Unsupervised Domain Adversarial Adaptive Regression Network for Cutting Force Prediction at Varying Spindle Speeds

The cutting force at varying spindle speeds produces different excitation frequencies, resulting in different computer numerical control (CNC) system responses. Due to the nonlinearity of the frequency response function, it is challenging to create a cutting force prediction model using the CNC inherent signals over the entire frequency range. A new nonparametric-estimation-based unsupervised domain adversarial regression (NE-UDAR) method is designed in this article for the cutting force prediction at varying spindle speeds. The dual-regression network is designed to fit the source domain mapping relationship after the feature extractor with shared parameters. The weight difference constraint and the consistency loss of the feature vector are proposed to extract the local consistency representation. NE-UDAR aligns the marginal distribution by performing adversarial training. The experimental outcomes demonstrate that NE-UDAR can achieve remarkable prediction accuracy at the new spindle speed, which outperforms the network trained directly with source data. The root mean square error is reduced by 65.7% and 39.2% for linear and circular trajectories, respectively, and the mean absolute percentage error is reduced by 61.8% and 35.8%, respectively. The prediction accuracy is comparable to the network trained using labeled data with the same spindle speed as the target domain.