Fault diagnosis of air handling units based on an MCNN-Transformer ensemble learning

• Published in: Journal of process control Vol. 154; p. 103526
• Main Authors: Xia, Yin, Zhang, Danhong, Liu, Chenyu, Cao, Zhiqiang, Su, Yixin, Chen, Yuhang
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.10.2025
• Subjects: AHU; Dempster–Shafer; Fault diagnosis; HVAC; MCNN-transformer; Fault diagnosis; HVAC; Dempster–Shafer; MCNN-transformer; AHU
• ISSN: 0959-1524
• DOI: 10.1016/j.jprocont.2025.103526

The Air Handling Units (AHU) in Heating Ventilation and Air Conditioning (HVAC) systems regulates air temperature and humidity to ensure indoor air quality and thermal comfort. Fault diagnosis of AHU is critical for reducing energy consumption and maintaining system performance. However, data noise and missing values introduce considerable uncertainty into AHU fault diagnosis, while most existing methods do not utilize time-series models and thus neglect the extraction of temporal features and the modeling of long-range dependencies. This limitation hinders the effective capture of fault evolution and long-term correlations, making it difficult to meet dynamic real-time requirements under complex operating conditions. To address these challenges, this paper proposes an ensemble learning framework that integrates Dempster–Shafer (DS) theory with a Multi-Channel Convolutional Neural Network and Transformer (MCNN-Transformer) model, aiming to enhance generalization and improve diagnostic performance. The DS theory combines the strengths of Random Forest, Pearson Correlation, and Mutual Information, effectively mitigating uncertainty and noise in fault feature data by fusing multi-source information. The MCNN-Transformer integrates multi-scale convolutional layers with a self-attention mechanism, enabling effective extraction of features across multiple temporal scales and modeling of long-range dependencies. Experimental results show that the proposed MCNN-Transformer framework achieves high efficiency and strong generalization capability, reaching a fault diagnosis accuracy of 99.2%, a precision of 0.992, a recall of 0.992, and an F1 score of 0.991, significantly outperforming traditional models. Moreover, the improved stability of the model’s accuracy curve further demonstrates its robustness. •DS fusion reduces sensor noise impact, lowers uncertainty, and improves feature reliability.•MCNN-Transformer extracts local dynamics and long-range temporal patterns for AHU data.•Integrated DS with MCNN–Transformer ensemble ensures robust multi-label fault diagnosis.•ASHRAE RP-1312 results show 99.2% accuracy, 0.992 precision, and 0.991 F1 outperforming baselines.