基于迁移学习的切削力神经网络预测模型优化策略

Optimization Strategy of Neural Network Prediction Model for Cutting Force Based on Transfer Learning

在实际生产中,生产条件的变更情况时常发生,重新训练一个神经网络预测模型的成本较为高昂。本文针对切削力预测的任务,结合迁移学习领域的理论和方法,研究了一种神经网络的训练方法。在训练神经网络模型时,使用一组相关但不完全相同的切削数据预训练一个网络模型;使用目标数据对该网络进行重训练,并在网络的优化目标中加入两组数据集的MMD距离,称为“迁移网络”。结果表明,与传统的BP神经网络相比,在一定条件下,迁移网络具有较为明显的性能优势。一方面,这意味着使用相同的实验样本,迁移网络的预测误差将得到控制;另一方面,当达到相同的预测误差时,迁移网络所需的实验样本数量将减少,能够有效的减少训练成本。

In actual production,the change of production conditions often occurs,so it is expensive to retrain a neural network prediction models.In this paper,aiming at the task of cutting force prediction,combined with the theory and method of transfer learning field,a neural network training method is studied.When training the neural network model,using a group of related but not identical cutting data to pre train a network model;using the target data to retrain the network,and add the MMD distance of two groups of data sets into the optimization goal of the network,which is called"transfer network".The results show that,compared with the traditional BP neural network,the transfer network has obvious performance advantages under certain conditions.On the one hand,this means that using the same experimental samples,the prediction error of the transfer network will be controlled;on the other hand,when the same prediction error is reached,the number of experimental samples required by the migration network will be reduced,which can effectively reduce the training cost.