Online evaluation of weld quality for Friction Stir Welding process

Mathur, Kartikey and Kennedy, Andrew (2024) Online evaluation of weld quality for Friction Stir Welding process. PhD thesis, Lancaster University.

Text (2024MathurPhD)
2024MathurPhD.pdf - Published Version
Available under License Creative Commons Attribution.
Download (10MB)

Abstract

Friction stir welding (FSW) and Friction stir processing (FSP) have been widely accepted by different manufacturing industries ranging from semiconductor to the shipbuilding industry. With advanced manufacturing capabilities, there is an increased demand to weld parts with complex geometry and curved surfaces through the FSW process. The single process parameter-based weld monitoring systems are limited by their dependence on prerequisite data and type of type of defect. This limits the applicability of the existing weld monitoring systems for different FSW platforms and setups. The existing literature highlights the potential of multi-process parameter monitoring for the detection of commonly occurring FSW defects and flaws. The presented study explores continuous acoustic emission monitoring methods for determining the weld quality for FSW joints. The novel acoustic emission (AE) based FSW monitoring system utilizes the ability of AE signals to capture the first level parameter of the FSW process yielding higher sensitivity to commonly observed FSW defects. The objective of this thesis is to develop a novel online weld monitoring system to detect the presence of commonly observed FSW defects and mitigate the need time time-consuming post weld inspections. The proposed AE-based FSW monitoring gives evidence of the detection of sub-surface voids, lack of penetration and lack of tool pressure from the tool shoulder. The defects and flaws are identified through patterns observed on the spectrograms obtained from the recorded AE signals of the FSW process. For the classification of welds according to the type of defects and flaws, convolutional neural networks (CNN) are utilized with spectrograms and mel-spectrograms as the input to the classification model. To obtain the optimized classification model, three parent CNN architectures are considered. This thesis presents a bespoke FSW CNN classification architecture that attained classification accuracy of 98% which highlights the potential of an integrated machine learning-AE online weld classification system for FSW