Usage
  • 48 views
  • 36 downloads

Machine Learning for Robust Tracking of Interface Level Inside a Primary Separation Vessel in the Presence of Occlusions and Noise

  • Author / Creator
    Amjad, Faraz
  • A Primary Separation Vessel (PSV), used in the oil sands industry, is an important process equipment, where Bitumen is separated from the oil sand using a density based separation process. The interface level between a bitumen rich layer (froth) and a layer that has moderate amounts of bitumen in it (middling), controls the efficiency of the separation process. Traditional sensors for detecting this interface level, like Differential Pressure (DP) cells, or nucleonic profilers, get easily clogged up due to the nature of the phases inside the PSV. Thus, in recent years, computer vision has been employed to track this interface level by making use of the sight glasses present on the PSV walls, which show the location of the interface level inside the tank. Most of the existing computer vision algorithms use manual feature extraction techniques, like frame differencing or edge detection, to infer the level of the interface. Due to the nature of the techniques used, the currently used algorithms are not able to deal with noise and occlusions well. In the present work, machine learning for image processing, namely Convolutional Neural Network (CNN), and its extension, Fully-Convolutional Networks (FCNs), are used for the task of tracking the interface level, with special focus on novel techniques to handle occlusions and noise. The thesis starts off with a more detailed description of the problem statement, followed by some basic introduction to computer vision and image processing in chapter 2. In chapter 3, an algorithm utilizing CNNs and state estimation through a Kalman filter, is proposed. A dynamic model of the PSV tank, obtained through the techniques of process identification is used to infer the level of the interface, when the image data of the PSV sight glass is not reliable (noise or obstructions). When image is reliable, CNNs give excellent performance and accuracy in tracking the interface. The inferred levels from the obtained dynamic model and the image data are combined through the Kalman filter. Recognizing that dynamic system models can sometimes be difficult to obtain for a gravity based process like the PSV, chapter 4 proposes a purely image-based novel algorithm utilizing FCNs with region growing. For accurate selection of seed points, required for region growing, a Gaussian Mixture Model (GMM) is also utilized in the proposed algorithm. This method gives similar accuracy to that of the Kalman filter based CNN, without requiring any dynamic model of the PSV, making it feasible to be easily applied on any industrial set-up of he PSV. Finally, in chapter 5, a method based on manual feature extraction and ensemble Extreme Learning Machines(ELMs) is proposed. Owing to its accurate tracking of the interface level in a scenario of no occlusions in images, the proposed approach can also be used for preliminary labelling of the images in the PSV unit for large data-sets, which is much easier and faster. Large labelled data-sets can then be used to train the more data hungry CNNs and FCNs, giving an even greater degree of accuracy.

  • Subjects / Keywords
  • Graduation date
    Spring 2021
  • Type of Item
    Thesis
  • Degree
    Master of Science
  • DOI
    https://doi.org/10.7939/r3-h5j7-js85
  • License
    This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.