Innovative techniques for industrial process modeling and monitoring
This research presents several innovations in industrial process modeling and monitoring with the purpose of better controlling the process. In semiconductor manufacturing industry, people show increased interested in thermal modeling of Low-Pressure Chemical Vapor Deposition (LPCVD) processes in order to understand the process better and get tighter control of film uniformity. Research in this area has resulted in several firstprinciples models. However, the common drawback of these models is that they are more academically-oriented than industrial-oriented, i.e., the intensive computation makes them very difficult to be applied online in order to meet the high-volume manufacturing needs. In this dissertation, a first principles transformed linear model is developed for the LPCVD process to address drawbacks of existing thermal models and facilitate its industrial implementation. The proposed model accurately predicts wafer temperatures using the furnace wall temperatures, and it can be solved using a direct algorithm in only a few seconds. The simplicity of the model form and the fast algorithm make the model desirable for real-time updating and control of industrial scale furnaces. In process industry, many control loops perform poorly due to reasons such as bad tuning or equipment problems. Among them, valve stiction is one of the most common equipment problems. Although there has been many attempts to understand and model valve stiction, those models are either physical models which are not practical to use, or empirical models but with rather complicated logic which make them difficult to understand and implement. In this work, a new valve stiction model is proposed with simple structure and straightforward logic which make it easy to implement. Furthermore, several published valve stiction detection techniques are reviewed. The inconsistency of Horch’s first method is theoretically analyzed and illustrated by a simulated example. A new valve stiction detection method is proposed based on curve- fitting for both self-regulating and integrating processes. The new method shows superior performance to other existing methods. Fault diagnosis plays an important role in supervision and maintenance of chemical processes in an effort to isolate the root cause once a fault is detected. The well known fault diagnosis approaches, i.e., contribution plots based on Principal Component Analysis (PCA) and Partial Least Squares (PLS) models, may not explicitly identify the cause of an abnormal condition, and sometimes may lead to incorrect conclusions. In this work, a new fault diagnosis method using fault directions in Fisher Discriminant Analysis (FDA) is developed in attempt to provide a better solution than the traditional contribution plot based on PCA. Besides, a new process monitoring method is proposed which consists of data pre-analysis, fault visualization and fault diagnosis. In both simulation example and a film industrial example, the contribution plots proposed based on fault directions in pair-wise FDA shows superior capability for fault diagnosis to the contribution plots method based on PCA. In today’s chemical industry, massive amount of data are easily available in computer controlled processes. But at the same time, the visualization of high dimensional data has been difficult. The dramatically increased computing power has not been utilized to improve the situation in industry. In this work, the commonly used visualization techniques, usually applied to relatively small static systems, are evaluated in the context of large dynamic systems. A general framework of hierarchical visualization is proposed and several multivariate visualization methods are developed in this work. The performance of PCA, PLS, Class Preserving Projection (CPP), FDA and two proposed approaches based on Support Vector Machines (SVM) are compared using an industrial data set.