FabricSpec-RAG: A knowledge graph-augmented Seq2Seq framework for quantitative analysis of complex textile blends...

Area of Science:

• Materials Science
• Analytical Chemistry
• Computer Science

Background:

• Quantitative analysis of complex textile blends using Near-Infrared (NIR) spectra is crucial for industrial recycling.
• Challenges include spectral overlap, compositional variability, and limitations of conventional deep learning models with rare compositions and transparency.
• Existing methods struggle with the nuances of diverse textile materials.

Purpose of the Study:

• To develop an advanced framework for accurate quantitative analysis of complex textile blends from NIR spectra.
• To overcome limitations of conventional deep learning models in handling rare compositions and providing transparency.
• To enhance the efficiency and reliability of industrial textile recycling processes.

Main Methods:

• Proposed FabricSpec-RAG, a Sequence-to-Sequence framework integrating a multi-scale convolutional encoder.
• Implemented a Retrieval-Augmented Generation (RAG) mechanism grounded in a hierarchical, evolving knowledge graph.
• Conducted extensive experiments comparing against chemometric, machine-learning, and deep-learning baselines.

Main Results:

• Achieved a Micro-F1 score of 0.9880 for component identification and a Mean Absolute Error (MAE) of 0.0026 for ratio prediction.
• Demonstrated superior performance compared to established baseline methods.
• Interpretability analysis confirmed the model's focus on chemically relevant spectral features.

Conclusions:

• FabricSpec-RAG offers a robust and interpretable solution for analyzing complex textile blends via NIR spectroscopy.
• The framework's continuous adaptation capability allows for incremental knowledge base updates without costly retraining.
• This advancement significantly supports sustainable industrial textile recycling by improving material analysis accuracy and adaptability.