Coal ash content detection method based on laser-induced breakdown spectroscopy data and machine learning
By combining laser-induced breakdown spectroscopy and machine learning, a non-destructive and rapid coal air-dry basis ash content detection model was constructed, which solved the problems of destructiveness and long cycle of traditional detection methods, realized in-situ rapid detection of coal ash content, and improved detection accuracy and efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHEASTERN UNIV CHINA
- Filing Date
- 2026-05-29
- Publication Date
- 2026-07-14
Smart Images
Figure CN122385579A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of coal quality testing technology, and in particular to a method for detecting coal ash content on a dry basis based on laser-induced breakdown spectral data and machine learning. Background Technology
[0002] Currently, the mainstream methods for detecting the ash content of coal on an air-dried basis in the industry include slow ashing, rapid ashing, and infrared ash analysis. Among them, slow ashing is the benchmark method for coal ash content determination and is widely used in national standards and specifications.
[0003] The core principle of the slow ashing method is to place the coal sample in a muffle furnace and gradually increase the temperature according to a prescribed procedure until the combustible components in the coal are completely burned off, and the inorganic minerals are converted into stable oxides. After multiple cooling, weighing, and constant weight determination, the air-dried ash content of the coal sample is calculated based on the mass of the residue. Although this method has the advantages of stable and accurate test results, its testing process is extremely lengthy, often requiring several hours for a single complete test. It necessitates strict control of the heating rate, ashing temperature, and holding time, and the operation steps are cumbersome and require a high degree of standardization from the laboratory personnel. Furthermore, the testing process can only be carried out in batches, resulting in low sample processing efficiency, which cannot meet the high-efficiency testing requirements of continuous production and real-time quality control in modern coal industry.
[0004] Rapid ashing is a commonly used method for rapid ash content detection in enterprise laboratories and quality inspection scenarios. Its core principle is to rapidly burn and ashing the coal sample in a high-temperature furnace by increasing the ashing temperature and shortening the heating time, and then calculating the ash content by weighing. Compared with slow ashing, this method significantly shortens the single test time, but it still has obvious limitations: First, the test is still destructive, and the coal sample is completely ineffective after high-temperature ashing and cannot be recovered and reused, resulting in high costs in scenarios such as determining scarce coal types and online batch testing; Second, the test relies on laboratory equipment such as large high-temperature furnaces and dryers, which are bulky and inconvenient to move, making it difficult to achieve in-situ rapid testing in coal mines, coal storage yards, coal blending workshops, etc.; Third, rapid ashing can easily lead to coal sample deflagration, ash splashing, or incomplete ashing, affecting the repeatability of the test results, requiring multiple parallel tests to ensure accuracy.
[0005] In addition to the inherent shortcomings of conventional detection methods, ash content classification still relies heavily on manual experience in coal mining, screening, and on-site coal blending. This method primarily depends on operators observing the coal's color, luster, and impurity content for rough classification, lacking quantitative testing data support. It suffers from significant subjective errors, low accuracy, and low efficiency, failing to meet the production requirements of intelligent sorting, refined coal blending, and clean and efficient utilization in modern coal industry. Therefore, there is a need for a rapid detection technology based on spectral analysis and intelligent algorithms to quickly and accurately output ash content data. Summary of the Invention
[0006] To address the aforementioned technical problems, this invention provides a method for detecting coal ash content on a dry basis based on laser-induced breakdown spectroscopy data and machine learning. The core objective is to overcome the technical shortcomings of existing traditional chemical detection methods, such as high destructiveness, long detection cycles, and inability to perform in-situ detection. This invention uses laser-induced breakdown spectroscopy as the data acquisition foundation, and constructs a complete detection system following standardized spectral data acquisition, spectral data preprocessing, prediction model construction and global intelligent optimization of model parameters, and model cross-validation training. Ultimately, this achieves non-destructive, rapid, in-situ, and accurate detection of coal ash content on a dry basis.
