GTRF: A General Tuples Recognition Framework Towards Deep Learning-Driven Structural Health Monitoring Adapted to Diverse Supervision Paradigms

Abstract

Leveraging the powerful capabilities of deep learning (DL) techniques, the DL-driven tuple recognition approach has successfully addressed numerous challenges in SHM by associating tuples with structural patterns. However, the robustness and generalizability of the model are significantly compromised due to limitations in their designated feature extraction strategies, network architectures, and supervision learning schemas. To address these issues, this study proposes a novel General Tuple Recognition Framework (GTRF) that supports supervised (SL), unsupervised (UL), and semi-supervised learning (SSL) paradigms. The present article provides a detailed explanation of the mechanism and workflow of the proposed GTRF. Equipped with sophisticated networks and innovative components, the GTRF demonstrates competence in various tuple recognition tasks across different learning paradigms. The validation experiments conducted in the field of SHM include vibration SL-recognition of a prototype skyscraper, damage UL-detection of a laboratory RC beam, and condition SSLassessment of a full-scale building model. To ensure the adaptability of diverse tuples, two commonly used data forms, namely acceleration measurements and piezoelectric signals, are employed in the experimental validations. The comprehensive results confirm the effectiveness and adaptability of the proposed GTRF. The flexible paradigm specialization, broad application, and potential for optimization make the proposed GTRF framework a promising prototype for bridging the gap between DL algorithm fusion and model integration across different learning paradigms.