PSO-LSSVM用于近红外光谱预测混合制浆材Klason木质素含量

Application of Particle Swarm Optimization Based on Least Square Support Vector Machine in Quantitative Analysis of Klason Lignin of Mixed Wood Using Near-infrared Spectroscopy

中文关键词:  近红外光谱  混合制浆材  Klason木质素  粒子群寻优算法  最小二乘支持向量机

Key Words:near-infrared spectroscopy  mixed pulping wood  Klason lignin  particle swarm optimization  least square support vector machine

中文摘要:

      为优化混合制浆材中Klason木质素含量的近红外分析模型，收集了5种常见制浆材的82个原木样品，将样品粉碎预处理后在便捷式近红外光谱仪上采集其近红外光谱信号，对原始光谱数据进行多元散射校正（MSC）预处理，利用粒子群寻优（PSO）算法对最小二乘支持向量机（LSSVM）算法中的参数进行优化，然后利用最优参数建立混合制浆材Klason木质素的LSSVM定量分析模型。将结果与偏最小二乘（PLS）和主成分降维后的BP神经网络（PCA-BPNN）算法进行比较。结果表明，PCA-BPNN和PSO-LSSVM模型均优于PLS模型，且PSO-LSSVM模型预测结果最优，预测结果的相关系数（R _v）最大为0.9857；预测标准差（RMSEP）为0.7498%，比PLS模型和PCA-BPNN模型分别降低了0.2767%和0.1455%；相对标准偏差（RPD）最大为5.6174，比PLS模型和PCA-BPNN模型分别提高了1.5144和0.9138；真实值与预测值间的绝对偏差（AD）范围最小，为0.0065%~1.8449%。

Abstract:

      To optimize the near-infrared analysis model of Klason lignin content of mixed wood, 82 samples from 5 common pulp woods were collected and prepared by grinding pretreatment, and their near-infrared spectrum signals were obtained via the portable spectrometer. The division of calibration set and testing set was followed by multiplication scatter correction (MSC). The parameters of least square support vector machines (LSSVM) were optimized by particle swarm optimization (PSO), and the selected parameters were used to establish LSSVM quantitative analysis model of Klason lignin content of mixed wood. The performance of PSO-LSSVM model was compared with partial least squares regression (PLS) model and principal component analysis combined with back propagation artificial neural network (PCA-BPNN) model. The calibration and validation results of PCA-BPNN and PSO-LSSVM were superior to PLS, and the PSO-LSSVM model showed the best performance. For PSO-LSSVM model, the correlation coefficients of validation (R _v) was 0.9857 which was the largest among the three models. The root mean square error of prediction (RMSEP) of PSO-LSSVM model was 0.7498%, which was 0.2767% and 0.1455% lower than PLS model and PCA-BPNN model, respectively. The relative standard deviation (RPD) was 5.6174 which was improved 1.5144 comparing to the PLS model and 0.9138 comparing to PCA-BPNN model. And the maximum and minimum absolute errors (AD) were 0.0065%~1.8449%, which were the least among the three models.