[0007] The technical means employed in this invention are as follows:
[0008] A method for detecting coal ash content on a dry basis based on laser-induced breakdown spectral data and machine learning includes: acquiring spectral data using a standardized five-point sampling method and performing data preprocessing; constructing a coal LIBS data feature extraction model and extracting global and local features from the spectral data using a multi-layer CNN network; introducing a Bayesian optimization algorithm to globally adaptively optimize the hyperparameters in the BLS model and constructing a BO-BLS coal ash content detection model; iteratively optimizing the BO-BLS coal ash content detection model using regularization and verifying the model performance using K-fold cross-validation.
[0009] Furthermore, the construction of the coal LIBS data feature extraction model specifically includes: constructing a 4-layer one-dimensional convolutional neural network, while introducing BatchNorm batch normalization and Dropout random deactivation regularization strategies to alleviate model overfitting and output a hierarchical dataset; setting up multiple independent CNN networks with the same structure but different initialization parameters and Dropout rates, inputting the preprocessed spectral data into the multiple networks respectively, and concatenating the outputs to generate CNN deep features; and using the partial least squares regression (PLS) method to compress the dimensions of the hierarchical dataset and the CNN deep features respectively, and extracting the feature components most correlated with coal ash content.
[0010] Furthermore, the construction of the BO-BLS coal air-dry basis ash content detection model specifically includes: encoding the number of feature nodes N1, the number of enhancement nodes N2, and the ridge regression regularization parameter λ of the BLS model into hyperparameter solution vectors. Initialize the algorithm hyperparameters and solution search space, define the legal search range of the hyperparameters to be optimized, and constrain the boundaries of hyperparameter values; Traverse the hyperparameter solution vectors of the initial sampling points, construct a BLS model based on the hyperparameters corresponding to each set of solution vectors, and after completing the model training, calculate the RMSE value of the model on the validation set as the objective function value corresponding to that set of hyperparameters; compare the objective function value of the current sampling point with the historical individual optimal solution, update the individual optimal hyperparameter combination, and at the same time select the global optimal hyperparameter combination with the smallest objective function value among all sampling points and update the global optimal solution; Based on existing sampling points and corresponding objective function values, a Gaussian process surrogate model is constructed using the Matérn kernel function. The mapping relationship between hyperparameter combinations and model prediction performance is fitted. The expected improvement is used as the acquisition function. By maximizing the acquisition function, the potentially optimal hyperparameter sampling points in the current iteration are selected. A BLS model is constructed based on the newly selected hyperparameter sampling points. After training, the RMSE (Recovery Mean Squared Error) on the validation set is calculated, which is the objective function value at each sampling point. Simultaneously, the change between this value and the current globally optimal objective function value is calculated. If ΔRMSE is less than the set minimum change threshold, or the number of iterations reaches the preset maximum number of iterations, the iterative optimization process is terminated; otherwise, iterative optimization continues. After the iteration terminates, the number of feature nodes N1, the number of enhancement nodes N2, and the regularization parameter λ corresponding to the obtained global optimal solution are used as the optimal hyperparameter combination of the BLS model to complete the construction of the BO-BLS coal air-dry basis ash content detection model.
[0011] Furthermore, when iteratively optimizing the BO-BLS coal air-dry basis ash content detection model, multiple regularization methods are adopted and an adaptive learning rate strategy is introduced to optimize the model; in the verification stage, the K-fold cross-validation method is used, and the average value of the evaluation index of K experiments is used as the performance index of the model.
[0012] Compared with the prior art, the present invention has the following advantages: This invention discloses a method for detecting the ash content of coal on an air-dry basis based on laser-induced breakdown spectroscopy and machine learning. The core objective is to overcome the technical shortcomings of existing traditional chemical detection methods, such as high destructiveness, long detection cycles, and inability to detect ash content in situ. This invention uses laser-induced breakdown spectroscopy as the data acquisition foundation, and constructs a complete detection system following standardized spectral data acquisition, spectral data preprocessing, prediction model construction and global intelligent optimization of model parameters, and model cross-validation training. Ultimately, this achieves non-destructive, rapid, in-situ, and accurate detection of the ash content of coal on an air-dry basis.
