E-Thesis 38 views 13 downloads
The application of neural network and statistical modelling to the scale-up and optimisation of a new adhesive tape manufacturing process. / J. D Grant-Abban
Swansea University Author: J. D, Grant-Abban
PDF | E-ThesisDownload (12.77MB)
This thesis is concerned with the successful application of statistical and neural modelling techniques to the efficient scale-up and optimisation of a new pressure- sensitive tape manufacturing facility. This thesis describes the generation of modelling data, the use of back propagation neural netw...
|Degree level:||Master of Philosophy|
No Tags, Be the first to tag this record!
This thesis is concerned with the successful application of statistical and neural modelling techniques to the efficient scale-up and optimisation of a new pressure- sensitive tape manufacturing facility. This thesis describes the generation of modelling data, the use of back propagation neural networks to model properties, and the optimisation of the process using the neural models. Modelling data was purposely generated in a series of trials, using the structure of Design of Experiments to cover the process envelope systematically and efficiently. A neural model, using data from pilot-scale process experiments, was used to influence the design of the full-scale processes. Once the full-scale processes were established, a more detailed 34 factor neural model was developed. This second neural model was validated and used to optimise and explore the full-scale process. The validation exercise detected anomalous data using statistical analysis. The anomalous data was subsequently quarantined and prevented from confusing the optimisation effort. The detailed understanding of the process came from running DoE's on the virtual process, represented by the second neural model. These DoE's provided useful contour plots and response surface visualisations of the 'black box' neural network model, giving insights that guided optimisation and indicated the potential for tape with significantly enhanced product properties.
College of Engineering