Computer Vision for Interface Level Detection in Oil Sands Primary Separation Cell

Abstract

In the bitumen extraction process, precise control of the froth-middling interface in the Primary Separation Cell (PSC) is critical for maximizing bitumen recovery. Traditional methods for monitoring this interface suffer from reliability issues due to sensor clogging and challenging process conditions. This thesis presents innovative computer vision techniques to enhance the accuracy and reliability of interface level estimation within PSCs. The research begins by detailing the problem followed by a review of computer vision and image processing principles, laying the groundwork for the methodologies employed in Chapter 2. The preceding chapter, Chapter 3, introduces a state-of-the-art image restoration algorithm paired with an image segmentation technique to refine interface level measurement. The approach employs a state-space model augmented with skewed-t distribution to handle image degradation, with parameter estimation facilitated by an expectation-maximization (EM) algorithm in conjunction with a robust Kalman filter (KF). The method's effectiveness is demonstrated through laboratory-scale experiments, where it surpasses existing models in estimating the interface level. In Chapter 4, the research advances with a novel framework designed to address the challenges posed by partially occluded PSC interface images. This includes the utilization of background subtraction and advanced autoencoder-based inpainting for image restoration. The developed framework integrates spatial and temporal analysis through Markov Random Field (MRF) segmentation and image differencing algorithms, respectively, augmented by an ARX model to capture process dynamics. The fusion of image-based observations with process models via dual Kalman filters results in a method that stands up to the rigors of industrial environments, outperforming traditional methods in accuracy and robustness. Finally, Chapter 5 applies the image processing methods developed in the thesis to tackle industrial challenges. These include addressing the fuzziness of the interface, sight glass stains, steam obstructions, camera view obstructions, and lighting variations. The techniques presented in this thesis signify a substantial improvement over current practices, promising enhanced control and recovery rates in oil sands bitumen extraction processes.