Fault diagnosis of HVAC system sensors: A method based on Box-Cox transformation and multi-model fusion

Sensor failures in HVAC systems can alter control rules, which can negatively impact indoor thermal comfort and lead to energy inefficiencies. Most existing fault diagnosis methods assume Gaussian or linear data distributions. However, operational data from HVAC systems often deviate from a Gaussian distribution, limiting the effectiveness of Gaussian-based models. To address this issue, we propose a Gaussian Naive Bayes-based fusion model that incorporates a Box-Cox transformation to convert non-Gaussian data into approximately Gaussian-distributed data. In addition, ensemble techniques such as Bagging and Boosting are applied to combine multiple Gaussian Naive Bayes models. The impact of the Box-Cox transformation on the performance of models with varying data distribution requirements under concurrent sensor faults is analyzed. The results show that the Box-Cox transformation improves the performance of the Gaussian ensemble model, as well as linear regression (LR) and support vector machine (SVM) models, to varying extents. However, it has minimal effect on decision tree (DT) models, which do not have specific distribution assumptions. The Gaussian ensemble model achieves the highest diagnostic accuracy, with an accuracy of 98.75 % and Hamming loss of only 0.0025. Furthermore, a model degradation analysis reveals that sensor failures significantly degrade the fault diagnosis model, with the greatest degradation observed in cases of complete sensor failure, where the degradation rate reaches 37.65 %. • Gaussian fusion method for concurrent fault diagnosis of sensors. • The Box-Cox transformation converts data into a Gaussian distribution. • Assessing the impact of sensor faults using model degradation rate. • Gaussian distribution enhances the robustness and stability.