Target-Driven Dynamic Weighted Joint Distribution Adaptation Network for Soft Sensing of Tumbler Strength in Sintering Process

Accurate soft sensing of tumbler strength plays a pivotal role for the normal and safe operation of the sintering process. Traditional deep learning-based soft sensing models assume that the training data and test data are in the same distribution, which is unrealistic for the sintering process. The distribution shift often occurs due to the fluctuations of raw materials and the changes of working conditions, which degrades the performance of soft sensing. However, most existing domain adaptation soft sensing researches fail to fully consider the conditional distribution adaptation and target-specific information. To overcome these shortcomings, in this article, we propose a target-driven dynamic weighted joint distribution adaptation method to learn domain-invariant and target-specific features to predict tumbler strength under distribution shift. Specifically, we propose a dynamic weighted distribution adaptation method to better align the marginal and conditional distributions by calculating the weights of their relative importance. In addition, we design a target-driven module to enable the model to extract target-specific features that are highly relevant to the label space of the target domain. Finally, experimental results on three datasets demonstrate that the proposed method outperforms the mainstream domain adaptation methods on the tumbler strength task.