文献类型：期刊文献

英文题名：Decouple to adaptive diagnosis: A dual-head open-set framework for cross-condition injection molding process

作者：Hu, Guangze[1];Lu, Jingyi[2];Ali, Husnain[1];Dong, Hao[1];Xu, Fangjian[1];Zhang, Zheng[1];Gao, Furong[1,3]

机构：[1]Hong Kong Univ Sci & Technol, Dept Chem & Biol Engn, Hong Kong, Peoples R China;[2]East China Univ Sci & Technol, MOE Key Lab Smart Mfg Energy Chem Proc, Shanghai 200237, Peoples R China;[3]Guangzhou HKUST Fok Ying Tung Res Inst, Guangzhou 511458, Peoples R China

年份：2026

卷号：166

外文期刊名：JOURNAL OF PROCESS CONTROL

收录：;EI(收录号：20263421338941);Scopus(收录号：2-s2.0-105047809439);WOS:【SCI-EXPANDED(收录号:WOS:001854656700001)】；

基金：The authors gratefully acknowledge the supportive research environment provided by the Hong Kong University of Science and Technology and the invaluable guidance of the Department of Chemical and Biological Engineering. This work was partially supported by the Hong Kong Research Grant Council under Project No. 16203322, the National Natural Science Foundation of China (Grant No. 62333010), and the RGC-NSFC Joint Research Scheme (N_HKUST628/22).

语种：英文

外文关键词：Batch processes; Cross-domain fault diagnosis; Domain shift; Out-of-distribution; Open-set domain adaptation

摘要：Injection molding demands highly reliable process diagnosis to ensure product yield and safety. However, the performance of fault diagnosis models often degrades due to domain shifts and unseen anomalies in real-world injection molding process, a problem known as Open-Set Domain Adaptation (OSDA). Current methods tightly couple domain alignment, causing severe feature entanglement. To address this, we propose DC-VOS (Decoupling CORAL and Virtual Outlier Synthesis), a novel dual-head framework isolating feature alignment from openset anomaly rejection. After normal condition anchoring to preprocess data, the alignment head utilizes secondorder covariance matching to bridge domain gaps without distorting known-class manifolds. Meanwhile, the rejection head generates virtual outliers and employs a Tree-structured Parzen Estimator to autonomously calibrate optimal decision boundaries. Real-world experiments demonstrate DC-VOS significantly outperforms existing baselines. It boosts the Harmonic Score by 8.7% to 79.90% cross-material and 9.5% to 75.21% crossmachine while achieving a 100% unknown anomaly rejection rate. Ultimately, DC-VOS offers a highly robust solution for complex open-set industrial diagnostics.