Fault diagnosis of air handling unit via combining probabilistic slow feature analysis and attention residual network

In the heating, ventilation and air conditioning (HVAC) system, the fault diagnosis of the air handling unit (AHU) is critical to ensure the proper operation of the whole system. The AHU system with complex feature variables is susceptible to noise in operation, which influences the diagnostic performance of the fault diagnosis approach. To further improve the fault diagnosis accuracy, this paper proposes an AHU fault diagnosis model based on probabilistic slow feature analysis (PSFA) and attention residual network (AResNet). Firstly, to suppress the influence of noise on fault diagnosis while dealing with the AHU dynamic temporal characteristics, the PSFA-based feature extraction method is proposed. Further, in order to focus on significant features and suppress unnecessary regional responses of the observed data, the AResNet is utilized to construct the fault diagnosis classifier. Before building the AResNet classifier, we adopt the data spatialization method to convert the process data into spatial grayscale images to achieve spatial placement of key features and enhance the spatial correlation characteristics between multiple features. Finally, detailed experiments and comparisons are made, in which three different intensities of noise are added to the experimental data provided by ASHRAE research project RP-1312. Experimental results show that the proposed PSFA-AResNet model has the best fault diagnosis performance at all three noise levels compared with other deep and shallow popular methods.

Automated summary

This paper proposes an AHU fault diagnosis model based on probabilistic slow feature analysis (PSFA) and attention residual network (AResNet) to improve the accuracy of fault diagnosis. The proposed model is built using the PSFA method and AResNet, and experiments are conducted on the experimental data with different noise levels. The results show that the proposed PSFA-AResNet model outperforms other popular methods in fault diagnosis performance under three different noise levels.