Research On Coal Quality Analysis Methods By Near Infrared Spectroscopy Based On Deep Learning

The industry-wide analysis of coal mainly includes the measurement of indicators such as moisture,ash,volatile matter and calorific value in coal,which is an important standard to measure the quality and practical value of coal utilization.Traditional coal quality analysis methods have problems such as complex operational processes,high human and material costs,destruction of sample structures,and single analysis objects,which are difficult to meet the new national demand for energy security and efficient utilization strategies.As an efficient and non-destructive detection method,nearinfrared spectroscopy technology has been widely used in the quantitative and qualitative analysis of substances.In view of this,based on near-infrared spectroscopy technology,this thesis studies a quantitative analysis model of coal quality based on multi task deep learning to realize rapid measurement of multiple industrial indicators of coal quality.The main research contents are as follows:Aiming at problems such as abnormal data and noise interference caused by environment,instruments,or operations during the spectrum acquisition process,abnormal sample removal and spectral data recovery for data sets are studied.Firstly,an iterative anomaly sample removal method based on the Pauta criterion and Euclidean distance is used to filter the sample set to obtain a clean modeling data set,Secondly,a variety of spectral information recovery methods are compared,among which the Savitzky-Golay convolution smoothing after the second derivative of the original spectrum is the best.The partial least squares regression method was used to validate the above method,and it was found that the root mean square error and mean absolute error of the predicted and labeled values for the calorific value decreased by 13.10%and 11.54%,respectively,and the correlation coefficient increased by 3.58%.Due to the complex representation of the relationship between coal sample spectra and its multiple industrial indicators,the existing analysis models have the problems of low prediction accuracy and poor robustness.In view of this,after analyzing the correlation between the four industrial indicators,this thesis proposes a Multi Task Attention Unet(MTA-Unet)based on parameter hard sharing.The model mainly includes a weight parameter sharing module and a multi task output module.The weight sharing module combines Unet,convolutional block attention module,and a multi scale feature fusion strategy;The multi task output module consists of four independent fully connected layers.In addition,the gradient normalization algorithm is used to optimize the gradient magnitude inundation problem caused by different training speeds between different tasks.The experimental results showed that the correlation coefficients between the predicted and labeled values of MTA-Unet for moisture,ash,volatile matter,and calorific value indicators were 0.9015,0.9538,0.8986,and 0.8884,respectively.Compared with single task learning,MTA-Unet increased by 1.84%,1.38%,5.21%,and 4.84%,respectively.And the effectiveness of each module and optimization algorithm was verified through ablation experiments.For the above methods,this thesis designs a near-infrared coal quality analysis software based on the software development toolbox of MATLAB,which mainly includes the functions of anomaly sample removal,pre-processing and unknown sample prediction.It combines the model method proposed in the thesis with the practical needs to promote the application of near-infrared spectroscopy technology in the industrial field.This thesis has 51 figures,11 tables and 92 references.