Researches On Fault Diagnosis For Process Industry With FDA/DPLS Methods

With the progress of modern industries in large-scale and integration, processes have become more and more complex and contain a large number of measured variables. Once faults occur, they will make casualties and significant economic losses. Therefore, the processing monitoring has been become one of the most active research areas in process control. If faults can be detected and diagnosed in time, the safety of process operation is improved, productivity loss is reduced and product quality is enhanced. Since the data-driven methods of process monitoring rely only on the historic process data and do not assume any form of model information, they have been paid more attention.In this thesis, some new methods of fault diagnosis for batch processes and continuous processes are proposed based on discriminant partial least squares (DPLS) and Fisher discriminant analysis (FDA).The main contributions are as follows:(1) A novel method, MPCA-MDPLS approach, is presented to diagnose faults for batch processes in order to detect new faults. Based on data collected from the plant during the normal operation and specific faults, MPCA model and MDPLS model are respectively constructed. Therefore, MPCA model can detect the unknown faults. The known faults detected by MPCA are further diagnosed by MDPLS model. Therefore, MPCA-MDPLS diagnosis method not only can diagnose the known faults, but also can recognize the unknown new faults. The method is proved to be effective and feasible by the application in diagnosing of an industrial streptomycin fermentation process.(2) A multi-model FDA method is developed to diagnosis faults for batch processes. The complete batch data are collected until the end of the batch operation so that the future unmeasured data must be estimated for MFDA in terms of on-line fault diagnosis. Although, many methods can be used to estimate and fill in the unmeasured data, the estimated values may not exactly follow the actual dynamic process behavior and they may lead to false detection and diagnosis. The proposed method not only avoids estimating or filling up the unknown part of the process variable trajectory deviations from the current time until the end, but also directly identifies the assignable cause of process abnormalities. Therefore, more accurate decisions are made via multi-model FDA for on-line fault diagnosis. The method isproved to be effective by the application to monitoring and diagnosing of a streptomycin fermentation process.(3) A recursive multi-model FDA (RMMFDA) method is represented in order to sufficiently utilize the finite information of faults and enhance the diagnostic performance. If there are not enough fault data in historical datasets, it makes high misclassification rate using multi-model FDA, which a model is build only based on a time slice at each time internal. The trait of the proposed method is that the training data sets are made of the current measured information and the past major discriminant information, not only the current information or the whole batch data. An industrial streptomycin fermentation process is used to test the performance of fault diagnosis of the proposed method. The results show that the RMMFDA method can further improve the diagnostic performance.(4) An improved method of fault diagnosis named as improved MFDA (IMFDA) is proposed. Batch data of measured variables are not complete until the end of the operation. Moreover, starting conditions and exotic environment of each batch run are different so that the total duration of the batches is not same. In order to overcome those drawbacks and enhance the diagnostic performance, IMFDA is presented. The diagnostic performance is tested by application to the simulated fed-batch penicillin fermentation. The results show that this method is feasible.(5) A mixture method of fault diagnosis, ICA-FDA, is proposed using independent component analysis (ICA) and Fisher discriminant analysis (FDA). A large number of process variables are measured in a chemical process, but this process is usually driven by...