Automated Extraction and Cross-Document Analysis of Aerospace Project Requirements Using LLMs

Kurzfassung

Requirements engineering remains a fundamental yet time-consuming activity for aerospace and other complex systems, where large project documents define mission- and safety-critical specifications. Traditional approaches are challenged by the scale, heterogeneity, and ambiguity of such documents, making it difficult to ensure traceability, consistency, and completeness. This paper introduces a human-in-the-loop requirements analysis pipeline that automates requirement extraction, classification, and verification for aerospace project documentation. The tool employs a hybrid architecture: a document-agnostic, Large Language Model (LLM)assisted parser induces Regular expressions (regex)-based rules to extract requirement blocks into a normalized database, while subsequent analysis pipelines evaluate the requirements from complementary perspectives. The structural pipeline identifies redundancy, traceability gaps, and document coverage issues; the semantic/logic pipeline leverages LLMs to assess ambiguity, conflicts, and requirement clarity. A human-in-the-loop review process validates and refines outputs for reliability in aerospace contexts. While the system is designed for on-premises execution to satisfy exportcontrol and privacy constraints, our experiments were conducted with cloud-hosted LLMs pending local deployment. An initial evaluation on a synthetic European Cooperation for Space Standardization (ECSS)-style corpus representative of spacecraft project documents demonstrates the feasibility of the approach, with strong performance in structured extraction and classification, and clear insights into current limitations of semantic reasoning. Across 225 requirements, leading models achieve Balanced Accuracy (BAcc) of 0.70-0.80 and we have 3% parse errors. The architecture provides a foundation for integration with formal modeling workflows and future verification tasks.