摘要
分析化学中传统的多元校正通常采用线性回归或人工神经网络算法。但线性回归不能适应实测数据或多或少的非线性,而人工神经网络又有过拟合弊病造成误差。为此我们提出用新发展的既能处理非线性数据,又能限制过拟合的支持向量机算法。本文首次提出导数光谱-支持向量回归法。该法用于NO_3^--NO_2^-体系的同时测定解得的浓度平均相对误差在±82%,明显好于ANN法(±9.15%)和线性回归法(±11.5%)。这表明支持向量机算法在分析化学的校正技术中是有用的。
Linear regression and artificial neural network are usually used in the multivariate calibration work in analytical chemistry. But linear regression is difficult to fit the nonlinearity of experimental data, while ANN method often exhibits overfitting. Both of these problems may lead to errors in computation. Therefore, a new method, support vector regression, which can fit nonlinear data and can depress overfitting at the same time, is first applied to multivariate calibration for derivative spectrum of NO3- - NO2- system. The relative analyzing errors are within ± 8.2% . It is lower than the error by ANN( ±9.15%) or linear regression( ±11.5%). So it appears that this new method is useful for calibration work in analytical chemistry.
出处
《计算机与应用化学》
CAS
CSCD
北大核心
2002年第6期752-754,共3页
Computers and Applied Chemistry
基金
国家自然科学基金(20175013)
上海市高校科技发展基金(01A17)