Research on the Health State Identification Method of Rotating Machinery Components Based on the Improved AlexNet Framework

Fault diagnosis of bearings is crucial. The traditional network fault diagnosis parameters are set artificially and randomly, resulting in poor diagnostic results. This paper is based on the hyperparameter optimization deep neural network for fault diagnosis, which significantly improves the accuracy of diagnosis. Firstly, the time-frequency image is obtained by a two-dimensional wavelet transform of the original vibration signal of the bearing; secondly, three lightweight networks are tested, and the hyperparameter optimization is performed by the progressive mesh method and the swarm optimization algorithm; lastly, the decision fusion is carried out by using the differences of the networks to obtain the integrated network, which is interpretable by the posterior through the class-activation mapping. The model was tested using a dataset from Western Reserve University and verified to have enhanced diagnostic performance.