Coal quality near-infrared rapid analysis method based on constraint learning and ADBN

Abstract

This invention discloses a rapid near-infrared analysis method for coal quality based on constrained learning and Adaptive Deep Belief Network (ADBN). First, a near-infrared spectrometer is used to acquire diffuse reflectance spectra of coal samples. Each sample is acquired 3-5 times and averaged to reduce random errors, and the true values ​​of corresponding coal quality parameters are determined. Next, the spectral data is preprocessed to eliminate noise, baseline drift, and scattering interference. A constrained learning algorithm is used to select feature bands strongly correlated with coal composition, constructing constrained spectral feature vectors. An Adaptive Deep Belief Network (ADBN) is then constructed. The spectral feature vectors are input into the model, incorporating physical constraints on coal quality parameters, spectral feature correlation constraints, and regularization constraints into the loss function. Finally, the trained model is used to predict the spectral composition of unknown coal samples, obtaining the coal composition content. This method effectively removes redundant bands and noise, solving the redundancy problem of traditional feature selection. ADBN can dynamically adjust the network structure to adapt to the inherent laws of coal quality, improving prediction accuracy and reliability.

Description

Technical Field

[0001] This invention relates to the technical field of the metallurgical industry, specifically to a rapid near-infrared analysis method for coal quality based on constraint learning and ADBN. Background Technology

[0002] As an indispensable fossil fuel in the global energy structure, coal's quality directly determines combustion efficiency, energy utilization level, and pollutant emission intensity. Accurate detection of coal quality parameters (such as moisture, ash, volatile matter, calorific value, and sulfur content) is a core prerequisite for the clean and efficient utilization of coal, safe management of industrial production, and trade settlement. With the growth of global energy demand and increasingly stringent environmental requirements, traditional coal quality analysis methods are no longer sufficient to meet the needs of rapid, accurate, and green testing in industrial settings. Developing efficient and reliable rapid coal quality analysis technologies has become an urgent requirement for the sustainable development of the coal industry.

[0003] Traditional coal quality analysis often employs offline laboratory testing methods, requiring complex sampling and sample preparation processes with analysis cycles lasting several hours. These methods are cumbersome and cannot achieve real-time online detection, making them unsuitable for the dynamic quality control needs of coal mining, power generation, and coal chemical industries. Various rapid detection technologies have emerged, among which X-ray fluorescence spectroscopy (XRF), laser-induced breakdown spectroscopy (LIBS), and microwave detection have been applied in specific scenarios. However, due to their technical principles, they still have the following limitations: XRF technology, based on elemental characteristic X-rays, lacks sufficient sensitivity for light elements with low atomic numbers (such as C and H), making it difficult to directly obtain information on organic components related to calorific value; LIBS technology uses laser-induced plasma for elemental analysis, and its signal is significantly affected by the coal sample matrix effect, physical morphology, and surface characteristics, requiring improvement in the stability of quantitative analysis; microwave technology mainly utilizes differences in dielectric properties for moisture detection, making it difficult to achieve simultaneous analysis of multiple coal quality parameters (such as ash content, volatile matter, and calorific value).

[0004] First, near-infrared spectral data of coal samples are characterized by high dimensionality and strong redundancy. While containing effective information strongly correlated with coal composition, they also contain a large amount of irrelevant bands and useless information such as noise, baseline drift, and scattering interference. Direct modeling increases computational complexity, prolongs training time, and can lead to overfitting, reducing prediction accuracy and generalization ability. Existing feature selection methods often employ unconstrained or simple constraint strategies, making it difficult to accurately select effective bands and effectively eliminate redundancy and interference, thus limiting model performance improvement.

