Method and device for calculating calorific value of coal

Abstract

The invention relates to the technical field of coal detection, in particular to a coal calorific value calculation method and device.The method comprises the steps that a raw coal sample of target coal is obtained, and spectral data of the raw coal sample is collected; preprocessing an original spectral vector corresponding to the spectral data to obtain a standardized spectral feature vector; and on the basis of the spectral feature vector and a pre-constructed prediction model, generating a local nonlinear feature and a global mapping relationship corresponding to the spectral feature vector so as to fuse the local nonlinear feature and the global mapping relationship, and generating the calorific value of the target coal. Therefore, the problems that in related technologies, a traditional calorimeter measuring method is long in time consumption, high in destructiveness and complex in operation, large-scale industrial production and real-time monitoring requirements are difficult to meet, and a modeling method based on a near infrared spectrum is prone to causing unstable prediction precision due to the characteristics of limited equipment performance, complex data characteristics and the like are solved; and a deep learning model used for spectrum modeling lacks interpretability to an internal mechanism.

Description

Technical Field

[0001] This application relates to the field of coal testing technology, and in particular to a method and apparatus for calculating the calorific value of coal. Background Technology

[0002] Among related technologies, coal calorific value detection technologies are mainly divided into two categories: one is the traditional calorimeter measurement method, which obtains accurate measurement values ​​through the complete combustion of coal samples according to the standard coal calorific value measurement method, and has high accuracy; the other is the modeling method based on near-infrared spectroscopy, which uses a portable near-infrared spectrometer to collect data and combines linear modeling or shallow machine learning models to predict the calorific value of coal, and has the advantages of being fast, non-destructive, and portable.

[0003] However, traditional calorimeter measurement methods suffer from problems such as long processing time, high destructiveness, and complex operation, making it difficult to meet the needs of large-scale industrial production and real-time monitoring. Near-infrared spectroscopy-based modeling methods are limited by the low resolution, insufficient signal-to-noise ratio, and high environmental sensitivity of portable devices. Furthermore, near-infrared spectral data itself has high dimensionality, strong collinearity, and nonlinearity, leading to unstable prediction accuracy. In addition, most deep learning models used for spectral modeling are "black box" models, lacking interpretability of internal mechanisms, which limits their promotion in practical industrial applications and urgently needs to be addressed. Summary of the Invention

[0004] This application provides a method and apparatus for calculating the calorific value of coal, in order to solve the problems in related technologies, such as the long time consumption, strong destructiveness, and complex operation of traditional calorimeter measurement methods, which are difficult to meet the needs of large-scale industrial production and real-time monitoring; the unstable prediction accuracy of near-infrared spectroscopy-based modeling methods due to equipment performance limitations and complex data characteristics; and the lack of interpretability of internal mechanisms in deep learning models used for spectral modeling, which limits their promotion in practical industrial applications.

[0005] The first aspect of this application provides a method for calculating the calorific value of coal, comprising the following steps: obtaining a raw coal sample of the target coal and collecting spectral data of the raw coal sample; preprocessing the original spectral vector corresponding to the spectral data to obtain a standardized spectral feature vector; and generating local nonlinear features and a global mapping relationship corresponding to the spectral feature vector based on the spectral feature vector and a pre-built prediction model, so as to fuse the local nonlinear features and the global mapping relationship to generate the calorific value of the target coal.

[0006] Optionally, in one embodiment of this application, the preprocessing of the original spectral vector corresponding to the spectral data to obtain a standardized spectral feature vector includes: calculating the mean and standard deviation of the spectral data based on the original spectral vector; and centering each reflectance intensity value in the original spectral vector to zero mean and scaling it to unit variance based on the mean and standard deviation to obtain the spectral feature vector.

[0007] Optionally, in one embodiment of this application, before generating the local nonlinear features and global mapping relationship corresponding to the spectral feature vector based on the spectral feature vector and the pre-built prediction model, the method further includes: constructing a radial basis function branch for capturing the local nonlinear features of the target spectral region in the spectral feature vector; constructing a linear branch for establishing the overall mapping relationship between the input features and the prediction target; constructing a single-layer prediction module in the pre-built prediction model based on the radial basis function branch and the linear branch; and sequentially superimposing multiple single-layer prediction modules to obtain the pre-built prediction model.

[0008] Optionally, in one embodiment of this application, generating the local nonlinear features and global mapping relationship corresponding to the spectral feature vector based on the spectral feature vector and a pre-built prediction model includes: inputting the spectral feature vector into the pre-built prediction model to locate the spectral intervals in the spectral feature vector that satisfy the target conditions based on the radial basis function branches in the pre-built prediction model; acquiring the local response of each spectral channel in the spectral feature vector corresponding to the spectral interval, and the local response of each spectral channel in other spectral intervals besides the spectral interval; and combining the local response and response weights corresponding to the spectral intervals and the local response and response weights corresponding to the other spectral intervals to obtain the local nonlinear features.

[0009] Optionally, in one embodiment of this application, generating the local nonlinear features and global mapping relationship corresponding to the spectral feature vector based on the spectral feature vector and the pre-built prediction model includes: inputting the spectral feature vector into the pre-built prediction model to establish a global mapping relationship between the spectral feature vector and the calorific value based on the linear branch in the pre-built prediction model.

[0010] Optionally, in one embodiment of this application, after calculating the calorific value of the target coal, the method further includes: identifying narrow-band spectral intervals in the spectral feature vector based on the pre-built prediction model; and establishing a mapping relationship between the narrow-band spectral intervals and the physicochemical absorption bands in the spectral data, so as to determine the interpretable results of the prediction model based on the mapping relationship.

[0011] A second aspect of this application provides a device for calculating the calorific value of coal, comprising: a data acquisition module for acquiring a raw coal sample of a target coal and acquiring spectral data of the raw coal sample; a processing module for preprocessing the original spectral vector corresponding to the spectral data to obtain a standardized spectral feature vector; and a calculation module for generating local nonlinear features and a global mapping relationship corresponding to the spectral feature vector based on the spectral feature vector and a pre-built prediction model, so as to fuse the local nonlinear features and the global mapping relationship to generate the calorific value of the target coal.

[0012] Optionally, in one embodiment of this application, the processing module includes: a calculation unit, configured to calculate the mean and standard deviation of the spectral data based on the original spectral vector; and a standardization unit, configured to center each reflectance intensity value in the original spectral vector to zero mean and scale it to unit variance based on the mean and the standard deviation, so as to obtain the spectral feature vector.

[0013] Optionally, in one embodiment of this application, it further includes: a first construction module, configured to construct a radial basis function branch for capturing local nonlinear features of the target spectral range in the spectral feature vector before calculating the calorific value of the target coal based on the spectral feature vector and the pre-constructed prediction model; a second construction module, configured to construct a linear branch for establishing an overall mapping relationship between the input features and the prediction target; a third construction module, configured to construct a single-layer prediction module in the pre-constructed prediction model based on the radial basis function branch and the linear branch; and a superposition module, configured to sequentially superimpose multiple single-layer prediction modules to obtain the pre-constructed prediction model.

