Transferring self-supervised pre-trained models for SHM data anomaly detection with scarce labeled data