[0005] Secondly, traditional modeling methods and conventional deep learning models (such as ordinary deep belief networks, DBNs) have significant shortcomings. Although DBNs, as models composed of multiple layers of restricted Boltzmann machines (RBMs), possess powerful high-dimensional data representation capabilities and unsupervised pre-training advantages, making them suitable for processing complex spectral data, conventional DBNs have a fixed network structure. The connections between hidden layer nodes do not consider the physicochemical correlations of coal composition and the correlations of spectral characteristics, lacking adaptive adjustment capabilities and making it difficult to adapt to the complex characteristics of coal spectral data and the inherent correlation laws of coal parameters. Summary of the Invention

[0006] The purpose of this invention is to provide a rapid near-infrared analysis method for coal quality based on constraint learning and ADBN, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A rapid near-infrared analysis method for coal quality based on constraint learning and ADBN includes:

[0009] (1) Collect near-infrared diffuse reflectance spectral data of coal samples. Near-infrared spectrometer was used for collection. Each coal sample was collected 3-5 times and the average value was taken to reduce random error. The true values ​​of coal quality parameters of the corresponding samples were determined and a spectral-parameter dataset was constructed.

[0010] (2) The spectral data is preprocessed to remove noise, baseline drift and scattering interference to obtain standardized spectral data;

[0011] (3) Based on the constrained learning algorithm, select the feature bands that are strongly correlated with the coal quality composition and construct the spectral feature vector under the constraint conditions;

[0012] (4) Construct an adaptive deep belief network (ADBN) model, define the output layer nodes as coal quality component indicators, and the connection relationship between hidden layer nodes is based on the physicochemical correlation and spectral feature correlation of coal quality components;

[0013] (5) Input the spectral feature vector into the ADBN model, introduce a constraint learning mechanism, integrate the physical constraints of coal quality parameters, the correlation constraints of spectral features and the model regularization constraints into the loss function of the ADBN model, optimize the model training process, iteratively optimize the network parameters through dynamic learning algorithm, and integrate the constraints to limit the output range of the model.

[0014] (6) Using the trained ADBN model, predict the near-infrared spectral data of unknown coal samples and output the coal composition content.

[0015] Preferably, the constraint learning algorithm includes the following steps:

[0016] (1) Introduce prior knowledge constraints, including the physicochemical range of coal composition, the known correlation between spectral bands and composition;

[0017] (2) Based on the prior knowledge constraints, the band selection algorithm is used to screen key feature bands and eliminate redundant bands. The band selection evaluation index is designed in combination with the constraints, while taking into account both correlation and constraint compliance, so as to avoid the selected bands satisfying the correlation but violating the physical constraints.

[0018] (3) Construct a constrained optimization objective function, combine the spectral data error with the constraint penalty term, and solve for the optimal feature subset through gradient descent or evolutionary algorithm.

[0019] Preferably, the dynamic learning algorithm of the ADBN model includes the following steps:

[0020] (1) Initialize the network structure, define the conditional probability distribution between nodes, and adopt a hierarchical initialization strategy. The input layer is initialized based on the spectral feature dimension, the hidden layer is initialized layer by layer based on the restricted Boltzmann machine (RBM), and the output layer is initialized in combination with the physical constraint range to improve the convergence speed of the model.

[0021] (2) Calculate the posterior probability of the node state based on the current batch of spectral data;

[0022] (3) Update network parameters based on posterior probability, including the connection weights between nodes and the conditional probability distribution parameters. Introduce momentum terms to optimize the update process, avoid parameter update oscillations, and improve training stability.

[0023] (4) An adaptive adjustment mechanism is introduced to dynamically adjust the learning rate or network structure complexity based on the model prediction error. The learning rate is dynamically adjusted using a cosine annealing strategy, and the network structure is optimized for complexity by adaptively adding or removing hidden layer nodes, taking into account both model accuracy and generalization ability.

[0024] Preferably, the constraints include at least one of the following:

[0025] (1) Physicochemical range constraints of coal composition, such as moisture content ∈ [0, 50%], ash content ∈ [0, 100%], volatile matter content ∈ [0, 60%], fixed carbon content ∈ [0, 100%], total sulfur content ∈ [0, 10%], calorific value ∈ [10, 35] MJ / kg;

[0026] (2) Logical constraints of spectral characteristic bands, such as the monotonic relationship between the intensity of absorption peaks in specific bands and the content of components, the intensity of hydroxyl absorption band (1900-2000nm) is positively correlated with water content, and the intensity of CH absorption band (2200-2400nm) is positively correlated with volatile content;

[0027] (3) Non-negativity constraints on model output ensure that the prediction results conform to the actual physical meaning.

