Fault Diagnosis Algorithms Of Chemical Process Based On Canonical Variate Analysis

Chemical processes become more complicated with continuous development of science and technology. Research on fault diagnosis algorithm plays a key role in ensuring the production safety, avoiding to damaging the process system and improving production quality of chemical process. Many scholars at home and abroad have improved fault diagnosis with many ways in order to achieve fast and accurate diagnosis effects. In general, actual chemical process has high complexity and large-scale features. Chemical process usually collects rich process variable data and fault diagnosis methods based on data-driven can directly analyze data which are collected by chemical processes to achieve the fault diagnosis for chemical process, an in-depth study of fault diagnosis based on data-driven not only has important theoretical significance, and also has broad prospects for actual production.This thesis analyzes and improves canonical variable analysis(CVA) algorithm, main workings are done as following several aspects:1. Because CVA algorithm cannot being well processing dimension reduction for high-dimensional data of chemical process, a fault diagnosis CVA algorithm of chemical process based on isometric mapping(ISOMAP) is proposed, At first, this algorithm uses ISOMAP algorithm of manifold learning to achieve realize nonlinear dimensionality reduction for initial data and maintain internal geometry structure of data. Then CVA is used to the extracted low dimensional data to obtain process state vector and monitoring statistics, the rationality of the algorithm used for fault diagnosis in chemical process is verified by the TE process simulation.2. Process data usually have time-varying characteristic, the fault diagnosis based on canonical variate analysis(CVA) cannot have an accuracy rate in the dynamic process system, so a moving window CVA algorithm is proposed. At first, the initial CVA model is set up and monitoring statistics is computed, then it is updated by calculating and updating data, the superiority of the proposed method is verified in Tennessee-Eastman process simulation.3. Usually, there are multiple operate modes in chemical process because of producing kinds of products. The traditional fault diagnosis methods for single operate mode are no longer applicable when they used to diagnose process of multiple operate models. therefore, a CVA algorithm based on Gaussian Mixture Model is proposed, first of all, history data of chemical process is decomposed to multiple Gaussian components by using Gaussian Mixture Model(GMM), then kernel canonical variate analysis(KCVA) algorithm is used to model for each Gaussian component and calculate the corresponding statistics for process monitoring. The effectiveness of the proposed method was verified in Tennessee-Eastman process simulation.