Research On Application Of Fault Diagnosis Of Water Chiller Based On Deep Learning-Evidence Theory

Building energy consumption in China’s total energy consumption accounted for a large proportion of the rapid growth,in the building energy consumption,HVAC system energy consumption accounted for about 40% to 50%.If air conditioning equipment operates with faults,it will generate a lot of additional energy consumption and shorten the service life of the equipment,so timely and effective diagnosis and troubleshooting is an effective means to ensure operation and energy saving.Thanks to the development of artificial intelligence technology,the fault diagnosis technology of HVAC chiller plant based on deep learning method is becoming a research hotspot,among which convolutional neural network is gradually becoming the main means of chiller plant fault diagnosis because of its good ability to deal with nonlinear and complex problems.In view of the difficulties in obtaining chiller failure data and the small amount of data samples,a one-dimensional convolutional neural network,which is more applicable to the situation,is studied as the research object,combined with the study of evidence theory,aiming to improve the performance index of chiller failure diagnosis.The specific work is as follows.(1)For the characteristics of nonlinear and non-Gaussian chiller operation data,a D1DCNN-BN diagnostic model containing two convolutional layers,two pooling layers,BN layers,Flatten layers and Dense layers is constructed based on the analysis and study of the respective characteristics of 1DCNN and 2DCNN,and a back propagation network(BPNN),long and short memory network(LSTM),and a one-dimensional convolutional model(1DCNN)as the control group.The simulation experiments used ASHRAR RP-1043 chiller fault data,and the experimental results showed that the diagnostic accuracy of the model was 98.70%,8.6% higher than LSTM and 2.77% higher than 1DCNN on 10,000 data samples of tiny fault levels,and the accuracy on 7 typical faults was higher than that of the control group,which proved the advantages of the model.(2)In view of the variable operating conditions of chillers and the high coupling of parameters between normal and system fault states,the study introduces the evidence theory,adds the evidence neural network layer constructed based on direct sequence theory to the previous model,and optimizes its relevant parameter settings,and proposes the 1DCNN-DS diagnosis model based on one-dimensional convolutional neural network and evidence theory.Under the same simulation experimental setup conditions,the model achieves 99.30% diagnostic accuracy on minor fault levels.Its system fault RL leakage rate and false alarm rate are reduced by 5.65% and 1.66%compared with D1DCNN-BN.(3)To verify the actual diagnosis effect of the method in this paper,a small air-cooled heat pump air conditioner fault diagnosis test bench is built,and the operation data acquisition,display and saving as well as the diagnosis and display of simulated faults are realized by the hybrid programming technology of C# and Python.Under the simulated minor fault level,the diagnosis accuracy of 1DCNN-DS model reaches 99.80% for 2000 data samples,which is significantly better than the rest of the control group.