[0028] Constraint learning mechanisms include:

[0029] Physical constraints: The physical range of coal quality parameters and the correlation between components are used as hard constraints to limit the rationality of the model output;

[0030] Feature constraints: Based on the correlation between spectral features and coal quality parameters, sparsity and selectivity constraints are imposed on the model feature extraction process;

[0031] Regularization constraints: Adding L1 / L2 regularization terms and dropout constraints to the loss function can suppress model overfitting and improve generalization ability.

[0032] Preferably, the preprocessing includes: using Savitzky-Golay smoothing, multivariate scattering correction (MSC), standard normal variable transformation (SNV) and first / second derivative combination processing to achieve noise reduction and baseline correction of spectral data.

[0033] Preferably, the coal composition indicators include at least one of moisture, ash, volatile matter, fixed carbon, total sulfur, and calorific value.

[0034] A rapid near-infrared analysis system for coal quality based on constrained learning and adaptive dynamic Bayesian networks includes:

[0035] (1) Spectral acquisition module, used to acquire near-infrared diffuse reflectance spectral data of coal samples. This module includes a near-infrared spectrometer, an automatic sample injection unit and a temperature control unit to avoid temperature fluctuations affecting the accuracy of spectral acquisition. The automatic sample injection unit enables continuous detection of coal samples and improves detection efficiency.

[0036] (2) Preprocessing module, used to denoise, correct baseline and normalize the spectral data, with built-in preprocessing parameter adaptive matching unit, which can automatically call the corresponding preprocessing parameters according to the coal sample type without manual intervention;

[0037] (3) Feature extraction module: Based on the constraint learning algorithm, feature bands are selected and feature vectors are constructed;

[0038] (4) ADBN modeling module, used for initializing, training and optimizing adaptive dynamic Bayesian network models;

[0039] (5) Prediction output module, which uses the trained ADBN model to output the coal composition content, including result display unit, abnormal alarm unit and data export unit. When the prediction result exceeds the physical constraint range, the abnormal alarm unit will automatically prompt, and the data export unit supports data export in multiple formats.

[0040] (6) Constraint Coordination Control Module, which integrates all constraints to achieve dynamic coordination of physical constraints, feature constraints and regularization constraints, monitors the satisfaction of constraints in real time, and triggers local fine-tuning of feature extraction module and ADBN modeling module when constraints are violated.

[0041] Preferably, the feature extraction module further includes:

[0042] (1) Prior knowledge base, storing the physicochemical range and spectral band correlation rules of coal composition;

[0043] (2) Constraint optimization unit, which constructs constraint conditions and filters feature bands based on the prior knowledge base.

[0044] Preferably, the ADBN modeling module further includes:

[0045] (1) Dynamic learning units, used to iteratively update network parameters;

[0046] (2) Adaptive adjustment unit, which dynamically adjusts the learning rate or network structure according to the prediction error;

[0047] (3) Constraint integration unit, which integrates the constraints described in claim 4 into the model training process.

[0048] Preferably, the system also includes a data storage module for storing historical spectral data, model training results and detection reports, and constraint configuration parameters, supporting online model updates and transfer learning. The system also adds a data encryption unit to ensure the security of the detection data, while supporting data backtracking and querying, which facilitates subsequent model optimization and problem investigation.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] 1. This invention introduces a constrained learning algorithm to screen feature bands in near-infrared spectral data of coal samples. This accurately removes redundant bands and noise interference that are irrelevant to the coal composition, effectively overcoming the feature redundancy problem caused by unconstrained or simple constraint strategies in traditional feature screening methods. It significantly reduces the model input dimension, reduces computational complexity, and improves the prediction accuracy and generalization ability of subsequent modeling. At the same time, combined with a high-resolution spectral acquisition and multiple repeated averaging preprocessing strategy, it reduces random errors and spectral interference from the source, laying the foundation for building a high-quality spectral-parameter dataset.

