Development of a Density Softsensor for a Mineral Grinding Plant

Abstract

This paper the design of a virtual sensor (softsensor) for the pulp density in a grinding plant is addressed. In particular the problem of the choice of the model parameters after the Sensor becomes unavailable is considered and a method for approaching this problem is proposed and tested. Freezing the values of the estimated softsensor model parameters at their value just before the sensor becomes unavailable, as has been currently done, is shown to be just a special case of a more general way to proceed in which the predicted parameters evolve from their values at the instant of failure to their mean values. Actual grinding plant data is used to test the method. An assumption of stationarity and ergodicity is required, atleast during a long enough time period to determine statistics of the estimated parameters with an acceptable variance using time averaging, before the sensor becomes unavaiable. A method for assessing the validity of the softsensor is proposed at the time when it is to replace the actual sensor based on the above assumption. Results supporting the developments are obtained when there is enough time to determine the statistical values involved, but inconclusive results appear when this is not the case, as should be expected.