Modeling the depth of penetration in laser welded Al 6061 using statistical and neural network approaches

This research attempted to evaluate the capibilities of modeling the depth of penetration in laser welded A1 6061 using both traditional statistical modeling tools and artificial neural networks. Aluminum 6061 was welded with a pulsed 400W Nd:YAG laser. The experiment consisted of measuring depth of penetration while three variable parameteres were varied: pulse energy, travel speed, and material thickness. The statistical modeling tools and artificial neural network tools attempted to model the depth of penetration behavior based on the experimental data. The statistical modeling tools examined were linear regression, nonlinear regression and local smoothing regression. The linear regression tool was mathematically the simplest modeling tool and the local smoothing regression tool was the most complicated, capable of modeling a complex surface. A number of feedforward backpropagation paradigm artificial neural networks were examined. The simplest architecture had no hidden layers and the most complex architecture had two hidden layers with 10 nodes in each hidden layer. It was concluded that both statistical modeling and artificial neural network tools could adequately model the depth of penetration in A1 6061. These conclusions were based on comparing the modeling surfaces wdth the experimental data surface and by the average RMS error between the models and the data. Among the three statistical modeling tools, only the local smoothing technique adequately modeled the data, while all the artificial neural networks with more than one hidden layer and 4 hidden nodes modeled the data well. However, the most complicated models appeared to have poor generalization capabilities. It was apparent that for extremely complicated surfaces an artificial neural network with many hidden layers and nodes could outperform even a local smoothing regression technique.