在激光-电弧复合焊接试验中,利用神经网络来模拟预测焊缝形貌时,由于众多的焊接工艺参数以及模型输出参数使得预测结果和试验真实形貌误差较大,而且模型对焊接参数变化不灵敏。因此,引入多种群遗传算法对神经网络进行初始权值和阈值逐代优化,可以避免未成熟的收敛问题,使得优化后的模型对焊缝形貌预测具有较高精度。同时,将模型的输入参数和输出参数转换,得到利用熔深、熔宽、余高三个形貌尺寸参数来预测焊接工艺参数的网络模型。综合上述模型得到一套由理想熔深、熔宽、余高的焊缝形貌尺寸对应的焊接工艺参数,进而模拟出具体的焊缝形貌曲线图的系统。通过优化模型预测的形貌和实际形貌对比,预测误差在5%以内,参数转化后,优化的预测模型误差在3%左右。结合两个预测模型的预测系统,最后得到的预测形貌和实际形貌误差在10%左右。此系统可以避免大量的试验,大大缩短复合焊接研究周期,对焊接工艺参数优化具有重要意义。

Due to the multiple technical parameters of welding and the outputs of neural networks ,when only using neural net-work the prediction of welding bead shape can’t well fit the real one. Also the neural model is insensitive to the change of welding parameters. Considering that, we introduce the genetic algorithm to optimize the neural networks’ weights and threshold gradually .With the optimized model ,the outcomes of the calculation of welding bead shape are pretty good. At the same time , we exchange the inputs and outputs to use the main dimensional parameters of welding bead(welding depth, welding width, welding reinforcement) to calculate the welding parameters. Then synthesizing the two models above ,we can get a system which can obtain the welding parameters according to the given welding depth, welding width and welding reinforcement. Then finally simulate the concrete welding bead curve. The err about the first predicting neural model is about 5%, and that for the transformed neural model is about 3%. The final integrated system’s err is about 10%. This system can reduce the times of experiments greatly and shorten the cycle of the hybrid welding research which will have a great influence on the optimization of technical parameters of welding.