[0014] Optionally, in one embodiment of this application, the calculation module includes: a positioning unit, configured to input the spectral feature vector into the pre-built prediction model to locate the spectral interval in the spectral feature vector that satisfies the target condition based on the radial basis function branch in the pre-built prediction model; an acquisition unit, configured to acquire the local response of each spectral channel in the spectral feature vector corresponding to the spectral interval, and the local response of each spectral channel in other spectral intervals besides the spectral interval; and a weighting unit, configured to combine the local response and response weight corresponding to the spectral interval and the local response and response weight corresponding to the other spectral intervals to obtain the local nonlinear feature.

[0015] Optionally, in one embodiment of this application, the calculation module includes: a building unit, configured to input the spectral feature vector into the pre-built prediction model to establish a global mapping relationship between the spectral feature vector and the calorific value based on the linear branch in the pre-built prediction model.

[0016] Optionally, in one embodiment of this application, it further includes: an identification module, configured to identify a narrow-band spectral range in the spectral feature vector based on the pre-built prediction model after calculating the calorific value of the target coal; and a determination module, configured to establish a mapping relationship between the narrow-band spectral range and the physicochemical absorption bands in the spectral data, so as to determine the interpretable result of the prediction model according to the mapping relationship.

[0017] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method for calculating the calorific value of coal as described in the above embodiments.

[0018] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for calculating the calorific value of coal.

[0019] A fifth aspect of this application provides a computer program product, including a computer program that, when executed, is used to implement the above-described method for calculating the calorific value of coal.

[0020] This application embodiment can collect and preprocess the spectral data of the target coal raw sample, and then calculate the calorific value of the target coal by combining it with a pre-constructed prediction model. This achieves rapid inference using the RBF-KAN network model without the need for coal sample combustion, avoiding the time-consuming and destructive nature of traditional calorimetric methods and meeting the needs of real-time industrial monitoring. The dual-branch structure of the radial basis function branch and linear branch in the RBF-KAN network model can balance local nonlinear feature capture with global stability, effectively improving the model's prediction accuracy and reliability. Furthermore, through visualization of two types of spectral domains, the physicochemical correspondence between model parameters and the near-infrared absorption band of coal can be established, realizing the interpretability of the RBF-KAN network model and enhancing its reliability and trustworthiness in industrial applications. This solves the problems in related technologies, such as the long time consumption, high destructiveness, and complex operation of traditional calorimeter measurement methods, which are difficult to meet the needs of large-scale industrial production and real-time monitoring; the unstable prediction accuracy of near-infrared spectroscopy-based modeling methods due to equipment performance limitations and complex data characteristics; and the lack of interpretability of internal mechanisms in deep learning models used for spectral modeling, which limits their promotion in practical industrial applications.

[0021] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0022] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a method for calculating the calorific value of coal according to an embodiment of this application; Figure 2 This is a schematic diagram of the overall framework of an RBF-KAN network model according to an embodiment of this application; Figure 3 This is a flowchart illustrating the training and testing process of an RBF-KAN network model according to an embodiment of this application. Figure 4 This is a cross-validation diagram of one embodiment of this application; Figure 5 This is a visualization of the radial basis weights according to an embodiment of this application; Figure 6 This is a visualization of the weighted radial basis function of one embodiment of this application; Figure 7 This is a schematic diagram of the structure of the coal calorific value calculation device provided according to the embodiments of this application; Figure 8 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application.

[0023] Figure label: 10-Coal calorific value calculation device: 100-Acquisition module, 200-Processing module and 300-Calculation module; 801-Memory, 802-Processor and 803-Communication interface. Detailed Implementation

[0024] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0025] The following describes a method and apparatus for calculating the calorific value of coal according to embodiments of this application, with reference to the accompanying drawings. Regarding the related technologies mentioned in the background section, traditional calorimeter measurement methods are time-consuming, destructive, and complex to operate, making them difficult to meet the needs of large-scale industrial production and real-time monitoring. Near-infrared spectroscopy-based modeling methods are prone to unstable prediction accuracy due to equipment performance limitations and complex data characteristics. Furthermore, deep learning models used for spectral modeling lack interpretability of their internal mechanisms, limiting their widespread application in practical industrial settings. This application provides a method for calculating the calorific value of coal. In this method, spectral data of a target coal sample can be collected and preprocessed, and then combined with a pre-constructed prediction model to calculate the calorific value of the target coal. This enables rapid inference using the RBF-KAN network model without the need for coal sample combustion, avoiding the time-consuming and destructive nature of traditional calorimetric methods and meeting the real-time monitoring needs of industry. The dual-branch structure of the radial basis function branch and linear branch in the RBF-KAN network model can balance local nonlinear feature capture with global stability, effectively improving the model's prediction accuracy and reliability. Furthermore, through visualization of two spectral domains, the physicochemical correspondence between model parameters and the near-infrared absorption band of coal can be established, realizing the interpretability of the RBF-KAN network model and enhancing its reliability and trustworthiness in industrial applications. This solves the problems in related technologies, such as the long time consumption, destructive nature, and complex operation of traditional calorimeter measurement methods, which are difficult to meet the needs of large-scale industrial production and real-time monitoring; the unstable prediction accuracy of near-infrared spectroscopy-based modeling methods due to equipment performance limitations and complex data characteristics; and the lack of interpretability of the internal mechanisms of deep learning models used for spectral modeling, which limits their promotion in practical industrial applications.

[0026] Specifically, Figure 1 A flowchart illustrating a method for calculating the calorific value of coal provided in an embodiment of this application.

[0027] like Figure 1As shown, the method for calculating the calorific value of coal includes the following steps: In step S101, a raw coal sample of the target coal is obtained, and the spectral data of the raw coal sample is collected.

[0028] It is understandable that "target coal" here refers to a specific coal object whose calorific value is to be tested; "raw coal sample" here refers to an unprocessed physical sample of coal collected from the mining, storage, or transportation of the target coal.

[0029] In some embodiments, this application can collect spectral data of raw coal samples in order to analyze the spectral data of the raw coal samples and generate calorific value analysis results of the target coal.

[0030] For example, embodiments of this application may, but are not limited to, selecting representative raw coal samples from the target coal. A representative raw coal sample here refers to a raw coal sample that accurately reflects the core attributes of the target coal, such as its overall material composition, particle size distribution, and calorific value, and is free from local bias. This sample may be obtained, but is not limited to, through the following methods: (1) Principle of full coverage of sampling locations: For the storage or stacking form of the target coal, a layered and point-based sampling method is adopted to avoid the limitations of single-point sampling. For example, if the target coal is in a stacked state, it is necessary to sample according to the pattern of "three layers of top, middle and bottom + multiple points of edge / center", such as selecting 2-3 sampling points at the top center, top edge, middle of the stack and bottom of the stack. If the target coal is in a transportation state, it is necessary to conduct interval sampling at different times and locations during the transportation process to ensure coverage of the entire transportation batch. (2) Sample particle size and mixing uniformity principle: Raw coal has different particle sizes such as lump coal and pulverized coal. The calorific value and chemical composition of coal with different particle sizes are different. It is necessary to ensure that the sample particle size distribution is consistent with the whole. That is, after collecting raw coal from each point, all sub-samples need to be mixed and reduced by quartering. The excess part is removed and the retained sample must include coal samples of different particle sizes. Avoid collecting only lump coal or only pulverized coal to prevent calorific value deviation due to single particle size. (3) Sample quantity and quality standards: The sample weight must meet the testing requirements, while avoiding accidental errors caused by insufficient sample quantity; for example, for the scenario of spectral detection + calorific value prediction, it is recommended that the total weight of a single sampling be no less than 5kg, and the weight of the sample used for testing after reduction should be no less than 500g to support multiple parallel tests; the sample must be kept dry and free of impurities, and avoid mixing in non-coal materials such as gangue and sawdust to prevent interference with spectral data and calorific value judgment.

