CAN-Net: A Multi-hidden Layer Attention Deep Learning Method for Surface Roughness Prediction During Abrasive Belt Grinding of Superalloy with Local Weights

Abstract

Nickel-based superalloys are widely employed in aerospace due to their excellent high-temperature strength, good oxidation resistance, and hot corrosion resistance. Abrasive belt grinding can effectively solve the problems of excessive residual stress and tool wear during the processing of superalloys. However, due to the grinding process being complex and changeable, and a wide range of affecting factors, the surface roughness prediction of abrasive belt grinding has become a challenging topic. In this study, a CAN-Net multi-hidden layer deep learning prediction model is established. The concatenate path is utilized to fuse local weights to optimize the intermediate weights of network training. To increase the predictability of the model, the attention mechanism is included to distribute the weights of the grinding parameters, and the impact of the attention mechanism on the prediction is then carefully analyzed. The results demonstrate that the CAN-Net network model has outstanding parameter flexibility and prediction accuracy, with accuracy reaching 0.984 and a correlation coefficient of 0.981 between the anticipated value and the true value.