Research On Near Infrared Spectral Data Processing Methods In Fuel Quality Detection

Combining artificial intelligence methods to achieve objective and efficient digital evaluation of fuel quality is an inevitable trend in the development of the energy industry.In recent years,the scarcity of petroleum resources and the increasingly serious environmental pollution have made high-precision control of fuel quality and the development and application of alcohol based fuels a hot topic of global concern.This article focuses on larger molecule diesel fuels as the main detection object,and proposes qualitative and quantitative analysis methods using artificial intelligence based on near-infrared（NIR）spectroscopy data,quality properties,grades,and chemical composition content data of diesel.In response to the limitations of high-dimensional complexity,weak useful signals,and severe spectral overlap in diesel NIR spectra,which result in low accuracy and difficulty in qualitative and quantitative detection,this article proposes specific solutions from the perspectives of Chemometrics,machine learning,ensemble learning,and deep learning.The purpose is to achieve digital and precise control of detection of various quality indicators,brands recognition,and type identification and alcohol content detection of diesel,providing a methodological basis for modern diesel testing and quality standard research.The research in intersect field of spectroscopy,artificial intelligence and energy detection science is of great practical significance.The main research work of this paper is as follows:（1）In order to solve the problem that conventional Chemometrics modeling methods are difficult to accurately mine the hidden essential information of high-dimensional complex spectra,a feature reduction strategy combining correlation coefficient（CC）and t-Distributed Stochastic Neighbor Embedding（t SNE）technology was proposed to detect the three basic physical properties of diesel,namely density,viscosity and freezing point.This research strategy combines feature selection and manifold learning technique to greatly reduce the number of features while ensuring that the effective feature information of the input Support Vector Regression（SVR）machine calibration model.Comparing the feature reduction rates of multiple models,and calculating the predicted root mean squared error（RMSE）and coefficient of determination（R²）,this combination strategy can provide a new method for the detection of physical properties of diesel and related fuels,which is of great significance to the development of NIR and Chemometrics.（2）In order to solve the problem of Chemometrics modeling methods that rely too much on the data quality and prior knowledge,lack of adaptive ability,and are difficult to effectively applied to the simultaneous detection of more properties of diesel,a strategy based on machine learning was proposed.This strategy can quantitatively detect multiple physical and chemical properties of diesel,including density,viscosity,freezing point,boiling point,Cetane number and total aromatics of diesel oil simultaneously.This method designed a new nonlinear Grey Wolf Optimization（IGWO）algorithm,as well as a hybrid algorithm of Differential Evolution and Grey Wolf Optimization（DEGWO）to solve the problem of the GWO algorithm falling into local optima when searching for the optimal parameters of the SVR machine.And improve the XY Co-occurrence Distance（SPXY）partition sample set method through weighting method to enhance spectral interpretability.Comparing multiple models and calculating the predicted RMSE,mean absolute percentage error（MAPE）and R² of each algorithm,the proposed model is simple,high precision,strong generalization ability,does not need any NIR knowledge and experience,and has strong practicability.It has the potential to become a more practical method in the field of industrial detection.（3）In order to address the problem of low recognition accuracy and poor stability of a single classification model under the influence of serious spectral overlap,collinearity and sample imbalance,the idea of Boosting ensemble learning was introduced into the NIR field to realize diesel brands recognition.Combining Tree-based feature selection and Synthetic Minority Oversampling Technology（SMOTE）,two ensemble learning diesel brands recognition models,namely e Xtreme Gradient Boosting（XGBoost）and its improved version of Light gradient boosting machine（Light GBM）,were constructed respectively.Comparing multiple models and calculating the classification accuracy,sensitivity,specificity,precision,and F₁ score of each algorithm,the proposed model has the best classification performance and strong stability.This study provides a new approach for multimodal recognition of NIR data with severe spectral overlap and common sample imbalance.（4）In order to solve the key technical problem that the more mature deep learning image processing methods can not be directly used for high-performance qualitative and quantitative analysis of one-dimensional NIR spectra,two mathematical methods,namely Gram Angle Field（GAF）and Markov Transition Field（MTF）in the frequency domain,were proposed to convert one-dimensional NIR spectra into two-dimensional images.Based on the obtained images,a two-dimensional deep convolutional neural network（CNN）was established to identify the categories of alcohols-diesel and detect the contents of methanol and ethanol.Comparing multiple models and calculating the classification and prediction indicators of each algorithm,the proposed method exhibits strong overall performance and high sensitivity in classification and prediction tasks.This research strategy provides the possibility for deep learning image processing to maximize its powerful advantages in the field of NIR data analysis,and has enormous application potential and prospects.