A coal quality detection method, a model construction method and a detection system

Abstract

The application provides a coal quality detection method, a model construction method and a detection system. The method comprises the following steps: S1, acquiring LIBS spectrum and NIRS spectrum of a coal sample, and a standard measured value of a coal quality index of each sample; S2, extracting a characteristic peak area of the LIBS spectrum based on a plurality of preset elements to obtain a LIBS characteristic peak area matrix; performing noise reduction and normalization on the NIRS spectrum to obtain a NIRS spectrum matrix; S3, respectively constructing a kernel function with an independent bandwidth parameter for the LIBS characteristic peak area matrix and the NIRS spectrum matrix; S4, weighting and fusing the two kernel functions based on a weight factor of adaptive learning to construct a hybrid kernel function; S5, taking the hybrid kernel function as a kernel function of any regression model, and optimizing a parameter vector by using a parameter optimization algorithm to obtain an optimal value of the parameter vector; the parameter vector comprises a weight factor, a regularization parameter of the regression model and a bandwidth parameter; and S6, inputting the optimal value of the parameter vector into the regression model to perform coal quality detection.

Description

Technical Field

[0001] This application belongs to the technical field of coal quality testing, and in particular relates to a coal quality testing method, a model construction method, and a testing system. Background Technology

[0002] In the context of new power system construction, the deep peak shaving and flexible operation of coal-fired power units urgently require real-time and accurate online monitoring of coal quality indicators. Laser-induced breakdown spectroscopy (LIBS), an atomic emission spectroscopy technique, can rapidly and non-destructively detect the elemental composition of coal; near-infrared spectroscopy (NIRS), a molecular absorption spectroscopy technique, is highly sensitive to hydrogen-containing groups and organic molecular structures. Theoretically, combining LIBS and NIRS can achieve physical complementarity between "atomic elemental information" and "molecular structure information," comprehensively improving the detection capability of key coal quality indicators.

[0003] Existing coal quality detection methods based on LIBS and NIRS dual-modal spectral data can be mainly classified into the following three categories: Feature-level simple splicing model fusion: For example, the published patent CN111044503A - a coal quality measurement method based on the fusion of laser-induced breakdown spectroscopy and near-infrared spectral information, first performs Decimal Scaling normalization on LIBS and NIRS spectra respectively, then directly splices the high-dimensional feature vectors end to end to form a single fused feature matrix, and finally inputs it into a PLS (partial least squares) regression model for multi-index detection of coal quality. However, this method directly splices the high-dimensional LIBS and NIRS features end to end. Since the dimensions and intensity ranges of LIBS and NIRS are very different, the spectral features with larger values ​​will mask the features with smaller values, i.e., variance suppression. In addition, forcibly putting two data with completely different physical properties into the feature space of the same machine learning model and sharing a single mapping parameter makes it impossible to perform optimal nonlinear mapping for the respective distribution characteristics of atomic and molecular spectra, which easily leads to the neglect of physical heterogeneity.

[0004] Post-processing and fusion of data layers: For example, the published patent CN120427566A (a spectral preprocessing method and related equipment for online coal quality detection using multispectral coupling) proposes a synchronous LIBS-NIRS acquisition device. This device first filters effective spectra using a standard deviation threshold, then uses an RF classifier to identify defocusing, PCA to extract principal components, and SHAP values ​​to iteratively eliminate defocusing-affecting components. Finally, it reconstructs the spectrum using inverse PCA and then uses Decimal Scaling to stitch and fuse the data before inputting it into a quantitative model. This approach solves some of the defocusing problems and achieves the acquisition of atomic and molecular information at the same measurement point, but the fusion still remains at the level of linear normalization and stitching.

[0005] Decision-level fusion: Decision-level schemes first construct independent detection models for LIBS and NIRS (often using SVM or RF), and then perform a weighted average of the output values. However, this method severs the deep nonlinear interaction between atomic and molecular information in the underlying high-dimensional space, resulting in the loss of feature cross-information.

[0006] Advanced deep learning fusion: Directly extracting multimodal features end-to-end within the network. In the coal quality field, high-precision physicochemical label samples require thousands of samples, resulting in high acquisition costs. Furthermore, under small sample conditions, deep learning models are prone to overfitting; simultaneously, their black-box nature leads to a lack of physical interpretability, failing to intuitively reflect the dependence of coal quality indicators on specific spectral peaks.

[0007] The aforementioned existing technologies suffer from several problems: the accuracy of heterogeneous spectral fusion is limited by modal differences, the average relative error is relatively high, and there is a lack of adaptability and interpretability for small samples. Summary of the Invention

[0008] To address the limitations of existing technologies, such as the accuracy of heterogeneous spectral fusion being constrained by modal differences, high average relative error, and lack of adaptability and interpretability for small samples, this application provides a coal quality detection method, comprising: S1. Obtain the LIBS spectrum, NIRS spectrum, and standard values ​​of coal quality indicators for multiple coal samples. LIBS spectrum is atomic emission spectroscopy; NIRS spectrum is molecular absorption spectroscopy; the standard values ​​of coal quality indicators for the samples are determined using standard chemical testing methods. These standard values ​​serve as sample labels, ensuring a one-to-one correspondence between the spectral data and the actual coal quality indicator characteristics.

[0009] S2. Preprocess the LIBS and NIRS spectra separately, including: S21. Normalize the total intensity of the LIBS spectrum. The purpose of total intensity normalization is to eliminate multiplicative interferences such as laser energy fluctuations, changes in focusing state, and differences in sample surface morphology, so that the spectra of different pulses and different measurement times are comparable, while retaining the relative concentration information between elements.

[0010] The characteristic peak areas of major elements in multiple pre-defined coal samples are extracted to obtain a LIBS characteristic peak area matrix. These elements can include C, H, O, Si, Al, Ca, Fe, Mg, etc. Extracting the characteristic peak area involves extracting physical features closely related to coal quality from the original thousands-dimensional spectrum, while simultaneously reducing dimensionality significantly. LIBS spectra contain atomic emission lines of each element, and the peak area of ​​each line is proportional to the number density of that element in the plasma, thus relating to the elemental content in the coal.

[0011] The LIBS feature peak area matrix is ​​subjected to min-max normalization along the column direction, mapping it to the [0,1] interval. Column-direction normalization eliminates dimensional and order-of-magnitude differences between feature peak areas of different elements. Since the spectral intensities of different elements vary significantly, directly using the original area values ​​would result in features with larger orders of magnitude dominating similarity calculations (such as Euclidean distance in the subsequent kernel function), while the contribution of features with smaller orders of magnitude would be suppressed. This double normalization before and after feature extraction ensures that the similarity matrix calculated by the subsequent kernel function better reflects the true coal quality differences, rather than spurious differences caused by measurement conditions or dimensions, thus improving adaptability to small sample training.

