中文摘要：

近红外光谱（NIRS）可以检测溶解于油中的水分含量,但油中水分较多时会散射而非吸收NIRS,从而引起较大误差。为此,筛选非离子型表面活性剂（Span-80）将含水油液稳定分散成小颗粒,利用其NIRS数据建立水分含量的支持向量回归模型。实验中油水稳定化将NIRS测定变压器油中水分含量的上限从传统的0.1%提升到1%（V/V）,通过应用连续投影算法,在511个NIRS变量中筛选出15个有效变量（占原变量的2.9%）,建立的支持向量回归模型对验证集的预测均方根误差为2.93%,相关系数为0.9944,相对分析误差为9.4732。

英文摘要：

Near infrared spectroscopy （NIRS） is capable of determining water contents in oils. However, too much moisture contents in the oils will scatter rather than absorb the NIRS. This may cause greater measurement error. For this reason, a nonionic surfactant （Span-80） was screened to make the water in the oils evenly dispersed into small droplets. The NIRS analysis was subsequently employed to build support vector regression （SVR） model of the water content. In this experiments, the upper limit of the water content determination was improved from the conventional 0. 1% to 1.0% （V/V） by the oil-water stabilization. Applying successive projection algorithm, 15 valid variables （2.9% of the original ones） from 511 NIRS variables were selected. With the proposed SVR model, the measurement precision criteria for the validation dataset were root mean squares error percentage 2.93 %, correlation coefficient 0. 9944, and relative percent derivation 9. 4732%.