Abstract: A grinding surface roughness measurement method was proposed based on SA-CNN for convenient and accurate prediction of roughness values on screw rotor surfaces after grinding. Through orthogonal experiments, the surface roughness values of screw rotors and corresponding surface images were obtained. After preprocessing including adaptive histogram equalization and unsharp masking, the images were used as training samples input into the SA-CNN model. The SA-CNN model was employed to predict the roughness values on the grinding surfaces of screw rotors and compared with the predictions of classical networks such as ResNet, AlexNet, VGG-16, basic CNN, and graph neural network (GNN). Experimental results show that the SA-CNN model achieves an average prediction accuracy of 95.24%, with an RMSE of 0.0706 μm and an MAPE of 7.4206%, outperforming the compared networks. Furthermore, the SA-CNN model exhibits fast convergence, high accuracy, and good robustness.

Key words: grinding, surface roughness, convolutional neural network(CNN), orthogonal experiment

摘要： 为便捷、准确地预测磨削后螺杆转子的表面粗糙度，提出了一种基于自注意力卷积神经网络（SA-CNN）的磨削曲面粗糙度测量方法。通过正交试验获得螺杆转子的表面粗糙度以及粗糙度数值对应位置的表面图像，图像经自适应直方图均衡化、反锐化掩蔽等预处理后作为训练样本输入SA-CNN模型中。采用SA-CNN模型对磨削后的螺杆转子表面粗糙度值进行预测，并与经典网络ResNet、AlexNet、VGG-16、基础CNN以及图神经网络GNN预测结果进行对比。试验结果表明，SA-CNN模型的平均预测精度达到95.24%，均方根误差(RMSE)为0.0706 μm，平均绝对百分比误差(MAPE)为7.4206%，均优于对比网络，且模型收敛较快，表现出较高的精度和良好的鲁棒性。

关键词: 磨削, 表面粗糙度, 卷积神经网络, 正交试验