Fisher Discriminant Analysis Based Semi- Supervised Fault Classification

As the important element of the process system engineering, process monitoring technology has the research significance and realistic application value to make the operation of processes safe and stable and improve the quality of products and other core objectives of modern industrial processes. Due to the continuous development of industrial control technology, the distributed control systems are wildly applied in the industrial process. In the industrial processes, large amount of process data were well collected. Therefore, the multivariate statistical analysis and pattern recognition based process monitoring has raised the interest of both academic and industrial fields and become an hotspot. In past two decades, there are lots of researches and applications in this field.However, traditional multivariate statistical analysis and pattern recognition methods do not fully consider the problems in the complex industrial process, for example, the training samples label missing, data imbalance of normal and faulty operation mode, nonlinear. After studying the existing supervised classification methods, the semi-supervised learning are introduced in the fault classification to address the above problems.(1) Describe the background and status quo of the process monitoring, data-driven based multivariate statistical analysis and pattern recognition process monitoring are discussed in detail. Aiming at the data characteristics of complex industrial process, introduce the fault classification methods combine with semi-supervised learning ideology.(2) To solve the overfitting problem of Fisher discriminant analysis when there are only small number samples available, propose a semi-supervised FDA by introduce the semi-supervised learning ideology. There are always a smaller number of faulty samples compared to the normal samples in industrial process, as a result, the faulty training samples for model building are quite rear. In addition, it need expert experiences and prior process knowledge to label the faulty samples into different types, which is time-consuming and costly, In this thesis, a novel SFDA based fault classification model is proposed, which can make use of the supervise information of labeled samples and the whole structure from the unlabeled samples simultaneously to avoid the overfitting problems under the small samples situation.(3) To solve the nonlinear problem of the processes, developed a novel KSFDA model by introducing Kernel Trick. A KSFDA based nonlinear fault classification method is proposed in this thesis. The k nearest neighbor and Bayesian classifier are applied on the feature extracted by KSFDA for the classification decision. It enhances the generalization capability of original semi-supervised classification and expand the range of its application.Finally, research results in this thesis are concluded, then the future work is discussed.