Abstract: Deep learning, particularly convolutional neural networks (CNN), has garnered significant attention in the field of building energy systems. In the context of fault diagnosis for air handling units (AHU), the effectiveness and applicability of CNN's diagnostic performance require further validation. Additionally, the lack of interpretability in CNN fault diagnosis models hinders their broader application in practical engineering. To address these issues, utilized the publicly available ASHRAE RP-1312 AHU fault data to develop a fault diagnosis model based on CNN, and employed the layer-wise relevance propagation (LRP) method to interpret the CNN model. The results demonstrated that the CNN-based diagnostic model exhibits good applicability, with an average diagnostic accuracy of 99.94%. The LRP method provides strong interpretability for the CNN model, and identifies the diagnosis mechanism of the model in the decision-making process. Finally, an in-depth analysis was conducted on the impact of model parameters such as the number of convolutional layers, learning rate and β parameter on the interpretation results.

Key words: Air handling unit, Convolutional neural network, Fault diagnosis, Layer-wise relevance propagation, Interpretation