FGDAE : a new machinery anomaly detection method towards complex operating conditions

Yan, Shen and Shao, Haidong and Min, Zhishan and Peng, Jiangji and Cai, Baoping and Liu, Bin (2023) FGDAE : a new machinery anomaly detection method towards complex operating conditions. Reliability Engineering and System Safety, 236. 109319. ISSN 0951-8320 (https://doi.org/10.1016/j.ress.2023.109319)

[thumbnail of Yan-etal-RESS-2023-FGDAE-a-new-machinery-anomaly-detection-method-towards-complex-operating-conditions]
Preview
 Text. Filename: Yan-etal-RESS-2023-FGDAE-a-new-machinery-anomaly-detection-method-towards-complex-operating-conditions.pdf
Accepted Author Manuscript
License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 logo
Download (2MB)| Preview

Abstract

Recent studies on machinery anomaly detection only based on normal data training models have yielded good results in improving operation reliability. However, most of the studies have problems such as limiting the detection task to a single operating condition and inadequate utilization of multi-channel information. To overcome the above deficiencies, this paper proposes a new machinery anomaly detection method called full graph dynamic autoencoder (FGDAE) towards complex operating conditions. First, a full connected graph (FCG) is developed to obtain the global structure information by establishing structural connections between every two channels. Subsequently, a graph adaptive autoencoder (GAAE) model is constructed to aggregate multi-perspective feature information between channels by adapting changes of the operating conditions and to reconstruct the information containing the essential features of normal data. Finally, a dynamic weight optimization (DWO) strategy is designed to guide the model learning the generalization features by flexibly adjusting the data reconstruction loss weights in each condition. The proposed method performs multi-condition anomaly detection under the challenge of training models with multi-condition unbalanced normal data and achieves better performance compared to other popular anomaly detection methods on the machinery datasets.

CORE (COnnecting REpositories)