diagnosis methods, the results show that this method can effectively identify various fault features and realize fault diagnosis.