[0012] S22. Denoise and normalize the NIRS spectrum to obtain the NIRS spectral matrix; S3. Based on radial basis functions, kernel functions with independent bandwidth parameters are constructed for the LIBS characteristic peak area matrix and NIRS spectral matrix, respectively, to provide a suitable nonlinear mapping scale for the heterogeneous LIBS characteristic peak area matrix and NIRS spectral matrix.

[0013] In machine learning, a kernel function is a method for calculating the inner product of vectors in a high-dimensional feature space. It allows for indirect high-dimensional nonlinear mapping within the original low-dimensional space, thus solving nonlinear problems. The similarity matrix between samples calculated by the kernel function is called the kernel matrix, and the space spanned by this matrix is ​​the kernel space. In the kernel space, data that was originally linearly inseparable may become linearly separable. This application chooses the Radial Basis Function (RBF) as the basis kernel because: RBF is a general-purpose kernel function capable of handling nonlinear relationships and requires less prior knowledge of the data; it has only one core parameter, the kernel bandwidth parameter, making optimization relatively simple; the feature space corresponding to RBF is infinite-dimensional, theoretically able to approximate arbitrarily complex decision boundaries, and its mathematical form is: ; in, c For kernel bandwidth parameters; x i , x j The first i The and the first j A sample vector.

[0014] S4. Based on the weighting factors obtained from adaptive learning, the two kernel functions are weighted and fused to construct a hybrid kernel function for the fusion of heterogeneous data.

[0015] S5. Using the mixed kernel function as the kernel function of any regression model, a parameter optimization algorithm is used to jointly optimize the parameter vector to obtain the optimal value of the parameter vector; where the parameters include weight factors, regularization parameters of the regression model, and kernel bandwidth parameters of the two independent kernel functions.

[0016] S6. Input the optimal value of the parameter vector and the LIBS and NIRS spectra of the coal to be tested into the regression model, and output the detection results of the coal quality indicators.

[0017] Furthermore, coal quality indicators include at least one of the following: calorific value, ash content, volatile matter, and carbon content.

[0018] Furthermore, existing technologies using a single kernel function to process bimodal spectral data cannot adapt to the physical heterogeneity of LIBS spectra (sharp, discrete atomic spectra) and NIRS spectra (broad, continuous molecular spectra), leading to heterogeneous data mapping mismatch, limited fusion accuracy, and a lack of personalized similarity measurement mechanisms for different spectral characteristics, thus failing to fully extract the effective information of each spectrum. In step S3 of this application, independent Gaussian radial basis function (RBF) kernels are constructed for the two types of spectra, and each is configured with independent kernel bandwidth parameters. The two kernel functions are the LIBS kernel function and the NIRS kernel function, respectively. The expression for the LIBS kernel function is: ; in, x i L , x j L These are the first two peaks in the LIBS characteristic peak area matrix. i The and the first j One sample vector; γ L The first kernel bandwidth parameter controls the similarity measurement scale of LIBS features in the kernel space.

[0019] The expression for the NIRS kernel function is: ; in, x i N , x j N These are the first two spectral elements in the NIRS spectral matrix. i The and the first j The feature vector of each sample; γ N This is the second kernel bandwidth parameter, adapted to the continuous smoothness characteristics of NIRS spectra.

[0020] Furthermore, the expression for the hybrid kernel function is: ; in, β As a weighting factor; x i , x j The first i The, the j The joint feature vector of each sample, encompassing all feature information from LIBS and NIRS. Weighting factors β The value is not fixed by humans, but is automatically learned by subsequent optimization algorithms based on the specific detection task.

[0021] In existing technologies, feature-level splicing fusion directly concatenates high-dimensional features end-to-end, ignoring variance suppression caused by physical heterogeneity; decision-level fusion severs the interaction of low-level features, leading to information loss; and there is a lack of fusion mechanisms that can adaptively adjust the contribution ratio of the two spectral types. This application fuses two independent kernel functions and weighting factors. β Its dynamic and adjustable characteristics enable the fusion strategy to adaptively adjust according to the characteristics of different coal quality indicators (such as ash content detection). β Prefers LIBS, heat output detection β By balancing the two spectra, deep fusion at the kernel level is achieved, preserving the independent physical properties of the two spectra while mining underlying feature interaction information through weighted fusion. β The dynamic adjustment capability enables the model to adaptively allocate the spectral contribution ratio for different coal quality indicators, avoid the dominance of a single spectrum, give full play to the complementary value of atomic and molecular information, and improve the fusion accuracy. β As the value approaches zero, the model becomes more reliant on molecular information from NIRS spectra. β As the value approaches 1, the model becomes more reliant on atomic information from the LIBS spectrum.

[0022] Furthermore, regarding the weighting factors β Apply boundary constraints, the boundary constraint being 0.1 ≤ β ≤0.9. This constraint range is based on experimental verification of the contribution of dual-spectral characteristics: through multi-index detection experiments on 56 coal samples, it was found that when... β When the value is <0.1, the contribution of LIBS atomic information is insufficient, and the detection accuracy of indicators dependent on elemental composition such as ash content and carbon content decreases significantly; when β When the value is greater than 0.9, the molecular information of NIRS is suppressed, and the error of indicators such as calorific value and volatile matter that depend on molecular structure increases. Therefore, this constraint range can ensure that both spectral features can effectively participate in model training and will not be dominated by a single spectrum.

[0023] Furthermore, in step S5, joint optimization is performed, including: S51. Initialize the candidate solution set of the parameter optimization algorithm, where each candidate solution represents a set of parameter vectors; S52. Substitute the parameter vector corresponding to each candidate solution into the regression model and calculate the value of the fitness function or objective function. S53. Based on the value of the fitness function or objective function, evaluate the quality of the current candidate solution set and record the current best candidate solution in history. S54. An iterative update strategy based on parameter optimization algorithm to update the candidate solution set or generate new candidate solutions; S55. Repeat steps S52 to S54 until the preset termination condition is met. Output the optimal candidate solution recorded when the termination condition is met as the optimal value of the parameter vector.

[0024] Furthermore, the fitness function is the root mean square error of cross-validation (RMSE). RMSECV The expression is: ; in, M This represents the number of samples in the cross-validation set; when using five-fold cross-validation, M It is 1 / 5 of the total number of samples in the training set; y m For the first m The true values ​​of coal quality indicators for each sample (i.e., the standard measured values ​​in the sample label); For the first m The measured values ​​of coal quality indicators for each sample.

