A RAG-enhanced multi-agent LLM/VLM framework for auditable and region-aware structural health monitoring of a historic masonry structure

This research-oriented project explores an advanced, multi-agent artificial intelligence framework designed for the non-destructive and interpretable structural health monitoring (SHM) of historic masonry structures, demonstrated on a minaret case study. The system integrates signal processing, computer vision, and natural language processing through a coordinated manager-orchestrator architecture. Vibration data from four sensors undergo a systematic pipeline—including cleaning, power spectral density (PSD) baseline characterisation, adaptive windowing, and robust normalisation—before anomaly screening via a 1D convolutional autoencoder. To localize damage, the framework employs relation-based autoencoders on sensor pairs alongside sensor-to-zone mapping.

Simultaneously, a Vision-Language Model (VLM) processes inspection images, while a Retrieval-Augmented Generation (RAG) system provides context from a structural memory package containing geometry, materials, baseline dynamics, and finite element notes. A three-agent decision layer (Agent A for LLM+RAG, Agent B for VLM+RAG, and Agent C for cross-modal fusion) reconciles conflicting evidence to deliver global and zone-level risk assessments. This framework offers M.Tech students a state-of-the-art methodology for combining deep learning anomaly detection with generative AI reasoning in structural engineering.

Message: Hi FE, I need help with “A RAG-enhanced multi-agent LLM/VLM framework for auditable and region-aware structural health monitoring of a historic masonry structure” in “Civil Engineering”