Modelling the shape of electron beam welding joints by neural networks

This article discusses the experimental results from multi-pool electron beam welding, with dynamic positioning of the electron beam (beam splitting) [1], resulting in the formation of two consecutive welding pools. The 12Cr18Ni10Ti stainless steel samples are welded with a change in the process parameters: the distance between the two electron beams (electron beam positions) and the ratio between the two mean electron beam powers, the frequency of the deflection signal, the beam current and the welding velocity. The focusing current is kept at a constant value. The weld cross-sections, experimentally obtained at different process parameters, are used to train, validate and test neural models. The accuracy of prediction of the shapes of the welds (the form of the molten pool) is discussed and compared with that of an estimated regression model.