Research On The Fault Diagnosis Of Heat Exchanger Station Based On Time Series Analysis Method

For the fault diagnosis of non-linear,time-varying and time-series control systems such as heat exchange stations,the traditional non-time series fault diagnosis method has the problem of inaccurate fault identification,and a fault diagnosis strategy based on time series analysis method is proposed.The main features of this kind of system are that there is strong non-linear relationship among the variables and the variables are coupled with each other.At the same time,the output characteristics of the system are not only related to the change of the input signal,but also to the action time and duration of the input signal.Therefore,the information transfer and time series relationship between variables should be fully considered in the fault diagnosis of this kind of system,and the fault diagnosis model should be built based on the information of the current moment and the historical moment.First of all,by setting up a real physical system,the small multi-input and output non-linear,time-varying,time-series devices are specially studied to simulate the occurrence of related faults and build a feature class fault database.To analyze the information transfer and time series characteristics between variables,the transfer entropy method is used to analyze the information transfer strength between related variables on the time axis,so as to obtain the maximum transfer time.Since the original time series data has noise disturbance,it is symbolized to reduce the disturbance,and the processed data is used for the use of the transfer entropy method.At the same time,in order to ensure that the selected data segments contain diagnostic information as reasonably as possible,the maximum length of time extracted by the transfer entropy is used as the size of the data window,so that the divided data segments to be detected can reasonably contain both the historical and the current information of different faults.Secondly,the Long Short-Term-Memory neural network(LSTM)fault diagnosis model is established for time series data segments divided by sliding window technology to identify known fault types in the fault library.By comparing and analyzing the diagnostic results of traditional RBF,BP,PNN non-time series neural network models,the results show that the overall recognition accuracy of time series analysis method is 93.7%,higher than other methods by 10%.The time series relationship between data should be fully considered in the fault diagnosis of such systems.Thirdly,given that the known failure database data does not contain all the failures of the system.To solve the problem of identifying unknown failure classes in the fault library,the Support Vector Data Description(SVDD)method is introduced,a hypersphere is constructed for the known failure data in the fault library,and an unknown failure detection and decision threshold is set for the detection of unknown failure classes in the model to reduce the error diagnosis of unknown failure classes by the diagnostic model.By retaining unknown fault data and cluster analysis,new fault classes are generated according to their data density characteristics,and then the fault library and diagnostic model are iteratively updated after definition,so as to construct the openness and iteration progressive characteristics of this diagnostic method.Finally,the SVDD-LSTM fault diagnosis model and the Cyber-Physical Systems(CPS)are established to realize the real-time on-line fault diagnosis of the operation process of the heat exchange station.The method can accurately identify the known fault types online and intelligently increase the anomaly detection of unknown faults,which is helpful to ensure the safe operation of heat exchange stations.It can also provide reference for the fault detection and diagnosis of such production processes.