An unsupervised dual-regression domain adversarial adaption network for tool wear prediction in multi-working conditions

Intelligent tool wear prediction networks are widely used regarding their great advantages in utilizing big data. Since datas from new working conditions is unlabeled, the network must transfer the wear knowledge learned from other conditions. However, the global marginal distribution discrepancy and local degradation stages variation for different working conditions result in low prediction accuracy or failure of the network. An unsupervised Dual-Regression Domain Adversarial Adaption Network (DR-DAN) is proposed, which mines the global and local consistency of degradation features between different working conditions and realizes the knowledge transfer. The proposed weight discrepancy restriction constructs a feature space for extracting local consistency representation. Furthermore, the predictive consistency loss is proposed to match accurate degradation stages without label supervised. To ensure the stability of adversarial training, Wasserstein distance is employed as the optimization function. Two experiments are carried out to demonstrate that DR-DAN has a better performance than other state-of-the-art methods.