用于紫外光谱水质分析的Boosting-偏最小二乘法

A Boosting-Partial Least Squares Method for Ultraviolet Spectroscopic Analysis of Water Quality

为提高水质参数总有机碳(TOC)的紫外吸收光谱分析的预测精度,提出一种基于Boosting理论的迭代式回归建模算法,并根据统计学习理论提出一种新的迭代停止判据,可有效防止过拟合,显著提高模型预测精度。为评估所提算法的性能,分别采用本算法和3种常用的光谱分析方法,即偏最小二乘、主成分回归和人工神经网络,对自行研制的紫外光谱水质分析仪实测的一组数据进行了建模和预测。计算结果表明:相对于其他3种方法,本算法具有生成的模型预测精度高的显著优势。

A novel iterative regression modeling method, boosting partial least squares （BPLS） , based on boosting theory was proposed. It was used for improving the prediction precision of ultraviolet （UV） spectral model of water quality parameter total organic carbon （TOC）. And a new stopping criterion of iterations according to statistical learning theory was developed for BPLS to effectively avoid overfitting. With the new stopping criterion, BPLS has the advantages of notable improvement of prediction performance and easy computation. To evaluate the performance of the proposed method, a set of TOC-UV spectral data was measured by a self-developed UV spectroscopic water quality analyzer, and this method and other three common approaches in spectral analysis, such as partial least squares, principal component regression and artificial neural network, were used in modeling and prediction experiments respectively. The experimental results show that BPLS has the better prediction precision compared with other three conventional methods.