[0031] (4) Sampling process traceability principle: When sampling, the source (mining area, coal seam), sampling time, sampling location, transportation batch and other information of the target coal should be recorded to ensure the traceability of the sample. If the subsequent test results are deviated, the problem can be investigated by tracing the sampling process.

[0032] For example, a power plant plans to test the calorific value of a batch of imported thermal coal. Sampling personnel can select 12 sampling points, including but not limited to the top center of the coal pile, the top two side edges, the middle of the pile, and the four corners of the bottom of the pile. 500g of raw coal is collected at each point. The 6kg subsamples are mixed and then reduced to 1kg using the quartering method. This sample contains different particle sizes such as lump coal and pulverized coal, and is free of gangue impurities, which is a representative raw coal sample.

[0033] Then, in this embodiment of the application, the raw coal sample can be processed into uniform coal powder with a particle size not exceeding 0.2 mm through crushing and multi-stage sieving. The coal powder is then placed in a glass sample box and gently pressed to level it. A portable near-infrared spectrometer is used to randomly select different points above the sample surface (for example, 5 cm above the sample surface, which can be determined by a person skilled in the art based on the optimal detection position of the near-infrared spectrometer. This embodiment of the application is only illustrative and does not impose any specific limitations) to collect diffuse reflectance spectral signals. The spectra collected from multiple points are averaged to obtain stable and representative single sample spectral data of the raw coal sample.

[0034] It should be noted that the specific principles for sampling raw coal samples can be set or adjusted by professionals in this field according to actual conditions and needs. The embodiments in this application are only illustrative and do not impose specific limitations.

[0035] The embodiments of this application can collect data on-site using a portable near-infrared spectrometer without the need for coal sample combustion, which facilitates data processing and calorific value analysis, avoids the time-consuming and destructive nature of traditional calorimetric methods, and thus effectively meets the real-time monitoring needs of industry, effectively improving the practicality and scalability of this application.

[0036] Step S102: Preprocess the original spectral vector corresponding to the spectral data to obtain a standardized spectral feature vector.

[0037] In some embodiments, this application may further preprocess the original spectral vector corresponding to the spectral data to obtain a standardized spectral feature vector.

[0038] In this context, the original spectral vector corresponding to the spectral data can be understood as a numerical vector obtained by arranging the spectral detection data of the raw coal sample in a one-dimensional structure according to wavelength order. Simply put, the original spectral vector is the digital and matrix-based representation of the spectral data.

[0039] Optionally, in one embodiment of this application, preprocessing the original spectral vector corresponding to the spectral data to obtain a standardized spectral feature vector includes: calculating the mean and standard deviation of the spectral data based on the original spectral vector; and centering each reflectance intensity value in the original spectral vector to zero mean and scaling it to unit variance based on the mean and standard deviation to obtain the spectral feature vector.

[0040] In actual implementation, when preprocessing the original spectral vectors corresponding to the spectral data, this application may, but is not limited to, use standard normal transformation (SNV) to obtain standardized spectral feature vectors.

[0041] For example, this application may first calculate the mean and standard deviation of the original spectral vector.

[0042] For example, the original spectral vector is: ,in, It is the number of wavelength points (e.g., 350 wavelength points). It is the wavelength point The reflection intensity value. This raw spectral vector represents the diffuse reflectance intensity of the raw coal sample at multiple wavelengths.

[0043] The mean value of the spectral data of the raw coal sample can be, but is not limited to, expressed as:

[0044] The reflection intensity values ​​of the raw coal sample at all wavelength points are summed and then divided by the total number of wavelength points.

[0045] The standard deviation of the spectral data of raw coal samples can be, but is not limited to, expressed as:

[0046] This involves calculating the degree of dispersion of the intensity values ​​at all wavelength points of a raw coal sample relative to its own mean.

[0047] Based on this mean and standard deviation, the embodiments of this application can convert each reflection intensity value in the original spectral vector. By centering to zero mean and scaling to unit variance, the spectral feature vector of the raw coal sample can be obtained.

[0048] For example, this application can perform a normalization transformation (centering and scaling) on ​​the intensity value at each wavelength point, that is, for each original intensity value in the original spectral vector ( From 1 to Perform the following calculations:

[0049] in, It is a centering process that subtracts the overall mean of the sample spectrum from the original value of each wavelength point, so that the mean of the new spectrum becomes 0, thereby eliminating the baseline shift caused by the physical state of the sample surface (such as particle size and packing density).

[0050] and This involves scaling each reflection intensity value after centering to unit variance, which means dividing all the centered reflection intensity values ​​by the standard deviation of the sample spectrum itself, so that the standard deviation of the new spectrum becomes 1, thereby eliminating the signal amplitude difference caused by different sample scattering levels.

[0051] After processing, the original spectral vector It can be converted into a new vector This new vector The mean is 0 and the standard deviation is 1.

[0052] It should be noted that the main purpose of preprocessing the original spectral vectors corresponding to the raw coal sample spectral data in this embodiment is to standardize the spectral data by eliminating scattering effects, thereby reducing the impact of physical interference on the model used in subsequent calculations. Therefore, the data preprocessing method of the original spectral vectors can also be set or adjusted by those skilled in the art according to the actual situation. This embodiment is only illustrative and does not impose specific limitations. For example, multivariate scattering correction (MSC) can be used instead of SNV preprocessing. Standardizing the spectral data by eliminating scattering effects can also reduce the impact of physical interference on the model and achieve the data preprocessing goal.

[0053] The embodiments of this application can normalize the original spectral vector corresponding to the spectral data of raw coal samples, and obtain the standardized feature vector by centering it to zero mean and scaling it to unit variance. This effectively eliminates the interference of baseline drift and scattering effects caused by uneven coal particle size and bulk density, significantly enhances the stability and comparability of the data, and lays a reliable data foundation for the subsequent establishment of a high-precision and highly generalizable prediction model.

[0054] Step S103: Based on the spectral feature vector and the pre-built prediction model, generate the local nonlinear features and global mapping relationship corresponding to the spectral feature vector, so as to fuse the local nonlinear features and global mapping relationship to generate the calorific value of the target coal.

[0055] In some embodiments, in order to achieve high-precision and rapid analysis of coal calorific value, this application may, but is not limited to, pre-construct a certain prediction model so that the spectral feature vector can be directly input into the pre-constructed prediction model to calculate and output the calorific value of the target coal.

[0056] Here, the pre-built prediction model refers to the pre-built overall mapping function model, which is composed of RBF-KAN modules stacked sequentially, consisting of radial basis functions (RBF) and Kolmogorov-Arnold (KAN) networks, hereinafter referred to as the RBF-KAN network model.

[0057] The embodiments of this application can achieve rapid, non-destructive, online detection of coal calorific value by acquiring and preprocessing spectra, combined with a pre-constructed RBF-KAN prediction model. This avoids the disadvantages of traditional calorimetry, such as long processing time and sample damage. It can be directly analyzed in real time on site, greatly improving detection efficiency and practicality, and can effectively meet the needs of real-time monitoring in industrial production.

