Sensor-based particle mass prediction of lightweight packaging waste using machine learning algorithms.

Area of Science:

• Waste Management and Recycling
• Machine Learning Applications
• Sensor Technology

Background:

• Sensor-based material flow characterization (SBMC) requires accurate particle mass data for advanced sorting plant applications.
• Inline measurement of individual particle masses is not feasible, necessitating predictive methods.
• Lightweight packaging (LWP) waste presents unique challenges for mass prediction due to design and disposal variations.

Purpose of the Study:

• To investigate the efficacy of machine learning (ML) algorithms for predicting the particle masses of LWP waste.
• To compare the performance of ML models against established reference methods for particle mass prediction (PMP).

Main Methods:

• A dataset of 3,830 LWP particles was created, including 3D laser triangulation (3DLT) and RGB images with known masses.
• Sixty-six shape measurements were extracted from the sensor data to train six ML models for PMP.
• ML models were evaluated against reference methods using mean particle masses and grammages, with feature selection applied.

Main Results:

• Particle masses exhibited significant variation across different material and size classes.
• All investigated ML models outperformed the two state-of-the-art reference models.
• A random forest regressor achieved the highest accuracy, with an R²-score of 0.763 ± 0.091 and a normalized mean absolute error of 0.243 ± 0.050.

Conclusions:

• Machine learning algorithms offer a promising and effective approach for the particle mass prediction (PMP) of LWP waste.
• ML-based PMP significantly enhances accuracy compared to traditional methods, improving SBMC capabilities.
• The developed ML models provide a robust solution for overcoming the challenges in predicting LWP waste particle masses.