Partial transfer learning network for data imputation and soft sensor under various operation conditions

• Published in: Journal of Central South University Vol. 30; no. 10; pp. 3395 - 3413
• Main Authors: Ren, Jia-yi, Chen, Xu, Zhao, Chun-hui
• Format: Journal Article
• Language: English
• Published: Changsha Central South University 01.10.2023; Springer Nature B.V
• Subjects: Engineering; Knowledge management; Learning; Metallic Materials; Missing data; Quality control; Sensors; 多工况; various operation conditions; 数据缺失; partial transfer; soft sensor; 软测量; 部分迁移; missing data; 维数不一致; dimension inconsistency
• ISSN: 2095-2899; 2227-5223
• DOI: 10.1007/s11771-023-5448-8

Soft sensor plays a key role in the safe operation of industrial processes and product quality control. In industrial processes, the switching of operation conditions may lead to distribution discrepancy and dimension inconsistency between the training data (source domain) and testing data (target domain), leading to the soft sensor model mismatch problem. In addition, the data may be incomplete because of sensor transmission failure, where the missing values may influence the accuracy of the soft sensor. This article introduces a partial transfer learning network (PTL-Net) for soft sensors under different operation conditions with missing data. First, the imputation and soft sensor modules are constructed for source domain data, where a compactness loss is designed to induce feature aggregation to alleviate the influence of abnormal features mapped from missing data. Then a partial transfer strategy is proposed to reduce the distribution discrepancy between the source and target data. Furthermore, the proposed strategy selects the common components between two domains for partial knowledge transfer rather than inheriting all the parameters directly, which can overcome the model mismatch problem. The effectiveness of PTL-Net is verified under the nuclear dataset and the three-phase flow process.