A virtual signal enhanced non-imputation method for anomaly detection of complex equipment with channel-missing data

View at publisher Summarise with VIRAT
Authors: Zijun Liu, Xinwei Zhang, Jinglong Chen, Shunyan Peng, Zongliang Xie
Cited by: 0
References: 60

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

Abstract The integration and complexity of key equipment, such as the liquid rocket engine (LRE), render anomaly detection (AD) with multi-source data paramount for safe and effective operation. However, in real-world conditions, sensor failures frequently lead to incomplete data, which significantly degrades model performance. Consequently, existing AD methods are often ineffective when applied to incomplete multi-source data. To address this challenge, we propose a virtual signal-based depthwise separate network. The proposed network can perform AD on incomplete data without relying on imputation that requires strong distribution assumptions. Specifically, a channel-independent embedding layer augmented with virtual signals is designed to mitigate the performance degradation caused by incomplete inputs. In particular, the virtual signal is defined as a set of learnable parameters to characterize the normative patterns of sensors. Subsequently, the combined embeddings enable the proposed model to perform intra-channel feature extraction and inter-channel feature fusion, respectively. Finally, a message-passing reconstruction head is introduced, which reconstructs the observed channels utilizing the structure information learned by virtual sensors. Extensive experiments on real LRE datasets have demonstrated that the proposed method outperforms existing methods in both complete and incomplete conditions. Specifically, the proposed method achieves an F 1-score of 0.9975 on complete data and maintains a robust F 1-score of 0.8510 even under a 90% missing rate, outperforming the best baseline model by approximately 10% in severe missing scenarios.

Fields of Study & Subjects

Fault Detection and Control Systems Anomaly Detection Techniques and Applications Machine Fault Diagnosis Techniques

Publication details

Journal / source
Measurement Science and Technology
Volume
37
Issue
12
Pages
126105–126105
Publication date
27 March 2026
Language
EN