[0025] Furthermore, the fitness function is a multi-objective weighted loss function. F fitness Calculated using the following formula: ; in, ; K The quantity of coal quality indicators; RMSECV k For the first k The root mean square error of each coal quality index on the cross-validation set RMSECV k Calculated using sample labels as the true values; w k This is a preset weighting coefficient for the k-th coal quality indicator, pre-assigned based on the precision requirements of the industrial site or the magnitude of the coal quality indicator values. For example, K=4, where the coal quality indicators are calorific value, ash content, volatile matter, and carbon content; in a coal-fired unit scenario, calorific value and ash content are core control indicators and can be set as calorific value weights. w 1 =0.4, gray weight w 2 =0.3, volatile matter weightw 3 =0.15, carbon content w 4 =0.15. If the numerical values ​​of coal quality indicators vary greatly (e.g., calorific value is in MJ / kg, carbon content is in %), the weights can be set as the reciprocal of the variance of each indicator to balance the impact of dimensional differences. By flexibly adapting the weight coefficients to the personalized needs of industrial scenarios, the detection accuracy of core indicators is ensured. Weight normalization avoids optimization bias caused by dimensional or priority differences, enabling the model to simultaneously improve the detection performance of all indicators.

[0026] Furthermore, in S2, the characteristic peak areas of several preset major elements are extracted from the LIBS spectrum, including: The center wavelengths of the characteristic spectral lines of each pre-defined major element were determined based on the NIST atomic spectral database. A half-window width is set for the characteristic spectral lines, and discrete data points within the window are extracted. The wavelength and intensity within the window are numerically integrated using the trapezoidal rule to obtain the area of ​​the characteristic peak. Further, in step S2, the NIRS spectrum is normalized using either maximum-minimum normalization or Z-score normalization.

[0027] Furthermore, the regression model is a kernel-based machine learning regression algorithm, such as Support Vector Regression (SVR), Kernel Ridge Regression (KRR), or Gaussian Process Regression (GPR). Support Vector Regression (SVR) achieves detection by finding the optimal regression hyperplane, exhibiting good adaptability to high-dimensional data and strong generalization ability. It is suitable for coal quality detection scenarios with moderate sample sizes (e.g., 50-200 samples) and high feature dimensions, making it the preferred model in this application. Kernel Ridge Regression (KRR) combines kernel methods with ridge regression, suppressing overfitting through regularization terms. Its computational complexity is lower than SVR, making it suitable for online detection scenarios with smaller sample sizes (e.g., 30-50 samples) and high computational speed requirements. Gaussian Process Regression (GPR), based on a Bayesian framework, can output the confidence interval of the detection results, making it suitable for scenarios with high uncertainty requirements (e.g., quality arbitration in coal trading). However, its computational complexity is higher, and its real-time performance is slightly weaker when the sample size is large. The core adaptability of these three regression models lies in their support for custom kernel functions, allowing direct input of the hybrid kernel matrix of this application without modifying the model structure.

[0028] This application also proposes a model construction method, including: S1. Obtain the LIBS and NIRS spectra of multiple coal samples; S2. Extract the characteristic peak area of ​​LIBS spectrum based on multiple pre-set coal sample elements to obtain the LIBS characteristic peak area matrix; denoise and normalize the NIRS spectrum to obtain the NIRS spectrum matrix. S3. Based on the radial basis function, construct kernel functions with independent bandwidth parameters for the LIBS characteristic peak area matrix and the NIRS spectral matrix, respectively; S4. Based on the weight factors obtained from adaptive learning, the two kernel functions are weighted and fused to construct a hybrid kernel function; S5. Using the mixed kernel function as the kernel function of any regression model, a parameter optimization algorithm is used to jointly optimize the parameter vector to obtain the optimal value of the parameter vector; wherein, the parameter vector includes the weight factor, the regularization parameter of the regression model, and the bandwidth parameters of the two kernel functions respectively.

[0029] Furthermore, step S5 is followed by a model training step: Obtain the standard measurement values ​​of coal quality indicators for each coal sample, and use the standard measurement values ​​as sample labels; Training and test data were separated from LIBS and NIRS spectra with sample labels. The training set feature vector is obtained by processing the spectrum in the training set through steps S2 to S4. The final mixing kernel matrix is ​​calculated on the training set feature vector. The selected kernel regression model is trained using the final mixing kernel matrix and the sample labels in the training set data as input. The regression model is support vector regression, kernel ridge regression or Gaussian process regression.

[0030] Furthermore, the model building method also includes a model validation step: After processing the spectra in the test set data through steps S2 to S4, the feature vector of the test set is obtained. The mixed kernel matrix between the feature vector of the test set and the feature vector of the training set is calculated. The mixed kernel matrix is ​​input into the trained kernel regression model, and the predicted value of coal quality index is output. The predicted value is compared with the sample label of the test set to evaluate the model performance.

[0031] This application also proposes a coal quality testing system, comprising: The acquisition module is used to acquire the LIBS spectrum, NIRS spectrum and standard measurement values ​​of coal quality indicators of multiple coal samples, and use the standard measurement values ​​as sample labels; The preprocessing module is used to extract the characteristic peak area of ​​LIBS spectra based on multiple preset coal sample elements to obtain the LIBS characteristic peak area matrix; and to denoise and normalize the NIRS spectra to obtain the NIRS spectral matrix. The kernel function construction module is used to construct kernel functions with independent bandwidth parameters for the LIBS characteristic peak area matrix and the NIRS spectral matrix, respectively, based on radial basis functions. The hybrid kernel function mapping module is used to construct a hybrid kernel function by weighting and fusing the two kernel functions obtained based on the weight factors of adaptive learning. The parameter optimization module is used to use the hybrid kernel function as the kernel function of any regression model, and to perform joint optimization of the parameter vector using a parameter optimization algorithm to obtain the optimal value of the parameter vector; wherein, the parameter vector includes weight factors, regularization parameters of the regression model, and bandwidth parameters of the two kernel functions respectively; The detection module inputs the optimal value of the parameter vector and the LIBS and NIRS spectra of the coal to be tested into the regression model, and outputs the detection results of coal quality indicators.

[0032] Furthermore, the kernel function construction module includes: The LIBS kernel function construction submodule is used to construct LIBS kernel functions, whose expressions are as follows: ; in, x i L , x j L These are the first two peaks in the LIBS characteristic peak area matrix. i The and the first j One sample vector; c L This refers to the bandwidth parameter of the first core. The NIRS kernel function construction submodule is used to construct NIRS kernel functions, whose expressions are as follows: ; in, x i N , x j N These are the first two spectral elements in the NIRS spectral matrix. i The and the first j The feature vector of each sample; γ N This refers to the bandwidth parameter of the second core.

