Determination of the total acid number (TAN) of used mineral oils in aviation engines by FTIR using regression models

- Predictive models are useful to determine TAN using IR from ashless dispersant oils for aviation.
- For dimension reduction, PCA and Independent component analysis (ICA) have been applied.
- A model based on non-linear regression by Support Vector Machines (SVM) with Gaussian kernel was implemented.
- A hybrid method considering group of bands as features was used for modelling which showed better performance.

Total acid number (TAN) has been considered an important indicator of the oil quality of used oils. TAN is determined by potentiometric titration, which is time-consuming and requires solvent. A more convenient approach to determine TAN is based on infrared (IR) spectral data and multivariate regression models. Predictive models for the determination of TAN using the IR data measured from ashless dispersant oils developed for aviation piston engines (SAE 50) have been developed. Different techniques, including Projection Pursuit Regression (PPR), Partial Least Square, Support Vector Machines, Linear Models and Random Forest (RF), have been used. The used methodology involved a five folder cross validation to derive the best model. Then a full error measure over the whole dataset was taken. A backward variable selection was used and 25 highly relevant variables were extracted. RF provided an acceptable modelling technology with grouped dataset predictions that allowed transformations to be performed that fitted the measured values. A hybrid method considering group of bands as features was used for modelling. An innovative mechanism for wider features selection based on genetic algorithm has been implemented. This method showed better performance than the results obtained using the other methodologies. RMSE and MAE values obtained in the validation were 0.759 and 0.359 for PPR model respectively.