University of Southampton
Neurofuzzy and supanova modelling of the processing-property relationships of aerospace al-alloys
Abstract
dc:description.abstract<p>Neurofuzzy networks and the SUPANOVA technique comprise two parsimonious adaptive modelling approaches, the former combining well established neural-type learning algorithms with the transparent knowledge representation of fuzzy systems, the latter emerging from recent advances in statistical learning theory and support vector methods for regression.</p><p>This thesis has investigated the performance of such techniques in modelling physical and tensile properties of Al-Mg-Li powder metallurgy and wrought Al-Zn-Mg-Cu alloy systems, from compositional and processing information. Prior system knowledge was employed in the form of physically motivated transformation and initialising (neurofuzzy) network structures.</p><p>By adapting their structure to infer the nature of the processing-property relationships contained in the data, both adaptive methods determined a number of non-linear dependencies, resulting in more appropriate models compared with multiple linear regression analyses. The data sets were seen to be representative of experimental, small and large scale processing conditions, which then reflected the different predictive performances exhibited by the adaptive methods.</p><p>Metallurgical understanding, conditioning and regression diagnostics, allowed a greater understanding of the statistical properties of the data, permitting an enhanced interpretation and validation of the models identified by the adaptive methods, understanding the empirical results in light of the representativeness of the set of input variables, sample sizes and data weaknesses characterising the data. Generally, the dependencies inferred by the adaptive methods were seen to be consistent with metallurgical understanding, a number of which suggested interesting interdependencies, particularly an interaction between the Magnesium content of Al-Zn-Mg-Cu alloys and the age-hardening behaviour. Overall, the results gave a clear indication of the benefits associated by performing statistical analyses on experimentally designed data sets and highlighted the problems of modelling observational data.</p>
Degree
thesis:*- Name dc:type.qualificationname
- Ph.D.
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- University of Southampton
- Year dc:date.issued
- 2000
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Femminella, Oliver Paul
Rights
- Language dc:language
- English, eng