Engine lubricating oil is the cornerstone to ensure the long-term and stable operation of automobile engines. Accurately evaluating various performance indicators of engine lubricating oil is an essential step in the entire process from production to use. Engine lubricating oil will deteriorate for a variety of reasons after being used for a while. The engine lubricating oil deterioration indicators can be expressed in terms of non-magnetic particulate matter concentration, metal filings content, pH value, viscosity, water content and so on. To detect water content in engine lubricating oil, the traditional detection methods have the disadvantages of complicated operation and poor timeliness. Terahertz has strong absorption of water and is suitable for analysing micro-water content in sample products. In this paper, the transmission coefficients of six engine oils with different water contents were used to obtain the absorption coefficient curve of 1.0~3.5 THz by the transmission terahertz time domain spectroscopy system. The spectroscopic data were preprocessed with Savitzky-Golay(SG).Then, the sample was divided into a calibration set and test set by the Kennard-Stone algorithm after rejecting the odd samples.The interval Partial Least Squares (iPLS), backward interval partial least squares (BiPLS), and synergy interval partial least squares (SiPLS) were used to screen their terahertz time-domain spectral characteristic spectral intervals. They were focusing on the impact of factors such as the number of intervals, the number of PLS components, the number of best principal factors, and the selection of intervals on the PLS model’s properties. It also models and analyzes lubricants with different water contents, compares and selects different models, and establishes an optimal quantitative analysis model. The modeling results indicate that the feature spectrum region filtering can improve modeling performance and reduce model complexity. The characteristic spectrum region screening algorithm eliminates the non-linear or irrelevant variables in the terahertz absorption coefficient spectrum of engine lubricants so that the modeling results can better express the relationship between the absorption coefficient spectrum and its water content. The results show that the optimal model for quantitative analysis of trace water content in generator lubricants was obtained with BiPLS method that separated the whole spectra into 26 intervals and selected ［18 10 4 3 8 12 5 11 24 13 16 21 2］ intervals. The number of major factors is 10. The BiPLS model had a root mean standard error of cross-validation (RMSECV)of 0.003 5 and root mean standard error of prediction(RMSEP) of 0.004 6. The correlation coefficient (r) of the correction set is 0.913 9, and the correlation coefficient (r) of the prediction set is 0.865 7. Overall, BiPLS method could accurately predict the water content of engine lubricants, and the experimental process is simple, the modeling and calculation speed is fast, and the effect is ideal, and it can be applied to the quantitative analysis of the water content of non-contact oil products.