(699b) Multivariate Image Analysis and Gaussian Process Regression Model Based Soft Sensor for Nickel Pellet Size Distribution Prediction

Authors: 
Chen, J., McMaster University
Yu, J., McMaster University
Zhang, Y., Vale



Particle size distribution is one of the key quality variables in material processing because the accurate measurement of size distribution is essential for designing advanced process control and real-time optimization strategies towards the best product quality and energy efficiency. Thus, the precise estimation and prediction of particle sizes are advantageous for improving product quality, yield and efficiency. Mechanical sieving is a traditional way to measure particle size distribution but cannot be implemented in an on-line fashion with high precision. Alternately, multivariate image analysis techniques provide quick and non-intrusive solutions to measure and estimate particle size distribution in a fast real-time manner. However, conventional image analysis based particle sizing methods are unable to deal with particle overlapping effect, which may result in inaccurate measurements of size distributions.

In this study, two novel video analysis based pellet sizing methods are developed for the online estimation and prediction of particle size distribution without any intrusive tests. These two approaches make use of the video frames and can largely eliminate the particle overlapping effect that cannot be readily conquered by the image analysis based pellet sizing methods. The videos of the free falling nickel pellets are first taken and the free falling tracks in the video frames are then utilized for estimating the diameters of different pellets. In the first proposed method, particle edge detection is designed to capture the boundary of the free falling tracks so as to identify the diameters of the free falling particles. In contrast, the second method is developed in the way of scanning the filtered gray-scale images row by row in order to obtain the gray-scale curves corresponding to different nickel pellets. Then the novel Gaussian process regression (GPR) models are built to decompose the multi-peak curves into various sub-curves and estimate the diameters of the corresponding pellets including the overlapped ones along the horizontal direction. Further a counting rule is established to estimate the pellet size distribution by successfully eliminating the pellet overlapping effect along the vertical direction.

The utility of the two video analysis based pellet sizing methods is examined and compared through the application to the real lab-scale video clips that capture the free falling movements of nickel pellets within the decomposer. With the accurate prediction of pellet size distribution, the amount of seed addition into pellet decomposers can be modeled and further controlled in an optimal way so that the best product quality in terms of pellet distribution can be achieved. Further, it becomes possible to dynamically optimize the pellet circulation throughout decomposer beds, which in turn results in not only fewer process interruptions but also improved unit availability. The computational results show that the pellet size distribution can be more accurately predicted from the second video analysis method due to the fact that this approach can better handle the overlapping effect of nickel pellets along both vertical and horizontal directions.