A Research On Fault Diagnosis Of Chiller Sensors Based On The Data-Temporal Attention Mechanism

In the air conditioning systems,chiller senor fault diagnosis has great significance for ensuring normal operation and energy-saving.According to the operation mechanism of the chiller system,this paper analyses and summarizes the complicated data characteristics existed in the sensors time series.It can be found that sensors time series exhibit dynamic data-temporal dependencies and are easily affected by external factors and control parameters.To further capture these complex data characteristics and to establish an efficient sensor fault diagnosis strategy,in this paper,we propose a datetemporal attention network based strategy for fault diagnosis of chiller sensors.The main contributions of this study are manifested as follows:(i)In order to fully capture the complex data characteristics existed in the chiller sensors time series,this study proposes a novel data-temporal attention network(DAN).The proposed DAN method adopts the traditional EDN model as the basic framework.Inspired by the attention mechanism,we develop the data attention mechanism and temporal attention mechanism to be separately embedded in the traditional encoder and decoder,which aims to characterize the dynamic data-temporal correlations between the chiller sensor time series.Considering the influences of external factors and control parameters,this paper incorporate these influential factors to form a fusion module in the decoder.(ii)To overcome the negative impacts on the sensor fault diagnosis performance caused by the PCA model,this study designs an effective DAN-based chiller sensor fault diagnosis strategy.Without any data dimensionality reduction operation,this proposed DAN-based strategy has better capability to learn the complex data characteristics in the large amount of chiller normal sensors time series.Therefore,the proposed DANbased strategy is more sensitive to the abnormal reaction reflected by the faulty sensor.At the same time,by taking the advantage of “end to end”structure in the proposed DAN model,this chiller sensor fault diagnosis strategy construct the reconstruction error vector to directly separate the specific faulty sensor from the fault-free sensors,which achieves the purpose of efficient sensor fault diagnosis.(iii)In the end,the experiments which adopt data sets from a real compression chiller platform are conducted,and detailed validations and comparisons are made.Meanwhile,the four typical sensor faults are introduced for validation,i.e.fixed biases,drifting fault,precision degradation and complete failure.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.Meanwhile,compared with the PCA-based chiller sensor fault diagnosis strategy,the proposed DAN model has the capacity of achieving the better sensor fault diagnosis performance,which verifies the stronger potential of the DAN model.Unlike the PCA model,the developed DAN model learns the data-temporal related information among a mounts of normal sensors time series via its novel structure components.In addition,the DAN-based chiller sensor fault diagnosis strategy utilizes the ”end-to-end” basic framework to directly identify the faulty sensor,which greatly enhances the chiller sensor fault diagnosis ratios.