Bridge report generation method and system based on multi-modal causality and reinforcement learning

By employing multimodal causal and reinforcement learning methods, the problems of discrepancies between images and text and missing causal reasoning in bridge inspection reports have been solved, achieving high efficiency, logical consistency, and self-evolution capabilities in the reports, resulting in a significant improvement in the quality of the generated reports.

CN122088460BActive Publication Date: 2026-09-15CHINA RAILWAY SOUTHWEST SCI RES INST CO LTD
3 Cites 0 Cited by

Patent Information

Application Number
CN202610453377.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-04-08
Publication Date
2026-09-15
Estimated Expiration
2046-04-08

AI Technical Summary

Technical Problem

Existing technologies for bridge inspection suffer from multimodal semantic gaps, lack of causal reasoning, and lack of self-evolution capabilities, resulting in frequent discrepancies between text and images and inappropriate maintenance recommendations in the generated bridge reports.

Method used

We employ multimodal causality and reinforcement learning methods, use a visual language model to verify the consistency between text and images, construct a causal knowledge graph of bridge defects for causal reasoning, and introduce a multi-agent architecture and an engineering domain reward model for self-evolutionary optimization.

Benefits of technology

It achieves consistency verification of text and images, integrity of causal logic, and self-evolution capabilities, significantly improving the accuracy and logical consistency of generated reports, and also significantly increasing report generation efficiency.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The application discloses a bridge report generation method and system based on multi-modal causality and reinforcement learning. The method comprises the following steps: acquiring disease images and text descriptions, extracting structured disease information from the text descriptions, generating multi-modal feature data and performing consistency verification of images and texts; based on the verified disease information, searching for standard provisions and historical cases, positioning the disease node in the causality knowledge graph and generating an inference chain by reverse tracing the root cause node; fusing the multi-modal feature data, the search results and the inference chain, generating an evaluation report draft through a large language model and performing logical verification; acquiring expert correction text input, inputting a reward model to calculate a feedback score, and optimizing the edge weight of the knowledge graph and the report generation strategy. The application also provides a system for implementing the above method. Through the consistency verification of images and texts, the causal reasoning and the reinforcement learning, the application effectively improves the accuracy, the interpretability and the compliance of the report.
Need to check novelty before this filing date? Find Prior Art