Design of Soft-Sensors Using Cluster Techniques

Cluster analysis is used to design grade soft-sensors in a rougher flotation plant and a ball mill overload soft-sensor. Soft-sensors are virtual sensors which either replace unavailable actual sensors or infer measurements for non existing sensors. In the case of the grade soft-sensors clustering serves to determine clusters of related operating points. To each cluster there corresponds an ARMAX model structure whose parameters are estimated using only the points in the corresponding cluster. Thus a much better correlation is found than if all the data points were used for a global model. The problem of identifying clusters when the measurement of the real sensor is not available is addressed. In the case of the overload soft-sensor, variables were selected which are related to overload, either from a phenomenological point of view or according to operator experience. Different clusters are formed, some of which are identified as determining an overload condition or a possibility of such condition. Each of these clusters has a set of operating conditions which suggests the corrective action to be taken when necessary. Both kinds of soft-sensors are tested with industrial plant data and promising preliminary results are obtained.