Edge-Deployed Generative Language-Based Retrieval for Aerospace Asset Health Management

Abstract

Aerospace asset health management increasingly relies on access to large volumes of maintenance documentation; however, operational environments are often constrained by limited connectivity and computational resources, restricting the use of cloud-based intelligence systems. This paper presents a fully offline, edge-deployable retrieval-augmented generation (RAG) framework for aerospace maintenance and prognostics using technical documentation. The framework integrates a locally hosted lightweight large language model with vector retrieval and cross-encoder reranking to support natural language querying of Airworthiness Directives (ADs) and maintenance records. Deployed on an NVIDIA Jetson Orin Nano 8GB device, the system performs document ingestion, indexing, retrieval, reranking, and response generation entirely on-device without cloud connectivity. Experimental evaluation on real aerospace maintenance documents demonstrates the ability to identify failure mechanisms, failure modes, root causes, affected components, and maintenance procedures described in FAA documentation. The cross-encoder reranking stage improves retrieval precision by refining semantically overlapping maintenance evidence prior to generation. The framework achieved an average fidelity score of 0.83, indicating that most generated responses remained grounded in retrieved FAA evidence. Across representative AD queries, the system achieved practical edge inference latency of approximately 4–10 seconds on the Jetson platform. The results demonstrate the feasibility of privacy-preserving, low-latency generative artificial intelligence for aerospace maintenance decision support on resource-constrained edge devices.