Plastic pollution is a global issue of increasing health concern, thus requiring innovative waste management. In particular, there is a need for advanced methods to identify and classify the different types of plastics. Near-infrared spectroscopy is currently operational in some waste-sorting facilities, yet remains challenging to discriminate different black plastics because black targets have low reflectance in some spectral regions. Here we used partial least squares discrimination analysis, soft independent modeling of class analogy, linear discriminant analysis and convolutional neural network to classify the plastics. We analyzed 159 plastic samples, including 84 black plastics, made of high impact polystyrene, acrylonitrile butadiene styrene, high-density polyethylene, polyethylene terephthalate, polyamide 66, polycarbonate and polypropylene. Results show that the convolutional neural network model yielded an accuracy up to 98%, whereas other models displayed accuracy of 57–70%. Overall, convolutional neural network analysis of infrared plastic data is promising to solve the bottleneck problem of black plastic discrimination.

塑料污染是一个日益引起健康问题的全球性问题，因此需要创新的废物管理。尤其需要先进的方法来识别和分类不同类型的塑料。近红外光谱目前在一些废物分类设施中运行，但由于黑色目标在某些光谱区域具有低反射率，因此区分不同的黑色塑料仍然具有挑战性。在这里，我们使用偏最小二乘判别分析、类类比的软独立建模、线性判别分析和卷积神经网络对塑料进行分类。我们分析了 159 个塑料样品，包括 84 个黑色塑料，由高抗冲聚苯乙烯、丙烯腈丁二烯苯乙烯、高密度聚乙烯、聚对苯二甲酸乙二醇酯、聚酰胺 66、聚碳酸酯和聚丙烯。结果表明，卷积神经网络模型的准确度高达 98%，而其他模型的准确度为 57-70%。总的来说，红外塑料数据的卷积神经网络分析有望解决黑色塑料识别的瓶颈问题。