With rapid urbanization and surging traffic volumes, traffic accident data have become high-dimensional, multi-source, heterogeneous, and spatiotemporally dynamic, posing challenges for traditional statistical methods and machine learning models to simultaneously account for data heterogeneity and nonlinear interactions. Knowledge graphs, by constructing structured semantic networks that integrate accident events, participants, environmental factors, and other multidimensional elements, inherently support multi-source information fusion and reasoning. In this study, following a top-down ontology design principle, we construct a California Traffic Accident Knowledge Graph (TAKG) encompassing over one hundred elements, and propose an enhanced TransD embedding model. Our model introduces entity–attribute projection vectors into the dynamic mapping mechanism to explicitly encode domain attributes, and designs a dual-limit scoring loss function to independently regulate the positive and negative sample boundaries. Experimental results demonstrate that our method significantly outperforms traditional translation-based models on the self-built TAKG as well as on the FB15K-237 and WN18RR benchmark datasets. This research provides a solid data foundation and algorithmic support for downstream traffic accident risk prediction and intelligent traffic safety management.