Research On Fault Detection And Diagnosis Of Variable Air Volume Air-Conditioning System Based On MCNN-BiLSTM Hybrid Neural Network

Under the background of energy green transformation strategy,building energy saving and consumption reduction has become the top priority at present.Heating,ventilation,and air conditioning(HVAC)systems are the main source of energy consumption in buildings,and fault of HVAC systems can not only affect the thermal comfort of indoor occupants,but also reduce the lifespan of equipment and increase energy consumption.Therefore,applying fault detection and diagnosis(FDD)technology to air conditioning systems can detect and find faults accurately and timely,reducing the "illness" of the system and promoting stable and operation efficient while reducing equipment operation and maintenance costs.Based on the ASHRAE RP-1312 datasets,fault detection and diagnosis research are focused on variable air volume air conditioning systems.To address the insufficient feature extraction of traditional FDD methods for fault data,a hybrid neural network model based on multi-scale convolution and bidirectional long short-term memory(MCNN-BiLSTM)is proposed.The model uses different sizes of convolution kernels in the MCNN neural network to extract feature information at different frequencies of input samples,followed by improved a LSTM neural network for time series feature learning,and finally the Softmax function for fault classification.Compared with three other FDD models as CNN-LSTM,CNN,and SSA-BP,the experimental results show that the MCNN-BiLSTM fault diagnosis model increases accuracy by 5%,7%,and 9%,respectively under summer working conditions.In order to solve the problem of insufficient fault data caused by the difficulty of introducing faults in an actual air conditioning system,a fault diagnosis method combining transfer learning and a convolutional neural network was proposed.Taking fault samples with sufficient data as the source domain and samples with insufficient data as the target domain,the source domain data and target domain training data were respectively input into the FDD model for feature extraction.On the basis of parameter sharing,the model parameters were fine-adjusted by the maximum mean difference method to complete the establishment of the migration learning model.The experimental results show that by comparing the performance of the FDD model before and after migration,the accuracy of a single fault is improved,and the accuracy of concurrent fault also achieved a good diagnostic effect.Based on the analysis of fault characteristic information,a fault pattern recognition database is established.Taking the variable air volume air conditioning system of a university in Tianjin as the platform,five kinds of single faults and five kinds of concurrent faults were simulated.According to expert knowledge,the fault identification system is established through qualitative and quantitative analysis of fault characteristics.When a fault occurs in the air conditioning system,the measurement parameter changes of the fault data are compared with the fault pattern identification database,and the preliminary identification of the fault type is completed according to the comparison results.