Fault diagnosis of the air handling unit via combining a cross-attention fused spatio-temporal parallel transformer with an enhanced ensemble adaptive classification strategy.

The effective fault diagnosis of the air handling unit (AHU) is critical for maintaining indoor environmental quality and reducing building energy consumption. However, existing fault diagnosis methods often fail to fully capture spatio-temporal features from AHU's operational data and struggle to maintain classification performance when different fault modes present distinct feature distributions. To address these challenges, this paper proposes a spatio-temporal feature extraction with an enhanced ensemble adaptive classification strategy, named STFE-SDTEL. Firstly, a spatio-temporal parallel transformer architecture is designed to simultaneously capture temporal dependencies and spatial features, where a parallel gated convolutional neural network branch is incorporated to enhance spatial feature extraction and suppress redundant information, thereby overcoming the limitations of traditional transformer in spatial feature modeling. In addition, to strengthen the interaction and fusion of temporal and spatial features, a bidirectional cross-attention mechanism is introduced, enabling the model to learn more discriminative spatio-temporal representations. Finally, to improve classification adaptability to heterogeneous feature distributions across different fault modes, an enhanced ensemble adaptive classification strategy is developed, which dynamically combines multiple soft decision tree outputs using adaptive weights. Experimental results on the ASHRAE RP-1312 dataset show that STFE-SDTEL achieves an overall diagnostic accuracy of 99.68%. Comparative results with several mainstream baseline models further demonstrate the superior diagnostic performance of the proposed method.