[0033] Furthermore, the hybrid kernel function mapping module includes: The weighted fusion submodule is used to construct the fusion kernel function according to the following expression: ; in, x i , x j They represent the first i The and the first j The joint feature vector of the samples; β This is the weighting factor.

[0034] Furthermore, the hybrid kernel function mapping module also includes: Boundary constraint submodule: Used to apply boundary constraints to the weighting factor β, where the boundary constraint is 0.1 ≤ β ≤0.9.

[0035] Furthermore, in the parameter optimization module, the expression used to calculate the fitness function is a multi-objective weighted loss function. F fitness Calculated using the following formula: ; K represents the number of coal quality indicators; RMSECV k For the first k The root mean square error of each coal quality index on the cross-validation set RMSECV k Calculated using sample labels as the true values; w k For the first k The preset weighting coefficients for each coal quality index are pre-allocated based on the accuracy requirements of the industrial site or the magnitude of the coal quality index values.

[0036] The above-mentioned technical solution of this application has at least the following beneficial technical effects: (1) This application abandons the traditional approach of forcing heterogeneous data to be spliced ​​in the original data space and innovatively integrates them at the kernel space level; it constructs independent kernel functions for LIBS spectral data and NIRS spectral data with different physical properties and assigns them independent kernel bandwidth parameters; it allows the model to learn the optimal metric scale and rules for atomic spectra (sharp, discrete) and molecular spectra (broad, continuous) when calculating sample similarity, which fundamentally solves the problem of single kernel function mapping mismatch and variance suppression caused by data heterogeneity.

[0037] (2) This application introduces a learnable explicit fusion weight factor, transforming the implicit, fixed feature contribution in existing technologies into an explicit, dynamically adjustable kernel contribution. This weight factor, along with the regressor regularization parameter and two independent kernel bandwidth parameters, is incorporated into the global joint optimization framework of the parameter optimization algorithm. This enables the model to adaptively learn the optimal contribution ratio of LIBS and NIRS spectral information during training based on specific detection tasks (such as calorific value and ash content), achieving a true intelligent fusion of physical perception and task-driven processing, and greatly enhancing the physical interpretability of the model.

[0038] (3) To address the pain points of scarce industrial field samples and high cost of obtaining high-precision labels, this application organically combines "independent feature dimensionality reduction, multi-view interval constraints (applying boundary restrictions to the weight factor β), and cross-validation optimization." This strategy, while retaining the deep nonlinear interaction capability of heterogeneous spectra, fully leverages the natural advantage of kernel regression models in minimizing structural risk under small sample conditions, forming a fusion paradigm with extremely strong anti-overfitting ability and high robustness. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of this application or in the conventional technology, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of the coal quality detection method based on weighted multi-core learning heterogeneous spectroscopy in this application.

[0041] Figure 2 This is a schematic diagram showing the results of detecting calorific value using LIBS & NIRS spectroscopy in this application and in the prior art; wherein, (a) corresponds to this application; and (b) corresponds to a simple splicing model in the prior art.

[0042] Figure 3 This is a schematic diagram of the results of ash content detection using LIBS & NIRS spectroscopy in this application and the prior art; wherein, (a) corresponds to this application; and (b) corresponds to a simple splicing model of the prior art.

[0043] Figure 4 This is a schematic diagram of the results of volatile matter detection using LIBS & NIRS spectroscopy in this application and the prior art; wherein, (a) corresponds to this application; and (b) corresponds to a simple splicing model of the prior art.

[0044] Figure 5 This is a schematic diagram of the results of carbon content detection using LIBS & NIRS spectroscopy in this application and in the prior art; wherein, (a) corresponds to this application; and (b) corresponds to a simple splicing model in the prior art. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details are presented in the various embodiments of this application to facilitate a better understanding of the application. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the various embodiments below is for ease of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined with and referenced by each other without contradiction.

[0046] This application provides a coal quality testing method. In this embodiment, the online testing of coal entering the furnace of a coal-fired power plant is used as the application scenario. For 56 representative coal samples, the method simultaneously detects four key coal quality indicators: calorific value, ash content, volatile matter, and carbon content. The specific method includes: S1. Select 56 representative coal samples.

[0047] An online detection system using dynamic particle flow spectral acquisition is employed. The core of this system lies in the coaxial, shared optical path of LIBS and NIRS: the laser beam is focused onto the coal powder flow after passing through a beam splitter, inducing plasma generation; simultaneously, the NIRS light source (halogen tungsten lamp) illuminates the coal powder flow through the same optical path, acquiring diffuse reflectance spectra. This ensures that both LIBS and NIRS spectra reflect coal quality information from the exact same measurement area, avoiding deviations introduced by different sampling locations.

[0048] For each sample, 200 LIBS single-pulse spectra (laser wavelength 1064 nm, pulse energy 50 mJ, delay time 1.5 μs) and 200 NIRS spectra (wavelength range 1000–2500 nm, resolution 8 nm) were continuously acquired under dynamic particle flow conditions. The 200 LIBS spectra for each coal sample were averaged to obtain a representative LIBS spectrum for that coal sample; similarly, the 200 NIRS spectra were averaged to obtain a representative NIRS spectrum. This averaging process effectively eliminates the influence of particle flow fluctuations and random noise.

[0049] Each sample was sent to the laboratory for chemical analysis according to the national standard (GB / T 212-2008 "Industrial Analysis Methods for Coal"), and the following four coal quality indicators were determined: calorific value (MJ / kg), ash content (%), volatile matter (%), and carbon content (%). The measurement results were the standard measured values ​​of the coal quality indicators and were used as sample labels for subsequent regression model training. Table 1 lists the statistical distribution of each coal quality indicator for 56 samples. It can be seen that each coal quality indicator has a wide range, which is conducive to training a model with strong generalization ability.

[0050] Table 1. Statistical distribution of coal quality indicators for 56 coal samples.

[0051] S2. In order to eliminate the dimensional differences between different spectroscopic techniques, LIBS and NIRS spectra were processed independently.

[0052] S21. Preprocessing of the LIBS spectrum includes total intensity normalization, extraction of characteristic peak area, and normalization of the characteristic peak area in the column direction. Specifically: To minimize non-physical intensity fluctuations, the total intensity of the LIBS spectrum is first normalized. For the i-th pixel in the LIBS spectrum, its normalized intensity value is calculated as follows: ; in, I raw ( i ) is the first i The original spectral intensity of each pixel data point sum ( I raw ) represents the sum of the intensities of all pixels in the LIBS full spectrum.

