Decision Fusion Method for Bearing Faults Classification Based on Wavelet Denoising and Dempster-Shafer Theory

Bearing is one of the fundamental tools in rotating machinery in which unexpected shutdown may occur by any fault. This paper addresses bearing fault detection and classification in induction motors. In this paper, vibration together with current sensor is utilized as a multi-sensor framework for high-accuracy bearing fault detection. Wavelet packet transform is used for each sensor as an effective preprocessing method for signal denoising. A universal threshold based on the Stein's Unbiased Risk Estimate method is established in the denoising algorithm. After extracting appropriate time-domain features from the denoised signal, a support vector machine is trained for the fault classification. Output vector of this classifier is represented as one of the four following categories: healthy mode, inner-race fault, outer-race fault, and double holes on outer race. A decision-level fusion based on Dempster-Shafer theory is used to fuse the decision of each classifier. One of the most prominent aspects of the proposed method is developing time-domain features analysis that requires less computational effort compared to the conventional fault detection methods that use frequency-domain features analysis. Moreover, motor parameters such as the number of slots and rotation speed are not required during implementation of the algorithm. Finally, experimental results assure the applicability of the proposed method.