Automated visual inspection of friction stir welds: a deep learning approach

Friction stir welding is a solid-state welding process. The technology is used in high-precision applications such as aerospace. Thus, monitoring the weld quality is highly relevant for detecting inaccurate welds. Various studies have shown a significant dependence of the weld quality on the welding speed and the rotational speed of the tool. Frequently, an unsuitable setting of these parameters can be detected by visually examining the resulting surface defects, such as increased flash formation or surface galling. The visual inspection for these defects is often conducted by humans and is therefore associated with increased costs and personnel allocation. In this work, a deep learning approach to automatically detect irregularities on the weld surface is introduced. A total of 112 welds with a total length of 18.4 metres were made to train and test of the artificial neural networks. Colour images of the welds were made using a digital camera, while images of the weld surface topography were made with a three-dimensional profilometer. The approach consisted of a two-step procedure. First, an object detector using a neural network localised the friction stir weld on the image. Second, a neural network classified the surface properties of the weld seam. The object detector localised the friction stir welds with an Intersection over Union up to 89.5%. The best result in classifying the surface properties was achieved by using the topography images. Here, a classification accuracy of 92.1% was reached by the DenseNet-121 convolutional neural network. The results will form the basis for the future development of a parameter optimization method for friction stir welding.