Leveraging Digital Twin Simulations to Pretrain Contrastive Encoders for Few-Shot Structural Damage Classification

Abstract

The classification of structural damage under severely limited labeled data conditions remains a persistent bottleneck in automated infrastructure inspection, where the cost and complexity of expert annotation constrain the scalability of conventional supervised learning pipelines. This paper proposes a framework that integrates digital twin (DT) simulation with contrastive self-supervised pretraining and prototypical few-shot classification to address this bottleneck. High-fidelity finite element models of reinforced concrete (RC) and steel frame structures are used to generate large-scale synthetic image datasets parameterized across a comprehensive range of damage types and severities. A SimCLR-inspired contrastive encoder is pretrained on this synthetic corpus without any human labels, learning transferable structural damage representations that bridge the simulation-to-reality gap. The pretrained encoder is subsequently adapted to real-world inspection imagery using prototypical networks operating in the five-shot and one-shot regimes. Experiments conducted on the PEER Hub ImageNet dataset and a collected bridge inspection corpus demonstrate that the proposed method achieves classification accuracy substantially exceeding ImageNet-pretrained baselines under equivalent few-shot conditions, validating the utility of DT-driven synthetic pretraining for structural health monitoring (SHM) applications.