ESTIMATING PARTICLE SIZE FRACTION PROPORTIONS – A MULTISCALE APPROACH USING MULTIPLE KERNEL LEARNING

摘要

Machine vision systems are increasingly being used in many monitoring and control tasks in thernmineral processing industry. In general, the necessary baseline information is obtained by processingrnimages acquired from process streams to provide textural characterisations that can be mapped tornoperating conditions or plant status. Different textural characterisation methods have been proposedrnthat emphasise various aspects of texture and include statistical, structural, and spectral methodsrnamong other. Each method typically has a number of parameters that may need optimal tuning,rnmaking the design of a fully automated control system a challenge. In an automated control system,rnthe ideal is for underlying algorithms to identify and use relevant features with minimal operatorrnintervention. Moreover, it may be necessary to combine image data with other heterogeneous datarnsources that are monitored.rnThis paper investigates a framework for combining different image data representations or viewsrnusing multiple kernel learning (MKL) in the monitoring and control of particle size distributionsrnof ore and rock material. Multiple kernel learning has recently been proposed in the context ofrnstatistical learning with kernel methods. In these methods, the choice of the kernel function isrncrucial in the learning algorithm and encodes prior knowledge in terms of similarity between pairsrnof observations. In MKL the goal is to learn an optimal kernel from the data. Such an approachrnfacilitates automatic model selection, integration of different data sources, variable selection, as wellrnas model interpretability, a property key in autonomous control. A learning framework for estimatingrnproportions of size fractions in feed material based on multiscale representation of image data isrnproposed. Using image data from coal particles, the proposed approach is shown to give improvedrnperformance in comparison to conventional approaches.