改进的随机蛙跳算法对农机润滑油污染浓度的近红外光谱检测研究

润滑油是农业机械正常作业的必要物资，农业机械发动机工作的动力性、安全性、经济性以及寿命与润滑油状况有着紧密联系。污染浓度作为油液的综合评价指标，常规的实验室检测耗时长、成本高，所以开发高效的润滑油污染浓度检测技术具有重要意义。提出了一种基于近红外光谱技术的农机润滑油污染浓度的检测方法，同时针对随机蛙跳(RF)特征波长选择算法中迭代次数大，结果再现性低等缺点，提出了一种迭代保留信息变量的随机蛙跳(IRIV-RF)特征波长选择算法。该算法一方面利用迭代保留信息变量(IRIV)算法提取出强信息变量和弱信息变量，将其作为RF算法中的初始变量集，消除初始变量集的随机性对结果再现性的影响。另一方面通过对变量按被选概率值由大到小正向排序后，从首个波长开始依次增加一个波长建立偏最小二乘回归(PLSR)模型，选择交叉验证均方根误差(RMSECV)值最小时的变量子集为特征波长，消除RF算法所提取的特征波长数量的不确定性。利用近红外光谱仪采集自行配制的101份不同污染浓度的农机润滑油原始光谱数据，选用三种不同的预处理方法分别对原始光谱进行处理，确定最佳的预处理方法为变量标准化(SNV)。在此基础上通过RF，IRIV和IRIV-RF三种算法分别对全谱进行特征波长选择，并建立PLSR模型。通过对全谱-PLSR，RF-PLSR，IRIV-PLSR以及IRIV-RF-PLSR模型的预测精度进行比较，结果表明，经过IRIV-RF算法提取特征波长后所建立的PLSR模型预测精度最高，预测相关系数(R_p)为0.965 7，预测均方根误差(RMSEP)为9.0584，显著提升了预测精度与运行效率，降低模型复杂程度。IRIV-RF是一种有效的特征波长选择算法，研究证明了近红外光谱联合改进的IRIV-RF算法检测农机润滑油污染浓度的可行性，为鉴定润滑油品质提供了一种新的思路。

Near-Infrared Spectroscopy Detection of Pollution Concentration of Agricultural Machinery Lubricating Oil Based on Improved Random Frog Algorithm

The use of lubricating oil is necessary for the normal operation of agricultural machinery. The power performance, safety, economy and life of agricultural machinery engines are closely related to the condition of lubricating oil. The pollution concentration is the comprehensive evaluation index of oil, routine laboratory testing takes a long time and costs a lot, so it is of great significance to develop efficient detection technology for lubricating oil pollution concentration. This paper takes agricultural machinery lubricating oil as the research object. A method for detecting pollution concentration of agricultural machinery lubricating oil based on near-infrared spectroscopy is proposed. At the same time, aiming at the shortcomings of the Random Frog (RF) feature wavelength selection algorithm, such as a large number of iterations and low reproducibility, and iteratively retains informative variables-Random Frog (IRIV-RF) feature wavelength selection algorithm is proposed. On the one hand, IRIV-RF uses the iteratively retains informative variables (IRIV) algorithm to filter the strong and weak information variables. It is used as the initial variable subset of RF to eliminate the effect of the randomness of the initial variable set on the reproducibility of the results. On the other hand, IRIV-RF builds a Partial least squares regression (PLSR) model by arranging the variables in descending order of the selected probability values and then adding one wavelength at a time, starting with the first. The variable subset with the minimum Root Mean Square Error of Cross Validation (RMSECV) value is selected as the characteristic wavelength to eliminate the uncertainty of the number of characteristic wavelengths extracted by the RF algorithm. The original spectrum data of 101 samples of agricultural machinery lubricating oil with different pollution concentrations are collected by near-infrared spectrometer. Three different pretreatment methods are used to process the original spectrum, and the optimal pretreatment method is Standard Normal Variate (SNV). On this basis, the characteristic wavelength of the whole spectrum is selected by RF, IRIV and IRIV-RF algorithms, and the PLSR model is established. By comparing the prediction accuracy of full-spectrum PLSR, RF-PLSR, IRIV-PLSR and IRIV-RF-PLSR models, the results show that the prediction accuracy of the PLSR model based on the IRIV-RF algorithm is the highest, the Correlation Coefficient of Prediction (R_p) is 0.965 7 and the Root Mean Square Error of Prediction (RMSEP) is 9.058 4. It significantly improves the prediction accuracy and operation efficiency, reducing the model’s complexity. It is proved that the proposed IRIV-RF algorithm is an effective characteristic wavelength selection algorithm, and the feasibility of near-infrared spectroscopy combined with the improved IRIV-RF algorithm to detect the pollution concentration of agricultural machinery lubricating oil is proved, which provides a new idea for identifying the quality of lubricating oil.