Multi-Label One-Dimensional Convolutional Neural Network Based on Infrared Spectroscopy for the Identification of Additives in Lubrication Oils

Lubricating oil additive is a chemical substance used to improve and enhance the original performance of oil products, endowing oil products with new characteristics and is widely used in various types of lubricating oils. Exploring effective methods for component analysis and substance identification of additives in lubricating oil has practical significance and theoretical value for the identification and evaluation of new oil quality, and the determination of using oil operation and maintenance plans. It has always been a research direction of concern for scholars. Focused on the issue of identifying additives in lubricating oil, polyalphaolefins (PAO40) and Polyol ester (PriEco3000) were mixed as base oil and four common commercial additives (MoDTC, ZDDP, L135 and L57) were blended in different proportions to prepare lubricating oil samples. Fourier transform infrared radiation (FT-IR) was employed to collect mid-infrared spectral information of the samples, which were then subjected to first and second derivative processing. The three spectra were integrated into a Lube oil feature matrix as the input of the model for displaying the feature information of lubricating oil FT-IR spectra to the greatest extent. Different from the traditional single-spectrum and single-label classification task, the identification of additives in lubricating oil essentially reflected the existence of multiple labels in a single FT-IR spectrum. Based on the idea of multi-label classification, a Multi-Label One-Dimensional Convolutional Neural Network (ML-1D-CNN) model was proposing. The model performed supervised feature selection and additive recognition on the input lubricating oil infrared spectra. Deep learning model that was correct and had a certain degree of generalization ability should be able to extract the FT-IR spectral features of each additive, and the effect of feature extraction was an important criterion for judging the correctness of the model. Gradient-weighted Class Activation Mapping (Grad-CAM) was introduced to observe the effect of the models in extracting features from input data. Experimental results revealled that the spectral features extracted by the proposed models corresponding well with the original spectral feature peaks. It was shown that the model could effectively and correctly study the FT-IR spectra of each additive. Whether on the validation set of the training process or on additional test set data, a comparison with three hybrid models PCA+PLS-DA, SPA+SVM and IGA+RF recently mentioned in the literature showed that ML-1D-CNN model performed exceptionally, achieving an recognition accuracy rate of 93.33% to 100% for individual additives and demonstrated strong synchronous recognition capabilities for similar additives. Furthermore, the ML-1D-CNN model maintained a satisfactory recognition rate for additives trained outside the model samples, showing a good generalization ability.