论文部分内容阅读
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。
In the laser-arc hybrid welding experiment, when using neural network to simulate the prediction of weld morphology, due to the large number of welding process parameters and model output parameters, the prediction results and the experimental real morphology error are large, and the model does not change the welding parameters Sensitive. Therefore, the introduction of multi-population genetic algorithm for neural network initial weight and threshold optimization, to avoid premature convergence problem, making the optimized model for the prediction of weld morphology with high accuracy. At the same time, the input parameters and the output parameters of the model are converted, and the network model of the welding process parameters is obtained by using three dimensional parameters of penetration depth, weld width and residual height. Based on the above model, a set of welding process parameters corresponding to the ideal welding depth, width and height of the weld is obtained, and then the system of the specific welding profile is simulated. The prediction error is less than 5% by comparing the morphology and the actual morphology predicted by the optimization model. After the parameters are transformed, the error of the prediction model is about 3%. Combined with the prediction system of two prediction models, the final prediction morphology and the actual shape error are about 10%. This system can avoid a large number of experiments, greatly shortening the research cycle of composite welding, and is of great significance to the optimization of welding process parameters.