Deep learning–based data anomaly detection for highway slope structural health monitoring: A comparative study

Abstract

Highway slope instability has a significant influence on traffic safety. However, there are many anomalies in slope SHM data, which is critical to timely warnings and safety assessments of slopes. In this paper, we carried out a multi-case comparative study of deep learning models to examine the recognition accuracy of anomalous data. First, our program collected the monitoring data from the Baihe and Lueyang slopes of the Shi-Tian expressway in Shaanxi province. Six categories of abnormal data were found using K-means clustering. Second, based on the time response images, the frequency response and Gramian Angular Field (GAF) images were superimposed to improve the model’s ability to identify six types of data anomalies. Third, we employed combo loss to tackle the data imbalance problem by incorporating dice loss and focal loss. Last, we conducted a comparative study of nine deep learning models to investigate the anomaly detection capability with the combo loss function. The results indicated that a combination of frequency response and GAF images can effectively improve the identification of abnormal data. The combo loss with equal weights significantly improved specific anomaly categories’ performance. In addition, the performance of ResNet50, EfficientNetB1, and Vision Mamba exhibited impressive classification accuracy, generalization ability, and computational efficiency. Our investigation has the potential to recognize slope SHM data abnormal types in various scenarios, further improving the precision of subsequent slope capability analysis.