Classification of recycled plastics using sparse and imbalanced spectral data and data augmentation by the generative...

Area of Science:

• Materials Science
• Data Science
• Environmental Science

Background:

• Accurate post-consumer plastic identification is crucial for sustainable recycling and circular economy initiatives.
• Recycled plastic spectra (FTIR) often suffer from noise, baseline shifts, and overlapping signals, leading to sparse and imbalanced datasets.
• These data complexities challenge conventional machine learning classifiers, increasing plastic misclassification rates.

Purpose of the Study:

• To investigate the effectiveness of data augmentation using generative adversarial networks (GANs) for improving polymer classification performance.
• To address challenges of data sparsity and class imbalance in spectral analysis of recycled plastics.
• To develop a robust method for online polymer classification systems.

Main Methods:

• Implemented a GAN framework with adversarial training and a classifier-guided feedback loop.
• Synthesized realistic, class-discriminative FTIR spectra for six common polymers: PE, PP, PS, PC, PET, and ABS.
• Trained multilayer perceptron classifiers on datasets with varying ratios of synthetic data.

Main Results:

• Optimal balanced accuracy of 96.2% was achieved with 50% synthetic data in the training set.
• Excessive synthetic data (>90%) led to degraded generalization performance.
• GAN-based augmentation significantly improved acrylonitrile butadiene styrene (ABS) classification accuracy, precision, and recall by 43%, 50%, and 33%, respectively, compared to no augmentation.

Conclusions:

• GAN-based data augmentation effectively mitigates data sparsity and class imbalance in spectral classification of common plastics.
• This approach provides a practical foundation for developing robust online polymer classification systems.
• Enhanced polymer identification supports high-performance recycling and a more sustainable circular economy.