Experimental study on performance assessments of HVAC cross-domain fault diagnosis methods oriented to incomplete data problems

Abstract

The cross-domain fault diagnosis (CDFD) method can provide accurate fault diagnosis models for HVAC systems in the case of incomplete labeled data. However, the relationship between the three key factors (i.e., the similarity between the source domain and the target domain, the availability of target domain labeled data, and the type of classifier) and the diagnosis accuracy of CDFD methods are still unclear. Therefore, this study sets 728 diagnostic scenarios to evaluate the influence of the above three key factors on accuracy. The CDFD methods involved include the direct prediction-based CDFD (DPFD) method, the feature transformation transfer learning-based CDFD (FTFD) method, and the pre-training and fine-tuning transfer learning-based CDFD (PFFD) method. The evaluation is based on an experimental HVAC system and its simulation model. There are three main findings. First, the weaker the similarity, the worse the accuracy, yet the sensitivity between similarity and accuracy is gradually weakened for DPFD, FTFD and PFFD methods. Secondly, expanding target domain data availability can improve the performance of FTFD and PFFD methods, the performance improvement is small in the strong similarity scenarios but significant in the weak similarity scenarios. Thirdly, for the DPFD and FTFD methods, the models obtained by different classifiers have significant performance differences. Specifically, shallow machine learning classifier has performance advantages over deep learning classifier, and the support vector machine is the most prominent. The insights obtained can provide practical guidance and ensure that the obtained models provide satisfactory accuracy.