An adaptive transfer fault detection method for rotary machine with multi-sensor information fusion

Abstract

Multi-sensor information fusion method has good performance in fault detection of rotary machine, in which each sensor information has made different contributions. The contribution of each sensor changes based on the working conditions of the machine, which can lead to a degradation in the performance of the transfer method when used in cross-domain mechanical fault detection. To solve this problem, an adaptive transfer fault detection method for rotary machine with multi-sensor information fusion is proposed. Firstly, multi-sensor data under different working conditions is collected, and features of different sensors are extracted by the corresponding deep learning model. Secondly, the multi-information interaction fusion network is designed to exchange sensor information and obtain fusion features. Then the fusion feature transfer model is proposed for cross-domain fault detection. Finally, the model is trained with the bearing dataset of the University of Paderborn. The results show that the transfer fault detection method with multi-sensor information fusion achieves state-of-the-art performances in cross-domain fault detection. It can adjust adaptively the contribution of each sensor information in the cross-domain fault detection.