Bearing Fault Diagnosis Method Based on Multi-sensor Feature Fusion Convolutional Neural Network

Abstract

AbstractMost existing deep learning models use a single sensor signal as input data, which makes them susceptible to external variables and cannot represent the operating state of a certain component. To achieve intelligent fault diagnosis of the rolling bearing in permanent magnet synchronous motors (PMSM) under cross-working conditions, a novel method based on a multi-sensor feature fusion convolutional neural network (MFFCN) is presented. The proposed model consists of a core network and two sub-networks to achieve information sharing and exchange. The deep features in multi-sensor signals are extracted by using dilated convolution blocks in the subnetwork. An attention mechanism is introduced in the core network to flexibly extract and fuse more effective features. The proposed model is verified by a fusion of vibration and current signal for bearing fault diagnosis of the PMSM. Experimental results show that the proposed model has higher diagnostic accuracy in fault diagnosis under various working conditions. In addition, the interpretability of the network model is improved through network visualization.KeywordsBearing fault diagnosisFeature fusionAttention mechanism