Dual Neural Extended Kalman Filtering Approach for Multirate Sensor Data Fusion with Industrial Applications

  • Author / Creator
    Wang, Jingyi
  • The Kalman filter algorithm and its variants have been widely applied to the multisensor data fusion problems to provide joint state estimation, which is more accurate than estimations from individual sensors. The performance of the Kalman filter based fusion relies on the accuracy of the models as well as process noise statistics. Deviations from correct system models and violations of noise assumptions may lead to unsatisfied sensor fusion results and even divergence.

    Two types of measurements are typically utilized to estimate process quality variables. One is frequent measurements, which are available at a fast and regular sampling rate but suffer from lower accuracy and higher measurement noises. The other type is infrequent measurements that are available at a slower sampling rate. The infrequent measurements, such as lab analysis results, have less availability but higher accuracy and are usually used as references to improve state estimation. The objective of this thesis is to develop new multirate sensor data fusion algorithms that can compensate for model inaccuracies and violations of noise assumption to improve the online sensor fusion performance. To fulfill this objective, a dual neural extended Kalman filter (DNEKF) algorithm is proposed by employing two neural networks to improve state estimation and output predictions. Using both frequent and infrequent measurements enables the DNEKF to provide more reliable training for the neural networks and hence to provide more robust and reliable sensor fusion results.

    Additionally, infrequent measurements are usually subject to irregular sampling rate and time-varying time delays. To address these problems while preserving the estimation accuracy, a fusion method that fuses frequent DNEKF estimates with infrequent estimates from the state model compensation NEKF (SNEKF) is proposed. In this approach, frequent and infrequent estimates are fused in the fusion center when the delayed infrequent measurements arrive. The weights and biases of the state model compensation neural network (SNN) are shared between the two synchronized estimation processes.

    In the primary separation cell (PSC) used for oil sands bitumen extraction, the interface level estimation is based on various sensors. Image processing based computer vision system, which uses a camera to capture sight glass vision frames, is considered to be the most accurate among these sensors. Although the accuracy of computer vision interface level estimation is high, its qualities are influenced by abnormalities, such as vision blocking, stains, and level transition between sight glasses. Under such abnormal scenarios, a sensor fusion strategy, which adaptively updates the fusion parameters, is proposed and integrated with the image processing based computer vision system.
    The performance of the proposed fault-tolerant multirate sensor fusion algorithms is demonstrated using numerical examples and case studies with industrial process data. The factory acceptance test (FAT) was conducted for the sensor fusion and computer vision integrated system in the computer process control (CPC) industrial research chair (IRC) lab under industrial environmental conditions and it demonstrated the improved estimation accuracy under various process abnormalities.

  • Subjects / Keywords
  • Graduation date
    Fall 2020
  • Type of Item
  • Degree
    Master of Science
  • DOI
  • License
    Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.