An Improved Graph Deviation Network for Chiller Fault Diagnosis by Integrating the Sparse Cointegration Analysis and the Convolutional Block Attention Mechanism

Abstract

Chiller fault diagnosis plays a crucial role in optimizing energy efficiency within heating, ventilation, and air conditioning (HVAC) systems. The non-stationary nature of chiller fault data presents a significant challenge, as conventional methodologies often fail to adequately capture the relationships between non-stationary variables. To address this limitation and enhance diagnostic accuracy, this paper proposes an improved graph deviation network for chiller fault diagnosis by integrating the sparse cointegration analysis and the convolutional block attention mechanism. First, in order to obtain sparse fault features in non-stationary operation, this paper adopts the sparse cointegration analysis method (SCA). Further, to augment the diagnosis accuracy, this paper proposes the improved graph deviation network (IGDN) to classify fault datasets, which is a combination of the output of a graph deviation network (GDN) with a convolutional block attention mechanism (CBAM). This novel architecture enables sequential evaluation of attention maps along independent temporal and spatial dimensions, followed by element-wise multiplication with input features for adaptive feature optimization. Finally, detailed experiments and comparisons are performed. Comparative analyses reveal that SCA outperforms alternative feature extraction algorithms in addressing the non-stationary characteristics of chiller systems. Furthermore, the IGDN exhibits superior fault diagnosis accuracy across various fault severity levels.