Multimode Industrial Process Modeling And Monitoring Based On Statistical Theory

For modern industrial processes whose scales and complexity increased continuously, it is important to establish effective process monitoring methods to guarantee the safety of production, enhance the product quality and economy benefit. For complex industrial processes, it is difficult to achieve their accurate mathematical model to describe the relationships of energy and mass. On the other hand, with the rapid development of computer technology, a large amount of process data have been sampled and collected. How to explore valuable information from these mass data becomes a practical research topic. In this context, statistical process monitoring approaches based on statistical theories emerge, and have gained intensive attentions and applications.Most of the research achievements in statistical process monitoring field are established based on the assumption that all the historical data and online samples come from single operating mode. However in real industrial applications, operating modes often shift from one to another due to the demands of markets, change of product specifications, or addition of new recipes. In these circumstances, the process data show clustering feature, and the existing monitoring methods build for single mode processes can not be directly planted to multimode processes. In the perspective of data distributions, this dissertation proposes corresponding statistical monitoring strategies for different types of multimode processes by integrating many pattern matching tools. With a novel monitoring index as the main line, the main contents of this dissertation are stated as follows:(1) Constructing a novel global monitoring index based on Mahalanobis distance for multimode processes. The traditional T2, SPE, Mahalonobis distance, and combined monitoring statistics are not suitable for multimode processes because the faulty symptoms caused by real faults will be buried in the variations caused by different operational modes. To overcome this deficiency, a novel integrated monitoring index based on Bayesian rules is proposed. The novel index not only carries local information from all the operating modes, but also contains global statistical information though posterior probabilities, and is able to monitoring dynamic samples from transient process. The effectiveness of the index is verified though a continuous stirred tank heater simulation process with five operating modes. It is also adopted by the other chapters of this dissertation.(2) The historical data from multimode processes usually contain lots of transitional samples, which is difficult to model accurately. To solve this problem, a fuzzy clustering based global modeling approach is proposed. Without identifying and rejecting transitional data in advance, training from all the historical data directly with the local distance of BID index as objective function, the new approach adopts G-K algorithm to train the FCM model, and introduced a PD index to determine the number of fuzzy clusters. This approach takes full advantage of the fuzzy boundaries and classifies the multimode data in a "soft" way, eliminating the influence of outliers and transitional samples, enhancing the robustness of the monitoring model. Compared with multiple model based methods, the proposed approach simplifies the modeling steps and reduces the false faulty alarm rate. The efficiency of the proposed method is verified though the TE benchmark with three operating modes.(3) A locality preserving projection (LPP) based dimensionality reduction and subspace modeling approach is proposed for large scale multimode processes with high-dimensional monitored data. For high-dimensional historical data, it is time-consuming to train fuzzy models directly. Besides, the model training procedure is prone to fail due to the correlations between variables, which may lead to singular matrix. Aiming at preserving neighborhood information of original data set, LPP is adopted to reduce the dimensionality of process data. In this way, the correlations between variables are eliminated, and meanwhile the clustering information in preserved in the subspace. The model established in LPP subspace has correspondence relationship with original multimode processes, and the monitoring results are easy to be explained.(4) An adaptive online modeling and monitoring approach based on Gaussian Mixture Model (GMM) is proposed for multimode processes with time-varying features. The Expectation Maximum algorithm is enhanced by incorporating Minimum Message Length criteria, thus the modeling procedures are upgraded into unsupervised pattern and no prior knowledge about the multimode process is needed. When implemented online, a moving window carrying the latest process information is utilized to adaptively update the mean and covariance of the target Gaussian component, which guarantees the GMMs contain the latest process information. Thus the monitoring model is accurate and able to discover abnormal events timely. Meanwhile, there is requirement adding a new operating mode online due to the demands of markets sometimes, even under this circumstance the proposed adaptive approach is still able to function effectively.(5) A hybrid unfolding based multiway Gaussian mixture model is proposed for modeling and monitoring batch processes with long running cycles. Starting with limited batches, the initial monitoring model is updated with data from latest batch. Considering the multway feature of batch process data, a hybrid unfolding method is implemented to pretreat data, thus the dynamic features from both batch-wise and variable-wise are preserved in the model, and the online monitoring procedure is also facilitated. A fed-batch penicillin fermentation process with multiple phases is utilized to demonstrate the effectiveness and sensitiveness of the proposed algorithm.