Particle size measurement, specially in product streams, is fundamental to control of a grinding circuit. In this paper, the design of a soft-sensor in an actual grinding plant is addressed, with the purpose of increasing the availability of such measurement. The system is designed to substitute the momentary unavailability of the real sensor, either because it has failed or it is in calibration or because it is time shared by several parallel grinding lines. An ARMAX model structure is determined using the stepwise regression method. Starting with a list of the presumably correlated candidate variables, the method selects, one by one and in a systematic manner, the variables that best model the measurement to be replaced. Not only the single measurements are included in the list of candidates, but also combinations of such measurements having physical significance. A phenomenological model has been built with the purpose of determining these combinations. It turns out that, in most of the cases, the components of combined measurements were selected by the method, instead of the single measurements. Actual grinding plant data is used to determine the model structure, to estimate the model parameters and to test the predictive capability of the soft-sensors developed. Good results are obtained during periods of up to 24 h of the particle size soft-sensor operation.