upipes/PipeNet-Expansion

What it solves

It addresses the problem of "domain shift" in Ground Penetrating Radar (GPR) subsurface diagnosis. Because GPR signals change based on local soil composition, water content, and road structure, a classifier trained in one geographical area often fails when applied to a new area. This project reduces the need for large amounts of labeled data in new areas by using environmental knowledge to adapt existing classifiers.

How it works

The system uses a combination of physics-aware knowledge and machine learning to bridge the gap between different environments:

1. Knowledge Generation: It uses Large Language Models (LLMs) to generate descriptions of how subsurface targets appear under different imaging conditions.
2. Adaptation: It employs a framework consisting of Physics-aware Knowledge Generation (PKG), Iterative State Refinement (ISR), and Causal Transformation Alignment (CTA) to adapt the weights of a classifier from an original area to a new area.
3. Prototype System: A full-stack application (Vue frontend, Django backend, MySQL database) allows users to configure area definitions, generate knowledge, train adapted models, and visualize results via activation maps.

Who it’s for

Researchers and engineers specializing in non-destructive infrastructure inspection, GPR signal processing, and cross-domain machine learning adaptation.

Highlights

• LLM-Guided Adaptation: Leverages LLMs to describe environmental impacts on GPR signals to guide model training.
• Cross-Area Transfer: Specifically designed to move diagnostic capabilities from one subsurface environment to another with minimal new labels.
• Full-Stack Prototype: Includes a complete system for managing the entire workflow from knowledge generation to model evaluation.
• Comprehensive Benchmarking: Includes implementations of multiple baseline comparison methods (e.g., DANN, ADDA, G2KD).