Wu Jiajun, Liu Xuejun, Qiao Hongchao, Zhao Yongjie, Hu Xianliang, Yang Yuqi, Zhao Jibin
Appl Opt. 2021 Apr 10;60(11):3114-3121. doi: 10.1364/AO.421431.
With the purpose of using the artificial neural network (ANN) method to predict the residual stresses induced by laser shock processing (LSP), the Ni-Cr-Fe-based precipitation-hardening superalloy GH4169 was selected as the experimental material in this work, and the experimental samples were treated by LSP with laser power densities of 4.24/, 7.07/, and 9.90/ and overlap rates of 10%, 30%, and 50%. The depth-wise residual stresses of experimental samples prior to and after LSP were taken according to the x-ray diffraction method and electrolytic-polished layer by layer. The ANN model for residual stress prediction was established, and the laser power density, overlap rate, and depth were set as input parameters, while residual stress was set as the output parameter. The residual stresses of untreated samples and those treated with laser power densities of 4.24/ and 9.90/ were selected as the training sets, and the data of experimental samples treated with a laser power density of 7.07/ were reserved as testing sets for validating the trained network. After LSP, beneficial stable compressive residual stresses were introduced in the material's near surface, and the overall maximum compressive residual stresses were formed on the top surface (surface residual stress). Depending on the LSP process parameters, the surface residual stresses ranged from -236 to -799, and the compressive residual stress depths of all treated samples were over 0.50 mm. According to the results obtained by ANN, the coefficient of determination of the training sets is 0.9948, which shows a good fitness for the training network. The of the testing sets is 0.9931, which is less than that of the training sets but still shows high accuracy. This work proves that the ANN method can be applied to predict the residual stress of metallic materials by LSP treatment with high accuracy and provides a guiding value for the optimization of the LSP process.
为了采用人工神经网络（ANN）方法预测激光冲击处理（LSP）引起的残余应力，本研究选用Ni-Cr-Fe基沉淀硬化高温合金GH4169作为实验材料，并对实验样品进行LSP处理，激光功率密度分别为4.24/、7.07/和9.90/，重叠率分别为10%、30%和50%。采用X射线衍射法和逐层电解抛光的方法测量了实验样品在LSP处理前后沿深度方向的残余应力。建立了残余应力预测的ANN模型，将激光功率密度、重叠率和深度设置为输入参数，将残余应力设置为输出参数。选取未处理样品以及激光功率密度为4.24/和9.90/处理后的样品的残余应力作为训练集，将激光功率密度为7.07/处理的实验样品数据作为测试集，用于验证训练后的网络。LSP处理后，在材料近表面引入了有益的稳定压缩残余应力，在顶面形成了整体最大压缩残余应力（表面残余应力）。根据LSP工艺参数的不同情况，表面残余应力范围为-236至-799，所有处理样品的压缩残余应力深度均超过0.50 mm。根据ANN得到的结果，训练集的决定系数为0.9948，表明训练网络具有良好的拟合度。测试集的决定系数为0.9931，虽小于训练集，但仍显示出较高的准确性。本研究证明，ANN方法可用于高精度预测LSP处理金属材料的残余应力，为优化LSP工艺提供了指导价值。