[0013] The present invention provides a method for detecting coal ash content on a dry basis based on laser-induced breakdown spectral data and machine learning. By using the Bayesian optimization algorithm (BO) to globally adaptively optimize the number of feature nodes, the number of enhancement nodes, and the ridge regression regularization parameters of the width learning system (BLS), it effectively avoids the problems of insufficient generalization ability, large fluctuations in prediction results, and difficulty in achieving global optimal detection accuracy of the traditional BLS model.
[0014] This invention relies on rapid detection technology based on spectral analysis and intelligent algorithms. It eliminates the need for complex sample pretreatment and chemical reagents, and can output ash content data in a short time. It can provide efficient and reliable technical support for coal sorting, coal blending and combustion, and quality grading, and is more in line with the development needs of modern clean and efficient utilization of coal. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 This is a flowchart of the coal air dry basis ash content detection method based on laser-induced breakdown spectral data and machine learning in this invention.
[0017] Figure 2 This is a flowchart of the optimization training procedure for the BO-BLS coal air-dried basis ash content detection model in this invention.
[0018] Figure 3 This is a scatter plot comparing the predicted and actual values of the CNN-BLS model on the test set in this embodiment of the invention.
[0019] Figure 4 This is a line graph showing the comparison between the predicted and actual values of the CNN-BLS model on the test set in this embodiment of the invention. Detailed Implementation
[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0023] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.
[0024] like Figure 1As shown, this invention provides a method for detecting the dry basis ash content of coal based on laser-induced breakdown spectral data and machine learning. The method includes: acquiring spectral data using a standardized five-point sampling method and performing data preprocessing. Specifically, in a preferred embodiment, the standardized five-point sampling method involves: uniformly selecting five sampling points on the surface of the coal sample; adjusting the height of the laser focusing lens to precisely focus a 1064nm wavelength laser onto the sample surface; and simultaneously adjusting the position of the fiber optic probe to ensure full collection of plasma characteristic light. For each sampling point, the automatic surface cleaning function of the spectrometer is triggered using ACCULIBS software to remove spectral distortion caused by dust, contaminants, etc., from the sample surface. Then, the sampling point undergoes five consecutive laser shocks, and the average of the five collected spectral data is taken as the representative spectrum of the sampling point. After completing the spectral acquisition of the five sampling points, the average of the representative spectra of the five sampling points is taken again as the LIBS spectral data of the coal sample. The sample used in this invention undergoes multiple processing steps to ultimately obtain 17-dimensional sample data.
[0025] During implementation, three types of samples were collected: coal, coal gangue, and surrounding rock. This ensured the geographical coverage and compositional diversity of the samples, thereby improving the generalization ability of the final detection model at the sample level. Spectral data acquisition was performed using an Ocean Insight MX2500+ spectrometer, equipped with 8 detection channels, covering a wavelength detection range of 199-1112nm, encompassing 15,527 spectral bands across the entire spectrum.
[0026] The data processing is performed in the order of wavelet transform denoising and principal component analysis (PCA) dimensionality reduction. The core objective is to eliminate noise interference and reduce data dimensionality while preserving the core spectral feature information, thereby improving the computational efficiency and detection accuracy of subsequent models. In a preferred embodiment of this invention, data preprocessing includes wavelet transform denoising, data augmentation, and standard normal variable transformation.
[0027] Wavelet transform denoising uses a comprehensive evaluation index Q based on signal-to-noise ratio (SNR) and relative peak error (REPV). SNR characterizes the relative noise content in the signal, while REPV characterizes the distortion of spectral characteristic peaks after denoising. The comprehensive evaluation index Q is obtained by normalizing 1 / SNR and REPV respectively and then summing them; a smaller Q value indicates better denoising performance. Through multiple sets of comparative experiments, this invention finally determined the optimal parameter combination for wavelet denoising: the heursure threshold is selected as the threshold function, sym6 is selected as the wavelet basis function, and the decomposition level is 4. Under this parameter combination, the SNR of the spectral data can reach 64.3508, the REPV is as low as 0.0018, and the comprehensive evaluation index Q is 0, achieving the optimal balance between noise removal and feature information preservation.
