Predictive Models for Quality Variables with Regular Sensors and Images

Abstract

In the process industry, certain quality variables cannot be measured regularly due to technical limitations or economic constraints. Consequently, the industry relies on laboratory analysis to measure such quality variables. However, laboratory analysis introduces long time-delays in obtaining measurements due to the need to collect representative samples, transport them, and conduct the analysis. The associated delay can pose challenges in terms of timely decision-making and real-time control. Therefore, fast-rate measurements in the process industry are crucial for real-time quality monitoring and control. By obtaining real-time data, industries can improve efficiency, meet customer expectations, and minimize risks associated with the quality variables. Data-driven predictive models, also known as soft sensors, have emerged as valuable tools in predicting quality variables in the process industry. The predictive models employ advanced data analysis techniques to predict the values of quality variables using different data sources including regularly measured variables by online sensors and visual information provided by video cameras. By leveraging historical data and identifying patterns, the models can provide fast-rate predictions of quality variables. This capability enables proactive decision-making and facilitates timely interventions to maintain product quality and process control, and contribute to process optimization. In this thesis, we develop predictive models based on both conventional sensor measurements and image data, each with its own advantages. The first two contributions are related to conventional sensors data and the following two contributions are concerned with image data as a means to develop predictive models. The main contributions of this thesis are listed as follows. The first contribution involves developing a statistical predictive model for quality variables with simultaneous consideration of time-varying time-delays, time-varying sample collection periods, and varying operating points. A non-parametric distribution is used to describe the distributions of the time-delays, sample collection periods, and switching of different operating modes, eliminating the need for prior knowledge about the distributions. Furthermore, the work enables online updating of the model parameters using a recursive Expectation-Maximization (EM) algorithm. Then, we extend the linear time invariant predictive model to a linear parameter varying (LPV) predictive model, and enhance robustness to outlying output observations through the use of t-distribution. Additionally, uncertainty of the unknown model parameters is estimated using variational Bayesian (VB) algorithm. Development of a computer vision model to predict quality variables is the third contribution. A modified Kalman filter is formulated to restore degraded images caused by factors like lighting conditions changes and camera noise. Additionally, to estimate the predictive model parameters, a robust-to-outlier EM algorithm is developed. The proposed model was validated on a tailing flotation process. In the last contribution, the development of a computer vision model that enables fast-rate prediction of quality variables is considered. To address the impact of environmental conditions like steam and lighting on images, an atuoencoder-based image inpainting algorithm is developed to fill in the missing regions in the images. The restored images, along with slow-rate sampled measurements, are then used in conjunction with the EM algorithm to construct an auto-regressive with eXogenous input (ARX) predictive model. All the proposed models in this thesis have been validated through experimental studies.