A study of neural network applications to aluminium manufacturing
thesis
posted on 2023-05-27, 00:26authored byFrost, FF
Process behaviour in the aluminium smelting industry is typically highly dynamic and unstable and involves non-linear, highly dimensional relationships among process parameters. Further, with the presence of noise associated with most of the measured parameters of the aluminium production technique, process modelling in the aluminium industry is often a complex task. However, the advancement of both knowledge and technique has resulted in significant changes to industrial processing techniques and process control methodologies. One such advancement is the development of artificial neural networks, which are a well-suited computational paradigm for use in monitoring and controlling complex dynamic processes. Neural networks offer a powerful mathematical technique for modelling, control and optimisation of dynamic processes that are developed using process data, without the need for a priori knowledge or understanding the associated scientific principles and underlying relationships among process parameters. Generally, when a neural network is initially trained for a particular task, some of the features of the training data will have no significant effect on the networks decision, while other features will be critical. In addition, there exist many networks for a particular task that may perform similarly, however, they may use different features of the training data to make their decision. This work presents an evaluation and empirical performance comparison of various neural networks in an important and actual application domain. Such studies are valuable to understanding the strengths and weaknesses of various problem solving models as well as the characteristics of various application domains. As neural networks are an advanced control technique that are often used as an opportunity to maximise corporate revenue, it becomes necessary to develop a set of selection criteria for selecting a particular neural network that produces optimum performance when applied to a specific application. Neural network selection can be completed based on economic considerations, such as cost associated with neural network accuracy, cost associated with measuring process parameters used as input variables in the model and cost associated with neural network computation time. In this work, evaluation of neural networks for three industrial applications, involving process modelling of reduction cells for aluminium production at Comalco Aluminium (Bell Bay) Limited, or CABBL, is completed. The performance of six distinct models of the neural network paradigm is assessed using specific assessment criteria. The decision of which neural network model is most suitable for a specific application is complex, requiring quantitative decision logic, particularly as the assessment criteria are not fundamentally of equal significance. e. It is shown that optimisation techniques are necessary to select an optimum neural network model for a specific application. While it is noted that available operations research techniques are capable of neural network optimisation and selection, such optimisation techniques are inappropriate for application in this instance. This work reports a systematic technique that optimises a neural network efficiently on command using precise mathematical models. It is shown in this work that the influence of each input parameter on prediction error is analysed to determine an optimum neural network model for each studied application. Moreover, while a feasible solution for the neural network model is identified in the first instance, an optimal solution is subsequently obtained and implemented to achieve maximum economic benefit. It is noted that the developed optimisation strategy is a unique and novel methodology for neural network optimisation and selection and has been carefully developed to facilitate its ease of application in industry.
History
Publication status
Unpublished
Rights statement
Copyright 2000 the author - The University is continuing to endeavour to trace the copyright owner(s) and in the meantime this item has been reproduced here in good faith. We would be pleased to hear from the copyright owner(s) This thesis has been made openly accessible except for the for zipped folder of publications which cannot be made available for copyright and proprietary reasons.