[0053] Raw LIBS spectra typically contain thousands of wavelengths, and directly using the entire spectrum as a feature would be computationally intensive. However, coal quality information is primarily related to the emission lines of specific elements. Therefore, a physical-driven feature reduction method is employed: extracting the characteristic peak areas of major elements in coal, such as C, H, O, Si, Al, Ca, Fe, and Mg. Specifically: Based on the NIST atomic spectral database, the characteristic spectral center wavelengths of the main coal elements, including C (247.86 nm), H (656.23 nm), O (777.19 nm), Si (288.16 nm), Al (308.22 nm, 309.27 nm), Ca (393.37 nm, 396.85 nm), Fe (404.58 nm, 438.35 nm), and Mg (279.55 nm, 280.27 nm), were pre-determined and denoted as . locs j ( j =1,2,...,N). For each characteristic spectral line, set the half-window width w (e.g., w=3) based on its center wavelength. i The first sample j One characteristic peak, and the wavelength range is selected. locs j -w , locs j +wFor discrete data points within the window, the wavelength and intensity within the window are numerically integrated using the trapezoidal rule. The calculation formula is as follows: ; in, l start , l end Corresponding to locs j -w , locs j +w ; A i,j For the first i The sample at the th j The integral area at each characteristic peak; l This represents the wavelength variable in the spectrum.

[0054] l k For the first in the window k Wavelength of discrete data points; I i ( l k )for l k The normalized spectral intensity corresponding to the discrete data points.

[0055] l k+1 For the first in the window k+1 Wavelength of discrete data points; I i ( l k+1 )for l k+1 The normalized spectral intensity corresponding to the discrete data points.

[0056] Through the above integral operation, the original high-dimensional LIBS spectrum of a single sample is reduced to an N-dimensional (192-dimensional feature vector in this embodiment) feature vector. X feat .

[0057] Finally, the extracted feature peak area matrix is ​​subjected to Min-Max Normalization along the column (feature) direction to map it to the interval [0, 1], as shown in the following formula: ; Where, max( X feat ) is the feature vector Xfeat The maximum value, min( X feat ) is the feature vector X feat The minimum value. After normalization, all features are mapped to the interval [0, 1], eliminating the influence of dimensions and giving each feature a relatively equal initial contribution opportunity in the model.

[0058] S22 NIRS spectra are mainly affected by hydrogen-containing groups, resulting in weak signals that are easily interfered with by sample particle size, optical path length variations, and surface scattering. Standard Normal Variation (SNV) ensures that the NIRS spectra of each sample have the same mean and standard deviation, with scattering effects mainly manifesting as translation and scaling. SNV denoising filtering is applied to the NIRS data, effectively eliminating baseline shifts and tilts caused by sample particle size inhomogeneity, optical path length variations, and surface scattering. Then, the denoised NIRS data is subjected to min-max normalization to obtain the matrix. .

[0059] In this embodiment, LIBS spectra were acquired using a grating spectrometer. The original LIBS spectra of the 56 samples had 56×8192-dimensional features, and after preprocessing, the characteristic peak area features of the main elements were retained as 56×192-dimensional features.

[0060] NIRS spectra were acquired using a 256-pixel infrared spectrometer. Due to the low signal-to-noise ratio at the two ends of the infrared spectrometer, approximately 177 effective data points from the middle portion (e.g., 1000-2500 nm) with a higher signal-to-noise ratio were typically extracted. The acquired NIRS spectra were 56×177 dimensional features, and remained 56×177 dimensional after preprocessing. Unlike the physical dimensionality reduction of LIBS spectra, the full NIRS spectrum itself contains crucial molecular structural information. These continuous and smooth data points, after SNV denoising, can be directly used for kernel function similarity calculations without further dimensionality reduction to ensure information integrity.

[0061] S3. Based on radial basis functions, construct kernel functions with independent bandwidth parameters for the LIBS characteristic peak area matrix and the NIRS spectral matrix, respectively. The purpose of this step is to transfer the fusion problem of LIBS and NIRS spectra from the original data space to the kernel space. By constructing independent kernel functions for the two heterogeneous data, the fusion of physical sensing can be achieved.

[0062] The two kernel functions are the LIBS kernel function and the NIRS kernel function; The expression for the LIBS kernel function is: ; in, x i L , x j L These are the first two peaks in the LIBS characteristic peak area matrix. i The and the first j Feature vectors of each sample; c L The first kernel bandwidth parameter controls the similarity measurement scale of LIBS features in the kernel space.

[0063] The expression for the NIRS kernel function is: ; in, x i N , x j N These are the first two spectral elements in the NIRS spectral matrix. i The and the first j Feature vectors of each sample; c N This is the second kernel bandwidth parameter, adapted to the continuous smoothness characteristics of NIRS spectra.

[0064] The reason for choosing RBF as the basic kernel is that it does not rely on prior data knowledge, only requires optimization of a single kernel bandwidth parameter, has high computational efficiency, and its feature space mapping capability can approximate complex nonlinear relationships, adapting to the nonlinear correlation between coal quality indicators and spectral characteristics. By configuring independent kernel functions and kernel bandwidth parameters for the two spectra respectively, physically-aware kernel space mapping is achieved. c The larger the value, the smaller the influence range of the radial basis function, and the more the model tends to fit each sample point (prone to overfitting). c The smaller the value, the larger the range of influence, and the smoother the model (which is prone to underfitting). c L It can be optimized for the discrete characteristics of atomic spectra. c N It can adapt to the continuous characteristics of molecular spectra, fundamentally solving the problem of mismatch in single kernel function mapping. At the same time, the nonlinear mapping capability of the RBF kernel function can fully explore the deep correlation between spectral features and coal quality indicators, laying the foundation for subsequent fusion.

[0065] S4. Introduce adaptively learnable weight factors to weight and fuse the two kernel functions obtained in step S3 to construct a hybrid kernel function; the expression of the hybrid kernel function is: ; in, β As a weighting factor; xi , x j The first i The, the j The joint feature vector of each sample, covering the feature information of LIBS and NIRS spectra.

[0066] Preferably, for weighting factors β Apply boundary constraints, the boundary constraint being 0.1 ≤ β ≤0.9, ensuring that information from both modalities can be effectively utilized.

[0067] To avoid bias in the distribution of the training and test sets, the Kennard-Stone (KS) algorithm is used for sample partitioning. The KS algorithm selects highly representative samples for the training set, ensuring that the training set covers the entire feature space. The specific steps are as follows: First, calculate the Euclidean distance matrix of all samples in the preprocessed feature space; then, select the two samples with the greatest distance to each other to enter the training set; subsequently, each time, select the sample with the largest minimum distance to the selected training set samples and add it until a preset ratio is reached. In this embodiment, 56 samples are divided into 45 training sets and 11 test sets at an 8:2 ratio. The training set is used for model training and hyperparameter optimization, while the test set is used for final model performance evaluation.

