A chemometric approach for FTIR-Raman-LIBS tri-modal spectral fusion: Transformer-based accurate identification of...

Plastic waste management is a key challenge for a sustainable circular economy. Traditional sorting technologies are inefficient, labor-intensive, and prone to errors, leading to increased landfill rates for recyclable materials. The rapid development of spectral technology and deep learning has opened up new avenues for addressing this issue. Different spectral types reveal distinct characteristics of polymers, and leveraging these features can significantly enhance sorting accuracy. The variety of deep learning models available presents a significant challenge in selecting the most suitable one. This study proposes an innovative chemometric method that, for the first time, applies the Transformer deep learning architecture-which has significant advantages for analyzing long sequences-to the fusion analysis and cross-modal generation of three spectral modalities: FTIR, Raman, and LIBS. It also introduces a genetic algorithm to optimize Transformer hyperparameters, enhancing model robustness and achieving high-precision classification of recyclable plastic polymers. Experimental results show: (1) Cross-modal data augmentation effectively addresses the issue of sample imbalance, improving classification accuracy by 3.84 %; (2) The multi-modal fusion strategy significantly captures spectral structural correlations, achieving an accuracy rate of 97.69 %, which is 9.74 % higher than the optimal single-modal model; (3) After hyperparameter genetic algorithm optimization, the model achieved an accuracy rate exceeding 99.23 % on an independent test set, demonstrating a significant advantage. These results indicate that this work provides a practical new paradigm for spectral analysis technology in the classification of recyclable plastics, holding direct value for advancing the application of analytical chemistry in sustainable technologies. More importantly, it also offers scalable intelligent tools for the application of spectral analysis in other fields.