An extended Kalman Filter is used to build an on-line estimator of total ore, fine ore and water hold-ups in a SAG mill. In addition, an equivalent grinding rate from coarse to fine ore, as well as the discharge rates of solids (fines) and water are also estimated as part of the adaptation procedure of the model used for the Kalman Filter. The measurements used in the Kalman Filter estimator are the total mill weight provided by load cells, and inferred discharge flows of fines and water from flow and density measurements performed downstream. The net ore and water feed to the SAG mill are used as inputs to the estimator, and also the proportion of fine ore in the feed. As an example of how to introduce robustness in the estimation in the face of sensor bias errors, an estimation of the load cell bias error is introduced in the estimator, thus providing self calibration. The estimator presented in this paper has been tested by simulation using a detailed SAG mill mechanistic dynamic model comprising 27 size fractions well fitted to a pilot mill. The maximum error in the estimation of the operational variables - total and fine ore as well as the mill contents under different operating conditions - is about ten per cent.