Induction Motor Fault Diagnosis Based on Multi-Sensor Fusion Under High Noise and Sensor Failure Condition

Abstract

Data-driven methods have gained great success in motor fault diagnosis. Most researches only use signals from a single sensor, which limits the diagnosis accuracy. Multi-sensor fusion methods have been studied in the past few years to enhance model performance. However, in real applications, high noise usually exists in the collected signals and sometimes some sensors may encounter unexpected failure, which will greatly influence the diagnosis accuracy. In this paper, an innovative fault diagnosis model based on multi-sensor fusion is proposed to solve the problems. The proposed model is divided into two parts: parallel physical signal denoising network and memorized credibility evidence theory. The parallel physical signal denoising network is composed of one-dimensional convolutional neural network and residual building block. The memorized credibility evidence theory is proposed based on Dempster-Shafer evidence theory, and the concept of memory credibility is introduced. Experiment on a real induction motor Multi-sensor fault dataset illustrates the superiority of proposed model compared with traditional data fusion algorithm, feature fusion algorithm and proposed model without memory credibility.