摘要
神经网络算法在金属零件加工表面粗糙度预测中有着广泛的运用,但是尚存网络初始化及梯度弥散这类“黑盒”问题。提出一种基于传统PSO-BP框架的改进模型,使用Xavier替代传统高斯分布初始化粒子群,在算法优化器方面使用最新的自适应矩限制取代随机梯度下降算法进行模型参数更新,经实际测试发现其训练集均方误差及其测试集绝对误差相对传统PSO-BP算法均有所降低,可以更有效担任预测任务。
Neural network algorithm is widely used in the prediction of the surface roughness of metal parts,but there are still"black box"problems such as network initialization.An improved model based on the traditional PSO-BP framework is proposed.Xavier is used to replace the traditional standard Gaussian distribution to initialize the particle swarm,and the latest adaptive moment limit is used to replace the random gradient descent algorithm in the algorithm optimizer to update the model parameters.The actual test shows that the mean square error of the training set and the absolute error of the test set are lower than those of the traditional PSO-BP algorithm To be more effective in forecasting tasks.
作者
史柏迪
庄曙东
韩祺
SHI Bai-di;ZHUANG Shu-dong;HAN Qi(School of Mechanical Engineering,HoHai University,Changzhou Jiangsu 213022,China;Mettler Toledo International Trade(Changzhou)Co.,Ltd.,Changzhou Jiangsu 213022,China;不详)
出处
《组合机床与自动化加工技术》
北大核心
2021年第2期30-33,38,共5页
Modular Machine Tool & Automatic Manufacturing Technique
基金
江苏省高校实验室研究会立项资助研究课题(GS2019YB18)
江苏省精密与微细制造技术重点实验室数学建模课题组(CZ520007812)
中央高校基本科研业务费(2018B44614)。
关键词
表面粗糙度
神经网络
算法优化器
粒子群优化
surface roughness
neural network
optimizer
particle swarm optimization
