A sudden fault detection network based on Time-sensitive gated recurrent units for bearings. (December 2021)

• Record Type: Journal Article
• Authors: Liu, Shuangjie; Shen, Changqing; Chen, Zaigang; Huang, Weiguo; Zhu, Zhongkui
• Is Part Of: Measurement. Volume 186(2021)
• Publication Date: 2021-12
• Subjects: Condition monitoring -- Bearing -- Sudden fault detection -- Gated Recurrent Unit; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2021.110214
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: A sudden fault detection model is proposed to track equipment health status. The sudden change of the bearing state is simulated by a new dataset slicing method. Gated recurrent units can update the node status and identify the sudden failure. The proposed network quickly responds to sudden failures in case studies. Abstract: Mechanical fault diagnosis is an indispensable part of the modern industrial production process. The application scenario of fault diagnosis according to a neural network from theory to practical application is worth exploring. Most existing work focuses on remaining life prediction and fault classification, treating these two steps independently. This study aims at coherent real-time monitoring of bearing status and proposes a gated recurrent unit -based fault monitoring structure to obtain timely response and preliminary classification of sudden faults. In the existing fault classification research, more attention is paid to the classification accuracy of the fault details. The time sequence law followed by the sudden fault and the short-term prediction of that fault are easily overlooked. The proposed method is trained by key-frames of bearing data. These data frames first pass through the feature extraction layer which consists of two layers of 1D convolution. Then, the reset gates and update gates of developed units keep the valid information at the last moment and update the unit state at that instant. A sudden fault will trigger the detection network, and the detected fault frame will be extracted for further classification by the independently trained fault classification network. During the test, a prediction of the moment of failure occurrence is directly obtained. When the status is judged to be faulty, the fault frame is directly extracted and used as the input of the classification network. Experimental results confirm that the network quickly responds to sudden failures under an operating environment, and the classification accuracy rate can stably reach more than 98%.