基于进化神经网络的磁粒研磨表面粗糙度预测方法研究

Research on Surface Roughness Prediction Method of Magnetic Abrasive Finishing Based on Evolutionary Neural Network

目的为实现磁粒研磨光整加工的表面粗糙度精准预测,提出一种遗传算法优化表面粗糙度BP神经网络的预测模型。方法将表面粗糙度作为预测的目标,影响磁粒研磨5052铝合金管内表面粗糙度的五个主要工艺参数作为神经网络的输入。合理设计正交试验,得到不同工艺参数配置下的表面粗糙度值,将其作为神经网络的输出。通过建立非线性预测模型,对比遗传算法优化后和传统BP神经网络的均方差与仿真时间,分析优化前后表面粗糙度的预测效果。结果通过试验数据建立了结构为5-11-1的神经网络,进化BP神经网络预测模型均方差为0.044,建模仿真时间为0.187 s,其平均相对误差率为13.2%。传统的BP神经网络预测模型均方差为0.231,建模仿真时间为1.840 s。结论通过遗传算法优化后的BP神经网络均方差更小,建模仿真时间更短,进化BP神经网络可以实现更为精准的预测,同时能够极大地避免传统BP神经网络易陷入局部极小值的弊端。

In order to achieve accurate prediction of surface roughness of magnetic abrasive finishing,the precision prediction model of surface roughness by BP neural network optimized by genetic algorithm was established.Taking the surface roughness as the target of prediction,five main process parameters affecting the inner surface quality of 5052 Al alloy tube by magnetic abrasive finishing as inputs of neural network.Through the orthogonal experimental design,the surface roughness under different process parameters was obtained as the output of the neural network.By establishing a nonlinear prediction model,the prediction effect of surface roughness before and after optimization was analyzed by comparing the mean square error and simulation time of the optimized genetic algorithm and the traditional BP neural network.Based on experimental data,the BP neural network with 5-11-1 topology structure is established.The mean square deviation of the prediction model is 0.044,the simulation time is 0.187 s,and the average relative error rate is 13.2%.The mean square deviation of the unoptimized BP neural network prediction model is 0.231,and the simulation time is 1.840 s.The mean square error of evolutionary BP neural network is smaller,and the modeling and simulation time is shorter,and more accurate prediction by evolutionary BP neural network can be achieved.What’s more,the BP neutral network can greatly avoid the disadvantage of traditional BP neural network easily falling into local minimum.