Statistical Modeling And Online Monitoring For Multiple Mode Processes

Recently, with the urgent requirement of multi-type and high-quality products of the market, the efficient production process of multiple products become emphasis in many industries. The process safety and reliability have been the focus of engineers’ attention. However, due to different manufacturing strategies or varying feedstock, the process covers multiple modes which have distinct process correlation characteristics. Considering the process high dimensionality, multi-operation, time-variant characteristics, and unknown mode duration, it is challenging to conduct statistical analysis and online application for multi-mode processes.Multi-mode industrial processes are fairly complex with rich data statistical characteristics. It should not only pay attention to develop statistical modeling and process monitoring for stable modes, but also take the transitional mode between stable modes into consideration. On the purpose of application, this dissertation develops a series of mode-based statistical modeling, process monitoring methods for multi-mode processes especially focusing on several key points, such as, data classification, mode identification, feature extraction, and so1. Mode identification. Mode identification is category division for data without mode information. Identifying the mode type is the most important for both offline modeling and online monitoring. For modeling data mode identification, a novel method based on the similarity of data characteristics and underlying process behaviors is proposed to distinguish stable mode and transitional mode. For online data mode identification, the right model is chosen based on current operating status and Mode Transformation Probability (MTP), which can digs out the empirical knowledge hidden in offline data.2. Statistical modeling for stable mode. In order to extract the underlying process correlation characteristics of stable mode, a mode-representative modeling method is developed to build proper statistical model for each stable mode. Process data is classified into three types:Gaussian distribution data, non-Gaussian distribution data, and the data containing both distributions. According to Gaussianity test of the data, proper multivariate statistical algorithm is selected automatically to extract features for each stable mode.3. Statistical modeling for transitional mode. Transitional mode is a dynamic process occurring when production operate mode changes. The dynamic performance reflects on both the varying variables and the changing variables correlation. The valuable information is extracted by analyzing variables variance, which can be expressed using differential data. At last, segmented modeling is applied to describe the characteristics of transitional mode and is proved to preform well.4. Modeling for new mode. Alterations of operating specifications or fluctuations in the external environment will lead to production in un-modeled new mode. In order to improve scalability and reliability of the monitoring system, a modeling algorithm for new mode of multi-mode process is proposed. New mode is distinguished from abnormal status on the basis of statistics indicators and expert knowledge. The initial model is established using limited new data combining the characteristics of history models. With increase of modeling data, the iterative model based on the initial model is performed to fit the process more accurately.5. System design and implementation. By combining the proposed mode identification and modeling methods, framework for multi-mode process monitoring system is set up with its primary functions specified. An experiment platform including mode identification, process monitoring, fault diagnosis, and history fault query is also implemented for continuous annealing line to verify and illustrate the proposed statistical methods.The proposed approaches considering both stable mode and transitional mode enrich the achievement of statistical modeling, online monitoring and fault diagnosis for complex multi-mode process. The simulations, experiments and applications of the proposed approaches based on experiment systems demonstrate the effectiveness of the present methods. The development of the experiment platform also has practical significance, since it can be used not only for continuous annealing process but also for other complex industrial processes.