[0058] The following section provides a further explanation of the construction and application process of the RBF-KAN network model in the embodiments of this application.

[0059] Optionally, in one embodiment of this application, before generating the local nonlinear features and global mapping relationship corresponding to the spectral feature vector based on the spectral feature vector and the pre-built prediction model, the method further includes: constructing a radial basis function branch for capturing the local nonlinear features of the target spectral region in the spectral feature vector; constructing a linear branch for establishing the overall mapping relationship between the input features and the prediction target; constructing a single-layer prediction module in the pre-built prediction model based on the radial basis function branch and the linear branch; and sequentially superimposing multiple single-layer prediction modules to obtain the pre-built prediction model.

[0060] Based on the descriptions of other embodiments, it will be understood that the pre-built prediction model in this application can be understood as an overall mapping function model consisting of RBF-KAN modules sequentially stacked, which are composed of radial basis functions (RBF) and Kolmogorov-Arnold (KAN) networks.

[0061] Therefore, the embodiments of this application can construct the RBF-KAN network model based on a two-branch structure of radial basis function branch + linear branch.

[0062] Among them, the radial basis function branch can capture the local nonlinear features of the target spectral interval in the spectral feature vector. That is, the radial basis function branch in this embodiment can adaptively focus on the target spectral interval in the spectral feature vector through learnable central parameters. The target spectral interval can be understood here as a spectral interval with rich information and a more significant contribution to the calculation of calorific value, thereby characterizing the local nonlinear features with rich information in the spectral feature vector.

[0063] The linear branch can establish an overall mapping relationship between input features and prediction targets (in this case, the calorific value of the coal sample), thereby maintaining global mapping stability.

[0064] Figure 2 This is a schematic diagram of the overall framework of an RBF-KAN network model according to an embodiment of this application. Figure 2 As shown, a single-layer prediction module (RBF-KAN module) in the prediction model can be constructed from radial basis function branches and linear branches. That is, the prediction model in this embodiment has multiple RBF-KAN modules. Each RBF-KAN module contains radial basis function branches, linear branches, and a normalization operator set after branch fusion to scale the module's output. In the output layer, this normalization operator can be, but is not limited to, replaced by an identity mapping, such as... Figure 2 The batch normalization layer / identity mapping layer shown.

[0065] Specifically, the spectral feature vector is processed by radial basis function branch and linear branch in the same layer, and the processing results are fused at the same layer. The fusion method is as follows: the outputs of the two branches are calculated in parallel for the input of the same layer, and element-wise additive fusion is performed in the same layer. Then, batch normalization is applied to the fusion result (the output layer is replaced by identity mapping) to obtain the final output representation of the layer. The final output representation of the layer can be used as the input feature vector of the next layer.

[0066] By sequentially stacking multiple RBF-KAN modules, the pre-built prediction model in this application embodiment can be obtained.

[0067] Specifically, the RBF-KAN network model in this application embodiment is composed of multiple layers of sequentially stacked RBF-KAN modules, wherein each RBF-KAN module in the same layer contains radial basis function branches and linear branches.

[0068] With the first Taking a layer as an example, the input dimension of this layer is The output dimension is Configuration for each input dimension Each of the radial basis functions has a corresponding central parameter. These are learnable parameters; the weight tensors corresponding to the radial basis function branches. Dimensions .

[0069] Linear branches may, but are not limited to, using weight matrices. Achieve cross-dimensional linear mapping and introduce certain activation functions, such as SiLU, ReLU, or GELU, into this branch to effectively transform input features while maintaining the model's nonlinear fitting ability, without affecting the overall mapping effect of the model.

[0070] The outputs of the radial basis function branch and the linear branch are fused element-wise additively within the same level and then batch normalized to represent the final output of that level.

[0071] The following provides a further explanation of the construction process of the RBF-KAN network model in the embodiments of this application.

[0072] Figure 3 This is a flowchart illustrating the training and testing process of an RBF-KAN network model according to an embodiment of this application. Figure 3 As shown: In the embodiments of this application, the learnable parameters of the RBF-KAN network model include, but are not limited to, the weights, biases, radial basis function centers, and batch normalized affine parameters of each layer, and the hyperparameters include, but are not limited to, the number of network layers, the width of each layer, the number of radial basis functions per dimension, and other configurations set before training.

[0073] Table 1 is a model learnable parameter table of one embodiment of this application, which can be represented as follows, but is not limited to: Table 1

[0074] In the embodiments of this application, the modeling data used are all derived from the near-infrared spectral test data of actual coal samples and their corresponding standard calorific value reference values.

[0075] For example, similar to the process of acquiring raw coal samples and their spectral data, when preparing training data, this application may, but is not limited to, select representative raw coal samples of various typical coals (e.g., covering the three basic categories of lignite, bituminous coal, and anthracite to ensure that the final model can cover the low, medium, and high full calorific value ranges). The coal samples of each type of coal are crushed and screened in multiple stages to obtain uniform coal powder with a particle size not exceeding 0.2 mm. The coal powder is placed in a glass sample box and gently pressed to level it. A portable near-infrared spectrometer is used to randomly select different points about 5 cm above the sample surface to collect diffuse reflectance spectral signals. The spectra collected from multiple points of the same coal sample are averaged to obtain stable and representative near-infrared spectral data of a single sample.

[0076] The portable near-infrared spectrometer can be, but is not limited to, a handheld near-infrared spectrometer with a spectral resolution of less than 10 nm and a wavelength range of 908.1 nm to 1676.2 nm. For each coal sample, 125 wavelength points are collected at uniform intervals within the above wavelength range, with a sampling interval (resolution) of approximately 6.1 nm (less than 10 nm), thereby forming the corresponding near-infrared spectral feature vector.

[0077] Finally, in this embodiment, the reference calorific value of each coal sample can be determined in a qualified laboratory to form a modeling dataset. The reference calorific value of each coal sample is determined in a qualified laboratory, and the determination method used conforms to the Chinese national standard GB / T 213-2008 "Method for Determination of Calorific Value of Coal" to ensure the accuracy and reliability of the calorific value data.

[0078] In this embodiment of the application, near-infrared spectral data are matched one-to-one with the corresponding calorific value reference values, and a dataset containing 671 coal samples is constructed. The calorific value of the coal samples ranges from 17.36 MJ / kg to 32.69 MJ / kg, covering low, medium and high calorific value ranges.

[0079] The aforementioned modeling dataset can be used for model training, validation, and testing. The collected near-infrared spectral data and the calorific reference values ​​measured by standard methods together constitute the sample pairs required for supervised learning, providing a data foundation for subsequent model training and performance evaluation.

[0080] Before training the model, the embodiments of this application may first preprocess the acquired raw near-infrared spectral data to reduce the influence of factors such as coal sample particle size differences, surface scattering effects and baseline drift on the modeling results.

[0081] In the embodiments of this application, each spectrum can be normalized sample by sample using the standard normal transformation method or the multivariate scattering correction method, that is, the mean of a single sample is subtracted from its spectral vector and divided by the standard deviation, thereby achieving centering and scale uniformity.