[0028] To address the overfitting issue in small-sample scenarios, data augmentation employs dynamic Gaussian noise, random scaling, polynomial baseline shift, and piecewise scaling to enhance spectral data. While preserving the intrinsic correlation between spectrum and ash content, the original 75 training samples are expanded to 825. In implementation, dynamic Gaussian noise is generated based on the sample's own standard deviation to avoid excessive perturbation. Random scaling applies a scaling factor independently to each sample to simulate spectral intensity fluctuations. Polynomial baseline shift introduces linear offset to simulate instrument baseline drift interference. Piecewise scaling divides the spectrum into two segments at the midpoint and scales them separately to simulate local frequency domain characteristic changes.
[0029] Standard normal variable transformation is used to eliminate interference caused by spectral baseline drift and scattering effects; RobustScaler robust normalization method is used to scale the data, eliminate baseline offset, and improve the data's resistance to outliers and robustness.
[0030] To address the issues of single CNN feature extraction easily filtering out key effective features related to ash content detection in the original spectrum, the mismatch between the dimensions of high-dimensional original spectra and low-dimensional CNN deep features preventing direct fusion, and insufficient feature representation capabilities, a coal LIBS data feature extraction model is constructed. This model utilizes a hierarchical CNN network to extract global and local features from the spectral data. Specifically, as a preferred embodiment of this invention, the construction of the coal LIBS data feature extraction model includes: constructing a 4-layer one-dimensional convolutional neural network, employing a hierarchical design that uses large convolutional kernels to capture global features and progressively narrows the receptive field to extract local features; introducing BatchNorm batch normalization and Dropout random deactivation regularization strategies to alleviate model overfitting; and outputting a 16-dimensional hierarchical dataset. Multiple independent CNN networks with identical structures but different initialization parameters and Dropout rates are set up. The preprocessed 15527-dimensional spectral data is input into three networks, and the outputs are concatenated to form a 48-dimensional dataset, generating CNN deep features and reducing random errors caused by single network initialization.
[0031] To address the issue of dimensionality mismatch between CNN deep features and original spectral features, which prevent direct concatenation, PLS-based feature dimensionality reduction was employed to compress the dimensions of both the hierarchical dataset and the CNN deep features, extracting the feature components most correlated with coal ash content. The 48-dimensional CNN deep features were reduced to 11 dimensions using PLS, serving as input for subsequent prediction models.
[0032] To address the issues of the Wide Learning System (BLS) relying on multiple rounds of manual tuning for the number of feature nodes, augmentation nodes, and ridge regression regularization parameters, which leads to insufficient feature representation, overfitting or underfitting, large fluctuations in results, inability to achieve global optimum detection accuracy, and extremely low parameter tuning efficiency, this invention innovatively introduces the Bayesian Optimization (BO) algorithm to globally adaptively optimize the three hyperparameters of the BLS model. This constructs the BO-BLS coal air-dry basis ash content detection model, achieving global optimum matching and automated tuning of the model's hyperparameters.
[0033] A Bayesian optimization algorithm is introduced to globally adaptively optimize the hyperparameters in the BLS model, thus constructing a BO-BLS coal air-dry basis ash content detection model. Specifically, as a preferred embodiment of this invention, the construction of the BO-BLS coal air-dry basis ash content detection model includes: The number of feature nodes N1, the number of augmentation nodes N2, and the ridge regression regularization parameter λ of the BLS model are encoded into hyperparameter solution vectors. The core optimization objective function of the Bayesian optimization (BO) algorithm is to minimize the root mean square error (RMSE) of the model's predictions on the validation set. A smaller objective function value indicates higher prediction accuracy and better generalization performance for the corresponding hyperparameter combination. The algorithm's hyperparameters and solution search space are initialized. The initial number of random sampling points for Bayesian optimization is set to 5, the maximum number of iterations to 12, and the minimum objective function change threshold. The legal search range for the hyperparameters to be optimized is defined as follows: number of feature nodes N1∈[10,50], number of augmentation nodes N2∈[10,40], and regularization parameter λ∈[10,50]. -6 10 - ²], constraining the boundaries of hyperparameter values; Traverse the hyperparameter solution vectors of the initial sampling points, construct a BLS model based on the hyperparameters corresponding to each set of solution vectors, and after completing the model training, calculate the RMSE value of the model on the validation set as the objective function value corresponding to that set of hyperparameters; compare the objective function value of the current sampling point with the historical individual optimal solution, update the individual optimal hyperparameter combination, and at the same time select the global optimal hyperparameter combination with the smallest objective function value among all sampling points, and update the global optimal solution.
