Autonomous Transportation System (ATS) is the leading orientation of a new round of transportation intelligent technology development in the world and safety is one of the core requirements of ATS. Currently, the safety assurance system of ATS is in its initial stage, which needs a systematic and scientific architecture for guidance. This paper aims to construct an autonomous safety assurance architecture based on the interaction of spatial-temporal whole-process traffic information to enhances the level of autonomous road safety assurance. Firstly, this paper gives a method for construct the architecture, including component generation, property labeling, inter-component relationships labeling, information flows generation, logical architecture construction and physical architecture construction. Secondly, this paper combines large language models (LLMs) and retrieval augmented generation (RAG) technique to construct a LLM fusion framework which is used to generate components of the autonomous road transportation safety assurance system. 100 components and 162 inter-component relationships are generated and they are visualized with Neo4j. Finally, this paper constructs the architecture from the logical and physical dimensions. The architecture provides functions of autonomous perception, autonomous decision and autonomous intervention, and provides architectural support for the comprehensive development of the autonomous road transportation safety assurance system.