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

An improved back propagation(BP) neural networks model was proposed based on the presented by Liu Guo-dong.With LabVIEW,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.