Use of Decision Trees for the Development of Decision Support Systems for the Control of Grinding Circuits

Grinding circuits can exhibit strong nonlinear behaviour, which may make automatic supervisory control difficult and, as a result, operators still play an important role in the control of many of these circuits. Since the experience among operators may be highly variable, control of grinding circuits may not be optimal and could benefit from automated decision support. This could be based on heuristics from process experts, but increasingly could also be derived from plant data. In this paper, the latter approach, based on the use of decision trees to develop rule-based decision support systems, is considered. The focus is on compact, easy to understand rules that are well supported by the data. The approach is demonstrated by means of an industrial case study. In the case study, the decision trees were not only able to capture operational heuristics in a compact intelligible format, but were also able to identify the most influential variables as reliably as more sophisticated models, such as random forests.

Application of control units VAZM-2U in ore grinding

This paper considers the results of research work aimed at developing control systems to control ore grinding processes that would be compatible with the control unit VAZM-2U developed by Soyuztsvetmetavtomatika. The underlying principle concerning the unit is that grinding of ores with different mineralogical compositions is governed by the same common regularities in correlation between the physical processes that develop in grinding circuits and the defining process parameters. A grinding mill is fed with ore that has varying physical and mechanical properties, and this can lead to accumulation of material in the mill. Indicators of the probable overload condition include mill vibration level and active power draw of the mill drive motor. The point at which the overload condition has arrived is determined by analyzing active power draw and reverse vibration trends. It is demonstrated that a mill overload condition may take place in those time intervals when both the vibration level and the active power draw of the mill motor fall. In this case the VAZM-2U unit calculates a correction command for the ore flow rate regulator, and this way the overload condition is overcome while the ore feed rate returns to the initial value. The VAZM-2U unit can also help reach the maximum output of the overflow product from a spiral classifier avoiding overgrinding, with the finest material being monitored. The unit can also determine the underflow flow rate in the spiral classifier while adjusting this parameter within a given range of allowable values. The underflow flow rate is estimated with the help of an adaptive mathematical model, which can be utilized in closed-loop grinding circuits that include classifiers. The ore grinding control algorithms implemented in the VAZM-2U unit can be modified to be applicable for milling and flotation control.

Dynamic Monitoring of Grinding Circuits by Use of Global Recurrence Plots and Convolutional Neural Networks

Reliable control of grinding circuits is critical to more efficient operation of concentrator plants. In many cases, operators still play a key role in the supervisory control of grinding circuits but are not always able to act timely to deal with disturbances, such as changes in the mill feed. Reliable process monitoring can play a major role in assisting operators to take more timely and reliable action. These monitoring systems need to be able to deal with what could be complex nonlinear dynamic behavior of comminution circuits. To this end, a dynamic process monitoring approach is proposed based on the use of convolutional neural networks. To take advantage of the availability of pretrained neural networks, the grinding circuit variables are treated as time series which can be converted into images. Features extracted from these networks are subsequently analyzed in a multivariate process monitoring framework with an underlying principal component model. Two variants of the approach based on convolutional neural networks are compared with dynamic principal component analysis on a simulated and real-world case studies. In the first variant, the pretrained neural network is used as a feature extractor without any further training. In the second variant, features are extracted following further training of the network in a synthetic binary classification problem designed to enhance the extracted features. The second approach yielded nominally better results than what could be obtained with dynamic principal component analysis and the approach using features extracted by transfer learning.