Fault diagnosis of multimodal feature fusion convolutional neural network based on differential evolution optimization

Abstract

To overcome the performance degradation caused by data scarcity in bearing fault diagnosis -where acquiring sufficient fault samples proves particularly challenging, an effective fault detection approach has been developed. To utilize fault feature information more effectively, we have presented the vibration signal of bearings in a multimodal manner. Two forms of data, time series data in one dimension and gray images in two dimensions, are used as sample data. According to the characteristics of these two kinds of data, two modules are used for targeted feature extraction, and obtain the local frequency, vibration mode characteristics and global dependence of the original vibration signal. A multimodal feature fusion convolutional neural network (AMFCNN) is constructed. Simultaneously, the differential evolution algorithm undergoes optimization to ascertain the optimal combination of key hyper-parameters for the model. Three datasets are used for comparative and variable operating condition experiments to validate this method. The experimental results show that AMFCNN achieves an average accuracy of 91.56 % with only 21.43 % of the data used, which is >13 % improvement over the unimodal approach. AMFCNN is constructed from the two starting points of optimizing the vibration signal data and improving the feature extraction ability of the model, effectively avoiding the problems of over fitting and insufficient feature extraction ability of most fault diagnosis models.