[0068] S5. Using the mixed kernel function as the kernel function of any regression model, a parameter optimization algorithm is used to jointly optimize the parameter vector to obtain the optimal value of the parameter vector; wherein, the parameter vector includes the weight factor, the regularization parameter of the regression model, and the bandwidth parameters of the two kernel functions respectively.

[0069] The regression model can be Support Vector Regression (SVR), Kernel Ridge Regression (KR), or Gaussian Process Regression (GSR). Taking SVR as an example, it uses a hybrid kernel matrix to find the optimal regression hyperplane in a high-dimensional feature space, thereby achieving accurate detection of coal quality indicators. Since the hybrid kernel matrix integrates the physical information of two spectra, the subsequent regressor only needs to learn based on this kernel matrix. The regularization parameter of SVR is the penalty factor C, which, along with other parameters, constitutes a parameter vector. .

[0070] In this embodiment, the parameter optimization algorithm adopts the particle swarm optimization (PSO) algorithm, which aims to minimize the prediction error and jointly optimize the parameter vector to obtain the optimal value of the parameter vector; wherein, in step S5, the joint optimization using the particle swarm optimization algorithm includes: Initialize the particle swarm, extract the position vector of the current particle, where each particle's position vector represents a set of parameter vectors, and initialize the velocity of each particle; taking Support Vector Regression (SVR) as an example, the specific settings are as follows: The population size of PSO is set to 30, and the maximum number of iterations is set to 50. The boundary conditions for the parameters to be optimized are defined as follows: penalty factor C ∈ [0.1, 1000], γ... L ∈[0.001,10], γ N ∈[0.001,10], β∈[0.1,0.9].

[0071] Substitute the parameter vector corresponding to each particle into the regression model to calculate the value of the fitness function; where the fitness function is a multi-objective weighted loss function. F fitness Calculated using the following formula: ; K represents the number of coal quality indicators; RMSECV k For the first k The root mean square error of each coal quality index on the cross-validation set RMSECV k Calculated using sample labels as the true values; w k For the first k The preset weighting coefficients for each coal quality index are pre-allocated based on the accuracy requirements of the industrial site or the magnitude of the coal quality index values.

[0072] Based on the value of the fitness function, update the individual historical best position of each particle and the global historical best position of the particle swarm; Update the velocity and position of each particle based on its individual historical best position and the global historical best position; Repeat steps S52 to S54 until the preset termination condition (convergence) is met, and output the global historical best position as the optimal value of the parameter vector.

[0073] S6. Configure the regression model using the optimal value of the parameter vector, and then input the optimal value of the parameter vector and the LIBS spectrum and NIRS spectrum of the coal to be tested into the regression model to output the detection results of coal quality indicators.

[0074] This application inputs the LIBS and NIRS spectra of the sample into a simple stitching detection model in the prior art, and compares and evaluates the detection results of the above embodiments of this application with the detection results of the simple stitching detection model: The simple concatenation detection model directly concatenates the preprocessed LIBS features (192-dimensional) and NIRS features (177-dimensional) into a 369-dimensional feature vector, and then performs Min-Max normalization. It is then optimized using the PSO-SVR framework, where the SVR kernel function employs a single RBF kernel (kernel parameters...). c The PSO search vector space is [C∈[0.1,1000], c∈[0.001,10]], the optimization objective is to minimize the RMSECV of conventional five-fold cross-validation. The simple concatenation model does not involve... β Weights and dual-core parameters.

[0075] Figure 2-Figure 5 The comparison of prediction results between this application and a simple stitching model is shown, where the vertical axis represents the model prediction value of the method in this application or the comparative method, and the horizontal axis represents the reference value (the standard measured value according to the national standard method). It can be seen that: Figure 2 Calorific value index: The predicted R value of the regression model in this application 2 The RMSEP reached 0.976, with a low RMSEP of 0.366 MJ / kg; in comparison, the RMSEP of the simple splicing model was much lower. 2 The value was 0.935, and the RMSEP was 0.609 MJ / kg, with all errors significantly reduced.

[0076] Figure 3 Ash index: In the regression model R of this application 2 The RMSEP reached 0.981, and was as low as 0.782%; while the RMSEP of the simple splicing model was... 2 It is only 0.967, while the RMSEP is as high as 1.061%.

[0077] Figure 4 Volatile matter index: Also under a balanced weighting of 0.5, the test R-value of the regression model in this application is... 2 The RMSEP reached 0.923, and was as low as 0.496%; the RMSEP of the simple splicing model was... 2 The value is 0.812, and the RMSEP is 0.775%.

[0078] Figure 5 Carbon content index: Test R of the regression model in this application 2 The RMSEP reached 0.980, with a low RMSEP of 0.880%; in comparison, the RMSEP of the simple splicing model was much lower. 2 The value was 0.939, and the RMSEP was as high as 1.527%.

[0079] The comparison of prediction results proves that the detection method proposed in this application can break through the variance suppression of heterogeneous information by the traditional splicing method. Through the multi-kernel learning model, it effectively achieves the proportional deep fusion of atomic features (LIBS) and molecular features (NIRS) in the bottom kernel space, thereby consistently improving the quantitative analysis accuracy and reliability of all key coal quality indicators.

[0080] This application also proposes a model construction method, including: S1. Obtain the LIBS and NIRS spectra of multiple coal samples; S2. Extract the characteristic peak area of ​​LIBS spectrum based on multiple pre-set coal sample elements to obtain the LIBS characteristic peak area matrix; denoise and normalize the NIRS spectrum to obtain the NIRS spectrum matrix. S3. Based on the radial basis function, construct kernel functions with independent bandwidth parameters for the LIBS characteristic peak area matrix and the NIRS spectral matrix, respectively; S4. Based on the weight factors obtained from adaptive learning, the two kernel functions are weighted and fused to construct a hybrid kernel function; S5. Using the mixed kernel function as the kernel function of any regression model, a parameter optimization algorithm is used to jointly optimize the parameter vector to obtain the optimal value of the parameter vector; wherein, the parameter vector includes the weight factor, the regularization parameter of the regression model, and the bandwidth parameters of the two kernel functions respectively.

