Roller Bearing Fault Diagnosis Using Deep Transfer Learning and Adaptive Weighting

Abstract

A fault diagnosis approach for roller bearings utilizing deep transfer learning and adaptive weighting is suggested to address the issue that extra fault state samples in the target domain data of roller bearings impair the fault diagnostic accuracy. CNN-LSTM is a network model proposed by Lecun et al., which has good performance in image processing and image processing. It can effectively apply predictive local perception of time series and weight sharing of CNN, which can greatly reduce the number of networks and improve the efficiency of model learning. The method first establishes a feature extraction module, and uses a deep convolutional neural network to map bearing samples to a high-dimensional feature space. Secondly, uses the transfer learning concept to design a weighted domain discriminator, and adaptively weights the samples; and finally, through the confrontation in the feature space, the bearing samples are classified. Training to increase the domain similarity of the healthy state samples shared by the target domain and the source domain. Then measuring the similarity between these samples based on the sample weight size, and setting a threshold to label the additional fault state samples of the target domains as unknown. The suggested technique is validated using gearbox bearing data, roller bearing data from Case Western Reserve University, and locomotive wheel bearing data. The diagnostic accuracy of the samples is less than 80%, suggesting that the suggested approach can successfully overcome the effects of extra fault state samples and diagnose roller bearing faults.