Determination of Water Content in Automobile Lubricant Using Near-Infrared Spectroscopy Improved by Machine Learning Analysis

The main objective of this paper is to determine the water content of automobile lubricant based on the near-infrared (NIR) spectra collected and to observe whether NIR spectroscopy could be used for predicting water content. Least square support vector machine (LS-SVM), back-propagation neural networks (BPNN) and Gaussian processes regression (GPR) were employed to develop prediction models. There were 150 samples for training set and test set, 6 inputs for one sample obtained by principle component analysis (PCA). LS-SVM models were developed with a grid search technique and RBF kernel function. The Levenberg-Marquardt algorithm was employed to optimize back-propagation neural network (BPNN) and models with 5 and 6 neurons in hidden layer were developed, respectively. The BPNN model with 5 neurons in hidden layer outperformed the one with 6 neurons. Three GPR models were built based on full data points (full GPR), subset of regressors (SR GPR) and subset of datasets (SD GPR), respectively, with Squared exponential (SE) covariance function. The full GPR outperformed SR GPR and SD GPR.The overall results indicted that the Gaussian processes model outperformed LS-SVM and BPNN model. GPR was an effective way for the regress prediction. NIR spectroscopy combined with PCA and GPR had the capability to determine the water content of automobile lubricant with high accuracy.