[0034] Based on existing sampling points and corresponding objective function values, a Gaussian process surrogate model is constructed using the Matérn kernel function. The mapping relationship between hyperparameter combinations and model prediction performance is fitted. The expected improvement is used as the acquisition function. By maximizing the acquisition function, the potentially optimal hyperparameter sampling points in the current iteration are selected, thus achieving a balance between global exploration and local development of the hyperparameter solution space.
[0035] A BLS model is constructed based on the newly selected hyperparameter sampling points. After training, the RMSE (Recovery Mean Squared Error) on the validation set is calculated, which is the objective function value at each sampling point. Simultaneously, the change between this value and the current globally optimal objective function value is calculated. If ΔRMSE is less than the set minimum change threshold, or the number of iterations reaches the preset maximum number of iterations of 12, the iterative optimization process is terminated; otherwise, iterative optimization continues.
[0036] After the iteration terminates, the number of feature nodes N1, the number of enhancement nodes N2, and the regularization parameter λ corresponding to the obtained global optimal solution are used as the optimal hyperparameter combination of the BLS model to complete the construction of the BO-BLS coal air-dry basis ash content detection model.
[0037] The BO-BLS coal air-dried basis ash content detection model was iteratively optimized using regularization, and its performance was verified using K-fold cross-validation. Specifically, in a preferred embodiment of this invention, multiple regularization methods were employed during the iterative optimization of the BO-BLS coal air-dried basis ash content detection model, and an adaptive learning rate strategy was introduced to optimize the model. To address the problems of overfitting, poor convergence, and incomplete performance evaluation during model training, this invention uses multiple regularization methods to enhance model performance. The model training parameters were set as follows: initial learning rate set to 0.001, learning rate decay factor set to 0.5, and weight decay coefficient set to... The loss function used is mean squared error (MSE), the training epochs of the CNN network are set to 100, and the training epochs of the BLS model are set to 50. In the validation phase, K-fold cross-validation was used. The 17 samples were randomly shuffled and divided into K groups. One group was selected as the validation set each time, and the remaining K-1 groups were used as the training set. The training and validation were repeated K times to ensure that each sample group had participated in both training and validation. The average value of the evaluation index from the K experiments was used as the performance index of the model.
[0038] Example In this embodiment, the root mean square error (RMSE) and linear regression coefficient (R) are used to evaluate the quality of the model.
[0039] The RMSE value measures the magnitude of the error between the model's predicted values and the actual values. Its calculation formula is:
[0040] The R-value measures the strength of the linear correlation between model predictions and actual values. Its calculation formula is:
[0041] Specifically, the CNN-BLS model achieved a test set RMSE as low as 1.6759 and a regression correlation coefficient R as high as 0.9815.
[0042] Figure 3 A scatter plot comparing the predicted and actual values of the CNN-BLS model on the test set can reflect the deviation between the predicted and actual values after training. Figure 4 A line graph comparing the predicted and actual values of the CNN-BLS model on the test set can reflect the deviation between the predicted and actual values after training.
[0043] This invention employs a Bayesian optimization (BO) algorithm to globally adaptively optimize the number of feature nodes, the number of augmentation nodes, and the ridge regression regularization parameters of a wide learning system (BLS). This effectively avoids the problems of insufficient generalization ability, large fluctuations in prediction results, and difficulty in achieving global optimum detection accuracy found in traditional BLS models. Experimental results show that on the same coal LIBS spectral dataset, the CNN-BLS model proposed in this invention outperforms the comparative models ELM, RF, and SVR in prediction accuracy. Specifically, the RMSE and R values of the traditional ELM model are 5.4591 and 0.8233, respectively; for the RF model, they are 1.8500 and 0.9764, respectively; and for the SVR model, they are 1.9628 and 0.9747, respectively. In contrast, the CNN-BLS model of this invention achieves RMSE and R values of 1.6759 and 0.9815, respectively, demonstrating a significant improvement in detection accuracy, a substantial reduction in error level, and further enhancement of goodness of fit.
