| Raising costs have shifted the emphasis of industrial plant operating polices from that of processing as much feed as possible to one of minimizing energy consumed per unit weight of metal produced. With the vast tonnages of raw grade materials being refined, comminution circuits for improving the control of this operation become attractive. The complexity of the grinding mill circuits results from feed disturbances such as frequently changing ore hardness and feed ore size distribution. The objective of improved control strategies is the optimal operation of mineral plants that reduces energy consumption, enhances throughput, and optimizes specific output.; In spite of advanced control technologies, there have been practical difficulties in applying these technologies to industrial processes because of the lack of adequate theoretical models. In the industrial process, the major disturbances in grinding mill circuits are variations in feed ore hardness, variations in feed ore size distribution, and variations in tonnage to be processed. The process is more susceptible to variations of feed ore hardness the other two.; The initial work shows the use of artificial neural networks for industrial mineral processing applications. The applications studied are industrial data and inferential model based grinding mill circuits. The focus then shifts to dynamic grinding mill circuits with optimal control (state space based control with LQG), model predictive control, and artificial neural network control. They are to demonstrate the breadth of applicability of the proposed system identification technique. A new method of control with artificial neural networks is related to incorporate process knowledge to the recurrent and back-propagation models.; Artificial neural networks are investigated in this research as a means of filling the need for new, accurate, multivariable, and nonlinear models. A novel method for a general nonlinear model predictive control scheme is laid out. The control strategies are in three areas. First, a dynamic model and on-line measurements are used to build a prediction of future output behavior expressed in terms of current and future manipulated input changes. Second, optimization is performed to find a sequence of input changes that minimizes a chosen measure of the output derivation from their respective plant values while satisfying all the given constraints. Third, the quality of prediction may improve as more measurements are collected, only the first of the calculated input sequence is implemented and the whole optimization is repeated at the next sampling time. Another important contribution is the incorporation of population balance model.; Finally, the new method of control with artificial neural networks where simplicity and speed are two attractive features of the control scheme, make it a promising method for solving mineral industrial control problems. |