[0082] Furthermore, embodiments of this application may, but are not limited to, use a training objective consisting of a loss function and regularization constraints to train the model parameter set. Optimize to obtain the optimal parameters. For the training dataset ,in Indicates the first Input features of each sample This represents the corresponding true calorific value, and the model applies this to the sample. The predicted output can be defined, but is not limited to, as follows:

[0083] The objective function for model training can be defined, but is not limited to, the following:

[0084] in, To smooth the L1 regression loss, and Let represent the weight sets of the radial basis function branches and the linear branches, respectively. and This is the regularization coefficient.

[0085] During training, the embodiments of this application may, but are not limited to, use adaptive gradient optimization algorithms (such as AdamW optimizer, RAdam optimizer) to iteratively update parameters. The learning rate gradually decays according to a preset number of rounds (the preset number of rounds (the preset number of training rounds) can be set or adjusted by those skilled in the art. This embodiment of the application is only for illustrative purposes and does not impose any specific restrictions). Gradient clipping is introduced in the backpropagation stage to limit the gradient norm.

[0086] By dividing the training set into training and validation sets, and using root mean square error (RMSE) and mean absolute error (MAE) as evaluation metrics on the validation set, the optimal parameter combination can be selected as the final model.

[0087] During model parameter optimization, an adaptive gradient optimization algorithm is used to minimize the objective function. In this embodiment, the RAdam (Rectified Adam) optimizer is preferably used. This algorithm introduces a variance correction mechanism based on Adam, which can effectively alleviate the instability of the adaptive learning rate in the early stages of training and improve the stability of the model convergence process. If the AdamW optimizer is used instead, the regularization effect can be enhanced through weight decay, and combined with learning rate decay and gradient pruning strategies, the stability of parameter optimization and model prediction accuracy can be guaranteed.

[0088] During training, the learning rate can be gradually decayed according to the preset training rounds to avoid excessive parameter updates in the later stages that could affect model convergence. At the same time, the gradient norm can be pruned during the backpropagation stage to limit the gradient magnitude, thereby preventing gradient explosion during deep network training and further enhancing the stability and reliability of the training process.

[0089] Furthermore, in this embodiment, the key hyperparameters of the model can be evaluated experimentally on a validation set before training to assess the prediction error index, and then explicitly set after verifying its rationality. For example, in this embodiment, regarding the network structure, the input layer dimension can be set to 125, corresponding to 125 wavelength points uniformly sampled within the 908.1 nm to 1676.2 nm band of the near-infrared spectrum; the network can adopt a three-layer sequentially stacked structure, with the output dimensions of each layer set to 128, 64, and 1 respectively. Specifically, the first hidden layer dimension is slightly higher than the input dimension to fully extract the overall spectral features; the second hidden layer dimension is reduced to 64 to compress redundant information and reduce model complexity; the output layer dimension is 1, used to achieve scalar regression prediction of coal calorific value.

[0090] In the radial basis function (RBF) branch, each input dimension can be configured with, but is not limited to, 10 RBF functions to characterize the local nonlinear variations of the spectrum in different numerical ranges. This value, in experiments, ensures local modeling capability while avoiding the risk of excessive parameter scaling and overfitting due to an excessive number of RBF functions. The scaling parameter of the RBF can be, but is not limited to, set to 1.0. This value matches the numerical distribution of the spectral data after SNV preprocessing, allowing the RBF functions to produce effective responses in local ranges without excessive smoothing.

[0091] In the linear branch, the SiLU activation function can be used, but is not limited to, to perform a nonlinear transformation on the linear mapping result. This activation function has continuous, smooth, and non-saturating characteristics, which helps maintain the stability of the global mapping relationship and avoids the gradient vanishing problem. Batch normalization is introduced in all intermediate layers to stabilize the feature distribution and accelerate model convergence. Normalization is no longer used in the output layer to directly output the prediction result.

[0092] Regarding the training parameter settings, the optimization algorithm used in this embodiment is mainly RAdam, and the initial value of the adaptive learning rate can be, but is not limited to, set to... This value strikes a balance between convergence speed and training stability in the experiment. The number of training epochs can be set to, but is not limited to, 200 epochs, and the learning rate should be gradually decayed according to the preset number of epochs during training to promote model convergence in the later stages.

[0093] The L1 regularization coefficients for both radial basis function branch weights and linear branch weights can be set to, but are not limited to, 1.0 to suppress redundant parameters and enhance model sparsity. During the backpropagation phase, the gradient norm can be pruned, and the pruning threshold can be set to, but is not limited to, 5.0 to prevent gradient explosion during deep network training.

[0094] Table 2 shows the experimental results of a five-fold cross-validation test according to an embodiment of this application, which can be represented as follows, but is not limited to: Table 2

[0095] Figure 4 This is a cross-validation diagram of one embodiment of this application. (In conjunction with...) Figure 4 As shown in Table 2, the RBF-KAN exhibits relatively stable performance across all folds: the average RMSE (root mean square error) is 1.35 MJ / kg, and the average MAE (mean absolute error) is 0.92 MJ / kg. The RMSE ranges from 1.29 to 1.45 MJ / kg (spanning 0.16 MJ / kg), and the MAE ranges from 0.88 to 0.97 MJ / kg (spanning 0.09 MJ / kg), indicating minimal fluctuations across folds and stable generalization.

[0096] In summary, the RBF-KAN network model in this application maintains a consistent error level under different data partitions and can reliably complete the regression mapping from spectrum to calorific value. This indicates that the hyperparameter combination achieves a good balance between prediction accuracy, training stability, and model complexity.

[0097] Furthermore, Figure 5 This is a visualization of the radial basis weights according to an embodiment of this application. Figure 6 This is a visualization of the weighted radial basis functions according to one embodiment of this application. Figure 5 and Figure 6 As shown, Figure 5 A weighted visualization of two typical channels with relatively strong wavelengths in the 1403-1422 nm band is presented; Figure 6 The display shows the response curves of the weighted radial basis functions in the 1409.8-1459.4 nm wavelength channel. Each subplot corresponds to a specific wavelength. The colored curves represent the local responses of multiple RBF basis functions at different input values ​​at that wavelength, and the red dashed lines represent the normalized zero points.

[0098] It can be seen that the RBF-KAN network model in this application embodiment only forms obvious peaks in a few input intervals, indicating that the RBF-KAN network model has a strong response to specific value ranges at these wavelengths, thereby adaptively focusing on the information-rich spectral intervals and characterizing local nonlinear features.

[0099] Furthermore, the weights learned by the RBF-KAN network model exhibit a significant non-uniform distribution across the wavelength dimension. Specifically, the weight response is significantly enhanced in certain key wavelength ranges (1403-1422 nm), while the weights approach zero in other bands where information contribution is less. This indicates that the prediction model constructed and validated in this embodiment can adaptively identify and focus on spectral ranges that contribute significantly to coal calorific value prediction during training, rather than performing indiscriminate modeling across the entire spectrum.

[0100] The dual-branch structure of the RBF-KAN network model in this embodiment can balance local nonlinear feature capture with global stability, thereby effectively improving the accuracy and robustness of model prediction. Furthermore, the dual-branch structure of the model in this embodiment is simple to construct, avoiding the cumbersome operations of manual feature selection and multi-step modeling required by traditional methods, reducing the risk of error accumulation, and improving the practicality and generalizability of the method.

[0101] After preprocessing the spectral data of the target coal raw sample to obtain the standardized spectral feature vector, it is input into the trained RBF-KAN network model. The scalar predicted calorific value of the target coal can be calculated through hierarchical mapping. It is extremely suitable for single-sample online inference and batch inference scenarios, realizing rapid and non-destructive analysis on site.

