Incremental transfer learning for monitoring fatigue crack sizes in welded joints based on strain variations

Abstract

This paper presents an incremental transfer learning (TL) approach for monitoring fatigue crack sizes in welded joints based on strain variations. The proposed method integrates a pretrained machine learning (ML) model, developed from a calibrated numerical database, with incremental TL via limited experimental data to enhance the accuracy of the crack size estimation. Compared with experimental measurements, the pretrained ML model shows escalating errors in the crack profile estimations as the fatigue crack evolves, while the proposed incremental TL approach provides an effective solution to mitigate this issue. Validation on one bending and three tensile cruciform welded specimens demonstrates the strong predictive performance of the proposed incremental TL approach. The study also reveals the effect of fine-tuning frequency on the accuracy of crack size prediction, providing practical guidance for model updating strategies during fatigue crack monitoring. For a practical scenario, where the strain is only available after the crack detection and the strain variation is referenced to the strain measured from a cracked joint, the proposed approach demonstrates accurate estimations of the subsequent crack growth. The success of this method underscores the high potential of cost-effective strain sensors in supporting crack diagnosis and structural health monitoring.