[0051] 2. This invention constructs an Adaptive Deep Belief Network (ADBN) model, breaking through the limitations of the fixed structure of conventional deep belief networks. It dynamically adjusts the connection relationships of hidden layer nodes based on the physicochemical correlation and spectral feature correlation of coal components, so that the network topology matches the inherent properties of coal. In the model training process, physical constraints of coal parameters, spectral feature correlation constraints, and model regularization constraints are all integrated into the loss function. Through dynamic learning algorithms, iterative optimization of network parameters and effective limitation of output range are achieved, significantly improving the model's ability to represent complex coal spectral data, prediction stability, and robustness. This solves the technical problem that traditional deep learning models cannot simultaneously take into account high-dimensional spectral features and the inherent correlation of coal parameters in near-infrared analysis of coal. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the method steps of the present invention;

[0053] Figure 2 This is a schematic diagram of the system flow of the present invention. Detailed Implementation

[0054] 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. 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.

[0055] Example 1:

[0056] Please see Figures 1-2 The present invention provides a technical solution:

[0057] A rapid near-infrared analysis method for coal quality based on constraint learning and ADBN includes:

[0058] (1) Collect near-infrared diffuse reflectance spectral data of coal samples. Near-infrared spectrometer was used for collection. Each coal sample was collected 3-5 times and the average value was taken to reduce random error. The true values ​​of coal quality parameters of the corresponding samples were determined and a spectral-parameter dataset was constructed.

[0059] (2) The spectral data is preprocessed to remove noise, baseline drift and scattering interference to obtain standardized spectral data. The preprocessing includes Savitzky-Golay smoothing, multivariate scattering correction (MSC), standard normal variable transformation (SNV) and first / second derivative combination processing to achieve noise reduction and baseline correction of the spectral data.

[0060] (3) Based on the constrained learning algorithm, select feature bands that are strongly correlated with coal composition and construct spectral feature vectors under constrained conditions. The constrained learning algorithm includes the following steps:

[0061] (1) Introduce prior knowledge constraints, including the physicochemical range of coal composition, the known correlation between spectral bands and composition;

[0062] It should be noted that although near-infrared spectral data has a high dimensionality, the evolutionary algorithm (such as a genetic algorithm) used in step (3) does not directly perform an unconstrained search on the original full-spectral dimension. This algorithm uses the prior knowledge constraints introduced in step (1) (such as focusing only on the band intervals corresponding to specific functional groups) to limit the search space to key regions in a low dimension, thereby greatly reducing invalid searches and ensuring that the feature selection process is completed within an acceptable range of computing resources and time, making it suitable for offline model training in industrial scenarios.

[0063] (2) Based on the prior knowledge constraints, the band selection algorithm is used to screen key feature bands and eliminate redundant bands. The band selection evaluation index is designed in combination with the constraints, while taking into account both correlation and constraint compliance, so as to avoid the selected bands satisfying the correlation but violating the physical constraints.

[0064] (3) Construct a constrained optimization objective function, combine the spectral data error with the constraint penalty term, and solve for the optimal feature subset through gradient descent or evolutionary algorithm;

[0065] (4) Construct an Adaptive Deep Belief Network (ADBN) model, defining the output layer nodes as coal quality component indicators, and the connection relationships between hidden layer nodes based on the physicochemical correlation and spectral feature correlation of coal quality components. The dynamic learning algorithm of the ADBN model includes the following steps:

[0066] (1) Initialize the network structure, define the conditional probability distribution between nodes, and adopt a hierarchical initialization strategy. The input layer is initialized based on the spectral feature dimension, the hidden layer is initialized layer by layer based on the restricted Boltzmann machine (RBM), and the output layer is initialized in combination with the physical constraint range to improve the convergence speed of the model.

[0067] (2) Calculate the posterior probability of the node state based on the current batch of spectral data;

[0068] (3) Update network parameters based on posterior probability, including the connection weights between nodes and the conditional probability distribution parameters. Introduce momentum terms to optimize the update process, avoid parameter update oscillations, and improve training stability.

[0069] (4) An adaptive adjustment mechanism is introduced to dynamically adjust the learning rate or network structure complexity based on the model prediction error. The learning rate is dynamically adjusted using a cosine annealing strategy, and the network structure is optimized for complexity by adaptively adding or removing hidden layer nodes, taking into account both model accuracy and generalization ability.

[0070] (5) Input the spectral feature vector into the ADBN model, introduce a constraint learning mechanism, integrate the physical constraints of coal quality parameters, the correlation constraints of spectral features and the model regularization constraints into the loss function of the ADBN model, optimize the model training process, iteratively optimize the network parameters through a dynamic learning algorithm, and integrate the constraints to limit the output range of the model. The constraints include at least one of the following:

[0071] (1) Physicochemical range constraints of coal composition content, such as moisture content ∈ [0, 50%], ash content ∈ [0, 100%], volatile matter content ∈ [0, 60%], fixed carbon content ∈ [0, 100%], total sulfur content ∈ [0, 10%], calorific value ∈ [10, 35] MJ / kg. The coal composition index includes at least one of moisture, ash, volatile matter, fixed carbon, total sulfur and calorific value. Based on the industrial analysis principle of coal quality analysis, an additive constraint is applied to the same coal sample, that is, the sum of the mass percentages of moisture, ash, volatile matter and fixed carbon satisfies 95% ≤ moisture + ash + volatile matter + fixed carbon ≤ 100%, to ensure that the prediction results conform to the physical logic of coal composition.

[0072] (2) Logical constraints of spectral characteristic bands, such as the monotonic relationship between the intensity of absorption peaks in specific bands and the content of components, the intensity of hydroxyl absorption band (1900-2000nm) is positively correlated with water content, and the intensity of CH absorption band (2200-2400nm) is positively correlated with volatile content;

[0073] (3) Non-negativity constraints on model output ensure that the prediction results conform to the actual physical meaning.

[0074] Constraint learning mechanisms include:

[0075] Physical constraints: The physical range of coal quality parameters and the correlation between components are used as hard constraints to limit the rationality of the model output;

[0076] Feature constraints: Based on the correlation between spectral features and coal quality parameters, sparsity and selectivity constraints are imposed on the model feature extraction process;

[0077] Regularization constraints: Adding L1 / L2 regularization terms and dropout constraints to the loss function can suppress model overfitting and improve generalization ability.

[0078] (6) Using the trained ADBN model, predict the near-infrared spectral data of unknown coal samples and output the coal composition content.

[0079] Example 2:

[0080] The difference between Example 2 and Example 1 is that:

[0081] A rapid near-infrared analysis system for coal quality based on constrained learning and adaptive dynamic Bayesian networks includes:

[0082] (1) Spectral acquisition module, used to acquire near-infrared diffuse reflectance spectral data of coal samples. This module includes a near-infrared spectrometer, an automatic sample injection unit and a temperature control unit to avoid temperature fluctuations affecting the accuracy of spectral acquisition. The automatic sample injection unit enables continuous detection of coal samples and improves detection efficiency.

[0083] (2) Preprocessing module, used to denoise, correct baseline and normalize the spectral data, with built-in preprocessing parameter adaptive matching unit, which can automatically call the corresponding preprocessing parameters according to the coal sample type without manual intervention;

[0084] (3) Feature extraction module, which filters feature bands and constructs feature vectors based on a constrained learning algorithm, the feature extraction module further comprising:

[0085] (1) Prior knowledge base, storing the physicochemical range and spectral band correlation rules of coal composition;

[0086] (2) Constraint optimization unit, which constructs constraint conditions and filters feature bands based on the prior knowledge base;

[0087] (4) An ADBN modeling module, used for initializing, training, and optimizing an adaptive dynamic Bayesian network model, wherein the ADBN modeling module further includes:

[0088] (1) Dynamic learning units, used to iteratively update network parameters;

[0089] (2) Adaptive adjustment unit, which dynamically adjusts the learning rate or network structure according to the prediction error;

[0090] (3) Constraint integration unit, which integrates the constraints described in claim 4 into the model training process;

[0091] (5) Prediction output module, which uses the trained ADBN model to output the coal composition content, including result display unit, abnormal alarm unit and data export unit. When the prediction result exceeds the physical constraint range, the abnormal alarm unit will automatically prompt, and the data export unit supports data export in multiple formats.

[0092] (6) Constraint Coordination Control Module, which integrates all constraints to achieve dynamic coordination of physical constraints, feature constraints and regularization constraints, monitors the satisfaction of constraints in real time, and triggers local fine-tuning of feature extraction module and ADBN modeling module when constraints are violated.

[0093] The data storage module is used to store historical spectral data, model training results and detection reports, and constraint configuration parameters, supporting online model updates and transfer learning. The system also adds a data encryption unit to ensure the security of detection data, while supporting data backtracking and querying, which facilitates subsequent model optimization and problem troubleshooting.

[0094] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A rapid near-infrared analysis method for coal quality based on constraint learning and ADBN, characterized in that, include: (1) Collect near-infrared diffuse reflectance spectral data of coal samples. Near-infrared spectrometer was used for collection. Each coal sample was collected 3-5 times and the average value was taken to reduce random error. The true values ​​of coal quality parameters of the corresponding samples were determined and a spectral-parameter dataset was constructed. (2) The spectral data is preprocessed to remove noise, baseline drift and scattering interference to obtain standardized spectral data; (3) Based on the constrained learning algorithm, select the feature bands that are strongly correlated with the coal quality composition and construct the spectral feature vector under the constraint conditions; (4) Construct an adaptive deep belief network (ADBN) model, define the output layer nodes as coal quality component indicators, and the connection relationship between hidden layer nodes is based on the physicochemical correlation and spectral feature correlation of coal quality components; (5) Input the spectral feature vector into the ADBN model, introduce a constraint learning mechanism, integrate the physical constraints of coal quality parameters, the correlation constraints of spectral features and the model regularization constraints into the loss function of the ADBN model, optimize the model training process, iteratively optimize the network parameters through dynamic learning algorithm, and integrate the constraints to limit the output range of the model. (6) Using the trained ADBN model, predict the near-infrared spectral data of unknown coal samples and output the coal composition content.

2. The near-infrared rapid analysis method for coal quality based on constraint learning and ADBN according to claim 1, characterized in that: The constraint learning algorithm includes the following steps: (1) Introduce prior knowledge constraints, including the physicochemical range of coal composition, the known correlation between spectral bands and composition; (2) Based on the prior knowledge constraints, the band selection algorithm is used to screen key feature bands and eliminate redundant bands. The band selection evaluation index is designed in combination with the constraints, while taking into account both correlation and constraint compliance, so as to avoid the selected bands satisfying the correlation but violating the physical constraints. (3) Construct a constrained optimization objective function, combine the spectral data error with the constraint penalty term, and solve for the optimal feature subset through gradient descent or evolutionary algorithm.

3. The near-infrared rapid analysis method for coal quality based on constraint learning and ADBN according to claim 1, characterized in that: The dynamic learning algorithm of the ADBN model includes the following steps: (1) Initialize the network structure, define the conditional probability distribution between nodes, and adopt a hierarchical initialization strategy. The input layer is initialized based on the spectral feature dimension, the hidden layer is initialized layer by layer based on the restricted Boltzmann machine (RBM), and the output layer is initialized in combination with the physical constraint range to improve the convergence speed of the model. (2) Calculate the posterior probability of the node state based on the current batch of spectral data; (3) Update network parameters based on posterior probability, including the connection weights between nodes and the conditional probability distribution parameters. Introduce momentum terms to optimize the update process, avoid parameter update oscillations, and improve training stability. (4) An adaptive adjustment mechanism is introduced to dynamically adjust the learning rate or network structure complexity based on the model prediction error. The learning rate is dynamically adjusted using a cosine annealing strategy, and the network structure is optimized for complexity by adaptively adding or removing hidden layer nodes, taking into account both model accuracy and generalization ability.