[0102] The embodiments of this application can deeply integrate the powerful ability to capture local nonlinear features with the stable ability to learn global mapping relationships through RBF branches and linear branches, so that the RBF-KAN network model can not only finely characterize the complex changes in key spectral ranges, but also grasp the overall trend, thereby achieving a balance between prediction accuracy and model robustness.

[0103] Optionally, in one embodiment of this application, based on the spectral feature vector and a pre-built prediction model, generating local nonlinear features and a global mapping relationship corresponding to the spectral feature vector includes: inputting the spectral feature vector into the pre-built prediction model to locate the spectral intervals in the spectral feature vector that satisfy the target conditions based on the radial basis function branches in the pre-built prediction model; collecting the local response of each spectral channel in the spectral feature vector in the spectral interval, and the local response of each spectral channel in other spectral intervals besides the spectral interval; and combining the local response and response weights corresponding to the spectral intervals with the local response and response weights corresponding to other spectral intervals to obtain the local nonlinear features.

[0104] As one possible implementation, the structural components of the radial basis function branch in the embodiments of this application include, but are not limited to, the following three points: (1) By spectral channel dimension The expanded set of RBF basis functions, for each input dimension For a fixed wavelength channel, configure on that wavelength channel One radial basis function; (2) Wherein, a single radial basis function (response) can be defined, but is not limited to, as: , in, Indicates the first wavelength channel One learning center; Denotes the scale (width) parameter of the corresponding radial basis function. Indicates the first The first in the layer The component, that is, the first component. Input eigenvalues ​​of the layer radial basis function branch.

[0105] (3) This represents a learnable weight tensor used to adjust the direction and strength of the contribution of the local response to the final prediction.

[0106] During training, under the influence of error backpropagation and sparse regularization constraints, the radial basis function centers can adaptively cluster in the numerical intervals that contribute significantly to calorific value prediction, thus forming a focus on a narrow band of information-rich intervals in the spectral domain. For example... Figure 5 As shown, the information-rich narrowband interval corresponds to Figure 5 The darker areas in the middle.

[0107] Specifically, learnable centers in radial basis functions It can adaptively locate the effective response range (the numerical range that contributes significantly to calorific value prediction), rather than being fixed in a uniform position beforehand, but continuously adjusted during backpropagation, when a certain wavelength channel... When the prediction of calorific value is more sensitive within a specific value range, the corresponding center They will cluster around that value range. For example... Figure 6 As shown, Figure 6 The differential clustering of RBF functions in different wavelength channels can be observed, with the center... They will cluster around the range of values ​​that are more sensitive to calorific value prediction.

[0108] Secondly, the Gaussian kernel in the radial basis functions has local response characteristics, meaning it only responds to input eigenvalues ​​when the input eigenvalues ​​are within the specified range. A significant response is only observed when the object is near the center; as it moves away from the center, the response decays rapidly.

[0109] Meanwhile, in this embodiment, multiple Gaussian radial basis functions anchored by learnable center parameters can be configured on each spectral channel to model the local nonlinear response of the channel's feature values ​​in different numerical ranges. The local responses of each channel are weighted and aggregated, and after normalization, the output is used as the local nonlinear feature representation of the layer.

[0110] Among them, the RBF branch is in the 1st The calculation of a layer can be, but is not limited to, represented as:

[0111] in, Denotes the batch normalization operator acting on the output of the radial basis function branch. This indicates the dimension of the input features for this layer. Indicated as used to regulate the first The first input dimension The radial basis function response to the first Learnable weight parameters for the contribution strength and direction of each output channel. This represents the local response value of the corresponding radial basis function.

[0112] In terms of computational structure, according to input dimension Expand, without dimension blending, assign each input dimension There are _radial basis functions_, and all responses are weighted and summed before a unified output is obtained. Each radial basis function response can be defined, but is not limited to, as follows: .

[0113] Therefore, in practical applications, this application can input the spectral feature vector into a pre-built prediction model. The radial basis function branch in the model will use the spectral feature vector as input features to locate the spectral interval in the spectral feature vector that satisfies the target condition.

[0114] Here, the target condition refers to the fact that the located spectral interval (essentially a numerical interval) is more sensitive to calorific value prediction and contributes more to calorific value calculation compared to other spectral intervals. The spectral interval that meets this condition will appear as a narrow band with richer information in the spectral domain.

[0115] Multiple Gaussian radial basis functions configured on each spectral channel, anchored by learnable center parameters, can collect the local response information of the eigenvalues ​​of each spectral channel in the spectral feature vector within the located numerical interval (spectral interval) that meets certain conditions, as well as the local response information of the eigenvalues ​​of each spectral channel in other numerical intervals (numerical intervals other than the located numerical interval). Local nonlinear response modeling is performed on the local response information of each spectral channel in each numerical interval. Finally, the local responses of each spectral channel in all spectral intervals are weighted and aggregated according to the corresponding response weights, and the output after normalization can be used as the local nonlinear feature representation of this layer.

[0116] After the local nonlinear features of this layer are fused with the output of the linear branch, they can be used as the input feature vector for the next layer, until all layers (all RBF-KAN modules) in the prediction model are traversed, and finally the calorific value prediction result of the target coal is generated.

[0117] It should be noted that the local nonlinearity in the embodiments of this application has two meanings: first, spectral locality, that is, in the RBF branch, the nonlinear effect of the RBF basis function family is limited to a single wavelength channel and does not depend on the global combination across channels (the global combination is the effect of the linear branch); second, numerical locality, that is, in a single RBF mapping function, the nonlinearity only occurs when the input feature is close to the center value. It is activated at specific times, and does not have a uniform effect across the entire numerical range.

[0118] The RBF branch in this embodiment can effectively utilize a Gaussian kernel function with a learnable center parameter to adaptively locate and focus on the key spectral value range that contributes the most to the calorific value prediction. This avoids uniformly processing all data and instead allows the RBF-KAN network model to autonomously focus on the local region with the richest information during training and application. This enables efficient and accurate modeling of subtle but crucial nonlinear changes in the spectrum, thereby effectively improving the model's prediction accuracy.

[0119] Optionally, in one embodiment of this application, generating local nonlinear features and a global mapping relationship corresponding to the spectral feature vector based on the spectral feature vector and a pre-built prediction model includes: inputting the spectral feature vector into the pre-built prediction model to establish a global mapping relationship between the spectral feature vector and the calorific value based on the linear branch in the pre-built prediction model.

[0120] In other embodiments, in order to maintain the global mapping stability between the spectral feature vector and the final output calorific value, the prediction model in this application also introduces a certain linear branch.

[0121] Specifically, the linear branch in this application embodiment can establish an overall trend relationship between the input (spectral feature vector) and the output (coal calorific value) by performing cross-wavelength linear weighting and smooth mapping on the full-spectrum spectral features. This allows the output of the prediction model to maintain a smooth response to continuous changes in the input spectrum, thereby suppressing the excessive influence of local noise and single-band anomalies on the prediction results and providing a stable global backbone for the overall mapping in the dual-branch fusion structure.

