Research On Industrial Process Fault Diagnosis Method Based On Statistical Theory

With the continuous expansion of production scale and complexity of the modern industrial process,the effective fault diagnosis method is the key to ensuring process safety and improving product quality.The method based on statistical theory has become one of the main methods in the field of fault diagnosis and developed rapidly,it has received widespread attention of the researchers.This method does not need the precise mathematical model for the system,and it mainly depends on normal process data to establish the statistical model that is used to describe the safety in the process of producing.However,due to the complexity and diversity of industrial process,the heavy use of sensors,and the diversity of market demands which lead the product specifications to change repeatedly,which cause the collected data has different distribution characteristics and the change of product specifications.This would limit the use of fault diagnosis method based on statistical theory and lead to an error diagnosis result.Therefore,in view of these features,how to deal with different properties of process data,propose effective process detection method and improve the accuracy of process monitoring are the key problems of fault diagnosis.When there are abnormalities or faults,it is necessary to diagnose and eliminate them.Therefore,how to locate and judge the abnormal variables,establish accurate "fault-sign" table which is used as a knowledge for fault evaluation and decision-making,realize the intelligentization of fault diagnosis system have become the most important parts of the research,and they are becoming the research focus in the field.On the basis of the in-depth understanding of the different process characteristics,data features as well as the traditional fault diagnosis methods based on statistical theory,a series of fault detection methods are researched.Besides,fault identification and location methods of distributed principal component analysis and data reconstruction are studied based on variable reasoning strategy,k-NN variable contribution analysis and so on,the goal of which are to identify abnormal variables preliminarily and accurately in industrial process,respectively.This dissertation work mainly manifests in the following several aspects:(1)In view of the non-Gaussian and nonlinear,as well as high dimension problems of industrial data,the concepts of diffusion maps and diffusion distance are introduced which belong to power system and used for nonlinear dimension reduction of process data,and a new fault detection method based on feature space k nearest neighbor diffusion distance is proposed on the basis of the diffusion mapping space.First,normal high-dimensional data sets are converted to a low-dimensional feature space by analyzing the insightful relationship between data points,and the data in feature space can represent major information of raw data.Subsequently,like the traditional k-NN method,the sum of k nearest neighbor diffusion distance is computed.Third,establishing k nearest neighbor fault detection model of normal samples and setting a threshold of normal process using the kernel density estimation method,they are applied to monitor industrial process.(2)In view of the multi-modal characteristics of industrial data,a novel multi-modal data processing method called weighted k neighbor standardization(WKNS)is proposed to address the multi-modal data problem.Unlike the traditional z-score method,the principle of WKNS is that each sample of normal data is standardized using weighted k neighbor standardization strategy,which can transform multi-modal data into an approximately unimodal or Gaussian distribution,can also erase the distribution characteristics which contained in multimodal data effectively and it can't destroy the relationship between the variables.Furthermore,combining WKNS with the method of PCA,a new fault detection approach called WKNS-PCA is developed and applied to detect process outliers.This method does not require process knowledge and multi-modal modeling,only a single model is required for multi-modal process monitoring.For the new sample,this method also doesn't need judge which model it is.This method can improve the fault detection accuracy effectively.(3)In view of the problem of abnormal variables identification when there are faults in industrial process,a fault diagnosis method based on distributed PCA and variable reasoning strategy is proposed.Before monitoring industrial process,the selection of relevant variables should be given in each local model,and then establishing the distributed model;secondly,the distributed PCA models should be built using traditional PCA method and calculate the control threshold of normal circumstances used to detect the new samples,and then determine the broken-down and normal sub models;finally,the initially judgment of abnormal variables can be given according to variable reasoning strategy and the traditional variable contribution plot method.This method can not only locate normal and fault variables effectively,but also give initiatory judgment for ambiguous variables.(4)In view of the problems of abnormal variables precise identification and how to establish"fault-sign" table when there are faults in industrial process,a novel abnormal variable identification method based on k-NN contribution analysis and data reconstruction is studied.Firstly,this method gives a detailed analysis of distance control indexes for each sample,assigns them to each variable,and establishes the contribution model of the normal variables.In order to ensure a greater contribution for abnormal variable,the feasibility of the method is verified from the angle of univariate and multivariate anomaly respectively;secondly,for the normal data,the contribution of each variable is set up used to "first time" identify abnormal variables.And then four data reconstruction methods based on k-NN are researched,which include average k-NN,k-INN,weighted k-NN and CNN.The principle and the reconstruction precision of these algorithms are analyzed and compared,and the advantage of CNN is validated.For fault samples,the contribution of each variable in the distance control index is calculated according to the contribution analysis method firstly,which is used to identify the abnormal variables "first time".And then abnormal variables can be reconstructed by CNN method,a new detection and"second time" identification processes should be implemented until all the abnormal variables are found.However,this method can ensure the accuracy of anomaly variable identification,and get the relationship between fault and variables,namely "fault-sign"table.