[0044] In terms of testing efficiency, coal-fired power plants currently typically use the rapid ash content method for ash content testing. The core common shortcomings are long testing cycles, low batch efficiency, and inability to achieve on-site real-time testing. The core time efficiency of each method is as follows.
[0045] The slow ashing method is the national standard arbitration method for coal ash content testing. Its core principle involves placing the coal sample in a muffle furnace and gradually heating it according to a prescribed procedure until all combustible components are completely burned off, and inorganic minerals are converted into stable oxides. After multiple cooling, weighing, and constant weight determinations, the air-dried ash content of the coal sample is calculated based on the mass of the residue. This method boasts the core advantages of stable and accurate test results and is a recognized arbitration-grade testing method in the industry. However, the testing process is extremely lengthy, typically requiring 3-4 hours for a single complete test. For complex coal samples or when constant weight is not met, the testing cycle can extend to over 5 hours. Strict control of the heating rate, ashing temperature, and holding time is required. The operation is cumbersome and demands high levels of expertise from laboratory personnel. Furthermore, the testing process can only be conducted in batches, resulting in low sample processing efficiency, which cannot meet the high-efficiency testing requirements of continuous production and real-time quality control in modern coal industries.
[0046] Rapid ashing is a commonly used method for rapid ash content detection in enterprise laboratories and quality inspection scenarios. Its core principle is to rapidly burn and ashing the coal sample in a high-temperature furnace by increasing the ashing temperature and shortening the heating time, and then calculating the ash content by weighing. Compared with slow ashing, this method significantly reduces the time for a single complete test to 1-1.5 hours, significantly improving detection efficiency. However, it still has obvious inherent limitations: First, the test is still destructive; the coal sample is completely ineffective after high-temperature ashing and cannot be recycled or reused, resulting in high costs in scenarios such as determining scarce coal types and online batch testing. Second, the test relies on large high-temperature furnaces, dryers, and other laboratory equipment, which are bulky and inconvenient to move, making it difficult to achieve in-situ rapid testing in coal mines, coal storage yards, coal blending workshops, etc. Third, rapid ashing can easily lead to coal sample deflagration, ash splashing, or incomplete ashing, affecting the repeatability of the test results, requiring multiple parallel tests to ensure accuracy.
[0047] Besides the conventional laboratory testing methods mentioned above, ash content classification still relies heavily on manual experience in coal mining, screening, and on-site coal blending. This method mainly relies on operators to roughly classify coal by observing its color, luster, and impurity content. Single-sample classification takes only a few seconds to a few minutes, offering strong on-site immediacy. However, it lacks quantitative testing data support, suffers from large subjective errors and low accuracy, and its testing efficiency is difficult to adapt to large-scale production scenarios. It cannot provide stable and reliable data support for intelligent sorting and refined coal blending, and thus fails to meet the production requirements of clean and efficient utilization in modern coal industry.
[0048] All traditional methods suffer from unavoidable rigid aging losses: sample grinding and preparation, as well as pretreatment for air-dry basis equilibration, must be completed, which takes a fixed 1-2 hours; on-site testing is not possible, and sample transportation adds several hours to several days of aging loss; batch testing is a serial operation, and efficiency decreases linearly with the increase of sample volume.
[0049] This invention completely breaks through the time bottleneck of traditional methods. It eliminates the need for sample preparation and chemical pretreatment. The entire process from spectral acquisition to output of detection results for a single sample can be completed within 5 minutes. Machine learning detection can achieve millisecond-level result analysis. With the help of portable equipment, on-site in-situ detection can be achieved, greatly improving detection efficiency.
[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for detecting coal ash content on a dry basis based on laser-induced breakdown spectral data and machine learning, characterized in that, include: Spectral data were acquired using a standardized five-point sampling method, and the data was preprocessed. A feature extraction model for coal LIBS data was constructed, and a multi-layer CNN network was used to extract global and local features from the spectral data. A Bayesian optimization algorithm is introduced to globally adaptively optimize the hyperparameters in the BLS model, and a BO-BLS coal air-dry basis ash content detection model is constructed. The BO-BLS coal air-dried basis ash content detection model was iteratively optimized using regularization, and the model performance was verified using the K-fold cross-validation method.