[0122] In each layer, the linear branch uses the exact same layer input as the radial basis function branch; they differ only in their processing method, not in their input source. The linear branch in the [missing information] layer... The computational structure of a layer can be represented, but is not limited to, as follows:

[0123] That is, linear branches can be in the order of the first... Layer output vector As input, by linearly weighting each input dimension and applying nonlinear activation and batch normalization, an overall mapping relationship between input features and prediction targets is established, and its output... It represents the global trend and serves as a globally stable term that is integrated with the radial basis function branch outputs at the same level.

[0124] In this embodiment, the linear branch can establish a smooth and stable overall mapping relationship from input to output by performing global linear weighting and nonlinear activation transformation on spectral features. This provides a reliable "global trend line" for the RBF-KAN network model, suppressing the excessive influence of local noise or anomalous bands and ensuring the continuity and stability of the model output. When fused with the RBF branch, it can act as a stabilizer to ensure that the global prediction does not deviate from physical common sense.

[0125] Optionally, in one embodiment of this application, after calculating the calorific value of the target coal, the method further includes: identifying narrow-band spectral intervals in the spectral feature vector based on a pre-built prediction model; and establishing a mapping relationship between the narrow-band spectral intervals and the physicochemical absorption bands in the spectral data, so as to determine the interpretable results of the prediction model based on the mapping relationship.

[0126] Based on the descriptions of other embodiments, it is understood that since most deep learning models used for spectral modeling are "black box" models, there is a lack of interpretability support between model prediction results and physicochemical absorption bands in the single modeling scenario of portable near-infrared spectroscopy.

[0127] As one possible implementation, this application also adjusts the RBF weights in the radial basis function branch. L1 regularization was applied, causing the weights of the wavelength × basis function pairs that do not contribute to approach 0. Only the narrow band range with high information content was retained, and the weights eventually showed a locally concentrated, strip-like enhanced distribution in the spectral domain.

[0128] Furthermore, through two types of spectral domain visualization, the embodiments of this application can achieve the interpretability of the RBF-KAN network model: first, visualization of the weighted radial basis functions, showing the local focusing mode of the function response in the wavelength dimension; second, visualization of the radial basis weights, presenting the distribution pattern of parameters in the "basis function × wavelength" plane, thereby identifying the narrow spectral range of interest of the model and establishing a correspondence with the known O–H, C–H and other physicochemical absorption bands in the near-infrared spectrum of coal.

[0129] In this context, the physicochemical absorption band refers to the characteristic signal range formed on the spectrum when photons within a specific wavelength range are absorbed by the target chemical bonds in the coal sample after near-infrared light irradiates the sample. It simultaneously encompasses molecular vibrational behavior at the physical level and chemical bond properties at the chemical level, serving as the core link between spectral characteristics and the chemical composition of coal.

[0130] Specifically, such as Figure 5 As shown, the RBF-KAN network model in this embodiment exhibits a significantly enhanced weight concentration region in the 1403-1422 nm band of the full-spectrum weight distribution, indicating that this narrow spectral range has a high contribution to the prediction of coal calorific value.

[0131] Further integration Figure 6 The visualization results of the radial basis function response show that at the corresponding wavelengths (such as 1409.8 nm, 1416.0 nm, and 1422.2 nm), multiple sets of radial basis functions form clear local response peaks in the input value space, and the peak positions of different basis functions are different, reflecting the model's ability to finely distinguish different spectral value ranges within this band.

[0132] This phenomenon indicates that, Figure 5 The high-weight narrowband identified is not a simple linear weighting result, but rather... Figure 6 The synergistic effect of multiple radial basis functions shown models the local nonlinear changes within the first-order overtone absorption band of OH, thereby effectively capturing the influence of moisture and hydroxyl state changes on the calorific value of coal. The two form a consistent and mutually reinforcing explanatory relationship at the levels of global weight distribution and local function morphology.

[0133] Alternatively, those skilled in the art can use wavelength contribution analysis instead of weighted radial basis function visualization. By calculating the contribution of each wavelength channel to the prediction results, key spectral ranges can be identified, thus achieving the same goal of model interpretability analysis.

[0134] This application embodiment can visualize the weighted radial basis functions and radial basis weights learned by the RBF-KAN network model, identify the narrow spectral ranges that the model focuses on, and establish a correlation between these ranges and the physicochemical absorption bands of known functional groups (such as OH and CH) in coal. This solves the problem of lack of interpretability caused by the "black box" characteristics of existing deep learning models, breaks the "black box" limitation, enhances the reliability and trustworthiness of the model in industrial applications, and has strong generalizability.

[0135] According to the coal calorific value calculation method proposed in this application, spectral data of the target coal raw sample can be collected and preprocessed, and then combined with a pre-constructed prediction model to calculate the coal calorific value. This achieves rapid inference using the RBF-KAN network model without the need for coal sample combustion, avoiding the time-consuming and destructive nature of traditional calorimetric methods, and meeting the needs of real-time industrial monitoring. The dual-branch structure of the radial basis function branch and linear branch in the RBF-KAN network model can balance local nonlinear feature capture with global stability, effectively improving the model's prediction accuracy and reliability. Furthermore, through visualization of two spectral domains, the physicochemical correspondence between model parameters and the near-infrared absorption band of coal can be established, realizing the interpretability of the RBF-KAN network model and enhancing its reliability and trustworthiness in industrial applications. This solves the problems in related technologies, such as the long time consumption, high destructiveness, and complex operation of traditional calorimeter measurement methods, which are difficult to meet the needs of large-scale industrial production and real-time monitoring; the unstable prediction accuracy of near-infrared spectroscopy-based modeling methods due to equipment performance limitations and complex data characteristics; and the lack of interpretability of internal mechanisms in deep learning models used for spectral modeling, which limits their promotion in practical industrial applications.

[0136] Next, the coal calorific value calculation apparatus according to the embodiments of this application is described with reference to the accompanying drawings.

[0137] Figure 7 This is a schematic diagram of the structure of the coal calorific value calculation device according to an embodiment of this application.

[0138] like Figure 7 As shown, the coal calorific value calculation device 10 includes: a data acquisition module 100, a processing module 200, and a calculation module 300.

[0139] The acquisition module 100 is used to acquire raw coal samples of the target coal and collect spectral data of the raw coal samples; the processing module 200 is used to preprocess the original spectral vectors corresponding to the spectral data to obtain standardized spectral feature vectors; the calculation module 300 is used to generate local nonlinear features and global mapping relationships corresponding to the spectral feature vectors based on the spectral feature vectors and a pre-built prediction model, so as to fuse the local nonlinear features and global mapping relationships to generate the calorific value of the target coal.

[0140] Optionally, in one embodiment of this application, the processing module 200 includes: a calculation unit for calculating the mean and standard deviation of spectral data based on the original spectral vector; and a standardization unit for centering each reflectance intensity value in the original spectral vector to zero mean and scaling it to unit variance based on the mean and standard deviation to obtain a spectral feature vector.

[0141] Optionally, in one embodiment of this application, it further includes: a first construction module, used to construct a radial basis function branch for capturing local nonlinear features of the target spectral range in the spectral feature vector before calculating the calorific value of the target coal based on the spectral feature vector and the pre-built prediction model; a second construction module, used to construct a linear branch for establishing the overall mapping relationship between the input features and the prediction target; a third construction module, used to construct a single-layer prediction module in the pre-built prediction model based on the radial basis function branch and the linear branch; and an overlay module, used to sequentially overlay multiple single-layer prediction modules to obtain the pre-built prediction model.

