A self-supervised learning-based approach for detection and classification of dam deformation monitoring abnormal data with imaging time series

Dam safety monitoring systems collect a significant amount of data, including numerous instances of abnormal data attributed to factors such as aging equipment and recording errors. The accurate detection and classification of abnormal data can effectively enhance the reliability of rockfill dam safety assessments based on monitoring data. Note that the scarcity of the abnormal data in actual monitoring datasets poses a significant challenge to existing data-driven anomaly detection studies. To address these issues, we develop a novel self-supervised learning-based framework for abnormal data detection and classification of rockfill dam deformation data. This framework includes an abnormal data detection method based on transformers and synthetic abnormal data. By analyzing and further modeling the real-world abnormal data, a criterion for synthesizing abnormal data is proposed to augment the scale of abnormal data. Additionally, we introduce an abnormal data classification method using imaging time series, which captures the multi-scale features of sequence data in a higher dimension by encoding it into image representations and employing a residual network (ResNet) for feature extraction. The effectiveness of the proposed approach is demonstrated through an engineering case study. The F1 scores for abnormal data detection and classification are 0.9722 and 0.9596, respectively, which surpass those of other conventional methods. The results demonstrate that the proposed approach achieves high-precision detection and classification of abnormal data, even under adverse conditions where abnormal data are sparse, thus ensuring reliable safety assessment of rockfill dams.