Automated safety interpretation of crack-aperture for concrete dams using mutation-aware visual-sensor fusion and knowledge-augmented reasoning

Automated safety interpretation of concrete-dam cracks remains challenging because crack-gauge records may be incomplete, abnormal changes are difficult to distinguish from normal evolution, and monitoring outputs are not readily translated into traceable decisions. This paper proposes a mutation-aware visual–sensor fusion and knowledge-augmented framework. A deep learning model, multiscale dynamic spatial YOLOv11n (MDS-YOLOv11n), extracts crack widths from images and aligns them with crack-gauge measurements to supplement missing observations. Autoformer models the fused sequence; a dynamic three-sigma (3 σ ) criterion identifies persistent deviations, and a Heaviside–exponential moving average (EMA) reconstructs mutation-induced increments. Graph retrieval-augmented generation (GraphRAG) connects mutation-analysis outputs with dam-crack safety knowledge. Validation on a concrete dam showed better prediction performance than comparative models; GraphRAG produced recommendations for intensified monitoring and sensor inspection, consistent with periodic inspection records. The framework establishes an automated pathway from multimodal physical observation to crack-state interpretation and knowledge-driven decision support for operation and maintenance.