摘要详情

ID / 提交时间

142 / 2024-08-31 16:07:12

标题

GDGC-AE: A New Approach to Mechanical Anomaly Detection Based on Graph Convolutional Networks and Autoencoders

关键字

Keywords—anomaly detection, graph convolutional network (GCN), global dynamic graph convolutional autoencoder (GDGC-AE), feature fusion

主题及专题

Special Sessions > SS05 Signal Decomposition-Integrated Intelligent Mechanical Fault Diagnosis and Remaining Useful Life Prediction

状态

全文录用

作者

ZhangMingzhe / Beijing University of Chemical Technology

SuZhengchang / Beijing University of Chemical Technology；China Nuclear Power Engineering Co.,LTD

HaoPengyuan / Beijing University of Chemical Technology

LinZesheng / Beijing University of Chemical Technology

WangHuaqing / Beijing university of chemical technology

SongLiuyang / Beijing university of chemical technology

摘要

For the mechanical anomaly detection task, the neural network model trained only on normal data has limitations in multi-working condition anomaly detection and multi-channel information correlation mining. In this paper, a global dynamic graph convolutional autoencoder (GDGC-AE) model based on Chebyshev convolution is proposed to cope with the above problems. Firstly, in terms of graph structure data generation, inter-channel graph data is generated based on graph transformer to extract global channel features. Secondly, dynamic graph convolutional autoencoder is proposed, which dynamically adjusts the feature weights of different edges in the graph structure data through the graph attention mechanism, so as to fully integrate the multi-view feature information and accurately reconstruct the representations reflecting the essential features of normal data when the working conditions change. The combination of the two effectively improves the anomaly detection ability and generalisation performance of the model under complex conditions. Finally, the validation results based on the bearing dataset of Politecnico di Torino, Italy, and the broken teeth dataset of TBD234V12 diesel engine in the laboratory show that the proposed model has an accuracy of 96.13% and 93.75% on the two datasets, respectively, with an excellent generalisation, which provides a more robust and accurate solution for the detection of industrial machinery faults.