An interpretable graph convolutional neural network based fault diagnosis method for building energy systems

Due to the fast-modeling speed and high accuracy, deep learning has attracted great interest in the field of fault diagnosis in building energy systems in recent years. However, the black-box nature makes deep learning models generally difficult to interpret. In order to compensate for the poor interpretability of deep learning models, this study proposed a fault diagnosis method based on interpretable graph neural network (GNN) suitable for building energy systems. The method is developed by following three main steps: (1) selecting NC-GNN as a fault diagnosis model for building energy systems and proposing a graph generation method applicable to the model, (2) developing an interpretation method based on InputXGradient for the NC-GNN, which is capable of outputting the importance of the node features and automatically locating the fault related features, (3) visualizing the results of model interpretation and validating by matching with expert knowledge and maintenance experience. Validation was performed using the public ASHRAE RP-1043 chiller fault data. The diagnosis results show that the proposed method has a diagnosis accuracy of over 96%. The interpretation results show that the method is capable of explaining the decision-making process of the model by identifying fault-discriminative features. For almost all seven faults, their fault-discriminative features were correctly identified.