Research On Neural Network-based FDD Method For AHU System

Heating,Ventilation and Air Conditioning(HVAC)is a control system that includes temperature,humidity,air cleanliness,and air circulation.As one of the most important devices in HVAC system,air handling unit(AHU)consists of multiple subsystems.Each hardware failure or controller error associated with these subsystems might result in the abnormality of AHU and thus affects the whole performance of HVAC system.In recent years,deep learning has developed rapidly,with artificial neural network(ANN)as its foundation.The back propagation neural network(BPNN)is used in the field of AHU fault detection and diagnosis due to its good classification and prediction abilities.However,more advanced artificial neural networks,such as convolutional neural networks(CNN),have not been fully studied and introduced in this field.Because of its superior feature extraction ability,CNN can learn and capture the abnormalities in the variables.Moreover,CNN possesses good robustness in dealing with the noise in the measurements of variables.Therefore,research on AHU fault detection and diagnosis using CNN is valuable.This thesis proposes a fault detection and diagnosis method for AHU system based on one-dimensional convolutional neural network(1-D CNN)and WaveCluster clustering analysis.Based on the operation of AHU system and the correlation between different devices,this thesis selects a key sub-system from AHU system,i.e.supply air temperature control loop,to implement and verify the proposed method.Based on in-depth study of the multivariate correlation caused by PID control in the supply air temperature control loop,four important variables are selected as the inspected variables.The feature extraction ability of 1-D CNN is used to capture the anomaly in a single inspected variable,and the anomaly detection result of each inspected variable are concatenated together to obtain a comprehensive anomaly detection result,which will be futher analyzed by WaveCluster clustering analysis.Then,a linear space is established,and the denoting clusters generated after the step of WaveCluster clustering are analyzed in this linear space to find the cause of failure.Finally,the T_c acquittal is applied to reduce false alarm.In this thesis,a 1-D CNN model is constructed based on Keras,a deep learning framework,and it is trained by a label feature sequence with time windows to capture the anomaly occurring in a single variable.Based on TRNSYS,a graphically based software environment,the supply air temperature control loop in the AHU system is simulated,providing simulation data that can be used to verify the proposed method.The validation results show that the proposed method can successfully detect and dia gnose four sensor faults in the supply air temperature control loop,including three monitoring variable sensor faults and one controlled variable sensor fault.The T_c acquittal effectively reduces false alarm ratio(FAR)and false diagnosis ratio(FDR),while missing diagnosis ratio(MDR)is slightly increased.A best trade-off between FAR and MDR can be obtained at T_c=0.85.The robustness test results show that the noise has little effect on MDR and FAR in the range of 6 dBm~13 dBm,which shows that the proposed method possess good noise immunity.