Statistical Process Monitoring Methods Based On Local-Global Structure Analysis

Process monitoring is an important part of the integrated automation technology for industrial process. On one hand, it reduces the process volatility, improves production quality, and avoids the occurrence of fatal accidents; on the other hand, it is a basis for the follow-up high-level automation technology, such as control, decision, and scheduling. Nowadays, since the industrial process is becoming more and more complex, multivariate statistical process control (MSPC) methods, which do not depend on the process model, are widely studied in both academic and industrial community.Traditional research works of MSPC, e.g., PCA, mainly focuses on global features in the variable space and sample space. However, there is not much attention on local features. To improve the performance of MSPC, we developed the concepts, frameworks and fault detection algorithms of local-global structure analysis approaches separately from two points of view:the variable space of data and the sample space. Specifically, the main work of this paper focuses on the following issues:(1) A local-global analysis framework in variable spaces is proposed. Different from the traditional methods, which consider a complete sample space and single model detection scheme, an independent component (IC) contribution based subspace partition and fault detection method is proposed. By dividing the original variable space and subspace detection modeling, the proposed method effectively improves the detection potential for small and latent fault. Thus, the advantage of local modeling on variable space is verified. Furthermore, the method is generalized with ensemble learning. Firstly, we showed the theory of global-local analysis framework with ensemble learning. In this framework, a novel ensemble learning approach called independent component subspace method (ICSM) is developed. ICSM, which is based on the distinguishing characters, raises a new ensemble strategy for fault detection, and constructs more efficient monitoring statistics. Thus, ICSM can describe the features in both global and local variable space. Experiment results show that the proposed framework performs better in fault detection.(2) A novel approach called Dynamic LPP (DLPP), which preserves the local sample space structure, is used to monitor dynamical processes. Compared with PCA, LPP is based on the local structure of sample space. Thus, LPP can effectively keep the latent manifold of data after dimensionality reduction and has better feature extraction ability. Therefore, the new DLPP method which combines LPP and dynamical matrix outperforms DPCA.(3) A local-global analysis framework in sample space is developed. There is abundant structure information of data in sample space. The extraction and preservation the structure information will greatly influence the performance of fault monitoring algorithms. Traditional PCA model only considers global structure in feature extraction, while manifold learning methods, e.g., LPP, mainly focus on local structure of data. We consider the merits and demerits of both PCA and LPP together, and built a dimensionality reduction framework to keep global-local information, which integrates PCA and LPP. A new optimization objective is constructed. The strategy and computation details to balance global-local structure are also given. The proposed method, which makes the latent space have similar local structure as well as global structure to the original measurement space, extracts more feature information from data. Moreover, Bayesian inference is introduced to make a strategy for fault identification. Studies on Tennessee Eastman (TE) process show that the new framework algorithm has better performance than PCA and LPP.(4) In local-global structure analysis framework, there are important issues to deal with noises and choose the number of latent space variables. Therefore, a fault detection method based on Bayesian PCA (BPCA) is proposed. In BPCA. hyper parameters are used to give a prior distribution for each direction of the mapping matrix. The distribution can be estimated by using expectation-maximization (EM) algorithm so that the probability of latent variables in different directions is given. Thus the latent variable can be determined automatically. Compared with the existing probabilistic PCA (PPCA) and factor analysis (FA) models, the new method not only determines the number of latent variables, but also avoids over-fitting problems and improves the performance of fault detection. The theory and fault detection efficiency are verified in TE process.At the end of this paper, we give a summarization of the whole paper and discuss the subject for further research.