基于EIS-CARS的微塑料FTIR特征光谱选择及ANN识别

Selection of FTIR Characteristic Spectra of Microplastics Based on EIS-CARS and ANN Recognition

摘要: 针对运用中红外光谱建立微塑料(MPs)识别的人工神经网络(ANN)建模效率和精度低的问题，提出了一种等间隔采样(EIS)耦合竞争自适应重加权采样算法(CARS)的特征光谱选择方法(EIS-CARS)。以20种典型MPs为研究对象，讨论了EIS-CARS与CARS选择特征光谱的差异性，并对比分析了特征光谱和全光谱条件下不同ANN模型对MPs的识别性能。试验结果表明：基于EIS-CARS算法共选择44个特征波数点，仅为CARS算法的33.85%，但是特征光谱的选择时间减少了97.25%，而且交互验证均方根误差(RMSECV)也略有下降；特征波数点主要集中在1802~1385 cm⁻¹、1307~1164 cm⁻¹、1073~956 cm⁻¹、774~500 cm⁻¹四个波数段，与MPs主要官能团的波数分布对应；基于特征光谱构建的单层神经网络(SNN)与多层神经网络(MNN)模型均表现出较快的收敛速率和较高的识别精度，其中MNN模型对待测MPs的正确识别率达到99.17%。基于EIS-CARS能够在不损失识别精度的情况下，显著提升了建模效率，特别是大幅缩减了计算成本高昂的超参数寻优时间。研究结果可为塑料的识别分类、研究MPs的环境行为及制定防治措施提供技术参考。

Abstract: Aiming at the low efficiency and accuracy of artificial neural network (ANN) modeling for microplastic (MP) identification using mid-infrared spectroscopy, this study proposes a novel feature spectral selection method called EIS-CARS, which combines equal interval sampling (EIS) with competitive adaptive reweighted sampling (CARS). Using 20 typical MPs as research objects, the differences between EIS-CARS and CARS in feature spectral selection were explored, and the recognition performance of various ANNs under full and feature spectra conditions was compared. Results show that EIS-CARS selects 44 feature wavenumbers, just 33.85% of CARS, cuts feature selection time by 97.25%, and slightly reduces RMSECV. The chosen feature wavenumbers mainly focus on four ranges: 1802~1385 cm⁻¹, 1307~1164 cm⁻¹, 1073~956 cm⁻¹, and 774~500 cm⁻¹, corresponding to the wavenumber distribution of MPs’ main functional groups. Both single neural network (SNN) and multilayer neural network (MNN) models built on the feature spectra show fast convergence and high recognition accuracy. The MNN model achieves a correct MP recognition rate of 99.17%. The ANN model based on EIS-CARS-selected feature spectra matches the full-spectrum ANN in correct recognition rate but with higher modeling efficiency, making it an excellent MP identification method. This study offers technical insights for plastic identification and classification, and is highly valuable for understanding MP environmental behavior and developing preventive measures.