Enhancing robustness of data-driven SHM models: adversarial training with circle loss