结合域对抗自适应的刀具磨损预测方法

数控加工中存在刀具几何误差及安装误差、刀具及工件材料性能的随机波动等因素，导致刀具之间的磨损过程与监测信号上存在较大差异的问题，使得刀具磨损值难以精确预测。为此，本文提出了一种结合域对抗自适应的多尺度分布式卷积长短时记忆网络模型（Multiscale time-distributed convolutional long short-term memory,MTDCLSTM）。将加工过程中采集到的多传感器信号作为模型输入，通过域分类器与预测器之间的对抗学习，提取出可有效表征刀具磨损且与域无关的多尺度时空特征，经预测器的非线性映射，实现对刀具磨损值的精确预测。实验结果表明，结合域对抗自适应的MTDCLSTM模型预测性能明显优于分布式卷积神经网络、长短时记忆网络、卷积神经网络与支持向量机模型。与基于迁移成分分析的支持向量回归模型相比，本文模型的均方根误差与平均绝对误差分别降低了59.8%和62.5%，决定系数提高了66.1%，可有效缩小刀具个体之间的差异，提高磨损值预测精度。

Prediction Method of Tool Wear Combined withDomain Adversarial Adaptation

In the CNC machining, due to the factors such as the geometric and the installation error of the cutting tool, the random variations in the material properties of the workpiece and the cutter, the wearing process and the monitoring signals between the individual tools have large differences, which makes the tool wear values to difficultly predict. To address this problem, a multiscale time-distributed convolutional long short-term memory model (MTDCLSTM) combined with domain adversarial adaptation is proposed. With the multi-sensor signals obtained in the machining as the model input, the model extract multiscale spatio-temporal features that can effectively characterize the tool wear and are independent of the domain through the adversarial learning between the domain classifier and the predictor. The accurate prediction value of the tool wear can be obtained by using the non-linear mapping of the predictor. Experimental results show that the prediction performance of the MTDCLSTM model combined with domain adversarial adaptation is significantly better than the time-distributed convolutional neural networks, long short-term memory neural networks model, convolutional neural networks and support vector machine models. Comparing with the support vector regression model based on transfer component analysis, the root mean square error and average absolute error of the present model were reduced by 59.8% and 62.5%, respectively, and the coefficient of determination was increased by 66.1%, which means the present model can effectively reduce the difference between the individual tools and improve the accuracy of the wear prediction.