An explainable one-dimensional convolutional neural networks based fault diagnosis method for building heating, ventilation and air conditioning systems

Abstract

Due to the frequently changed outdoor weather conditions and indoor requirements, heating, ventilation and air conditioning (HVAC) experiences faulty operations inevitably throughout its lifespan. Therefore, it is important to monitor and diagnose HVAC fault operations. Recently, deep learning methods have attracted more attentions for their guarantee of better diagnosis performance under various system configurations and operating conditions. However, these methods are black-box models which though highly accurate for fault diagnosis but are extremely hard to explain. To overcome the disadvantage of poor interpretability of deep learning black-box models, this study therefore proposes a novel explainable deep learning based fault diagnosis method that is suitable for HVACs. To maintain HVAC operational information and variable locations of all chiller input data samples, proposed method is established with three characteristics: 1) the pooling layer is excluded, 2) the size of convolution filter kernel is set as 1, and 3) use softsign as an activation function. Considering the resulting impacts of HVAC faults on system operating variables, a new Absolute Gradient-weighted Class Activation Mapping (Grad-Absolute-CAM) method is proposed to visualize the fault diagnosis criteria and make the model explainable by providing the fault-discriminative information. The proposed method is validated using fault experimental dataset of a typical building HVAC system (i.e., chiller) from the ASHRAE research project 1043 (RP-1043). The fault diagnosis accuracy is over 98.5% for seven chiller faults. Results indicates that it is capable of interpreting the model work mechanism by activation feature maps and explaining the fault diagnosis criteria by Grad-Absolute-CAM.