Kuo Liu

Single source domain transfer learning has been used in the field of tool wear states monitoring. However, this method ignores the case of transfer learning from multiple source domains and it is unreasonable to transfer knowledge from multiple source domains to the target domain with equal weights. Therefore, this article proposes a new multi-source unsupervised domain adaptation method based on domain similarity (SBMUDA) for tool wear states identification. This paper utilizes cosine distance to assess the similarity between each source domain and target domain. Based on domain similarity, the feature distance of each source and target domain and classification loss of the source domains are weighted to enhance attention to source domains with high similarity and reduce the negative transfer damage. Moreover, a multi-source domain model with weight sharing is established. By integrating the advantages of multiple source domains, more classification knowledge is transferred to the target domain, enhancing the model’s generalization ability. The milling experimental results show that the classification accuracy of multi-source domain transfer learning is significantly higher than that of the optimal single source domain, with a maximum improvement of 4.4%. Compared with other methods, this method has the highest average accuracy of 98.7% and the smallest standard deviation.

Mechanism-driven model (MDM) plays an important role for time-varying thermal error (TVTE) compensation. For complex working conditions of machine tools, it is needed to ensure that the mechanism-driven model is always accurate in long time. Therefore, the online evolutionary strategy (OES) for MDM was proposed. The model accuracy can be recognized based on the consistency between the calculated value and the measured value of the intermediate dependent variable. The parameters can be updated online. Furthermore, the mechanism-driven model was established based on the OES. The update effect of the model parameters was verified. The compensation accuracy of Model 1 is degraded from 11.8 μ m to 18.6 μ m, and Model 2 is degraded from 11.8 μ m to 14.6 μ m. After online evolution, the compensation accuracy is improved to 12.5 μ m for Model 1 and to 10.4 μ m for Model 2. It is shown that the OES can effectively correct the calculation accuracy of TVTE model.