A hybrid sensor fault detection and diagnosis method for air-handling unit based on multivariate analysis merged with deep learning

Abstract

This paper proposes a novel air-handling unit (AHU) sensor fault detection and diagnosis (FDD) method by utilizing multivariate analysis merged with deep learning. In reality, sensor measurements in AHU systems are affected by outdoor air temperature, which results in poor detection performance of existing data-driven methods. To overcome this difficulty, a robust canonical correlation analysis (RCCA) is firstly proposed by removing the effect of outdoor air temperature, which is realized by performing an orthogonal decomposition of process variables. The better detection performance is delivered by using data from an orthogonal subspace of outdoor air temperature. Then, with the aid of the proposed detection method, a RCCA-based fault bank is constructed based on the principle of parity space. A neural network-based diagnosis method is proposed by means of the RCCA-based fault bank, which reduces the influence of noises and thus faults are easily diagnosed compared with traditional neural network-based methods. The proposed method is purely data-driven, and thus it is easily used for FDD in real systems. Finally, the effectiveness of the hybrid method is verified using experimental data from ASHRAE RP-1312. Results show that the proposed method is superior to the state-of-the-art methods, and the diagnosis performance is significantly improved by using the deep learning method with the aid of multivariate analysis.