在线近红外光谱预处理对废旧纺织品定性识别的影响

Influence of online near-infrared spectroscopy preprocessing on qualitative identification of waste textiles

将近红外光谱分析技术与一阶导数、离散小波变换、标准正态变换、多元散射校正、 S-G平滑、移动平均平滑、均值中心化和最大最小归一化8种预处理方法相结合,采用其单一及组合的方法,对聚酯、锦纶、腈纶、棉、毛、真丝、聚酯/棉、聚酯/锦纶、聚酯/氨纶、聚酯/毛、真丝/棉、锦纶/氨纶和特殊类共计13类织物的3620个近红外谱图进行预处理,并建立基于深度卷积神经网络的废旧纺织品定性识别模型。依据定性模型的识别准确率,探讨出适宜该类数据集的谱图预处理方法。研究结果表明,采用S-G平滑、均值中心化+S-G平滑和标准正态变换+S-G平滑的方法预处理后,所得模型的识别准确率均在96%以上。将此3种模型与未经预处理的原模型分别导入“分拣装置”中,对未参与建模的280个样品进行成分识别检验,经预处理优化后的模型识别准确率均高于原模型的89.6%。且均值中心化+S-G平滑预处理后,模型的识别准确率最高达96.8%,识别、分拣时间小于2 s。因此,对原模型样本的近红外光谱数据进行预处理可极大提高模型的识别准确率,为废旧纺织品的在线高效识别与自动分拣提供新方法。

Near-infrared spectral analysis technology has been combined with 8 spectral preprocessing methods including First Derivative,Discreet Wavelet Transform,Standard Normal Variate,Multiplicative Scatter Correction,S-G Smoothing,Moving Average Smoothing,Mean Centering and Max-Min Scaling.The single pretreatment and their combinations methods were used to optimize and pretreat 3620 near-infrared spectra of 13types of fabrics,including polyester,nylon,acrylic,cotton,wool,silk,polyester/cotton,polyester/nylon,polyester/spandex,polyester/wool,silk/cotton,nylon/spandex and special types.The qualitative identification model of waste textiles based on deep convolution neural network was established.According to the recognition accuracy of the qualitative model,the suitable spectral preprocessing method for this kind of data set has been discussed.The results showed that the recognition accuracy of the optimized model trained was above 96%after preprocessing with S-G Smoothing,Mean Centering+S-G Smoothing and Standard Normal Variate+S-G Smoothing.The three models and the original model were respectively introduced into the"sorting device",then 280 samples that did not participate in the modeling were identified and tested,and the recognition accuracy of the model after preprocessing optimization was higher than 89.6%of the original model.Among them,after Mean Centering+S-G Smoothing preprocessing,the actual recognition accuracy of the model was the highest of 96.8%,and identification and sorting time was less than 2 s.Therefore,preprocessing the near-infrared spectral data of the original model samples can greatly improve the recognition accuracy of the model,and provide innovative technologies for efficient online identification and automatic sorting of waste textiles.