Prediction of Cutting Force for Different Tools Based on Transfer Learning and Neural Networks

Processing large and complex parts requires various tools, a number of which need to be replaced during the process. However, the variable dynamic characteristics of different cutting tools can significantly affect the accuracy of cutting force modeling and prediction. A data-driven approach utilizing neural networks and transfer learning is proposed for predicting milling forces across various tools. Initially, cutting data from milling experiments using different tools and machining parameters are collected to form a dataset. The source tool contains a comprehensive set of process parameters data, whereas the target tool includes a small number of labeled and test groups. Afterwards, the input data of the source and target tools are fed into an autoencoder with a maximum mean discrepancy loss function to reduce the marginal distribution discrepancy. Furthermore, affine transformation is performed to generate pseudo-labels, thereby augmenting the dataset and providing coarse data for the target tool. Finally, the TrAdaBoost.R2 algorithm is applied to establish the cutting force prediction model specific to the target tool. The training set of which includes a combination of pseudo-data and a small amount of target tool marked data, enabling accurate prediction of the cutting forces for the target tool's unlabeled data. Detailed experimental validation is performed on five-axis machine tools to verify the accuracy and effectiveness of the designed methodology. Comparison results show that prediction accuracy improved by more than 50%, 35%, and 65% compared with network trained directly with source domain data, models trained directly with TrAdaBoost.R2 algorithm, and network trained with small amounts of data from the target tool, respectively, which showcase the superiority of the proposed model.