Accuracy improvement in plastics classification by laser-induced breakdown spectroscopy based on a residual network

Area of Science:

• Analytical Chemistry
• Materials Science
• Computer Science

Background:

• Plastic pollution poses a significant threat to ecosystems globally.
• Effective plastic recycling relies on accurate and automated material classification.
• Existing classification methods often lack the precision required for comprehensive recycling.

Purpose of the Study:

• To develop an accurate and efficient automated plastic classification system.
• To leverage residual networks and laser-induced breakdown spectroscopy (LIBS) for enhanced classification.
• To evaluate the method's performance, particularly in few-shot learning scenarios and against interference from additives.

Main Methods:

• Utilized a residual network architecture for plastic classification.
• Employed laser-induced breakdown spectroscopy (LIBS) to obtain spectral data.
• Compressed LIBS spectral data using a continuous wavelet-based peak searching algorithm.
• Transformed spectral data into characteristic images for network training and validation.
• Tested classification accuracy with few-shot learning (1 training image) and cross-manufacturer datasets.

Main Results:

• Achieved 100% classification accuracy in few-shot learning scenarios (1 training image).
• Demonstrated significantly higher accuracy compared to conventional algorithms (BP, kNN, SVM).
• Obtained 73.34% accuracy when testing against plastics with varying additives from different manufacturers.
• Confirmed the residual network's superior performance in plastic classification.

Conclusions:

• The proposed residual network method offers a highly accurate approach for plastic classification using LIBS.
• This technique shows great potential for improving the efficiency and effectiveness of plastic recycling industries.
• The method contributes to pollution mitigation efforts by enabling better plastic waste management.