Interpretation of convolutional neural network-based building HVAC fault diagnosis model using improved layer-wise relevance propagation

•Propose feature-level CNN HVAC FD model interpretation method by ImLRP.•CNN model achieves 96.87% accuracy for classifying seven chiller faults.•Use new metric relevance difference to identify CNN chiller FD criteria and misdiagnosis.•Discuss influencing factors: CNN filter size, LRP parameter, relevance score introduction. Convolutional neural networks (CNNs) have been widely utili sed for fault diagnosis (FD) in building heating, ventilation, and air conditioning (HVAC) systems. Despite achieving high accuracy in many HVAC FD tasks, misdiagnosis still occurs. As a black-box model, the CNN FD model and its diagnostic mechanism and decision-making process are opaque, making it difficult for HVAC operators and managers to trust it. To address this, this study proposes an improved Layer-wise Relevance Propagation (ImLRP) method for interpreting CNN FD models in HVACs.The proposed method addresses the issue of preserving positive/negative information from HVAC inputs by adopting a Softsign activation function in the CNN. The feature-matching issue is addressed by excluding pooling layers from the CNN. ImLRP evaluates the contribution of each neuron in the network to the output decision by assigning a relevance score to each neuron in each layer during the backpropagation of the feedforward transmission process. The relevance score difference, a new metric, is used to obtain the net impact of HVAC faults. The proposed method was validated using RP-1043 chiller fault experiment data, which showed a CNN FD accuracy of 96%. Both correct-diagnosis and misdiagnosis were interpreted at the feature variable level, and the study also discussed the influence of the CNN model parameter, ImLRP parameter, and the relevance score difference on the results.