[0081] Furthermore, step S5 is followed by a model training step: Obtain the standard measurement values ​​of coal quality indicators for each coal sample, and use the standard measurement values ​​as sample labels; Training and test data were separated from LIBS and NIRS spectra with sample labels. The training set feature vector is obtained by processing the spectrum in the training set through steps S2 to S4. The final mixing kernel matrix is ​​calculated on the training set feature vector. The selected kernel regression model is trained using the final mixing kernel matrix and the sample labels in the training set data as input. The regression model is support vector regression, kernel ridge regression or Gaussian process regression.

[0082] Furthermore, the model building method also includes a model validation step: After processing the spectra in the test set data through steps S2 to S4, the feature vector of the test set is obtained. The mixed kernel matrix between the feature vector of the test set and the feature vector of the training set is calculated. The mixed kernel matrix is ​​input into the trained kernel regression model, and the predicted value of coal quality index is output. The predicted value is compared with the sample label of the test set to evaluate the model performance.

[0083] This application also proposes a coal quality testing system, comprising: The acquisition module is used to acquire the LIBS spectrum, NIRS spectrum and standard measurement values ​​of coal quality indicators of multiple coal samples, and use the standard measurement values ​​as sample labels; The preprocessing module is used to extract the characteristic peak area of ​​LIBS spectra based on multiple preset coal sample elements to obtain the LIBS characteristic peak area matrix; and to denoise and normalize the NIRS spectra to obtain the NIRS spectral matrix. The kernel function construction module is used to construct kernel functions with independent bandwidth parameters for the LIBS characteristic peak area matrix and the NIRS spectral matrix, respectively, based on radial basis functions. The hybrid kernel function mapping module is used to construct a hybrid kernel function by weighting and fusing the two kernel functions obtained based on the weight factors of adaptive learning. The parameter optimization module is used to use the hybrid kernel function as the kernel function of any regression model, and to perform joint optimization of the parameter vector using a parameter optimization algorithm to obtain the optimal value of the parameter vector; wherein, the parameter vector includes weight factors, regularization parameters of the regression model, and bandwidth parameters of the two kernel functions respectively; The detection module is used to input the optimal value of the parameter vector and the LIBS and NIRS spectra of the coal to be tested into the regression model, and output the detection results of coal quality indicators.

[0084] Furthermore, the kernel function construction module includes: The LIBS kernel function construction submodule is used to construct LIBS kernel functions, whose expressions are as follows: ; in, x i L , x j L These are the first two peaks in the LIBS characteristic peak area matrix. i The and the first j One sample vector; c L This refers to the bandwidth parameter of the first core. The NIRS kernel function construction submodule is used to construct NIRS kernel functions, whose expressions are as follows: ; in, x i N , x j N These are the first two spectral elements in the NIRS spectral matrix. i The and the first j The feature vector of each sample; γ N This refers to the bandwidth parameter of the second core.

[0085] Furthermore, the hybrid kernel function mapping module includes: The weighted fusion submodule is used to construct the fusion kernel function according to the following expression: ; in, x i , x j They represent the first i The and the first j The joint feature vector of the samples;β This is the weighting factor.

[0086] Furthermore, the hybrid kernel function mapping module also includes: Boundary constraint submodule: Used to apply boundary constraints to the weighting factor β, where the boundary constraint is 0.1 ≤ β ≤0.9.

[0087] Furthermore, in the parameter optimization module, the expression used to calculate the fitness function is a multi-objective weighted loss function. F fitness Calculated using the following formula: ; K represents the number of coal quality indicators; RMSECV k For the first k The root mean square error of each coal quality index on the cross-validation set RMSECV k Calculated using sample labels as the true values; w k For the first k The preset weighting coefficients for each coal quality index are pre-allocated based on the accuracy requirements of the industrial site or the magnitude of the coal quality index values.

[0088] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of this application and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of this application should be included within the protection scope of this application. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A method of coal quality detection, characterized by, include: S1. Obtain the LIBS and NIRS spectra of multiple coal samples, as well as the standard measurement values ​​of coal quality indicators for each sample. S2. Extract the characteristic peak area of ​​LIBS spectrum based on multiple pre-set coal sample elements to obtain the LIBS characteristic peak area matrix; denoise and normalize the NIRS spectrum to obtain the NIRS spectrum matrix. S3. Based on the radial basis function, construct kernel functions with independent bandwidth parameters for the LIBS characteristic peak area matrix and the NIRS spectral matrix, respectively. The two kernel functions are the LIBS kernel function and the NIRS kernel function. The expression for the LIBS kernel function is: ； wherein, x i L , x j L are the first and second sample vectors in the LIBS feature peak area matrix, respectively; i j γ L is a first kernel bandwidth parameter;​​ The expression for the NIRS kernel function is: ； in, x i N , x j N These are the first two spectral elements in the NIRS spectral matrix. i The and the first j The feature vector of each sample; γ N This refers to the bandwidth parameter of the second core. S4. Based on the weight factors obtained from adaptive learning, the two kernel functions are weighted and fused to construct a hybrid kernel function. The expression of the hybrid kernel function is as follows: ； in, x i , x j They represent the first i The and the first j The joint feature vector of the samples; β As a weighting factor; S5. Using the mixed kernel function as the kernel function of any regression model, a parameter optimization algorithm is used to jointly optimize the parameter vector to obtain the optimal value of the parameter vector; wherein, the parameter vector includes the weight factor, the regularization parameter of the regression model, and the bandwidth parameters of the two kernel functions respectively; S6. Input the optimal value of the parameter vector and the LIBS and NIRS spectra of the coal to be tested into the regression model, and output the detection results of the coal quality indicators.

2. The coal quality testing method according to claim 1, characterized in that, Weighting factors β Apply boundary constraints, the boundary constraint being 0.1 ≤ β ≤0.

9.

3. The coal quality testing method according to claim 1, characterized in that, Step S5 involves joint optimization, including: S51. Initialize the candidate solution set of the parameter optimization algorithm, where each candidate solution represents a set of parameter vectors; S52. Substitute the parameter vector corresponding to each candidate solution into the regression model and calculate the value of the fitness function or objective function. S53. Based on the value of the fitness function or objective function, evaluate the quality of the current candidate solution set and record the current best candidate solution in history. S54. An iterative update strategy based on parameter optimization algorithm to update the candidate solution set or generate new candidate solutions; S55. Repeat steps S52 to S54 until the preset termination condition is met. Output the optimal candidate solution recorded when the termination condition is met as the optimal value of the parameter vector.

4. The coal quality testing method according to claim 3, characterized in that, The fitness function is a multi-objective weighted loss function. F fitness Calculated using the following formula: ； K The quantity of coal quality indicators; RMSECV k For the first k The root mean square error of each coal quality index on the cross-validation set; w k For the first k The preset weighting coefficients for each coal quality index are pre-allocated based on the accuracy requirements of the industrial site or the magnitude of the coal quality index values.

