基于BP人工神经网络的钢轨交流闪光焊焊接接头质量预测

Quality prediction of alternating current flash butt welding of rail based on improved back propagation neural network

对刘国东等提出的BP(误差反向传播)神经网络归一化模型进行了改进,得到了适合钢轨交流闪光焊落锤质量预测的BP神经网络归一化模型。基于LabV iew开发软件编制了高速采集软件。采集了U71Mn钢轨焊接工艺正交试验的焊接电流、焊接电压和动立柱的位移,并从中提取加速烧化前一阶段的闪光率、能量输入、焊接时间和烧化量等质量特征量作为BP神经网络预测模型的输入量。建立了输入层单元数为5、隐含层单元数为14的BP神经网络焊接接头落锤质量的预测模型;以正交设计工艺试验的27个焊接接头中的17个作为训练样本,对预测模型进行训练。以余下的10个作为检验样本,采用将训练后的预测模型进行预测,预测准确率达到90%。

An improved back propagation (BP) neural networks model was proposed based on the presented by Liu Guo-dong. With Lab-VIEW, a high speed sampling software was programmed, and by sampling the welding current, voltage and displacement of welding procedure orthogonal methodology experiment of U71Mn rail with high frequency, the weld quality characteristic values were obtained, which were the percentage of the flashing time of which is before the accelerated flashing stage, the percentage of the flashing time of the accelerated flashing stage, the power input of weld, the welding time and the flashed length of rail, as input data of the rail weld impacted quality BP neural network prediction model. The prediction model contained 5 units in the input layer, 14 units in the hidden layer. The prediction accuracy of the model trained with 17 samples of 27 samples designed by adopting orthogonal methodology was 90% using the other 10 samples.