Artificial neural networks - an aid to welding induced ship plate distortion?

A preliminary study on the potential application of artificial neural networks in welded structures was expanded to metal inert gas welding of steel plates of grades D and DH 36. The main controllable variables were plate thickness, steel grade, plate cutting process, and heat input. A series of welded plates of each grade was manufactured, covering plate thicknesses of 6 and 8 mm. The topography of each welded plate was evaluated after tacking the plates together and after welding, allowing the actual distortion to be calculated. It was established that a multilayer perceptron network architecture configuration accurately represented the distortion for the 6 mm thickness plate, and for the 8 mm thickness plate after treatment of the data. The data generated were used to develop the PREDICTOR software package, which allows a distortion prediction to be produced, and to carry out a sensitivity analysis. Heat input was found to be the most sensitive factor related to distortion, with carbon content of the plates, yield/tensile strength ratio, carbon equivalent, and steel grade also having significant effects. Some test plates were modelled using finite element method software packages: the initially poor agreement was improved via the addition of significant detail, but the finite element model by its nature will normally predict symmetrical distortion from a symmetric weld, whereas the artificial neural network model developed was capable of predicting the asymmetric distortion observed in reality.