摘要:
针对目前卷烟生产线缺乏快速检测烟丝掺配比例方法的问题,利用高光谱成像技术和机器学习方法,采集不同掺配比例的混合烟丝光谱数据,探讨单一及组合预处理方法对模型构建效果的影响。采用偏最小二乘回归(PLSR)和支持向量机回归(SVR)建立回归模型,并利用最小角回归(LARS)、连续投影算法(SPA)、竞争性自适应重采样(CARS)及遗传算法(GA)进行特征波段选择,建立简化模型。结果表明:不同预处理方法的单一或组合使用均会影响回归模型的精度,其中小波变换与SG滤波联用(Wave+SG)方法相比原始数据将平均绝对百分比误差(MAPE)降低了1.2%;基于Wave+SG-GA-PLSR建立的回归模型表现最佳,其在两组分掺配中的训练集和测试集MAPE分别为1.415%和1.531%;基于两组分掺配建立的方法同样适用于多组分掺配,在三组分和四组分的预测集中,MAPE均低于8.3615%。高光谱成像技术结合机器学习方法能够较为准确地预测混合烟丝中各组分的掺配比例,可为烟丝掺配均匀性的在线监测及卷烟生产过程的质量控制提供参考。
Abstract:
No rapid method existed for detecting blend ratios on production lines, hyperspectral imaging technology and machine learning methods were used to collect spectral data from mixed tobacco with different blend ratios. The effects of single and combined preprocessing techniques on model performance were explored. Regression models were established using partial least squares regression ( PLSR) and support vector machine regression ( SVR). Feature wavelength selection was performed with least angle regression (LARS), successive projections algorithm (SPA), competitive adaptive reweighted sampling ( CARS), and genetic algorithm ( GA) to build simplified models. The results showed that preprocessing methods, either individually or combined, affected model accuracy. The combined wavelet transform and SG filtering ( Wave + SG) method reduced mean absolute percentage error (MAPE) by 1. 2 percentage points compared to raw data. The Wave+SG-GA-PLSR model performed best, with MAPE of 1. 415% and 1. 531% for the training and test sets of two-component blends, respectively. This method was also applicable to multi-component blends, with MAPE in three-component and four-component blends below 8. 3615%. Hyperspectral imaging combined with machine learning accurately predicted the proportions of components in mixed tobacco, providing a reference for online monitoring and quality control in cigarette production.
点击查看大图
下载:
