| The use of artificial neural network (ANN) technology for predicting the heat-affected zone (HAZ) hardness and fusion zone (FL) cross-sectional weld shape is presented in this thesis. In particular, the backpropagation network (BPN) method is applied to these problems. Two learning enhancements for BPN termed modified dynamic and step-declining are proposed and tested together with the traditional methods available for this purpose. Finally, the BPN method is adapted to solve the inverse weld shape problem of estimating welding conditions to provide a given weld shape.;An extensive data base of measured bead-on-plate HAZ hardness values was assembled from the literature for "training" and testing. The data is primarily associated with the submerged arc (SA) welding process because among other things, it is widely recognized in the industry to be the best controlled and most reliable welding technique. Nonetheless, the database also includes hardness measurements from research investigations where gas metal arc welding (GMAW) and the shielded metal arc (SMA) process were used.;To predict HAZ hardness from the input welding conditions (voltage, current, wire travel speed, plate temperature and plate thickness) and steel chemical composition, the calculation of an intermediate cooling time is necessary. A BPN is presented for this purpose. It is then combined with the hardness network to generate hardness values directly from the input welding conditions and steel chemical composition.;The data for the weld shape problem was kindly provided by Mr. Jack Pacey from the measurements of an investigation in the Northern College (Kirkland Lake, Ontario). Various critical dimensions (bead width, height, penetration, "bay length", deposit area and fusion area) were extracted from the traced weld shapes. BPNs were constructed and trained to predict these parameters from the welding conditions (current, voltage, wire travel speed and plate thickness). From these, a shape is approximated by fitting a semi-ellipse (area ;The advantages of the ANN technology versus traditional regression methods used for these problems are discussed. In general, ANN provides practitioners with a more flexible, more reliable tool than regression methods. It can be extended to include welding experience as it is gained. In addition, various assumptions that are necessary when physical modelling is used are implicit in the ANN method. |
