In the mineral processing industry, flotation constitutes the most important technique for extraction of the valuable minerals from the ore. This process involves both chemical and physical parts, the theory of which are not yet completely understood. Recently, extensive research has been performed to increase the efficiency of the flotation stage, bearing in mind the potential for increased profit this would imply for the plant. Typically, a flotation control system is built in a hierarchical manner, where a supervisory system is generating set points to local basic loop controllers. It is then assumed that these local control loops are working perfectly well, an assumption which is very often violated. The reason for this can to a great extent be found in the complex nature of the flotation process with strong nonlinearities, strong interactions between process states, mechanical problems etc. The commonly used control strategy, with distributed PI or PID controllers, generally performs badly under such circumstances, regardless of how they are tuned. In this thesis, the control of pulp levels in a flotation bank consisting of cascade coupled flotation cells is considered. The existing PI-based SISO strategy is replaced by multivariable controllers showing a significant increase in performance. A related problem is the functional status of the control valves between the tanks, which strongly affects the control performance actually achieved. Therefore, as a second part of this project, a method to supervise a linearized valve model is proposed. Control performance, in terms of the deviation of the estimated parameters from the nominal ones, is discussed for one of the proposed multivariable control strategies. The other subject discussed in the thesis is the area of recursive estimation. A new method to avoid windup in case of poor process excitation is suggested. The proposed algorithm is computationally simple and, in comparison to many other algorithms proposed in the literature, intuitive to tune since one directly decides the accuracy of the parameter estimates. The last property is highly desirable in \eg fault diagnosis, where one often wants to keep a constant false alarm rate (FAR).
Luleå: Luleå tekniska universitet, 2002. , 95 p.