Abstract:
Today, production facilities employ rigorous approach to process control, mon itoring and fault detection strategies, due to increased competition in the industry, and more severe manufacturing restrictions and safety concerns. A large number of process variables are measured online during operation, while quality variables, which may not be available online at the same rate as process variables need to be predicted using soft sensors. The traditional approach to data driven soft sensor design is based on statistical learning methods, such as, global ARX modeling, PLS and PCR. In this thesis, it is aimed to address the issues of multicollinearity and redundancy in process data, and concept drift in data driven soft sensor design. Experiments are performed on two synthetic datasets obtained from steady state and dynamic simulations, and one dataset comprised of dynamic industrial data. Incorporating feature selection and ensemble modeling into PLS and ANN models is shown to handle multicollinearity and redundancy in process data, and stabilize learners. RVM, an embedded feature selection method, yields superior prediction performance than PLS, under both virtual and real concept drift. RVM is also observed to handle redundancy in predictor space more effectively. Several RVM-based adaptive learning algorithms are developed to cope with concept drift. Adaptive window sizing in moving window models is shown to improve predictions, and the best overall performance is achieved in window size adjustment via explicit concept drift detection. Combining MW and JITL models in ensemble learning is suggested to further increase prediction accuracy, and it is ob served to yield the best predictive performance among all algorithms. The suggested adaptive learners are shown to outperform conventional methods from the literature on both synthetic and real data, while complying with time limits of online prediction.
