| Stabilizer is an important component of rocket propellant.By absorbing acid and nitrogen oxides generated when nitrate absorbs water or decomposes by heat,it can effectively slow down the speed of autocatalytic decomposition of propellant,delay the aging process,and ensure its storage stability.Reliable monitoring of stabilizer content is essential to accurately understand propellant aging and estimate useful life.Due to the current detection methods of stabilizer content mainly have the problems of cumbersome operation,time-consuming and labor-intensive,and low accuracy.Therefore,there is an urgent need to find a non-destructive,rapid,effective and accurate method for the detection of stabilizers.Terahertz time-domain spectroscopy(THz-TDS)technology has attracted much attention in the field of non-destructive testing because it has almost no ionization damage to the measured material and is extremely sensitive to the molecular vibration of the material.In this paper,N-methyl-4-nitroaniline(MNA)stabilizer is taken as the research object,THz-TDStechnology is used as the detection technology,and machine learning algorithm is used as data analysis methodto study the quantitative detection method of MNA stabilizer concentration.The relevant research contents and achievementsare as follows:(1)Terahertz spectroscopy analysis of MNA stabilizers.The time-domain spectra of 25MNA samples with different contents were obtained by THz-TDS system,and the absorption coefficient spectra of MNA were obtained by using the optical parameter extraction model.The experimental results show that MNA has two characteristic absorption peaks in the0.2?2.0THz frequency band,which are located at 1.296THz and 1.765THz respectively.And the R~2 of the linear fitting curve between the absorption coefficient at the peak and the MNA content reached more than 0.99,showing a good linear correlation,which provided a reliable basis for the subsequent quantitative study of MNA concentration.(2)In order to make the quantitative model have better prediction accuracy,it is necessary to preprocess the spectral data and select the characteristic data.Wavelet transform,MSC,SNV,S-G smoothing,z-score normalization and Min-max normalization were used to preprocess and analyze the terahertz spectral data.Random sampling,KS algorithm and SPXY algorithm were used for sample set division experiments.The research results show that the use of S-G smoothing Min-max normalization and KS algorithm for processing is more beneficial to the quantitative model prediction of stabilizers.In addition,BP,SVR and ELM were used to establish training models for time-domain and frequency-domain spectral data and simulate and predict,in order to explore the characteristic data for quantitative modeling of MNA content.The results show that the absorption coefficient data in the 0.8?2.0THz frequency band are more suitable as the characteristic data for quantitative modeling of MNA content.(3)Using BP,GRNN,SVR,ELM and XGBoost analysis methods to achieve internal and external quantitative analysis of the two groups of MNA sample sets.The results showed that the ELM model had the best prediction effect when it was verified internally,its RMSE and R~2respectively are 0.00033 and 0.9963.When externally verified,the SVR model has the best prediction effect,and its RMSE and R~2 are 0.00099 and 0.9956,respectively.ELM shows better learning rate and generalization performance.In addition,the experimental results of using the GA and PSO algorithms to optimize the quantitative model and apply it to the prediction of MNA content show that the GA algorithm performs better than the PSO algorithm in the prediction performance improvement effect of these five models.Among all the optimization models,the maximum deviation,RMSE and R~2 of the test set of the GA-ELM model are 0.0243%,0.00012 and 0.9995,respectively,which had the best prediction accuracy and provided a reliable reference for efficient,rapid and accurate quantitative detection of stabilizer content. |
PDF Full Text Request