Research On Sensor Fault Diagnosis Method Of Central Conditioning System Based On Wavelet Neural Network

Facing "Two carbon" idea and the overall trend,central air conditioning systems as the mainstream air treatment equipment,it plays an indispensable role in construction,transportation,medical treatment,textile,food,tobacco and other industries and scenes.As the most basic component of the central air conditioning system,sensors are distributed in the whole central air conditioning system.If they fail,the normal operation of the central air conditioning will be affected,which will not only reduce indoor air quality,but also lead to increased carbon emissions and energy consumption,and even cause safety accidents.Therefore,fault diagnosis for its working state and error compensation for fault sensors can help the staff to timely understand the specific operating status of the equipment and strive for a certain decision time after the occurrence of a fault,reduce the loss caused by the fault.In this thesis,HY-31 C sensor of central air conditioning experimental platform is taken as the research object,and the fault diagnosis method of central air conditioning sensor based on Wavelet Neural Network(WNN)is deeply studied aiming at four typical faults of sensor complete failure,deviation,drift and precision reduction.The research work mainly includes the following parts:(1)Develop the central air conditioning sensor signal acquisition system and simulate fault injectionThe cause and mechanism of typical sensor faults are studied and analyzed,the HY-31 C type central air conditioning experimental platform using MCGS configuration of embedded software development platform,the corresponding visual model interface and complete failure data acquisition system and four typical faults of sensors are simulated on the platform,The various variable signals of the central air conditioning sensor after normal operation and failure are collected.(2)Research on sensor data denoising method based on dual denoisingA dual noise reduction method is proposed to solve the problems of residual noise and inadequate adaptability in existing noise reduction methods.For the residual Noise of signals in Complete EEMD with Adaptive Noise(CEEMDAN),the modal estimation was replaced by the local mean estimation,and the k-order modes were extracted by the local mean of signals.Then the correlation coefficients are calculated to screen out the modal components with more noise so as to retain as much effective information as possible.For the false mode that appears in the early stage of Decomposition,Singular Value Decomposition(SVD)was used to reducing noise again and the denoising order was determined by difference spectrum,which achieves the purpose of secondary denoising and serves as the data signal for subsequent fault diagnosis.(3)The sensor fault diagnosis method based on adaptive fuzzy index and WNN is studiedFault diagnosis includes abnormal signal detection and fault classification.This thesis improved the cross mutation probability of Genetic Algorithm(GA)and population screening method to speed up the convergence,optimized the weight threshold of wavelet neural network,and established the wavelet predictor and classifier model respectively for fault detection and classification.The traditional threshold fault detection method is seriously disturbed by noise and difficult to set.According to the sensor fault characteristics and dimensionless index of time domain parameters,fuzzy threshold index is constructed as the control limit of abnormal signal detection and the optimized wavelet predictor model is used for fault detection.For each sensor in the system variable coupling effect and composite fault accuracy is not high question,multi-source information fusion based on wavelet neural network sensor fault diagnosis methods,the fractal dimension and energy characteristics of multisource information are calculated and input into the classifier respectively to obtain the classification results.At last,by time domain adaptive weighted matrix believe decision fusion diagnosis results are obtained.(4)Study the actual data of the central air conditioning experimental systemThe linear and nonlinear errors of sensors can not be compensated accurately and rapidly in existing methods.A multi-stage error compensation method was proposed based on Ordinary Least Squares(OLS)and WNN.The least square method requires fewer samples and has a good effect on linear error compensation.It is used for quick preliminary compensation of fault data and quick restoration to normal state.Finally,Levenberg-Marquardt(LM)algorithm was used to optimize the WNN to accelerate the convergence speed,and the data after the initial compensation was compensated in the second step to restore the measurement accuracy of the sensor.