| Composition detection is the key link of iron and steel production process control.The existing composition detection is generally completed in the physical and chemical room,which requires cooling,grinding and other steps for high-temperature samples.The whole analysis process is long,and the guidance for the smelting process lags.Laser-Induced Breakdown Spectroscopy(LIBS)is an emerging technology for material composition detection with low sample preparation requirements and fast detection speed,which has great potential for application in the composition detection of iron and steel production process,but still needs to break through some key technical bottlenecks.It is mainly reflected in the following aspects:1)The rapidity of the detection process requires that the number of LIBS spectral excitations and measurements be as few as possible.The traditional multiple averaging method has limitations,so a new method is needed to reduce the fluctuations of the spectral data during LIBS measurements.2)If the detection speed wants to be improved,the direct analysis of high-temperature samples is the way to go,but there is a significant difference between the LIBS spectrum at high temperature and the results obtained at room temperature,and the temperature of samples will change over time during the measurement process.It is very difficult to train regression models by obtaining high-temperature standard samples of different temperatures in engineering applications.Thus,new regression modeling methods must be researched to solve the problem of sample composition analysis in high-temperature and variable-temperature environments.This paper conducts the studies based on spectral correction and model optimization to improve the rapidity,accuracy and reproducibility of LIBS quantitative analysis in high-temperature environment and provides technical support for the application of LIBS in real industrial environment.The main research contents of this paper are as follows.(1)For the problem of large spectral fluctuation of LIBS multiple measurements,a method to enhance the LIBS quantitative analysis repeatability via spectral correction based on probability distribution is proposed.Based on the method of probability distribution spectral correction,a weight correction model is established between the spectral intensities with probabilitiesμ-2σ,μ-σ,μ,μ+σ and μ+2σ at the analytical line and the corresponding average spectral intensities to convert the one-shot measured spectrum to the average spectrum measured multiple times.The proposed method addresses the poor repeatability of LIBS quantitative analysis and provides the possibility for detecting the steel composition by a few measurements.Taking gradient material samples as the analysis object,the experimental results show that the average relative standard deviations of the 20 measured spectra for the quantitative analysis of Cr,Co,Fe and Ni elements using the method of this paper were reduced from 8.68%,9.23%,6.37%and 9.06%to 3.32%,4.18%,3.23%and 3.47%,respectively,compared with those without the method of this paper.(2)To address the problem that the single regression model is prone to overfitting due to the large fluctuation of spectra,resulting in decreased prediction accuracy of the model,a segmented LIBS quantitative analysis method based on kernel space classification is proposed.Based on the idea of segmental modeling,according to the number of modeling samples and the distribution of concentration intervals of testing samples,the concentration intervals of testing samples were evenly divided into multiple intervals and the regression model was established.A Gaussian kernel function based on the Marxian distance is used to map the volatile and not easily discriminated spectral data to a high-dimensional space for discrimination,and a classification model is built in the high-dimensional space.During the test,the measured spectra are mapped to a high-dimensional space for classification and then a suitable regression model is selected for concentration prediction.This method lays the foundation for the LIBS quantitative analysis in high-temperature environments.Taking alloy steel samples as the analysis object,the average absolute errors of the samples are less than 0.02%,0.02%,0.005%and 0.005%in the quantitative analysis of Mn,C,P and S elements LIBS,respectively.(3)For the problems of difficulty in obtaining high-temperature standard samples,the LIBS quantitative analysis method at high temperature based on feature transfer learning is proposed.Based on the feature transfer learning method,the spectral data at room temperature and the spectral data at high temperature are mapped to the kernel Hilbert space,and then a regression model is built by the room-temperature spectra in the high-dimensional space.This method does not require the concentration information of high-temperature samples in the modeling process.Taking alloy steel samples as the analysis object,the experimental results show that the average relative errors of the LIBS quantitative analysis of Cr,Mn and Ni are 7.09%,7.71%and 9.68%,respectively.Compared to the method based on linear model mapping,the average relative error of Cr element quantitative analysis is reduced from 23.29%to 7.09%.Compared with instance-based transfer learning method,only the spectral information of high-temperature samples is required for training the regression model,and no concentration information is needed,which has better practicality for the detection of substance composition of high-temperature samples in industrial sites.(4)To address the problem of inaccurate quantitative analysis caused by the changing temperature of samples,an LIBS quantitative analysis method for samples with changing temperature based on functional data analysis is proposed.Based on the method of functional mathematical modeling,the measured spectral data at a finite number of different temperatures are used to establish a functional relationship between the LIBS spectrum and the sample temperature.The background intensity of the LIBS spectrum in the near-infrared band is used to characterize the temperature information in this work,which reduces the additional temperature measurement means.Taking alloy steel samples as the analysis object,the experimental results show that the average relative errors of Cr,Mn and Ni are all less than 10%and most of them are less than 5%at 670℃,810℃ and 945℃.This work provides a useful exploration for the LIBS technology application in the rapid detection field of iron and steel metallurgical production.The research methods and ideas of the paper also have broad application prospects in the non-ferrous smelting field such as copper and aluminum. |
PDF Full Text Request