Sensor-Based Sorting of Post-Consumer Plastic Packaging

Abstract

Starting from the 1950s, plastic has found its way into many aspects of life; from packaging to transportation and construction. The annual global plastic production was estimated to be 330 million metric tons in 2017 and will be doubled within 20 years if the current growth rate persists. In Europe, packaging accounts for 40% of the annual plastic production. To decrease dependence on fossil fuels and the related emissions and reduce the negative impact of plastics that enter the environment, the EU targets to recycle 55% of plastic packaging material by 2025. A steady production of large quantities of high-grade plastic recyclate is essential to sell the recycled material on the market and substitute virgin plastics.
But, achieving this high grade is difficult due to the wide range of polymers, including additives and fillers. This research focuses on the sorting step of post-consumer packaging waste in a Material Recovery Facility (MRF) that takes place before recycling.

In MRF's, Near Infrared (NIR) sensor-based sorting systems are state of the art to create high-grade sorting products. NIR's use reflection spectra to identify and distinguish a range of materials including different types of polymers. However, while sorting, False Calls (FC) occur when particles are incorrectly classified leading to a loss of valuable material or contamination of the product. Thus, this research aims to increase the understanding of false calls of NIR sensor systems, contributing to higher quality sorting products.

To reach this goal, possible causes of FC's and their dependencies on the feedrate, feed composition and feeding mechanism were identified. These causes were arranged into a framework consisting of six types of FC's. A distinction was made between errors that are working range related and those that are not. Particles outside the working range may be classified incorrectly even when fed in a monolayer and under perfect circumstances due to the properties of the particle.

The created framework was applied to a NIR in an MRF in the Netherlands which was tasked to separate non-plastic particles and PVC from plastic particles. Experiments were performed to single out the probability of each type of error to occur for 13 material classes. Next, the test results were implemented in a statistical model and a Monte Carlo simulation was performed. In this research, the varying input values were used to imitate the changing feed composition. Three scenarios of adjusted feed characteristics were analysed using the statistical model.

It was concluded that to optimize the sorting performance of the analysed NIR, all types of False Calls should be tackled. The feed of the analysed NIR contained a large share of particles outside the working range. So, even if the particles were to be fed in a monolayer the average grade of the product is expected to be around 85%. The combination of the experiments and the statistical model allows for an effective evaluation of the sorting performance and a clear indication of the most problematic aspects of the feed characteristics.