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

The quantitative analysis of complex textile blends from Near-Infrared (NIR) spectra is critical for industrial recycling but is hampered by strong spectral overlap and compositional variability. Conventional deep-learning models often struggle on rare compositions and lack sufficient transparency. To address these limitations, we propose FabricSpec-RAG, a Sequence-to-Sequence framework that couples a multi-scale convolutional encoder with a Retrieval-Augmented Generation (RAG) mechanism, grounding predictions in a hierarchical, evolving knowledge graph. Extensive experiments show that the framework compares favourably with representative chemometric, machine-learning and deep-learning baselines, achieving a Micro-F1 score of 0.9880 in component identification and a Mean Absolute Error (MAE) of 0.0026 in ratio prediction. Interpretability analysis confirms that the model attends to chemically valid spectral features. In addition, the system is designed for continuous adaptation, allowing the knowledge base to be updated incrementally as new materials appear, without expensive retraining.