A novel data-temporal attention network based strategy for fault diagnosis of chiller sensors

Abstract

In the air-cooled chiller system, sensor fault diagnosis has great significance for ensuring normal operation. However, according to the operating mechanism of the chiller system, sensors’ readings exhibit dynamical data-temporal dependencies and are easily affected by external factors and control parameters. To further capture the data characteristics existed in the sensors time series, in this paper, we propose a novel data-temporal attention network (DAN) for the chiller sensor fault diagnosis. Based on the conventional encoder-decoder network (EDN), the proposed DAN model is firstly built by adding three novel parts: the first one is a data attention mechanism embedded in the encoder, which is used to capture the dynamic data correlation between different sensors; the second part is a temporal attention mechanism, which is employed in the decoder to model the dynamic time-dependencies among the sensors time series; considering the influence of external factors and control parameters, the third part is a fusion module to incorporate these influential factors from different domains. Thereafter, we design a specific chiller sensor fault diagnosis strategy using the proposed DAN model. The sensor fault diagnosis strategy uses only normal sequences for training and learns to reconstruct normal time series behaviors, and then determines the fault threshold of each chiller sensor, and finally identifies the specific sensor fault by comparing the absolute reconstruction error vector with the fault threshold vector. In the end, the experiments which adopt data sets from a real air-cooled chiller platform are conducted, and detailed comparisons are made. Various magnitudes of fixed biases are introduced into eleven sensors for validation. Experimental results reveal that the sensor fault diagnosis strategy with the proposed DAN model achieves the best training and fault diagnosis performance compared with its variants and the traditional EDN model. Especially for the sensor fault diagnosis performance, comparison results demonstrate that the proposed DAN model is more sensitive to the small biases than the other contrast models and has better robustness for impacts on fault sensor in the reconstruction of the fault-free sensors.