A fault detection and diagnosis method based on DWVMD-PCA-LightGBM for multi-sensor faults at VAV terminals in office areas of industrial buildings

In variable air volume (VAV) air-conditioning systems, sensor faults can cause control strategy failures, leading to increased energy consumption and reduced operational efficiency. To enable effective detection and classification of sensor faults, this study proposes a novel fault detection approach that integrates signal denoising, feature extraction, and fault classification. The method first employs an improved dynamic weighted variational mode decomposition (DWVMD) to denoise the raw sensor signals, effectively suppressing the interference of measurement noise during model training. Subsequently, principal component analysis (PCA) is used to reduce the dimensionality of the extracted multivariate feature vectors and perform initial fault detection. Finally, based on the PCA-transformed features and their corresponding labeled fault types, a light gradient boosting machine (LightGBM) classifier is trained to accurately identify both single-source fault (SSF) and multi-source fault (MSF). The study is based on real operational data from the VAV system in the office area of an industrial building, covering 7 types of SSFs and 6 types of MSFs. Comparative analysis was performed against five other diagnostic methods. The results indicate that the proposed DWVMD-PCA-LightGBM method demonstrates strong robustness and high accuracy in sensor fault detection, achieving diagnostic accuracies between 93.8% and 100%, with a false alarm rate (FAR) of only 0.07% and a fault detection rate (FDR) of 99.13%. Compared with the five benchmark methods, the proposed approach consistently delivers superior performance under complex operating conditions.