Th34 - Sorting Through the Fibers: Non-Destructive Classification of Textile Blends using Fiber Optic Reflectance Spectroscopy and Multivariate Statistical Analysis

Textiles are chemically diverse and commonly classified as natural (cotton, silk, wool), regenerated (viscose, rayon, cellulose acetate), or synthetic (polyester, polyamide, polyacrylic). Most garments today consist of tightly woven blends of these fiber types. Accurate fiber identification is essential in fields like fabric recycling, quality control, international trade, cultural heritage, and forensic science, where knowing the exact fiber type directly affects how a material is processed, conserved, handled, or analyzed. Conventional methods, such as burn tests, chemical solubility, and polarized light microscopy, require sampling, specialized expertise, and often fail to identify blends.

This study presents a novel, non-destructive approach using Fiber Optic Reflectance Spectroscopy (FORS) combined with machine learning to enable rapid classification of textile blends. FORS collects reflected light from samples in the UV-VIS-NIR range (350–2500 nm), providing rich spectral data without damaging the sample. When paired with chemometric modeling, such as Principal Component Analysis (PCA) and Partial Least Squares Discriminant Analysis (PLS-DA), FORS shows promise as a high-throughput technique for fiber classification.

Preliminary results using laboratory-dyed, certified reference fabrics demonstrate that the method can differentiate fiber types independent of dye chemistry. Spectral preprocessing techniques, including Savitzky-Golay smoothing and Normalization, improved class separation. PLS-DA models successfully classified major textile categories (natural vs. synthetic). More importantly, identification of 8 different textiles was obtained offering both high predictive accuracy and chemical interpretability as validated by a Cohen’s Kappa of 0.82. However, some classes remain indistinguishable due to chemical similarity which results in failing to separate cotton and linen or viscose and rayon.

Our findings represent a meaningful step toward integrating FORS and machine learning into textile sorting workflows. Compared to more common FTIR-ATR applications in polymer recycling, this method offers a broader spectral range and non-invasive benefits suitable for complex textile matrices. Future work will expand the model to address multi-label classification, allowing for the identification of multiple fiber types in a single sample, and will explore regression models to estimate blend ratios.