基于回归分析和遗传算法优化的BP神经网络熔覆层形貌预测  

Prediction of Cladding Layer Morphology Based on BP Neural Network Optimized by Regression Analysis and Genetic Algorithm

作　　者：杨思瑞 白海清[1,2] 鲍骏 任礼 李超凡 Yang Sirui;Bai Haiqing;Bao Jun;Ren Li;Li Chaofan(School of Mechanical Engineering,Shaanxi University of Technology,Hanzhong 723001,Shaanxi,China;Shaanxi Key Laboratory of Industrial Automation,Hanzhong 723001,Shaanxi,China)

机构地区：[1]陕西理工大学机械工程学院,陕西汉中723001 [2]陕西省工业自动化重点实验室,陕西汉中723001

出　　处：《激光与光电子学进展》2022年第21期141-149,共9页Laser & Optoelectronics Progress

基　　金：陕西理工大学研究生创新基金项目(SLGYCX2124);陕西省技术创新引导项目(2021QFY05-03)。

摘　　要：针对激光熔覆过程中熔覆层形貌难以控制的问题,以45钢和Fe45分别作为基材和熔覆粉末,设计3因素5水平的试验方案,并测量熔覆层的宏观尺寸。首先,利用遗传算法(GA)对反向传播(BP)神经网络的初值进行优化,建立了GABP神经网络模型,以激光工艺参数为输入、熔覆层形貌为输出进行了训练和测试,分析其预测精度。然后,分别以回归分析、BP神经网络和GA-BP神经网络三种方法建立预测模型,并与实际得到的熔覆层形貌进行比较。结果表明,通过遗传算法优化的GA-BP神经网络模型与实际的误差约为3%,BP神经网络模型与实际误差为7.38%,而回归分析预测模型预测误差最大可达到15.5%。经比较可知,GA-BP神经网络预测模型的结果与实际最为接近。故GA-BP神经网络预测模型对提高熔覆层质量具有一定的指导价值。The experiment employs 45 steel and Fe45 as the base material and cladding powder,respectively,to develop three factors and five levels of the test scheme,and evaluate the cladding layer’s macroscopic size,to solve the challenge that the cladding layer morphology is difficult to control in a laser cladding process,First,the backpropagation(BP)neural network’s initial value was optimized using a genetic algorithm(GA),and GABP neural network model was developed.The laser process parameters were taken as input and cladding layer morphology as output to train and test,and the prediction accuracy was examined.Second,the prediction model was developed using regression analysis,BP neural network,and GABP neural network,and compared with the actual cladding layer morphology.The findings demonstrate that the GABP neural network model’s error optimized by the genetic algorithm was about 3%,the maximum error of the BP neural network prediction model was 7.38%,and the maximum error of the regression analysis prediction model is 15.5%.It can be seen from the comparison that the results of the GABP neural network prediction model are the closest to the actual.Thus,the GABP neural network prediction model has a certain regulating value for enhancing the quality of the cladding layer.

关 键 词：激光技术 激光熔覆 遗传算法-反向传播神经网络 回归分析 形貌预测 

分 类 号：TG174.4[金属学及工艺—金属表面处理]