A Parameter-Optimized Mobilenet-v2 Using SSA and Its Application in Fault Diagnosis of Bearing

Traditional feature extraction methods have poor autonomous adaptation capabilities and lack universality in the face of large batches of data to obtain network models with optimal extraction capabilities. To improve each network's feature extraction capability and diagnostic accuracy, this paper proposes optimizing the mobilenet-v2 network framework using the Salp Swarm Algorithm (SSA). Firstly, the wavelet time-frequency transform is used to process the vibration signal from the CWRU-bearing data set and the wavelet time-frequency map is used as the input sample. Afterward, the root means square error (MSE) from the network training is used as the fitness function, and the optimal learning rate and a number of batch learning of Mobilenet-v2 are searched for using the bottle sheath swarm algorithm to find the optimal combination of parameters to minimise the error. Finally, combined with the powerful adaptive feature extraction and non-linear mapping capabilities of deep learning, the optimal parameters obtained from the search are input into the network to construct the best diagnostic model and test the data. The 99.32% correct rate was obtained through multiple tests on the sample data. Compared with the grey wolf optimization algorithm and the sparrow optimization algorithm, the iterative convergence converged faster and with higher accuracy.