4. The near-infrared rapid analysis method for coal quality based on constraint learning and ADBN according to claim 1, characterized in that: The constraints include at least one of the following: (1) Physicochemical range constraints of coal composition, such as moisture content ∈ [0, 50%], ash content ∈ [0, 100%], volatile matter content ∈ [0, 60%], fixed carbon content ∈ [0, 100%], total sulfur content ∈ [0, 10%], calorific value ∈ [10, 35] MJ / kg; (2) Logical constraints of spectral characteristic bands, such as the monotonic relationship between the intensity of absorption peaks in specific bands and the content of components, the intensity of hydroxyl absorption band (1900-2000nm) is positively correlated with water content, and the intensity of CH absorption band (2200-2400nm) is positively correlated with volatile content; (3) Non-negativity constraints on model output ensure that the prediction results conform to the actual physical meaning; Constraint learning mechanisms include: Physical constraints: The physical range of coal quality parameters and the correlation between components are used as hard constraints to limit the rationality of the model output; Feature constraints: Based on the correlation between spectral features and coal quality parameters, sparsity and selectivity constraints are imposed on the model feature extraction process; Regularization constraints: Adding L1 / L2 regularization terms and dropout constraints to the loss function can suppress model overfitting and improve generalization ability.

5. The near-infrared rapid analysis method for coal quality based on constraint learning and ADBN according to claim 1, characterized in that: The preprocessing includes Savitzky-Golay smoothing, multivariate scattering correction (MSC), standard normal variable transformation (SNV), and a combination of first / second derivatives to achieve noise reduction and baseline correction of the spectral data.

6. The near-infrared rapid analysis method for coal quality based on constraint learning and ADBN according to claim 1, characterized in that: The coal quality composition indicators include at least one of the following: moisture, ash, volatile matter, fixed carbon, total sulfur, and calorific value.

7. A rapid near-infrared analysis system for coal quality based on constrained learning and adaptive dynamic Bayesian networks, characterized in that, include: (1) Spectral acquisition module, used to acquire near-infrared diffuse reflectance spectral data of coal samples. This module includes a near-infrared spectrometer, an automatic sample injection unit and a temperature control unit to avoid temperature fluctuations affecting the accuracy of spectral acquisition. The automatic sample injection unit enables continuous detection of coal samples and improves detection efficiency. (2) Preprocessing module, used to denoise, correct baseline and normalize the spectral data, with built-in preprocessing parameter adaptive matching unit, which can automatically call the corresponding preprocessing parameters according to the coal sample type without manual intervention; (3) Feature extraction module: Based on the constraint learning algorithm, feature bands are selected and feature vectors are constructed; (4) ADBN modeling module, used for initializing, training and optimizing adaptive dynamic Bayesian network models; (5) Prediction output module, which uses the trained ADBN model to output the coal composition content, including result display unit, abnormal alarm unit and data export unit. When the prediction result exceeds the physical constraint range, the abnormal alarm unit will automatically prompt, and the data export unit supports data export in multiple formats. (6) Constraint Coordination Control Module, which integrates all constraints to achieve dynamic coordination of physical constraints, feature constraints and regularization constraints, monitors the satisfaction of constraints in real time, and triggers local fine-tuning of feature extraction module and ADBN modeling module when constraints are violated.

8. The system according to claim 7, characterized in that, The feature extraction module further includes: (1) Prior knowledge base, storing the physicochemical range and spectral band correlation rules of coal composition; (2) Constraint optimization unit, which constructs constraint conditions and filters feature bands based on the prior knowledge base.

9. The system according to claim 7, characterized in that, The ADBN modeling module further includes: (1) Dynamic learning units, used to iteratively update network parameters; (2) Adaptive adjustment unit, which dynamically adjusts the learning rate or network structure according to the prediction error; (3) Constraint integration unit, which integrates the constraints described in claim 4 into the model training process.

10. The system according to claim 7, characterized in that, It also includes a data storage module for storing historical spectral data, model training results and detection reports, and constraint configuration parameters, supporting online model updates and transfer learning; The system also adds a data encryption unit to ensure the security of the detection data, while supporting data backtracking and querying, which facilitates subsequent model optimization and problem investigation.

Images