2. The method for detecting coal ash content on a dry basis based on laser-induced breakdown spectral data and machine learning according to claim 1, characterized in that, The construction of the coal LIBS data feature extraction model specifically includes: A four-layer one-dimensional convolutional neural network was constructed, and BatchNorm batch normalization and Dropout random deactivation regularization strategies were introduced to alleviate model overfitting and output a hierarchical dataset. Multiple independent CNN networks with the same structure but different initialization parameters and Dropout rates were set up. The preprocessed spectral data was input into the multiple networks respectively, and the outputs were concatenated to generate CNN deep features.
3. The method for detecting coal ash content on a dry basis based on laser-induced breakdown spectral data and machine learning according to claim 2, characterized in that, The coal LIBS data feature extraction model uses the partial least squares regression (PLS) method to compress the dimensions of the hierarchical dataset and the CNN deep features, respectively, and extracts the feature components that are most correlated with the coal ash content.
4. The method for detecting coal ash content on a dry basis based on laser-induced breakdown spectral data and machine learning according to claim 3, characterized in that, The construction of the BO-BLS coal air-dried basis ash content detection model specifically includes: The number of feature nodes N1, the number of augmentation nodes N2, and the ridge regression regularization parameter λ of the BLS model are encoded into hyperparameter solution vectors. Initialize the algorithm's hyperparameters and solution search space, define the legal search range of the hyperparameters to be optimized, and constrain the boundaries of hyperparameter values.
5. The method for detecting coal ash content on a dry basis based on laser-induced breakdown spectral data and machine learning according to claim 4, characterized in that, The BLS model iterates through the hyperparameter solution vectors of the initial sampling points, constructs a BLS model based on the hyperparameters corresponding to each set of solution vectors, and after completing the model training, calculates the RMSE value of the model on the validation set as the objective function value corresponding to the hyperparameters; compares the objective function value of the current sampling point with the historical individual optimal solution, updates the individual optimal hyperparameter combination, and at the same time selects the global optimal hyperparameter combination with the smallest objective function value among all sampling points, and updates the global optimal solution.
6. The method for detecting coal ash content on a dry basis based on laser-induced breakdown spectral data and machine learning according to claim 5, characterized in that, Based on the global optimal solution and the existing sampling points and corresponding objective function values, a Gaussian process surrogate model is constructed using the Matérn kernel function. The mapping relationship between the hyperparameter combination and the model prediction performance is fitted. The expected improvement is used as the sampling function. By maximizing the sampling function, the potentially optimal hyperparameter sampling points in the current iteration are selected.
7. The method for detecting coal ash content on a dry basis based on laser-induced breakdown spectral data and machine learning according to claim 6, characterized in that, Based on the newly selected hyperparameter sampling points and combining them with the potential optimal hyperparameter sampling points, a BLS model is constructed. After training, the RMSE (Return on Validation) on the validation set is calculated, which is the objective function value of the sampling points. At the same time, the change between this value and the current global optimal objective function value is calculated. ; If ΔRMSE is less than the set minimum change threshold, or the number of iterations reaches the preset maximum number of iterations, the iterative optimization process is terminated; otherwise, iterative optimization continues.
8. The method for detecting coal ash content on a dry basis based on laser-induced breakdown spectral data and machine learning according to claim 7, characterized in that, After the iterative optimization process terminates, the number of feature nodes N1, the number of enhancement nodes N2, and the regularization parameter λ corresponding to the obtained global optimal solution are used as the optimal hyperparameter combination of the BLS model to complete the construction of the BO-BLS coal air-dry basis ash content detection model.
9. The method for detecting coal ash content on a dry basis based on laser-induced breakdown spectral data and machine learning according to claim 1, characterized in that, When iteratively optimizing the BO-BLS coal air-dry basis ash content detection model, multiple regularization methods are adopted and an adaptive learning rate strategy is introduced to optimize the model. In the verification stage, the K-fold cross-validation method is used, and the average value of the evaluation index of K experiments is used as the performance index of the model.