In-process Monitoring and Data Analysis for Quality Control of Friction Stir Weldin

The objective of this paper is to develop a method to predict the formation of discontinuity and its location and size during friction stir welding of aluminum alloys. The advantages of friction stir welding are significant, including the superior welding quality, energy savings, and inherent cost-effectiveness as compared to other traditional welding techniques. However, for some application where high-reliability is required, the need for significant weld inspection can increase the total cost significantly. A new approach to weld inspection, where the location and size of voids can be obtained by a prediction model, can reduce the cost drastically. In this paper, a supervised machine learning technique was employed to derive the prediction model. Resultant forces were measured and analyzed via wavelet transform, while the void location and size were measured through CT scan and image-processing of the data. The supervised machine learning algorithm via an artificial neural network trained the model using these two sets of data. The algorithms adaptively increase the accuracy of their prediction as the number of weld samples available for learning increase. The result shows that the trained model accurately predicted the void location and size. Accordingly, this proposed approach will significantly reduce the need for costly post-process inspection by identifying locations in the weld where a non-tolerable size defect has been predicted by the prediction model.