Effects of various information scenarios on layer-wise relevance propagation-based interpretable convolutional neural networks for air handling unit fault diagnosis

Abstract

Deep learning (DL), especially convolutional neural networks (CNNs), has been widely applied in air handling unit (AHU) fault diagnosis (FD). However, its application faces two major challenges. Firstly, the accessibility of operational state variables for AHU systems is limited in practical, and the effectiveness and applicability of existing DL methods for diagnosis require further validation. Secondly, the interpretability performance of DL models under various information scenarios needs further exploration. To address these challenges, this study utilized publicly available ASHRAE RP-1312 AHU fault data and employed CNNs to construct three FD models under three various information scenarios. Furthermore, the layer-wise relevance propagation (LRP) method was used to interpret and explain the effects of these three various information scenarios on the CNN models. An R-threshold was proposed to systematically differentiate diagnostic criteria, which further elucidates the intrinsic reasons behind correct and incorrect decisions made by the models. The results showed that the CNN-based diagnostic models demonstrated good applicability under the three various information scenarios, with an average diagnostic accuracy of 98.55%. The LRP method provided good interpretation and explanation for understanding the decision mechanism of CNN models for the unlimited information scenarios. For the very limited information scenario, since the variables are restricted, although LRP can reveal key variables in the model’s decision-making process, these key variables have certain limitations in terms of data and physical explanations for further improving the model’s interpretation. Finally, an in-depth analysis of model parameters—such as the number of convolutional layers, learning rate, β parameters, and training set size—was conducted to examine their impact on the interpretative results. This study contributes to clarifying the effects of various information scenarios on the diagnostic performance and interpretability of LRP-based CNN models for AHU FD, which helps provide improved reliability of DL models in practical applications.