基于塑料近红外光谱的判别分类研究

Discriminative Classification of Plastics Based on Near-infrared Spectra

中文关键词: 塑料分类  近红外  高光谱成像  主成分分析  聚类分析

英文关键词: classification of plastics  NIR  hyperspectral imaging  PCA  cluster analysis

中文摘要:

      目的 为了回收可用于不同物品包装的塑料，对不同塑料种类进行识别分类。方法 首先采集PP、PET、HDPE、TPE、PLA、PBT、TPU、POM-M90、PPO-GF20NC、TPB、PPS、ABS、PPO(natural)、SAN、POM-F20、PPO(white)16种塑料的近红外光谱数据，其次针对光谱数据采集时存在的噪声问题，使用SG平滑滤波进行了光谱数据预处理，之后利用主成分分析算法进行光谱数据降维，减少待处理数据量，最后分别运用无监督聚类K-means算法和监督聚类极大似然估计、Fisher判别式以及光谱角算法建立4类分类模型。结果 K-means算法可以将PPO-GF20N、PLA和PPO(本色)与其他塑料粒子区分开，准确率分别是100%、100%以及80%;Fisher判别式和极大似然估计法对POM-M90和POM-F20的识别准确率为93%，其他塑料粒子识别准确率均为100%;光谱角算法对PET的识别准确率为80%，POM-F20的识别准确率为47%，其余粒子的识别准确率均大于90%。结论 上述机器学习算法结合近红外光谱成像技术建立分类模型可为常见塑料的鉴别研究提供参考。

英文摘要:

      The work aims to identify and classify different types of plastics, in order to recover plastics that can be used to pack different items. Firstly, the near-infrared spectral data of 16 kinds of plastics including PP, PET, HDPE, TPE, PLA, PBT, TPU, POM-M90, PPO-GF20NC, TPB, PPS, ABS, PPO (natural colour), SAN, POM-F20 and PPO (white colour) were collected. Then, for the problem of noise in spectral data collection, the spectral data were pre-processed by the SG smoothing filtering, followed by dimensionality reduction of the spectral data with the principal component analysis algorithm to reduce the amount of data to be processed, and finally the four-class classification model was established by the K-means algorithm for unsupervised clustering and the great likelihood estimation for supervised clustering, the Fisher discriminant, and the spectral angle algorithm, respectively. The K-means algorithm could distinguish PPO-GF20N, PLA and PPO (native colour) from other plastic particles with an accuracy of 100%, 100%, and 80%, respectively. Fisher's discriminant and great likelihood estimation had an accuracy of 93% for the recognition of POM-M90 and POM-F20, and 100% for the recognition of all other plastic particles. Spectral angle algorithm had a recognition accuracy of 80% for PET, 47% for POM-F20, and an accuracy greater than 90% for the rest of the particles. The above machine learning algorithm combined with near-infrared spectral imaging technology can be used to establish a classification model, providing a reference for the identification research of common plastics.