Digital twin-assisted modular fast switching technology of assembly line

Abstract Aiming at the demand of digital twin-driven modular rapid switching of assembly line, this paper proposes an ontology fusion method based on Stacking ensemble learning (SEL) to solve the accuracy and robustness problems of multi-source heterogeneous knowledge matching. By integrating random forest (RF), support vector machine (SVM) and gradient lifting decision tree (GBDT) as primary learners, this paper uses logistic regression (LR) as meta-learner to construct a fusion model, and combines vocabulary (Levenshtein distance), semantic (WordNet) and structural (subtree kernel) three-dimensional features to optimize the matching process. The verification of the test shows that in the German/American/Japanese trilingual data set (VDI 2860, NIST, JRAIA), the model achieves an accuracy rate of 0.934, 0.928 F1-Score and 0.962 AUC-ROC, which is more than 2.1% higher than baseline models such as BERTMap. At the same time, under 20% noise interference, the performance of the model only decreases by 8.2% (the average decrease of the control group is 24.9%), and the matching delay is 142 ms, which meets the real-time needs of the industry. In addition, the Chinese-English-German accuracy rate of the model’s cross-language task is 0.919 (2.1% higher than that of BERT-Multilingual), but the Japanese adaptability is weak (0.907). To sum up, this model significantly improves the efficiency of assembly line knowledge fusion through feature collaboration and integration strategies, and its industrial practicality contributions are outstanding. In the future, it needs to be combined with FPGA acceleration and domain dictionary to optimize computing efficiency and cross-language generalization capabilities.