.png){kind=icon}
.png){kind=icon}
Return
Multiple space transfer learning based on maximizing mean variance differences for soft sensor modeling
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.130975.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The scarcity of labeled data significantly affects the effectiveness of industrial soft sensing. Domain adaptation can transfer rich label information from the source domain to the sparsely labeled target domain. However, existing domain adaptation methods typically align source and target domain features into a single feature space, which may negatively impact the adaptation performance of soft sensor models. Therefore, the multiple space transfer learning based on maximizing mean variance discrepancy (MMVD-MSTL) is proposed. First, source and target domain data are mapped into a spatio-temporal-frequency feature space based on feature-disentangled encoder. Then, the maximizing mean–variance discrepancy is designed to align disentangled features distributions across multiple spaces between the source and target domains. Finally, the cycle adversarial loss constrains feature distributions in both domains, establishing a reciprocal feature relationship. Comparative experiments on public datasets and real-world industrial process data demonstrate that MMVD-MSTL outperforms state-of-the-art domain adaptation soft sensor models, showing superior adaptation performance in the target domain.
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W