Abstract: Monitoring the stability of tailings pond slopes and accurately locating potential slip surfaces are key aspects of mine safety production. The traditional sliding surface positioning method has problems such as poor data continuity, insufficient ability to capture spatiotemporal heterogeneity, and weak adaptability to complex working conditions（such as seepage stress coupling）, making it difficult to meet the requirements of dynamic real-time warning. Therefore, a dynamic positioning technology for potential slip surfaces in tailings ponds based on YOLOX_s is proposed. At the level of multi-source data fusion, a unified encoding framework for time-series numerical, spatial distribution, and visual monitoring data has been constructed. Attention mechanism is introduced for dynamic weighted fusion, generating feature maps that can comprehensively reflect the spatiotemporal heterogeneity characteristics of slip surfaces. This fundamentally solves the problem of effective integration and deep utilization of multi-source heterogeneous data. At the level of dynamic positioning models, the lightweight object detection model YOLOX has been introduced for the first time into the field of tailings pond slip surface localization. By utilizing its CSPDarknet backbone network and spatial pyramid pooling structure, it adaptively extracts multi-scale channels and spatial features, significantly enhancing its ability to recognize fuzzy boundaries and small-scale evolution features of slip surfaces. At the engineering applicability level, through an end-to-end deep learning architecture, data fusion, feature extraction, and localization regression are integrated to achieve real-time and automated interpretation of slip surface positions. To verify the effectiveness of the proposed method, a practical tailings pond project is used as a case study for testing and analysis. Case tests have shown that the dynamic spatial positioning entropy of this technology is less than 0.5, and the spatiotemporal consistency index is higher than 0.95. It can accurately and stably locate potential slip surfaces in tailings ponds and is suitable for long-term safety monitoring. It can provide a new effective way to improve the real-time and accuracy of tailings pond safety monitoring and early warning.