Vibration Signal Used Motor Broken Rotor Bar Identification Based on Attention-CNN

Aiming at the problems that the vibration signal-based broken rotor bar diagnosis technique relies on manual feature selection with poor generalization ability and conventional convolution neural network(CNN) models that ignore sequence information in automatic feature extraction of time-series signals, an Attention-CNN network model is proposed using the Attention mechanism for the metric ability of local features in the overall expression. Firstly, attention is assigned on the original signal by Attention, secondly, the network is constructed by combining CNN for feature extraction, then the particle swarm optimization(PSO) algorithm is used to perform network hyperparameter search and train the broken rotor bar recognition model, and finally, the model is evaluated from both overall and local aspects. The experimental results show that the proposed recognition model can reach the traditional diagnosis level and has a higher generalization capability than existing methods, and is more suitable for motor broken rotor bar recognition through vibration signals.