Integrating Clustering Analysis For Fault Detection And Classification

Process safety, product quality and environmental protection are the core objectives of the modern industrial processes. As the important part and one of the key technologies in process automation systems, process monitoring has an important and realistic role in meeting these objectives. With the wide application of the distributed control systems in the industrial processes and the rapid development of computer storage technologies, a large number of process data can be well collected and stored. While rich in data, process knowledge is often lacking due to insufficient expertise and lack of practice tools. Multivariate statistical analysis and pattern recognition based data-driven process monitoring has become a research hotspot, and many research results and industrial applications have been obtained in the past decade, to address this challenge.Traditional multivariate statistical analysis and pattern recognition methods do not fully address the complex problems in the industrial process, such as labeling training samples, data imbalance of different faults, nonlinear, multiple operating modes, transition process, and so on. After studying existing methods, integrating clustering analysis and analyzing data preprocessing, feature extraction and pattern classification methods, by integrating clustering analysis, several efficient and effective fault detection and classification methods are proposed to address these problems.(1) To solve the problem of fault detection when training samples of the industrial process is impure, a new kernel principal component analysis (KPCA) algorithm integrating Fisher discriminant analysis-possibilistic c-means clustering-(FDA-PCMC) is presented. FDA is first applied to extract feature and pre-classify the training samples, and then PCMC is applied to cluster the training samples effectively. This means that a hybrid learning including classification and clustering is firstly presented to purify the training samples, and then, KPCA is applied for nonlinear fault detection.(2) Normal process is the majority pattern in an industrial process, while faulty patterns are the minority ones. Moreover, there will be different number of samples periods for each fault. Classifiers generally pay more attention to the majority pattern, while the learning concern should be the minority ones, which makes kernel Fisher discriminant analysis (KFDA) perform poorly when the data sets have an imbalance problem. In the thesis, a novel imbalance modified kernel Fisher discriminant analysis (IM-KFDA) approach named inductive bias KFDA is proposed. To handle the imbalance problem, a novel regular weighted matrix is incorporated into the minimum Euclid distance based pattern classification rule.(3) Research on simultaneous form and serial form based process monitoring. First of all, serial form and simultaneous form of pattern recognition based fault detection and classification are discussed. And then, under the serial form, a novel fault detection and classification method based on principal component analysis-support vector data description (PCA-SVDD) is proposed. Moreover, after discussing the relationship and difference between PCA and KPCA, under serial form and simultaneous form, two novel fault detection and classification approaches integrating KPCA and SVDD are proposed subsequently. At last, a performance assessment rule based on the monitoring system overall loss is presented. Beside of considering the fault detection and classification performance of the classifiers, the mis-classification costs in both of the fault detection and classification stages have also been considered.(4)Research on fault classification based on feature discrimination subspace. FDA and KFDA based feature extraction can project the original data space into the feature discrimination linear and nonlinear subspace, respectively. In this thesis, first of all, the importance of supervised learning for feature extraction is discussed. And then, FDA based feature extraction, Fisher linear classification and SVDD based nonlinear classification are proposed for faulty mode isolation. The cascade and series parallel combination forms are also proposed when incorporating FDA and SVDD. Finally, Gaussian mixture model (GMM) and k-nearest neighbor (kNN) classifiers based fault classification on KFDA feature subspace are proposed, and performance discussions of parameter and non-parameter classifiers are subsequently presented.(5) To accomplish multimode process identification and fault detection, a new bilayer ensemble clustering approach incorporating moving window is proposed. After two-step independent component analysis-principal component analysis (ICA-PCA) based feature extraction, k-ICA-PCA models based bilayer ensemble clustering are proposed for multimode process mode construction, and then, adjoined multiple ICA-PCA models are presented for multimode process identification and fault detection.(6) To solve the fault detection problem of transition process, a dynamic ensemble clustering integrated transition process monitoring approach is presented. To obtain the dynamic and non-Gaussian feature of the transition process data and to label the transition process, an ensemble clustering approach is proposed for transition process analysis, and a new dynamic k-ICA-PCA model is also accomplished. And then, PCA based feature extraction and SVDD based pattern classification are used for transition process monitoring.Finally, research results in this thesis are concluded, and then the future work is discussed.