Xu Zhiying, Zhao Xingyu, Peng Xinying, Wang Kai, Wang Kedong, Zhao Nan, Li Jiaming, Zhang Qingmao, Yan Xueqing, Zhu Kun
State Key Laboratory of Nuclear Physics and Technology, and Key Laboratory of HEDP of the Ministry of Education, CAPT, Peking University, Beijing, 100871, China; Guangdong Institute of Laser Plasma Accelerator Technology, 510540, China.
Guangdong Provincial Key Laboratory of Nanophotonic Functional Materials and Devices, School of Information and Optoelectronic Science and Engineering, South China Normal University, Guangzhou, 510006, China; State Key Laboratory of Optic Information Physics and Technologies, South China Normal University, Guangzhou, 510006, China.
Chemosphere. 2025 Jun;378:144412. doi: 10.1016/j.chemosphere.2025.144412. Epub 2025 Apr 15.
Efficient recycling of plastics is critical for environmental sustainability. In this work, an efficient and anti-interference method for plastic classification based on one-shot learning and laser-induced breakdown spectroscopy (LIBS) was proposed. A residual neural network model with full-spectrum training (ResNet-FST) was developed based on convolutional neural networks, achieving an accuracy of 99.65 % in one-shot learning classification. A multi-parameter peak search algorithm was employed to extract key spectral features, and a linear residual classification model with peak auto-search (LRC-PAS) was developed to further enhance efficiency. The number of residual blocks and neurons was optimized to 2 and 80, respectively. Compared with ResNet-FST, LRC-PAS significantly improved classification efficiency. The mechanism underlying the spectral interference caused by plastic additives in LRC-PAS was elucidated. The anti-interference of additives in LRC-PAS was achieved with high accuracy. The results demonstrated that the proposed method achieves highly efficient and anti-interference classification of plastics, demonstrating great potential for real-time classification in the recycling industry.
塑料的高效回收对于环境可持续性至关重要。在这项工作中，提出了一种基于一次性学习和激光诱导击穿光谱（LIBS）的高效且抗干扰的塑料分类方法。基于卷积神经网络开发了一种全光谱训练的残差神经网络模型（ResNet-FST），在一次性学习分类中达到了99.65%的准确率。采用多参数峰值搜索算法提取关键光谱特征，并开发了一种具有峰值自动搜索功能的线性残差分类模型（LRC-PAS）以进一步提高效率。残差块和神经元的数量分别优化为2和80。与ResNet-FST相比，LRC-PAS显著提高了分类效率。阐明了LRC-PAS中塑料添加剂引起光谱干扰的机制。LRC-PAS中添加剂的抗干扰实现了高精度。结果表明，所提出的方法实现了塑料的高效且抗干扰分类，在回收行业的实时分类中显示出巨大潜力。
