A semi-supervised spatiotemporal graph convolutional network with dynamic weighting for cross-condition chiller fault diagnosis

Wang, Guilong, Jin, Tao, Zhang, Xingxing, Rezgui, Yacine ORCID: https://orcid.org/0000-0002-5711-8400 and Li, Yu 2026. A semi-supervised spatiotemporal graph convolutional network with dynamic weighting for cross-condition chiller fault diagnosis. Engineering Applications of Artificial Intelligence 179 , 115188. 10.1016/j.engappai.2026.115188

PDF - Accepted Post-Print Version Available under License Creative Commons Attribution. Download (2MB)

Abstract

Amidst the growing emphasis on building energy system optimization and carbon neutrality goals, efficient fault diagnosis for chillers is of paramount importance. However, the scarcity of labeled data and the variability of operating conditions significantly constrain the generalization capability of existing diagnostic methods. Conventional strategies, such as transfer learning, semi-supervised learning, and data augmentation, have been explored to alleviate these challenges, but further improvement is still needed under complex and varying operating conditions. To address this issue, this paper proposes a novel semi-supervised Dynamic Weight Adaptive Graph Convolutional Network (DWAGCN), in which both labeled and unlabeled samples are jointly embedded into a graph structure, and neighbor aggregation weights are dynamically generated based on node features. By deeply integrating DWAGCN with a Long Short-Term Memory (LSTM) network, an end-to-end spatiotemporal feature learning framework is constructed, enabling the synergistic capture of temporal dynamics and spatial structures in system operational data. Experimental results demonstrate the strong robustness of the proposed method, with a diagnostic accuracy consistently exceeding 93% across multiple fault types and severity levels, outperforming the compared baseline models. Furthermore, under cross-conditions, the proposed semi-supervised transfer learning framework achieves an accuracy ranging from 75.39% to 85.00% with only 10% of the target-domain data, demonstrating strong generalization capability and providing an effective solution for intelligent diagnostics in data scarcity energy systems.

Item Type:	Article
Date Type:	Publication
Status:	Published
Schools:	Schools > Engineering
Additional Information:	RRS applied
Publisher:	Elsevier
ISSN:	0952-1976
Date of First Compliant Deposit:	1 June 2026
Date of Acceptance:	20 May 2026
Last Modified:	01 Jun 2026 09:30
URI:	https://orca.cardiff.ac.uk/id/eprint/187286

Actions (repository staff only)

Edit Item