Statistical Process Monitoring Via Hidden Markov Model

Advances in distributed control system (DCS) and measurement technology facilitate the collection of a large amount of process data. In contrast to the model-based approaches where a priori knowledge about the process is needed, the data-driven methods only require the large amount of process data. Many on-line monitoring schemes based on process data have been described and applied in the processes with multiple inputs and faults. In principle, the data collected from the process contains a full description of the operating status of the plant at any time. The challenge is to have a handy and powerful method to extract and interpret the important and relevant information from process data. It not only needs to detect process abnormality, but also needs to isolate the source of the abnormality, which is called fault detection and isolation (FDI). The FDI schemes based on pattern recognition typically consist of feature extraction and fault classification. First, the inherent characteristics are extracted from process data. Second, a classifier is used to match process features with known operational behavior, and identify the existing process abnormality.In this thesis, wavelet transform and principal component analysis (PCA) are used to improve the performance of feature extraction. Hidden Markov model (HMM) is utilized to identify faults. Then some new monitoring algorithms are proposed. The main contributions are as follows:1. A two-step feature extraction approach combining wavelet transform and PCA is presented. Wavelet transform provides a compact, information-rich expression of process data through a set of coefficients that carry localized transient information of process operating condition. PCA is used to reduce the dimension of correlated coefficients in an optimal way. The wavelet transform and PCA based method can extract more valuable information from process data than PCA.2. HMM is employed to build the statistical model of the feature sequences. The non-Gaussian properties of the feature sequences can be characterized by a mixture Gaussian distribution. And the serial correlations in the sequences can be described bythe state transition of hidden Markov model. Case studies from CSTR illustrate that the inherent characteristics of process data can be accurately modeled by HMM.3. A HMM based fault identification scheme is introduced. First, a set of HMMs are trained with historical process data to construct a database. Each HMM corresponds to a fault. Then wavelet transform and principal component analysis are utilized to capture the inherent characteristics from process test data. Finally, hidden Markov model is used for pattern comparison and classification of various process operation conditions. The proposed method is tested on the Tennessee Eastman process, and the results show that it is able to correctly classify the faults.4. An integrated framework for process monitoring and fault diagnosis is presented. Process data are assumed to be independent and identically distributed (IID) in most traditional statistical process monitoring methods. As a double stochastic model, HMM not only captures the serial correlations in the data, but also considers the non-Gaussian distribution of process data. So the HMM based approach does not require such IID assumption. A fixed-length moving time window is employed to track process dynamic data. It contains a certain number of data samples at each time and facilitates the correct fault detection and identification. Case studies from the Tennessee Eastman process illustrate that the proposed approach is better than the traditional PCA method.5. An on-line fault diagnosis scheme using HMM and variable-length moving time window is presented. It is always difficult to select the proper length for the fixed-length moving time window. Thus a variable-length moving time window is presented. The main advantage of this method is that the length of the window changes with time. When process abnormality is detected, the variable-length moving time window is used to track dynamic data. Over time, the moving window has a longer length. The window contains more valuable information reflecting process abnormal operation, which helps to correctly and quickly identify process abnormality. The Tennessee Eastman process is used for demonstration, and the results show that the proposed approach performs good fault diagnosis capabilities.Finally, the dissertation is concluded with a summary and some remain challenges.