Laser Induced Breakdown Spectroscopy (LIBS) has great potential in rapid analysis of coal quality due to the unique advantages of no complex sample pretreatment, simultaneous analysis of multiple elements and fast detection. However, due to the deviation in the precision of instrumentation, there are signal intensity difference and wavelength shift in spectra collected from different instruments. As a result, the effective quantitative analysis model based on the master instrument cannot be applied to the slave instrument, which hinders the application of LIBS. Therefore, this paper proposed a method combining wavelength shift self-calibration with feature transfer learning to improve the applicability of quantitative analysis model of coal quality. The wavelength shift between master and slave instruments was corrected by the wavelength shift self-calibration method based on the standard deviation value of characteristic peak intensity. Then, the transfer learning method based on Kernel Principal Component Analysis (KPCA) and Piecewise Direct Standardization (PDS) was used to fit the spectral features between slave instrument and master instrument. Finally, the quantitative analysis model of coal quality was established by the random forest (RF). Furthermore, the Competitive Adaptive Reweighted Sampling (CARS) feature selection method was used to select the input of model. As a result, the proposed CARS-KPCA-PDS method was able to improve the adaptability of quantitative models across different LIBS Systems. Compared with the RF model without transfer learning, the mean absolute error (MAEP) of CARS-KPCA-PDS model in predicting calorific value, carbon content and ash content were reduced by 58.32 %, 71.67 % and 77.48 %. The results demonstrated that the proposed method could improve the applicability of quantitative analysis model to different instruments and reduce the modeling cost.
扫码关注我们
求助内容:
应助结果提醒方式:
