Near-infrared coal quality quantitative analysis model construction method

Abstract

The invention relates to a near-infrared coal quality quantitative analysis model construction method, and belongs to the technical field of coal quality analysis. In order to solve the problems that a traditional coal quality analysis method is tedious in operation, time-consuming, destructive and difficult to meet rapid nondestructive detection requirements, the invention aims to realize rapid, accurate and quantitative detection and analysis of moisture, ash and volatile components in coal by combining a near infrared spectrum analysis technology with a deep learning algorithm. The method comprises the following specific steps: collecting and preparing a coal sample according to the national standard, and measuring moisture, ash and volatile components; the method comprises the following steps: acquiring spectral data by using a near infrared spectrometer, marking the spectral data, performing smoothing, multiplicative scatter correction and normalization processing, and dividing a training set and a test set according to a Kennard-Stone algorithm; and constructing a cascaded integrated deep learning prediction model of CNN-LSTM, and performing training, evaluation and optimization to predict coal quality indexes. According to the method, data are obtained based on a strict national standard experiment method, the accuracy of quantitative analysis of coal quality and the robustness of the model are improved by combining the near infrared spectrum technology with the deep learning model, and powerful support is provided for quality control and reasonable utilization of resources in the coal industry.

Description

Technical Field

[0001] The present invention relates to the technical field of coal quality analysis, and in particular to a near-infrared coal quality quantitative analysis model construction method for measuring moisture, ash and volatile matter in coal. Background Art

[0002] Coal occupies an important position in the energy field, and its quality parameters such as moisture, ash, volatile matter, etc. play a key role in coal utilization efficiency, combustion performance, processing technology and environmental impact. Although traditional analytical methods for determining moisture, ash and volatile matter in coal have a certain degree of accuracy, they are cumbersome, time-consuming and labor-intensive, and usually require sample destruction, which makes it difficult to meet the needs of rapid and non-destructive testing. Near-infrared spectroscopy analysis technology provides a highly potential alternative method for coal quality analysis with its advantages of rapidity, non-destructiveness, simultaneous determination of multiple components and low cost. However, near-infrared spectral data are complex and nonlinear, and are easily interfered by factors such as noise and baseline drift, which leads to challenges in accurately extracting effective information and building high-precision prediction models. The BP neural network and SVM algorithms commonly used in the current coal quality analysis field have limitations, such as being prone to falling into local optimality, affecting performance and generalization, and having high computational costs for large-scale data training. The convolutional neural network CNN has demonstrated excellent feature extraction capabilities in fields such as data processing, and the long short-term memory network LSTM is good at processing long-term dependencies in sequence data. Cascading the two can complement each other's advantages, comprehensively extract coal quality data features, improve model performance, accuracy and adaptability, enhance robustness, and effectively process its spatiotemporal characteristics. It is expected to overcome the shortcomings of existing technologies and achieve efficient analysis and accurate prediction of near-infrared coal quality data. However, there is currently a lack of a complete and effective model construction method for moisture, ash, and volatile matter analysis in coal. Summary of the invention

[0003] The present invention aims to provide a near-infrared coal quality quantitative analysis model construction method to achieve rapid, accurate, non-destructive quantitative analysis of moisture, ash, and volatile matter in coal, so as to overcome the limitations of traditional analysis methods. In order to achieve the above purpose, the technical solution of the present invention is as follows:

[0004] A near-infrared coal quality quantitative analysis model construction method comprises the following steps:

[0005] Acquisition and preprocessing of training samples: According to the requirements of the national standard GB-T 212-2008, the moisture, ash and volatile matter of the coal samples are measured, and the spectral data of the training coal samples are obtained by using a spectrometer. The obtained spectral data is preprocessed to obtain several training samples;

[0006] Construct a deep learning prediction model based on CNN-LSTM cascade integration;

[0007] Acquisition and preprocessing of test samples; Use a spectrometer to obtain spectral data of test coal samples, scan the test samples 30 times and take the average value; Use the same spectral data preprocessing method as the training samples to obtain the test samples;

[0008] The preprocessed test sample spectral data is input into the deep learning model to complete the prediction of the moisture, ash and volatile matter content of the coal in the test sample.

[0009] The acquisition and preprocessing of the training samples and the test samples include the following steps:

[0010] According to the requirements of the national standard GB-T 212-2008, representative coal samples are collected from the coal samples to be tested. After the collected coal samples are crushed, they are screened using appropriate screening equipment to ensure that the coal sample particle size is less than 0.2mm. The screened coal samples are fully mixed to ensure the uniformity of the coal samples and reduce the impact of uneven coal samples on subsequent analysis results;

[0011] According to the requirements of the national standard GB-T 212-2008, measure the moisture, ash and volatile matter of coal samples in sequence;

[0012] Spectral data collection and averaging: Use a near-infrared spectrometer to collect near-infrared spectral data of the above-mentioned coal samples that have been processed and measured for moisture, ash, and volatile content. Each coal sample is collected 30 times and the average value is taken. At the same time, the corresponding moisture, ash, and volatile content values ​​are used as markers for spectral data to establish a corresponding relationship between spectral data and coal quality component content, providing an accurate basis for subsequent model training and prediction;

[0013] Smooth the spectral data; use the Savitzky-Golay convolution smoothing method to smooth the collected average spectral data. The specific operation is to establish a filter function by using the least squares fitting coefficient, and perform polynomial least squares fitting on the wavelength point data in the moving window. Suppose the wavelength point data in the moving window is X j , where j = in to i + n, i is the center point position, and the expression of binomial fitting is in is the fitting value, a0, a1, a2 are the binomial coefficients. The least square method is used to solve the coefficients of the undetermined binomial equation, that is:

[0014]

[0015] make Solving the system of equations simultaneously can obtain the binomial coefficients. This step can effectively remove random high-frequency errors and improve the smoothness and signal-to-noise ratio of spectral data.

[0016] The spectral data is subjected to multiplicative scatter correction; the smoothed spectral data is subjected to multiplicative scatter correction to correct the scattering effect caused by factors such as uneven distribution of coal sample particles. Specifically, the spectrum A(λ) of the entire unknown coal sample is transformed into the spectrum A0(λ) of the hypothetical reference particle size, and the values ​​of α and β are specified according to the least squares method, and the estimated values ​​of the two factors are set as α` and β`, respectively, according to the formula:

[0017] A(λ)=α0A0(λ)+β+e(λ)

[0018] Get the transformation formula:

[0019]

[0020] Obtain the spectral data of a` and β`, using the average spectrum of all coal samples as:

[0021]

[0022] Among them A i represents the spectrum of the i-th sample, A is the spectral data of the modeling set, and α and β are obtained by least squares regression.

[0023] Normalize the spectral data; use the vector normalization method to normalize the spectral data after multiplicative scattering correction. Take a spectrum, and its data is expressed as:

[0024] x(1×m)

[0025] The vector normalization algorithm formula is:

[0026]

[0027] in m is the wavelength number. This method can eliminate the spectral changes caused by factors such as optical path differences, making the spectral data of different coal samples more comparable.

[0028] The preprocessed data is divided into training set and test set in a ratio of 8:2 using the Kennard-Stone algorithm to provide a data basis for subsequent model construction, training and evaluation;

[0029] The neural network used in the present invention is an integrated deep learning model CNN-LSTM including a convolutional neural network CNN and a long short-term memory network LSTM; model construction and training include the following steps:

[0030] The first feature extraction stage; the data passes through the input layer-c1 layer-s2 layer; the preprocessed spectral data is input, the c1 convolution layer uses 32 convolution kernels with a size of 3×N, N is the input data dimension, the moving step is 2, and the activation function uses the ReLU function to slide on the input data for convolution operation to extract local features in the input spectral data; the s2 subsampling layer performs dimensionality reduction processing on the features extracted by the convolution layer, compresses the feature data output by the convolution layer, reduces the data volume, and retains key feature information;

[0031] The second feature extraction stage, s2-s4 layers; after the subsampling layer processing, it enters the c3 convolution layer; it uses the feature data after dimensionality reduction processing in the s2 layer as input, and uses the convolution kernel again to perform convolution operations to further extract more complex and abstract features; in the s4 pooling layer, the data dimension is further reduced to reduce the computational complexity and prevent overfitting;

[0032] The third feature extraction stage is the Flatten layer-LSTM layer. The edge flattening layer adjusts the shape of the data processed by the s4 pooling layer to adapt it to the input requirements of the subsequent LSTM layer. The LSTM layer expects the input data to be in the form of a one-dimensional vector. The edge flattening layer realizes the data format conversion by flattening the multi-dimensional data into a one-dimensional vector, ensuring that the data can be smoothly transmitted and processed between different layers. The LSTM layer sets 6 hidden units to capture the long-term dependencies in the data.

[0033] In the analysis and prediction stage, the fully connected layer-output layer; finally, three fully connected layers are connected, which contain 64, 16 and 3 neuron nodes respectively, and the activation function uses the ReLU function; the fully connected layer integrates the features output by the LSTM layer, and maps the extracted features to the final prediction results through the fully connected relationship between neurons, so as to realize the prediction of the component content;

[0034] The model evaluation and optimization of the present invention comprises the following steps:

[0035] Use the test set data to evaluate the trained model; calculate the coefficient of determination R between the predicted value and the true value 2 , root mean square error RMSE, mean absolute error MAE three evaluation indicators; these indicators are used to comprehensively evaluate the accuracy, stability and generalization ability of the model;

[0036] Optimize the model based on the evaluation results; adjust the number of convolution kernels and size parameters of some convolutional layers of CNN, the number of hidden units in the LSTM layer, and the number of neurons in the fully connected layer; or use regularization technology to prevent overfitting and improve model performance through continuous optimization;

[0037] Beneficial Effects

[0038] The present invention obtains accurate data of moisture, ash and volatile matter in coal samples in strict accordance with the national standard GB-T 212-2008 experimental method and establishes the relationship between spectral data and component content. Compared with the prior art, the deep learning model based on CNN-LSTM cascade integration greatly improves the accuracy and model robustness of coal quality quantitative analysis. The data acquisition step ensures the representativeness of the sample and the reliability of the data. The preprocessing step effectively reduces noise interference, corrects scattering effects and enhances data comparability. The model construction and optimization give full play to the advantages of the deep learning algorithm, can accurately predict coal quality components, overcome many shortcomings of traditional coal quality analysis methods, and provide strong technical support for quality control and rational resource utilization in the coal industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The process framework diagram of the present invention is DETAILED DESCRIPTION

[0040] In order to have a further understanding and recognition of the structural features and the effects achieved by the present invention, a preferred embodiment and accompanying drawings are used for detailed description as follows:

[0041] like Figure 1 As shown, the method for constructing a near-infrared coal quality quantitative analysis model of the present invention comprises the following steps:

[0042] 1. Coal sample collection and preparation: According to the national standard GB-T 212-2008, representative coal samples are collected from the coal samples to be tested; the collected coal samples are crushed using appropriate crushing equipment, sieved to ensure that the coal sample particle size is less than 0.2mm, and the sieved coal samples are fully mixed to ensure the uniformity of the coal samples and reduce the impact of uneven coal samples on subsequent analysis results.

[0043] 2. Specifically, the moisture content of the present invention is determined by air drying method. Accurately weigh 1g (accurate to 0.0002g) of coal sample into an ash dish that has been pre-burned to constant weight. Place the ash dish into an air drying oven and dry it at 105-110°C for 1-1.5h. After taking it out, cool it in the air for 5min, put it into a dryer to cool it to room temperature and weigh it. According to the formula:

[0044]

[0045] M ad is the moisture content on air-dried basis, m is the mass of coal sample before drying, m1 is the mass of coal sample after drying, and the moisture content on air-dried basis is calculated.

[0046] 3. Specifically, the ash content determination method of the present invention is to weigh 1g (accurate to 0.0002g) of air-dried coal sample into an ash dish that has been pre-burned to constant weight, place the ash dish into a muffle furnace with a temperature not exceeding 100°C, open the furnace door, slowly heat the coal sample to 500°C within 15-20 minutes, and keep it at this temperature for 30 minutes. Then continue to heat it to 815±10°C, close the furnace door and burn it at this temperature for 1 hour. After the burning is completed, take out the ash dish and cool it in the air for 5 minutes, put it in a dryer to cool to room temperature and weigh it. According to the formula:

[0047]

[0048] Among them A ad is the ash content on air-dried basis, m2 is the total mass of the ash pan and ash after combustion, m3 is the mass of the ash pan, and m is the mass of the coal sample. Calculate the ash content on air-dried basis.

[0049] 4. Specifically, the volatile matter determination method of the present invention is to weigh 1g (accurate to 0.0002g) of air-dried coal sample into a covered porcelain crucible, place the crucible in a muffle furnace preheated to 900±10°C, quickly close the furnace door and heat for 7 minutes. After heating, take out the crucible and cool it in the air for 5 minutes, put it in a desiccator to cool to room temperature and then weigh it. According to the formula:

[0050]

[0051] Where V ad is the volatile matter content on an air-dried basis, m4 is the total mass of the crucible and residue after burning, m is the mass of the coal sample, and M ad is the moisture content on an air-dry basis, and the volatile matter content on an air-dry basis is calculated.

[0052] 5. Specifically, the spectral data collection and averaging of the present invention uses a near-infrared spectrometer to collect the near-infrared spectral data of the above-mentioned coal samples that have been processed and measured for moisture, ash, and volatile content. Each coal sample is collected 30 times and the average value is taken. The corresponding moisture, ash, and volatile content values ​​are used as markers for the spectral data, and the corresponding relationship between the spectral data and the coal quality component content is established to provide an accurate basis for subsequent model training and prediction. During the collection process, ensure that the wavelength range of the spectrometer covers the near-infrared characteristic absorption band of coal, the spectral resolution meets the accuracy requirements for coal quality component analysis, and maintain the stability and repeatability of the instrument to reduce data errors caused by instrument factors.

[0053] 6. Specifically, after the spectrum data is collected, it is first smoothed, and the average spectrum data collected is smoothed using the Savitzky-Golay convolution smoothing method. The filter function is established by using the least squares fitting coefficient, and the wavelength point data in the moving window are fitted with a polynomial least squares fit. Suppose the wavelength point data in the moving window is X j , where j = in to i + n, i is the center point position, and the expression of binomial fitting is in is the fitting value, a0, a1, a2 are the binomial coefficients. The least square method is used to solve the coefficients of the undetermined binomial equation, that is:

[0054]

[0055] make

[0056] Solving the system of equations simultaneously yields the binomial coefficients.

[0057] 7. Specifically, the smoothed spectral data is then subjected to multiplicative scatter correction to correct for the scattering effects caused by factors such as uneven distribution of coal sample particles. Specifically, the spectrum A(λ) of the entire unknown coal sample is transformed into the spectrum A0(λ) of the hypothetical reference particle size, and the values ​​of α and β are specified according to the least squares method, and the estimated values ​​of the two factors are set to α` and β`, respectively, according to the formula:

[0058] A(λ)=α0A0(λ)+β+e(λ)

[0059] Get the transformation formula:

[0060]

[0061] Obtain the spectral data of α` and β`, using the average spectrum of all coal samples as:

[0062]

[0063] Among them A i represents the spectrum of the i-th sample, A is the spectral data of the modeling set, and α and β are obtained by least squares regression.

[0064] 8. Specifically, after the multiplicative scatter correction, the vector normalization method is used to normalize the spectral data after the multiplicative scatter correction. Take a spectrum, whose data is expressed as:

[0065] x(1×m)

[0066] The vector normalization algorithm formula is:

[0067]

[0068] in m is the number of wavelengths.

[0069] 9. Specifically, the preprocessed data is divided into training set and test set in a ratio of 8:2 using the Kennard-Stone algorithm.

[0070] 10. Specifically, the model is constructed by first inputting the preprocessed spectral data. The c1 convolution layer uses 32 convolution kernels with a size of 3×N, where N is the dimension of the input data and a moving step of 2. The activation function uses the ReLU function to slide on the input data for convolution operations to extract local features from the input spectral data. The s2 subsampling layer performs dimensionality reduction on the features extracted by the convolution layer and compresses the feature data output by the convolution layer to reduce the amount of data while retaining key feature information.

[0071] 11. Specifically, after the subsampling layer is processed, it enters the c3 convolution layer, which uses the feature data after dimensionality reduction processing in the previous layer (s2 layer) as input, and uses the convolution kernel to perform convolution operations again to further extract more complex and abstract features; in the s4 pooling layer, the data dimension is further reduced, the computational complexity is reduced, and overfitting is prevented.

[0072] 12. Specifically, the Flatten Layer adjusts the shape of the data after being processed by the s4 pooling layer so that it can adapt to the input requirements of the subsequent LSTM layer. The LSTM layer expects the input data to be in the form of a one-dimensional vector. The Flatten Layer realizes the data format conversion by flattening the multi-dimensional data into a one-dimensional vector, ensuring that the data can be smoothly transmitted and processed between different layers; the LSTM layer sets 6 hidden units to capture the long-term dependencies in the data.

[0073] 13. Specifically, the last three fully connected layers contain 64, 16 and 3 neuron nodes respectively, and the activation function uses the ReLU function. It integrates the features output by the LSTM layer, maps the extracted features to the final prediction results through the fully connected relationship between neurons, and realizes the prediction of moisture, ash and volatile matter.

[0074] 14. Specifically, after the model is established, the training set data is input into the constructed CNN-LSTM model for training. The mean square error MSE is used as the loss function for model training, and the Adam optimizer is used to optimize the model parameters. The learning rate is set to 0.001 and the number of iterations is 800. Through multiple iterations of training, the model converges and learns the complex mapping relationship between spectral data and coal moisture, ash, and volatile matter.

[0075] 15. Specifically, after the model training is completed, the preprocessed test sample spectral data is input into the trained deep learning model, and the model will output the predicted values ​​of the moisture, ash, and volatile content of the coal in the test sample. The trained model is evaluated using the test set data to calculate the determination coefficient R between the predicted value and the true value. 2 , root mean square error RMSE, and mean absolute error MAE.

[0076] These indicators are used to comprehensively evaluate the accuracy, stability, and generalization ability of the model.

[0077] Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and the specification only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention to be protected. The present invention includes but is not limited to the detection of moisture, ash, and volatile matter in coal. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A near-infrared coal quality quantitative analysis model construction method, characterized in that: The following steps are involved: (1) Acquisition and preprocessing of training samples: According to the requirements of the national standard GB-T 212-2008, the moisture, ash and volatile matter of the coal sample are measured, and the spectral data of the training coal sample is obtained using a spectrometer. The obtained spectral data is preprocessed to obtain a number of training samples; (2) Construct a deep learning prediction model based on CNN-LSTM cascade integration; (3) Acquisition and preprocessing of test samples: Use a spectrometer to obtain spectral data of the test coal sample, and scan the test sample 30 times to obtain the average value; The same spectral data preprocessing method as the training sample is used for the test sample to obtain the test sample; (4) The preprocessed test sample spectral data is input into the deep learning model to predict the moisture, ash, and volatile matter contents of the coal in the test sample.

2. The method for constructing a near-infrared coal quality quantitative analysis model according to claim 1, characterized in that: The acquisition and preprocessing of the training samples include the following steps: (1) According to the requirements of the national standard GB-T 212-2008, collect representative coal samples from the coal samples to be tested. After the collected coal samples are crushed, they are screened using appropriate screening equipment to ensure that the coal sample particle size is less than 0.2 mm. The screened coal samples are fully mixed to ensure the uniformity of the coal samples and reduce the impact of the uneven coal samples on the subsequent analysis results; (2) According to the requirements of the national standard GB-T 212-2008, measure the moisture, ash and volatile matter of the coal sample in sequence; (3) Spectral data collection and averaging; (4) Smoothing the spectral data; (5) Perform multiplicative scatter correction on the spectral data; (6) Normalizing the spectral data; (7) Divide the samples into training set and test set.

3. The method for constructing a near-infrared coal quality quantitative analysis model according to claim 2, characterized in that: The spectrum collection and averaging comprises the following steps: (1) using a near infrared spectrometer to collect near infrared spectral data of the coal sample that has been processed and has had its moisture, ash, and volatile content determined, Each coal sample was collected 30 times and the average value was taken; (2) The corresponding moisture, ash, and volatile matter content values ​​are used as markers for the spectral data, and the corresponding relationship between the spectral data and the coal quality component content is established, providing an accurate basis for subsequent model training and prediction.

4. The method for constructing a near-infrared coal quality quantitative analysis model according to claim 2, characterized in that: The spectral data is smoothed using the Savitzky-Golay convolution smoothing method, which includes the following steps: (1) A filter function is established by using the least squares fitting coefficients, and a polynomial least squares fitting is performed on the wavelength point data within the moving window. Assume that the wavelength point data in the moving window is X j , where j = in to i + n, i is the center point position, j represents the index of the wavelength point data, and the expression of binomial fitting is: in is the fitting value, a0, a1, a2 are the binomial equation coefficients, and λ represents the wavelength value; (2) The least squares method is used to solve the coefficient ε of the undetermined binomial equation, that is: make Solving the system of equations simultaneously gives the binomial coefficients.

5. The method for constructing a near-infrared coal quality quantitative analysis model according to claim 2, characterized in that: The spectral data is subjected to multiplicative scatter correction, comprising the following steps: (1) The spectrum A(λ) of the entire unknown coal sample is transformed into the spectrum A0(λ) of the hypothetical reference particle size. The values ​​of α and β are specified according to the least squares method. The estimated values ​​of the two factors are set as α` and β`, respectively. The formula is: A(λ)=α0A0(λ)+β+e(λ) Get the transformation formula: (2) Obtain the spectral data of α` and β`, and use the average spectrum of all coal samples as: Among them A i represents the spectrum of the i-th sample, and A is the spectral data of the modeling set; (3) Finally, α and β are obtained through least squares regression.

6. A near-infrared coal quality quantitative analysis model construction method according to claim 2, characterized in that: The spectral data is normalized, comprising the following steps: (1) Take a spectrum, whose data is expressed as: x(1×m) (2) The vector normalization algorithm formula is: in m is the number of wavelengths.

7. The method for constructing a near-infrared coal quality quantitative analysis model according to claim 2, characterized in that: The spectral data are divided into a training set and a test set, and the Kennard-Stone algorithm is used to divide the data into a training set and a test set in a ratio of 8:2, providing a data basis for subsequent model construction, training and evaluation.

8. The method for constructing a near-infrared coal quality quantitative analysis model according to claim 1, characterized in that: The deep learning prediction model construction includes the following steps: (1) Construct a cascaded integrated deep learning model CNN-LSTM that includes a convolutional neural network (CNN) and a long short-term memory (LSTM) network. Its network structure includes an input layer, multiple convolutional layers, a pooling layer, an edge smoothing layer, an LSTM layer, a fully connected layer, and an output layer. (2) In the first feature extraction stage, the data passes through the input layer-c1 layer-s2 layer; the c1 convolution layer uses 32 convolution kernels with a size of 3×N, where N is the latitude of the input data and the moving step is 2. The activation function uses the ReLU function to slide on the input data for convolution operation. The s2 subsampling layer compresses the feature data output by the convolution layer to reduce the amount of data while retaining key feature information; (3) The second feature extraction stage, s2-s4 layers; after the subsampling layer is processed, it enters the c3 convolution layer and uses the convolution kernel to perform convolution operations again to further extract more complex features. The s4 pooling layer further reduces the data dimension, reduces the computational complexity, and prevents overfitting; (4) The third feature extraction stage, Flatten layer-LSTM layer; the edge flattening layer adjusts the shape of the data processed by the s4 pooling layer to make it adapt to the input requirements of the subsequent LSTM layer. The LSTM layer sets 6 hidden units to capture the long-term dependencies in the data; (5) Analysis and prediction stage, fully connected layer-output layer; finally, three fully connected layers are connected, which contain 64, 16 and 3 neuron nodes respectively. The activation function uses the ReLU function to integrate the features output by the LSTM layer. Through the fully connected relationship between neurons, the extracted features are mapped to the final prediction results to achieve the prediction of moisture, ash and volatile matter content; (6) Model prediction and evaluation stage: input the preprocessed test sample spectral data into the trained deep learning model to predict the moisture, ash, and volatile content of the coal in the test sample; use the test set data to evaluate the trained model and calculate the determination coefficient R between the predicted value and the true value 2 The three evaluation indicators of RMSE, mean absolute error (MAE) are used to comprehensively evaluate the accuracy, stability and generalization ability of the model.

9. The method for constructing a near-infrared coal quality quantitative analysis model according to claim 3, characterized in that During the coal sampling process, it should be ensured that the collected coal samples are representative and can reflect the overall characteristics of the coal to be tested. The operations of collecting, preparing and processing coal samples should comply with the relevant regulations and standard operating procedures in the national standard GB-T 212-2008 to ensure the accuracy and reliability of the experimental data.

10. The method for constructing a near-infrared coal quality quantitative analysis model according to claim 4, characterized in that When collecting spectral data, the near-infrared spectrometer should operate in a stable working state, its wavelength range should cover the near-infrared characteristic absorption band of coal, the spectral resolution should meet the accuracy requirements for coal quality component analysis, and the stability and repeatability of the instrument should be maintained during the collection process to reduce data errors caused by instrument factors.

11. The method for constructing a near-infrared coal quality quantitative analysis model according to claim 9, characterized in that During the model construction process, the parameter settings of the convolutional layer and pooling layer, the number of hidden units in the LSTM layer, and the number of neurons and activation function settings in the fully connected layer can all be adjusted and optimized according to the actual data characteristics and model performance to improve the model's feature extraction capabilities and ability to process sequence data.

Images