Nana Han ^1,2 Mohamad Ismail Mohd Ashraf ^1* Mingzhi Li ^3* Wending Miao ^4* Qiang Hu ^4* Liyuan Duan ¹ Jintao Tang ^5*

• ¹ School of Civil Engineering, University of Science Malaysia (USM), Nibong Tebal, Malaysia
• ² GuiZhou Equipment Manufacturing Polytechnic, Guiyang, China
• ³ Guangxi Communications Design Group Co. Ltd, Nanning, China
• ⁴ Guizhou transportation planning survey & design academe CO.LTD, Guiyang, China
• ⁵ Yunnan Transportation Planning and Design Institute, Kunming, Yunnan Province, China

Over the past few decades, China has vigorously advanced its strategy to build a powerful transportation network, constructing and maintaining a series of slope engineering projects. However, as time has passed, major safety incidents caused by slope failures have occurred frequently, making the need for automated identification of failure events during the operational phase of slopes an urgent issue. This paper integrates rainfall, surface displacement, and vertical displacement monitoring data of slopes, and proposes an automatic failure mode identification method based on deep convolutional autoencoder technology. The proposed model is trained using monitoring data from the normal operational phase of the slope, automatically extracting the features of normal monitoring data to reconstruct the original data. The trained model is then applied to the task of structural anomaly detection. Using the characteristic that reconstruction errors for failure mode samples are significantly higher than for normal samples, this method enables automatic identification of anomalous modes. Taking a specific slope as a case study, where on May 24, 2024, the displacement development rate in some areas increased significantly, eventually leading to collapse, the proposed model accurately identified the time and evolution of the landslide. The validation results show that the model can effectively distinguish previously unseen abnormal modes and holds significant practical value for identifying similar structural anomalies.