Modelling and control of an autogenous mill using a state space methodology and neural networks

[摘要] ENGLISH ABSTRACT:Metallurgical processes are often high dimensional and non-linear making themdifficult to understand, model and control. Whereas the human eye has extensivelybeen used in discerning temporal patterns in historical process data from theseprocesses, the systematic study of such data has only recently come to the forefront.This resulted predominantly from the inadequacy of previously used linear techniquesand the computational power required when analysing the non-linear dynamicsunderlying these systems. Furthermore, owing to the recent progress made withregard to the identification of non-linear systems and the increased availability ofcomputational power, the application of non-linear modelling techniques for thedevelopment of neural network models to be used in advanced control systems hasbecome a potential alternative to operator experience.The objective of this study was the development ofa non-linear, dynamic model of anautogenous mill for use in an advanced control system. This was accomplishedthrough system identification, modelling and prediction, and application to control.For system identification, the attractor was reconstructed based on Taken's theoremmaking use of both the Method Of Delays and singular spectrum analysis. Modellingconsisted of the development of multi-layer perceptron neural network, radial basisfunction neural network, and support vector machine models for the prediction of thepower drawn by an autogenous mill. The best model was subsequently selected andvalidated through its application to control. This was accomplished by means ofdeveloping a neurocontroller, which was tested under simulation.Initial inspection of the process data to be modelled indicated that it contained aconsiderable amount noise. However, using the method of surrogate data, it wasfound that the time series representing the power drawn by the autogenous mill clearlyexhibited deterministic character, making it suitable for predictive modelling. It wassubsequently found that, when using the data for attractor reconstruction, a connectionexisted between the embedding strategy used, the quality of the reconstructedattractor, and the quality of the resulting model. Owing to the high degree of noise inthe data it was found that the singular spectrum analysis embeddings resulted in better quality reconstructed attractors that covered a larger part of the state space whencompared to the method of delays embeddings; the data embedded using singularspectrum analysis also resulting in the development of better quality models.From a modelling perspective it was found that the multi-layer perceptron neuralnetwork models generally performed the best; a multi-layer perceptron neural networkmodel having an appropriately embedded multi-dimensional input spaceoutperforming all the other developed models with regard to free-run predictionsuccess. However, none of the non-linear models performed significantly better thanthe ARX model with regard to one-step prediction results (based on the R2 statistic);the one-step predictions having a prediction interval of 30 seconds. In general thebest model was a multi-layer perceptron neural network model having an input spaceconsisting of the FAG mill power (XI), the FAG mill load (X2), the FAG mill coarseore feed rate (X3), the FAG mill fine ore feed rate (X4), the FAG mill inlet water flowrate (X7) and the FAG mill discharge flow rates (X9, XIO).Since the accuracy of any neural network model is highly dependent on its trainingdata, a process model diagnostic system was developed to accompany the processmodel. Linear principal component analysis was used for this purposes and theresulting diagnostic system was successfully used for data validation. One of themodels developed during this research was also successfully used for the developmentof a neurocontroller, proving its possible use in an advanced control system.

[发布机构] Stellenbosch University