[0142] Optionally, in one embodiment of this application, the calculation module 300 includes: a positioning unit, used to input the spectral feature vector into a pre-built prediction model to locate the spectral intervals in the spectral feature vector that satisfy the target conditions based on the radial basis function branches in the pre-built prediction model; an acquisition unit, used to acquire the local response of each spectral channel in the spectral feature vector corresponding to the spectral interval, and the local response of each spectral channel in other spectral intervals besides the spectral interval; and a weighting unit, used to combine the local response and response weights corresponding to the spectral intervals with the local responses and response weights corresponding to other spectral intervals to obtain local nonlinear features.

[0143] Optionally, in one embodiment of this application, the calculation module 300 includes: a building unit, used to input the spectral feature vector into a pre-built prediction model to establish a global mapping relationship between the spectral feature vector and the calorific value based on the linear branch in the pre-built prediction model.

[0144] Optionally, in one embodiment of this application, it further includes: an identification module, used to identify narrow-band spectral intervals in the spectral feature vector based on a pre-built prediction model after calculating the calorific value of the target coal; and a determination module, used to establish a mapping relationship between the narrow-band spectral intervals and the physicochemical absorption bands in the spectral data, so as to determine the interpretable results of the prediction model based on the mapping relationship.

[0145] It should be noted that the foregoing explanation of the coal calorific value calculation method embodiment also applies to the coal calorific value calculation device of this embodiment, and will not be repeated here.

[0146] The coal calorific value calculation device proposed in this application can collect and preprocess the spectral data of the target coal raw sample, and calculate the coal calorific value of the target coal by combining it with a pre-constructed prediction model. This achieves rapid inference using the RBF-KAN network model without the need for coal sample combustion, avoiding the time-consuming and destructive nature of traditional calorimetric methods and meeting the needs of real-time industrial monitoring. The dual-branch structure of the radial basis function branch and linear branch in the RBF-KAN network model can balance local nonlinear feature capture with global stability, effectively improving the model's prediction accuracy and reliability. Furthermore, through visualization of two types of spectral domains, the physicochemical correspondence between model parameters and the near-infrared absorption band of coal can be established, realizing the interpretability of the RBF-KAN network model and enhancing its reliability and trustworthiness in industrial applications. This solves the problems in related technologies, such as the long time consumption, high destructiveness, and complex operation of traditional calorimeter measurement methods, which are difficult to meet the needs of large-scale industrial production and real-time monitoring; the unstable prediction accuracy of near-infrared spectroscopy-based modeling methods due to equipment performance limitations and complex data characteristics; and the lack of interpretability of internal mechanisms in deep learning models used for spectral modeling, which limits their promotion in practical industrial applications.

[0147] Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: The memory 801, the processor 802, and the computer program stored on the memory 801 and capable of running on the processor 802.

[0148] When the processor 802 executes the program, it implements the method for calculating the calorific value of coal provided in the above embodiments.

[0149] Furthermore, electronic devices also include: Communication interface 803 is used for communication between memory 801 and processor 802.

[0150] The memory 801 is used to store computer programs that can run on the processor 802.

[0151] The memory 801 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0152] If the memory 801, processor 802, and communication interface 803 are implemented independently, then the communication interface 803, memory 801, and processor 802 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized into address buses, data buses, control buses, etc. For ease of representation, Figure 8 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0153] Optionally, in a specific implementation, if the memory 801, processor 802, and communication interface 803 are integrated on a single chip, then the memory 801, processor 802, and communication interface 803 can communicate with each other through an internal interface.

[0154] The processor 802 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0155] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method for calculating the calorific value of coal.

[0156] This application also provides a computer program product, including a computer program that can run computer instructions. When the computer instructions are executed by a processor, they implement the method for calculating the calorific value of coal provided in this application.

[0157] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0158] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0159] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0160] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0161] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0162] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0163] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0164] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for calculating the calorific value of coal, characterized in that, Includes the following steps: Obtain raw coal samples of the target coal and collect spectral data of the raw coal samples; The original spectral vector corresponding to the spectral data is preprocessed to obtain a standardized spectral feature vector; Based on the spectral feature vector and the pre-built prediction model, local nonlinear features and global mapping relationships corresponding to the spectral feature vector are generated, and the local nonlinear features and global mapping relationships are fused to generate the calorific value of the target coal.

2. The method according to claim 1, characterized in that, The preprocessing of the original spectral vector corresponding to the spectral data to obtain a standardized spectral feature vector includes: Based on the original spectral vector, calculate the mean and standard deviation of the spectral data; Based on the mean and the standard deviation, each reflection intensity value in the original spectral vector is centered to zero mean and scaled to unit variance to obtain the spectral feature vector.

3. The method according to claim 1, characterized in that, Before generating the local nonlinear features and global mapping relationship corresponding to the spectral feature vector based on the spectral feature vector and the pre-built prediction model, the method further includes: Construct a radial basis function branch to capture the local nonlinear features of the target spectral region in the spectral feature vector; Construct a linear branch to establish the overall mapping relationship between input features and prediction targets; Based on the radial basis function branch and the linear branch, a single-layer prediction module in the pre-built prediction model is constructed; Multiple single-layer prediction modules are sequentially superimposed to obtain the pre-built prediction model.

4. The method according to claim 3, characterized in that, The step of generating local nonlinear features and global mapping relationships corresponding to the spectral feature vector based on the spectral feature vector and the pre-built prediction model includes: The spectral feature vector is input into the pre-built prediction model to locate the spectral interval in the spectral feature vector that satisfies the target condition based on the radial basis function branch in the pre-built prediction model. The local response of each spectral channel in the spectral feature vector in the spectral interval is collected, as well as the local response of each spectral channel in other spectral intervals besides the spectral interval. The local nonlinear characteristics are obtained by combining the local response and response weights corresponding to the spectral range and the local response and response weights corresponding to the other spectral ranges.

5. The method according to claim 3, characterized in that, The step of generating local nonlinear features and global mapping relationships corresponding to the spectral feature vector based on the spectral feature vector and the pre-built prediction model includes: The spectral feature vector is input into the pre-built prediction model to establish a global mapping relationship between the spectral feature vector and the calorific value based on the linear branch in the pre-built prediction model.

6. The method according to claim 1, characterized in that, After calculating the calorific value of the target coal, the method further includes: Based on the pre-built prediction model, narrow-band spectral regions in the spectral feature vector are identified; A mapping relationship is established between the narrow-band spectral range and the physicochemical absorption bands in the spectral data, so as to determine the interpretable results of the prediction model based on the mapping relationship.

7. A device for calculating the calorific value of coal, characterized in that, include: The acquisition module is used to acquire raw coal samples of the target coal and to acquire the spectral data of the raw coal samples; The processing module is used to preprocess the original spectral vector corresponding to the spectral data to obtain a standardized spectral feature vector; The calculation module is used to generate local nonlinear features and global mapping relationships corresponding to the spectral feature vector based on the spectral feature vector and the pre-built prediction model, so as to fuse the local nonlinear features and the global mapping relationships to generate the calorific value of the target coal.

8. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the method for calculating the calorific value of coal as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the method for calculating the calorific value of coal as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed, it is used to implement the method for calculating the calorific value of coal as described in any one of claims 1-6.

Images