针对水-乙二醇抗燃液压液(HFC)目前常用的水分含量和粘度测量方法存在操作复杂, 耗时长, 测量成本较高等问题, 研究了将中红外光谱分析法应用于HFC水分含量和粘度测量。 以85个实际在用油样为样本集, 采用传统的电量法和毛细管法分别测量水分含量和运动粘度; 使用衰减全反射液体池(ATR)采集各样本的中红外光谱并进行光谱校正、 背景扣除、 Savitzky-Golay(SG)平滑处理和基线校正等预处理。 水分含量测量模型构建过程中, 利用学生化残差-杠杆值法寻找强影响点, 判定为异常样本并加以剔除; 利用马氏距离SPXY法划分建模样本集和验证样本集。 根据Beer-Lambert定律, 水分含量和红外光谱吸光度成正比, 采用线性方法来构建水分含量和吸光度红外光谱之间的关系模型。 同时为提高模型的鲁棒性和泛化能力需要消除光谱中的无用信息、 冗余信息和噪声, 因此采用后向间隔偏最小二乘法(BiPLS)进行水分含量测量特征谱区优选, 并建立线性校正分析模型。 结果表明剩余3 526~3 354, 1 790~1 618, 3 005~2 660和1 096~924 cm^-1波段时获得最小的交叉验证均方根误差(RMSECV)。 特征波段优选结果可以理解为, 根据3 526~3 354 cm^-1波段预测HFC体系中所有OH的量, 减去根据3 005~2 660和1 096~924 cm^-1波段换算得到的乙二醇中的OH, 进而计算得到水分含量, 为进一步提高精度, 将水中OH变角振动吸收峰对应的1 790~1 618 cm^-1波段作为水分含量测量的辅助波段。 利用独立验证样本进行模型检验, 结果表明所建立的线性模型相关系数( r)为0.989 5, 预测均方根误差(RMSEP)为0.405 2, 满足HFC中水分含量测量精度需求。 粘度测量模型构建过程中, 利用马氏距离法寻找离群样本, 判定为异常样本并加以剔除, 利用马氏距离SPXY法进行建模样本集和验证样本集划分。 粘度是一个与多重因素有复杂关系的理化指标, 和红外光谱之间表现为非线性关系。 基于主成分分析结合BP神经网络法(PCA-BPNN)构建非线性粘度校正分析模型, 提取累积贡献率为95.12%的前10个主成分作为输入, 测得的粘度值为输出, 创建单隐层BP神经网络并进行训练, 训练结果相关系数 r为0.996 8。 利用独立验证样本进行模型检验, 结果表明所建立的非线性模型 r为0.984 3, RMSEP为0.615 1, 满足HFC粘度测量精度需求, 且优于采用BiPLS线性模型获得的结果。 确定了可用于HFC中水分含量测量的四个中红外波段, 可为窄带红外光谱分析法或其他类似研究提供依据。 研究结果表明中红外光谱分析法结合BiPLS和PCA-BPNN分析法可分别应用于HFC水分含量和粘度测量, 精度符合监测需求, 为HFC中分含量和粘度测量提供了一种新的测试方法, 相比于传统测量方法具有快速、 无损、 低成本的特点。

For water glycol fire-resistant hydraulic fluids (HFC), the commonly used methods for measuring the moisture content and viscosity present problems such as complicated operation, long time-consuming, high measurement cost, etc. The application of mid-infrared spectroscopy to HFC moisture content and viscosity measurement was studied. The sample set consisted of 85 actual in-use oil samples. The moisture content and kinematic viscosity were measured by traditional coulometric method and capillary method. The mid-infrared spectra of samples was collected using an ATR liquid pool. Spectra pretreatments such as spectra correction, background subtraction, Savitzky-Golay (SG) smoothing and baseline correction were performed. During the process of constructing the moisture content measurement model, strong influence points were searched by student residuals-leverage method, which were determined as abnormal samples and removed. Mahalanobis distance SPXY method was used to divide the sample set into modeled samples and verification samples. According to the Beer-Lambert law, the moisture content was directly proportional to the infrared absorbance. So a linear method should be used to construct the relationship between the moisture content and the infrared absorbance. At the same time, the no information variables, redundant information and noise in the spectrum need to be eliminated to improve the model robustness and generalization ability. By the backward interval partial least squares method (BiPLS), the moisture content measurement feature bands were optimized, and a linear calibration analysis model was established. The results showed that the minimum root mean square error of cross validation (RMSECV) was obtained when the bands of 3 526~3 354, 1 790~1 618, 3 005~2 660 and 1 096~924 cm^-1 remained. The preferred feature bands can be explained as that, and the OH amount in moisture was calculated by subtracting the OH amount in ethylene glycol converted by the 3 005~2 660 and 1 096~924 cm^-1 bands from the all OH amount in the HFC system converted by the 3 526~3 354 cm^-1 band, and then the moisture content was calculated. In order to improve the accuracy, the 1 790~1 618 cm^-1 band, corresponding to the moisture OH swing vibration absorption peak, was used as an auxiliary wave for measuring the moisture content. The independent verification samples were used to test the model. The results showed that the correlation coefficient ( r) of the established linear model was 0.989 5 and the root mean square error of prediction (RMSEP) was 0.405 2, which met the accuracy requirement for moisture content measurement in HFC. During the process of constructing the viscosity measurement model, the outlier samples were searched by Mahalanobis distance method, which were determined as abnormal samples and removed. Mahalanobis distance SPXY method was used to divide the sample set into modeled samples and verification samples. Viscosity was a physicochemical index that had complex relationships with multiple factors, and was nonlinear to the infrared spectrum. Based on principal component analysis combined with BP neural network method (PCA-BPNN), a nonlinear viscosity correction analysis model was established. The first 10 principal components with a cumulative contribution rate of 95.12% were extracted as input, and the measured viscosity values were treated as output. A single hidden layer BP neural network was created and trained. The training result correlation coefficient r was 0.996 8. The independent verification samples were used to test the model. The results showed that the r of the established nonlinear model was 0.984 3 and the RMSEP was 0.615 1, which met the accuracy requirement for HFC viscosity measurement and were superior to the ones obtained by the BiPLS linear model. Four mid-infrared bands were identified and can be used for moisture content measurement in HFC, which could provide basis for narrowband infrared spectroscopy or other similar studies. The research results showed that the mid-infrared spectroscopy combined with BiPLS and PCA-BPNN analysis method can be applied to the measurement of moisture content and viscosity of HFC, and the accuracy meets the monitoring requirements. Compared with the traditional measurement methods, a new test method is provided with the features of being fast, non-destructive and low-cost.