Gaussian Process Adaptive Soft Sensors and their Applications in Inferential Control Systems

building. This research reviews the use of this technique
as an adaptive soft sensor building method. It investigates different model structures, addresses issues associated with this technique, introduces Gaussian
process-based soft sensors in inferential control, and proposes a methodology to enhance the reliability of the introduced inferential control system. These are achieved by conducting various case studies and empirical experiments on real and artificial data retrieved from real and simulated industrial processes. The comparative case studies conducted on various Gaussian process model structures revealed that the Matern class covariance functions outperform the predominantly used squared exponential functions, particularly in clean and properly pre-processed data sets. The results show the plausibility of Gaussian processes in building adaptive soft sensors, particularly those based on windowing techniques. Specifically, empirical results have revealed that the prediction accuracy of the sensor can be improved by considering window-updating criteria. The research results have also shown that the size of raw data used for soft sensor development can be significantly reduced while preserving the informative properties of the data. This results in a significant reduction in the associated computational cost of Gaussian process-based models. Simulated results have also revealed that an adaptive soft sensor with a high prediction capability can be integrated with Proportional Integral controllers to build inferential control systems. The robustness and reliability of such a control system can be enhanced using a hybrid Gaussian process and kernel Principle Component Analysis-based method. This allows the implementation of the control system in the industrial process during both healthy and faulty operating conditions.