5. The coal quality testing method according to claim 1, characterized in that, S2 extracts the characteristic peak areas of LIBS spectra based on multiple preset coal sample elements, including: The center wavelength of the characteristic spectral lines of each element was determined based on the NIST atomic spectral database; A half-window width is set for the characteristic spectral lines, and discrete data points within the window are extracted. The wavelength and intensity within the window are numerically integrated using the trapezoidal rule to obtain the area of ​​the characteristic peak.

6. The coal quality testing method according to claim 1, characterized in that, Step S2, obtaining the LIBS characteristic peak area matrix, also includes: Before extracting the characteristic peak area of ​​the LIBS spectrum, the total intensity of the LIBS spectrum is normalized. After extracting the characteristic peak area of ​​the LIBS spectrum, the characteristic peak area is normalized by maximum and minimum.

7. The coal quality testing method according to claim 1, characterized in that, In step S2, the NIRS spectrum is normalized using either maximum-minimum normalization or Z-score normalization.

8. The coal quality testing method according to claim 1, characterized in that, Coal quality indicators include at least one of the following: calorific value, ash content, volatile matter, and carbon content.

9. A model construction method, characterized in that, include: S1. Obtain the LIBS and NIRS spectra of multiple coal samples; S2. Extract the characteristic peak area of ​​LIBS spectrum based on multiple pre-set coal sample elements to obtain the LIBS characteristic peak area matrix; denoise and normalize the NIRS spectrum to obtain the NIRS spectrum matrix. S3. Based on the radial basis function, construct kernel functions with independent bandwidth parameters for the LIBS characteristic peak area matrix and the NIRS spectral matrix, respectively. The two kernel functions are the LIBS kernel function and the NIRS kernel function. The expression for the LIBS kernel function is: ； in, x i L , x j L These are the first two peaks in the LIBS characteristic peak area matrix. i The and the first j One sample vector; γ L This refers to the bandwidth parameter of the first core. The expression for the NIRS kernel function is: ； in, x i N , x j N These are the first two spectral elements in the NIRS spectral matrix. i The and the first j The feature vector of each sample; γ N This refers to the bandwidth parameter of the second core. S4. Based on the weight factors obtained from adaptive learning, the two kernel functions are weighted and fused to construct a hybrid kernel function. The expression of the hybrid kernel function is as follows: ； in, x i , x j They represent the first i The and the first j The joint feature vector of the samples; β As a weighting factor; S5. Using the mixed kernel function as the kernel function of any regression model, a parameter optimization algorithm is used to jointly optimize the parameter vector to obtain the optimal value of the parameter vector; wherein, the parameter vector includes the weight factor, the regularization parameter of the regression model, and the bandwidth parameters of the two kernel functions respectively.

10. The model construction method according to claim 9, characterized in that, Step S5 is followed by a model training step: Obtain the standard measurement values ​​of coal quality indicators for each coal sample, and use the standard measurement values ​​as sample labels; Training and test data were separated from LIBS and NIRS spectra with sample labels. The training set feature vector is obtained by processing the spectrum in the training set through steps S2 to S4. The final mixing kernel matrix is ​​calculated on the training set feature vector. The selected kernel regression model is trained using the final mixing kernel matrix and the sample labels in the training set data as input. The regression model is support vector regression, kernel ridge regression or Gaussian process regression.

11. The model construction method according to claim 10, characterized in that, It also includes a model validation step: After processing the spectra in the test set data through steps S2 to S4, the feature vector of the test set is obtained. The mixed kernel matrix between the feature vector of the test set and the feature vector of the training set is calculated. The mixed kernel matrix is ​​input into the trained kernel regression model, and the predicted value of coal quality index is output. The predicted value is compared with the sample label of the test set to evaluate the model performance.

12. A coal quality testing system, characterized in that, include: The acquisition module is used to acquire the LIBS spectrum, NIRS spectrum, and standard measurement values ​​of coal quality indicators for each coal sample, and to use the measurement values ​​as sample labels. The preprocessing module is used to extract the characteristic peak area of ​​LIBS spectra based on multiple preset coal sample elements to obtain the LIBS characteristic peak area matrix; and to denoise and normalize the NIRS spectra to obtain the NIRS spectral matrix. The kernel function construction module is used to construct kernel functions with independent bandwidth parameters for the LIBS characteristic peak area matrix and the NIRS spectral matrix, respectively, based on radial basis functions. The kernel function construction module includes: The LIBS kernel function construction submodule is used to construct LIBS kernel functions, whose expressions are as follows: ； in, x i L , x j L These are the first two peaks in the LIBS characteristic peak area matrix. i The and the first j One sample vector; γ L This refers to the bandwidth parameter of the first core. The NIRS kernel function construction submodule is used to construct NIRS kernel functions, whose expressions are as follows: ； in, x i N , x j N These are the first two spectral elements in the NIRS spectral matrix. i The and the first j The feature vector of each sample; γ N This refers to the bandwidth parameter of the second core. The hybrid kernel function mapping module is used to weight and fuse the two kernel functions obtained based on the adaptive learning weight factors to construct a hybrid kernel function. The hybrid kernel function mapping module includes: The weighted fusion submodule is used to construct the fusion kernel function according to the following expression: ； in, x i , x j They represent the first i The and the first j The joint feature vector of the samples; β As a weighting factor; The parameter optimization module is used to use the hybrid kernel function as the kernel function of any regression model, and to perform joint optimization of the parameter vector using a parameter optimization algorithm to obtain the optimal value of the parameter vector; wherein, the parameter vector includes weight factors, regularization parameters of the regression model, and bandwidth parameters of the two kernel functions respectively; The detection module is used to input the optimal value of the parameter vector and the LIBS and NIRS spectra of the coal to be tested into the regression model, and output the detection results of coal quality indicators.

13. The coal quality testing system according to claim 12, characterized in that, The hybrid kernel function mapping module also includes: Boundary constraint submodule: used for weighting factors β Apply boundary constraints, the boundary constraint being 0.1 ≤ β ≤0.

9.

14. The coal quality testing system according to claim 12, characterized in that, In the parameter optimization module, the expression used to calculate the fitness function is a multi-objective weighted loss function. F fitness Calculated using the following formula: ； K The quantity of coal quality indicators; RMSECV k For the first k The root mean square error of each coal quality index on the cross-validation set; w k For the first k The preset weighting coefficients for each coal quality index are pre-allocated based on the accuracy requirements of the industrial site or the magnitude of the coal quality index values.

Images