Show simple item record

dc.contributor.authorSaptoro, Agus
dc.contributor.supervisorProf. Moses Oludayo Tadé
dc.contributor.supervisorAssoc. Prof. Hari Babu Vuthaluru

ANN technology exploded into the world of process modelling and control in the late 1980’s. The technology shows great promise and is seen as a technology that could provide models for most systems without the need to understand the fundamental behaviour or relationships among the process variables. Today, ANN applications have been applied successfully in a number of areas of process modelling and control, with the best-established applications being in the area of inferential measurements or soft sensors.Unfortunately, ‘the free lunch did not have much meat’. Overtime, people focused more on the true capabilities and power of ANN, the ability to model nonlinear relationships in data without having to define the form of the nonlinearity. However, there is often a tendency to merely plug in the data, turn the ANN training software on, and blindly accept the results. This is probably inevitable since, to date, there are no textbooks or scientific journal papers providing an integrated and systematic approach for ANN model development addressing pre-modelling, training and postmodelling stages. Therefore, addressing issues in those three phases of ANN model development is essential to support and to improve further applications of ANN technology in the area of process modelling and control.The model development issues in pre-modelling and training phases were addressed by reviewing current practice and existing techniques. For each issue, a novel method was proposed to improve the performance of ANN models. The new approaches were tested in a variety of benchmarking studies using artificial samples and coal property datasets from power station boilers.The research work in the post-modelling stage analysis which emphasises on taking the lid off black box model, proposes a novel technique to extract knowledge from the models and simultaneously obtain better understanding of the process. Postmodelling phase issues were addressed thoroughly including construction of prediction limit, sensitivity analysis and development of mathematical representation of the trained ANN model.Confidence intervals of the ANN models were analysed to construct the prediction boundary of the model. This analysis provides useful information related to interpolation and extrapolation of the model. It also highlighted how good the ANN models can be used for extrapolation purposes.An effort based on sensitivity analysis of hidden layers is also proposed to understand the behaviours of the ANN models. Using this technique, knowledge and information are retrieved from the developed models. A comparative study of the proposed techniques and the current practice was also presented.The last topic addressed in this thesis is knowledge extraction of ANN models using mathematical analysis of the hidden layers. The proposed analysis is applied in order to open the black box of the ANN models and is implemented to simulated and real historical plant data so that useful information from those data and better understanding of the process are obtained.All in all, efforts have been made in this thesis to minimise the use of abstract mathematical language and in some cases, simplify the language so that ANN modelling theory can be understood by a wider range of audience, especially the new practitioners in ANN based modelling and control. It is hoped that the insight provided in the dissertation will provide an integrated approach to pre-modelling, training and post-modelling stages of ANN models. This ‘new guideline’ of ANN model development is unique and beneficial, providing a systematic framework for the preparation, design, evaluation and implementation of ANN models in process modelling and control in particular and prediction / forecasting tool in general.

dc.publisherCurtin University
dc.subjectartificial neural networks
dc.subjectan integrated approach
dc.subjectopening the black box
dc.titleAn integrated approach to artificial neural network based process modelling
curtin.departmentDepartment of Chemical Engineering
curtin.accessStatusOpen access

Files in this item


This item appears in the following Collection(s)

Show simple item record