| Abstrakt: |
In the process of industrial production, products with different specifications (i.e., the difference in geometry, process conditions, and machine conditions, etc.) have different quality data distributions, which lead to a decrease in the accuracy of traditional data-driven quality prediction models that require the same quality data distribution. At the same time, due to economic cost factors, obtaining a large amount of accumulated data for different specifications is difficult, and the re-modeling data accumulation of multiple cross-specifications is insufficient. In order to solve the quality prediction problem of production with different data distributions and poor data accumulation, we use the deep transfer learning (DTL) method with unsupervised dynamic domain adaptation (DDA) to transfer the domain invariant features learned from labeled specification products (source domain) to other unlabeled new specification products (target domain). In order to improve the success rate of cross-domain quality prediction, the Wasserstein distance adapter is designed to match appropriate source domain samples and target domain samples to build multiple transfer tasks that are suitable for transfer. At the same time, the dynamic distribution adaptation and dynamic adversarial adaptation are combined to extract the domain invariant features to improve the adaptability of the prediction model for products with new specifications (e.g., size difference) and unlabeled and limited quality data. Finally, a comprehensive experiment is carried out using the actual production data of products with different specifications. The experimental results show that compared with the traditional non-transfer deep learning methods, the MAE, RMSE, and R2 of the proposed DTL method are improved by 18.26%, 16.66%, and 22.48% respectively. Compared with other transfer methods, the MAE, RMSE, and R2 of the DTL proposed in this paper are improved by 10.45%, 10.96%, and 9.72% respectively. [ABSTRACT FROM AUTHOR] |
