Linear Dynamic System Model Based Process Monitoring

Process monitoring plays a significant role in ensuring the safety of operations, improving the quality of products, reducing the losses of faults, and enhancing the competitiveness of enterprises in the international market. Since the rapid development of computer information technology has brought era of Big Data, data based process monitoring methods have become popular in process industries recently. However, faced with more and more complex characteristics of process data, traditional process monitoring methods have inherent limitations because of their considerations of only a single characteristic usually. This thesis introduces the linear dynamic system model (LDSM), a typical dynamic Bayesian network model, for process monitoring, which can efficiently deal with both dynamic and uncertain features of process data. The main contributions are summarized as follows:(1) In order to capture both dynamic and stochastic feature of process data, this thesis proposes a switching LDSM based approach for fault classification and unknown fault detection. A novel and convenient learning algorithm is developed for parameter estimation of the switching LDSM, and the Gaussian Sum Filtering method is introduced for online fault classification. In addition, considering the previously unknown dynamical behavior that may occur in the industrial process occasionally, a switching LDSM based threshold statistic is defined for unknown fault detection.(2) Taking essential process information included in quality variables into consideration, a supervised LDSM based approach for fault detection and a switching supervised LDSM based approach for fault classification are proposed. This thesis expands the LDSM into the supervised LDSM, and figures out the detailed derivation of building a supervised LDSM via the expectation maximization (EM) algorithm. A new Gaussian Sum Filtering method is developed to calculate the distribution of latent variables. And then T2 monitoring statistics is set up for online fault detection. Besides, while quality variables are available, a switching supervised LDSM is introduced for a better performance of fault classification.Finally, the thesis draws an conclusion of the research results, and forecasts the further study of LDSM in the field of process monitoring.