为提高硬脆材料微结构的加工效率和精度,需要预测微磨具的不确定性磨损.基于微磨具在位视觉磨损检测和聚类分析,提出基于遗传算法的反向神经网络(genetic algorithm back propagation,GA-BP)模型.选取微磨具磨头截面面积损失量为指标,以表征微磨具不确定性磨损特征.利用K-均值聚类算法划分微磨具磨损状态阶段.最后构建以主轴转速、进给率、微槽深度、磨削长度和微磨具初始截面面积为输入层神经元,以磨头截面面积损失量预测值为输出层的GA-BP神经网络模型.设计不同工艺参数条件下的单晶硅微槽微细磨削实验,基于自搭建的机器视觉系统在位测量微磨具的磨头截面面积磨损量.将实验测得的微磨具磨损量作为训练数据,与传统高斯过程回归预测模型对比,验证GA-BP神经网络模型的有效性和准确性.结果表明,GA-BP神经网络模型能够实现不同工艺参数和不同磨削长度下的微磨具磨损预测,比传统高斯过程回归预测模型具有更高预测精度,平均误差精度达到5%,可以实现微磨具磨损阶段状态预测.
Wear prediction of micro-grinding tool based on GA-BP neural network
An intelligent tool wear prediction model has been proposed for the micro-grinding tool,optimized using a genetic algorithm(GA)based BP neural network.The GA-BP prediction model is applied with in-situ tool wear detec-tion to obtain training set data and combines cluster analysis to divide the tool wear stages.To represent the uncertainty in wear characteristics,the loss of cross-sectional area of the micro-grinding tool has been selected as an index to evalu-ate tool wear loss.The K-means clustering algorithm is used to cluster and analyze the tool wear stages under different process parameters.The GA-BP neural network includes five neurons in the input layer:rotating speed,feed rate,cut-ting depth,grinding length,and the initial cross-sectional area of the tool.The output layer neuron predicts the loss of the tool's cross-sectional area.To validate the method,a series of micro-grinding experiments were performed under dif-ferent parameters for the micro-groove array of monocrystalline silicon.The loss of the tool's cross-sectional area was measured by a self-made visual inspection system,providing learning samples for the prediction model.The predicted results of the GA-BP neural network model were compared with the traditional Gaussian process regression method.The results show that the GA-BP neural network model can correctly predict tool wear loss and identify wear stages un-der different process parameters and grinding lengths.It has higher prediction accuracy during the self-learning process,with an average error of 5%.
国家科技期刊平台
NSTL
万方数据
