油中含水量近红外光谱具有复杂性、非线性和难以用明确数学模型表达的特点.利用无信息变量消除法提取有效波长, 采用模糊C均值聚类算法得出输入空间的划分和聚类中心, 结合递推最小二乘法辨识后件参量, 建立了润滑油的近红外光谱与含水量的Takagi-Sugeno模糊模型.将该辨识算法与偏振最小二乘法模型进行比较, 并对实测数据进行了验证.结果显示: 经无信息变量消除法提取34个特征波长建立的Takagi-Sugeno模型能够精确地反映出润滑油光谱数据与含水量的关系, 其对验证集样本进行预测的相关系数和均方根误差分别为0.964 6和1.531 2×10－4, 获得了满意的预测准确度.实验结果验证了应用光谱技术检测油中含水量的可行性, 同时也为油中其他污染物的在线监测提供了新方法.

The model of near infrared spectroscopy of water content in oil is complex, nonlinear, and difficult to be specified with mathematical methods. Uninformative variables elimination was applied to the extraction of effective wavelengths, and fuzzy C-means clustering algorithm was applied to obtain the input space location and the clustering center. By identifying the consequent parameters with recursive least-squares method, Takagi-Sugeno, a fuzzy model of near infrared spectroscopy of water content in oil, was established. This identification algorithm was compared with Partial Least Squares model and tested by experimental data. The results indicate: the Takagi-Sugeno model, constructed by a total of 34 variables selected by uninformative variables elimination, can accurately reflect the relation between near infrared spectral data of oil and moisture content; the correlation coefficient the model predicted for the samples from validation set is 0.964 6 and the root of mean square error is 1.531 2×10－4, which are satisfactory. The experimental results verify that it is feasible to detect the water content in oil by means of near infrared spectroscopy, which also offers a new alternative approach for the on-line monitoring of other contamination content in oil.