A Multidomain Spatiotemporal Deep Convolutional Autoencoder for Health Indicator Construction in Rotating Machinery

View at publisher Summarise with VIRAT
Authors: Yuwei Duan, Tao Wei, Man Li, Xiangang Cao
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Rotating machinery is widely used in industrial production. Constructing an accurate health indicator (HI) for it can reveal the degradation process and capture the health status, which is of great significance. However, most existing HI construction methods have problems such as a single form of sample input, poor nonlinear expression ability, difficulty in extracting degenerative features and capturing early signs of degradation. To overcome these issues, this paper proposes a multi-domain omni-dimensional spatiotemporal collaborative deep convolutional autoencoder (MOSCDCAE) for automatically constructing the HI of rotating machinery. The multi-domain two-dimensional weight samples of the vibration signal are introduced to enhance the expression of signal features. A HI construction deep convolutional autoencoder network integrating omni-dimensional dynamic convolution (ODC), bidirectional gated recurrent unit (BiGRU), and attention mechanism (AM) is proposed to generate high-performance HI. This network incorporates newly proposed spatial, temporal, and fusion modules, which can deeply explore the internal deep relationships of the data. On this basis, an exponential function that fits the root mean square (RMS) is proposed, and a new loss function is constructed to enhance the HI’s comprehensive performance. The validity of the method is verified by the standard bearing dataset, and the proposed method improves the comprehensive score by an average of 0.1253 compared with other methods. On the reducer accelerated degradation dataset, this method improves the comprehensive score by an average of 0.1465 compared with other methods. It is also able to detect early degradation points earlier and predict remaining useful life more precisely.

Fields of Study & Subjects

Machine Fault Diagnosis Techniques Reliability and Maintenance Optimization Fault Detection and Control Systems

Publication details

Journal / source
IEEE Sensors Journal
Volume
26
Issue
10
Pages
15629–15642
Publication date
27 March 2026
Language
EN