Prediction and optimisation of weld features using computational models is becoming increasingly important for industries as it can save a significant amount of time and money involved in unnecessary testing. However, the large number of input and output parameters involved in the welding process makes it difficult to predict the properties of the weldments before they are physically manufactured. Tungsten Inert Gas (TIG) welding process is commonly used in the high value manufacturing (HVM) sector such as the aerospace, nuclear, oil and gas, etc. The expensive materials used in these sectors limit the number of trials and errors that can be done before obtaining the optimum parameters for welding. Computational models that can predict all the properties of the weldments, including the geometrical, microstructural and the mechanical features, and optimise the input parameters to obtain the desired properties, even before the welding commences will have a significant benefit in terms of the cost and time involved in welding components. Inthepresentstudy,thinsheetsof304LstainlesssteelwereweldedusingtheTIGwelding process. With a focus on aerospace industry, the thickness of the sheets considered were 0.7 mm, 0.9 mm, 1.5 mm, 2.0 mm and 2.4 mm. Sheets thinner than 1.5 mm were welded autogenously,whilethosethickerthanandincluding1.5mmwereweldedheterogeneously. Several tests were performed on the welds to obtain nine geometrical features, two microstructural features and five mechanical properties (including the fatigue parameters such as the Kt and Kf), which can then be used for the development of the desired computational models. During testing, it was observed that the conventional Ferrite Number (FN) scale used for the measurement of retainedδ-ferrite in the welds is not applicable to thin sheet welds and gave trends that contradict the literature when plotted against the cooling rate. A new scale – Ferrite Density Number (FDN) – is proposed to replace the conventionally used FN scale. The development of artificial intelligence has provided several algorithms that can be used for the establishment of computational models. Among the most widely used models are the artificial neural networks (ANNs). Although these can be effectively used for the prediction of a few features of the welds, it was found that other algorithms such as the support vector machines (SVMs) are required to support the ANNs for prediction of such a large number (16 in this case) of features. Consequently, combining several algorithms for the prediction was found to significantly improve the accuracy of the models. Similarly, the optimisation of the process parameters to obtain the desired properties of the welds is challenging since the use of existing evolutionary algorithms such as the genetic algorithm (GA), particle swarm optimisation (PSO), simulated annealing (SA) either leads to a large error or requires excessively large computation effort making their application impractical. A robust algorithm combing the global search operator from the above algorithms and a local search algorithm using the Nelder-Mead simplex search was developed that reduced the computation effort required for optimisation by a factor of 100. All the models were validated through experiments for the 16 features mentioned above.
|Date of Award||2020|
|Supervisor||Steve Jones (Supervisor)|