SCLF: Self-Contrastive Learning Framework for Hydrogen Leakage Traceability

Le résumé fourni par la source

Hydrogen leakage traceability is crucial to ensuring hydrogen safety in the whole process of hydrogen production-storage–transport-application. However, due to the high cost of dismantling and testing, it is impractical to collect and label leakage locations for different leakage conditions. Traditional supervised learning methods lack generalization performance due to the difficulty of obtaining effective training on a limited number of labeled hydrogen leakage samples. Therefore, a self-contrastive learning framework (SCLF) is proposed in this study. Upstream of SCLF, a time-frequency fusion representation method based on self-contrastive learning (CL) is proposed to enhance domain-independent feature extraction, and a pretrained model (TF-FRM) is developed. Downstream of SCLF, a multirulebase fuzzy locator (MR-FL) based on fuzzy logic is proposed, which can locate the position of hydrogen leakage accurately by the representation based on TF-FRM. Finally, the proposed SCLF is tested on a real sensor signal, and its interpretability is also analyzed. Compared with other models, the SCLF demonstrates superior generalization, accuracy, and interpretability with limited supervised samples, achieving an average$F1$score of 95.59%.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.