Data-Driven Fault Detection And Diagnosis For Chillers

As a core component of HVAC system,chillers lay the foundation for the efficient and stable operation of the entire system.However,during the actual operation of the chillers,due to factors such as improper component selection,maintenance personnel errors and low welding quality,as well as its complex structure and poor operating environment,many different faults may occur in the chillers.When the chillers are in faulty operation for a long time,it will not only adversely affect the efficacy of the HVAC system and reduce indoor comfort,but also shorten the life of the equipment and cause energy waste.Therefore,the research on chillers fault diagnosis has important practical significance.With the development of the data collection technologies,researchers gradually tend to adopt the intelligent data-based fault diagnosis methods for the chillers.Such methods can not only improve the diagnostic efficiency,but also reduce the human investment.This thesis takes the chiller as the research object,according to the data-based method,and analyzes and studies from three aspects: the feature extraction,the data imbalance and the fault diagnosis modeling.The innovations and contributions of this thesis are as follows:（1）Aiming at the dynamic coupling characteristics of the chiller data,a feature enhancement technology is constructed by using the encoder-decoder network（EDN）and statistical pooling method to extract more informative feature data.In the feature enhancement technology,to obtain the information reflecting the independent changes of each variable,the long short-term memory network（LSTM）based EDN is used to calculate the residual feature.Meanwhile,to extract the information revealing the dynamic coupling characteristics among different variables,the original data are divided into the data blocks of different scales,and the statistical pooling operation is performed on each data block to obtain the statistical feature.Experimental results show that the feature enhancement technology can analyze the chiller data in multiple aspects,and the extracted feature data are helpful for the fault diagnosis model to distinguish different fault patterns.（2）In order to solve the data imbalance problem of the chillers,the fault data are effectively augmented by the stable synthetic minority oversampling technique（SSMOTE）.The SSMOTE method adopts the k-nearest neighbor method to obtain the nearest neighbor samples of each sample.Then,the samples with the same category as the most neighbor samples are selected by the setting threshold.Through the above approach,we can divide the stable feature space of each data category.Finally,the linear interpolation method is employed in each stable feature space to synthesize new fault samples.Experimental results show that the SSMOTE method ensures that new fault samples with lower overlap are generated in the feature space far from the distribution boundary.It is beneficial to improve the training effect of the fault diagnosis model.（3）To further improve the performance of the chiller fault diagnosis model,an STCN model is constructed through the skip connection based self-attention（SSA）mechanism and the temporal convolutional network（TCN）.With the advantages of its structure such as the dilated convolution and residual connection,the STCN model not only achieves efficient fault diagnosis,but also reduces the number of the model parameters.By encapsulating the SSA mechanism layer into each residual block,the STCN model is enabled to dynamically focus on different subsets of the input in different situations.Experimental results show that the STCN method not only enables the model to achieve efficient utilization of the information resources,but also improves the fault diagnosis performance of the model.