Coal calorific value on-line detection method based on xrf and nir feature level fusion

Abstract

The present application relates to the technical field of coal detection, in particular to a coal calorific value online detection method based on XRF and NIR feature level fusion, comprising: synchronously collecting X fluorescence spectrum data and near-infrared spectrum data of a coal sample, respectively preprocessing to extract an organic component feature set containing moisture and hydrogen-containing functional group related features, and an element component feature set containing inorganic mineral element and organic element related features, performing feature level fusion on the two types of feature sets to generate a fusion feature vector, inputting the fusion feature vector into an improved partial least squares regression model with adaptive weight adjustment according to coal sample surface laser topography data, and finally outputting a coal calorific value prediction value. The method effectively makes up for the one-sidedness of single spectrum detection, eliminates redundant information, reduces the deviation caused by sample surface topography differences, and improves the comprehensiveness, stability and accuracy of coal calorific value online detection.

Description

Technical Field

[0001] This invention relates to the field of coal detection technology, and in particular to an online detection method for coal calorific value based on the fusion of XRF and NIR characteristics. Background Technology

[0002] The calorific value of coal is a core indicator for evaluating coal quality and determining its application value. Online detection technology is crucial for quality control during coal production, transportation, and utilization. Currently, online detection of coal calorific value mainly employs single-spectral detection technologies. X-ray fluorescence spectroscopy (XRF) primarily focuses on the analysis of inorganic and organic elements in coal, obtaining elemental composition information by detecting the intensity and position of characteristic peaks, and thus indirectly correlating it with calorific value. Near-infrared spectroscopy (NIR) primarily targets the organic components in coal, reflecting information such as moisture and hydrogen-containing functional groups through characteristic band response values, thereby estimating calorific value. Furthermore, some technologies attempt to simply fuse the two spectral data, often at the data-level level, without specifically extracting and integrating the core features of the two types of spectra.

[0003] Single-spectral detection techniques have limitations. XRF detection cannot fully capture the influence of coal's organic components on calorific value, while near-infrared spectroscopy struggles to account for the effects of inorganic mineral elements, leading to potential biases in the results. Simple data-level fusion fails to screen key features of both spectral types, resulting in redundant information and failing to fully leverage the complementary advantages of dual spectra. Furthermore, existing detection models often employ conventional partial least squares regression models with fixed weights, neglecting the impact of coal sample surface morphology differences on spectral detection and calorific value prediction. This leads to significant fluctuations in detection accuracy for coal samples with different surface conditions, making it difficult to meet the precision and stability requirements of online detection. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an online detection method for coal calorific value based on the fusion of XRF and NIR characteristic levels.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an online detection method for coal calorific value based on the fusion of XRF and NIR characteristic levels, comprising:

[0006] Simultaneously collect X-ray fluorescence spectral data and near-infrared spectral data of coal samples to form a dual-spectral raw dataset;

[0007] The X-ray fluorescence spectral data and the near-infrared spectral data are preprocessed respectively to obtain the calibrated X-ray fluorescence spectral vector and the noise-reduced near-infrared spectral vector.

[0008] Feature extraction is performed on the noise-reduced near-infrared spectral vector to obtain a coal organic component feature set, which includes characteristic band response values ​​related to moisture and hydrogen-containing functional groups.

[0009] Feature extraction is performed on the calibrated X-ray fluorescence spectral vector to obtain a coal elemental composition feature set, which includes characteristic peak intensity and position information related to inorganic mineral elements and organic elements.

[0010] The feature-level fusion of the coal organic component feature set and the coal elemental component feature set is performed to generate a fused feature vector;

[0011] The fused feature vector is input into the improved partial least squares regression model, which adaptively adjusts the weights based on the laser morphology data of the coal sample surface.

[0012] The predicted value of coal calorific value is calculated and output using the improved partial least squares regression model.

[0013] As a further aspect of the present invention, the simultaneous acquisition of X-ray fluorescence spectral data and near-infrared spectral data of the coal sample constitutes a dual-spectral raw dataset, including:

[0014] The coal sample is controlled to pass through the online detection area at a constant rate;

[0015] When the coal sample passes through the online detection area, the coal sample is irradiated with the excitation source of an X-ray fluorescence spectrometer, and the X-ray fluorescence signal generated after the coal sample is excited is received, thereby acquiring continuous X-ray fluorescence spectral time-series data.

[0016] Simultaneously, the same coal sample is irradiated with the light source of a near-infrared spectrometer, and the near-infrared light signals reflected or transmitted by the coal sample are received to obtain continuous near-infrared spectral time-series data.

[0017] A laser profilometer is used to scan the coal sample passing through the online detection area, and laser morphology point cloud data of the coal sample surface is acquired simultaneously.

[0018] Align and bind the corresponding segments extracted from the X-ray fluorescence spectral time series data, the corresponding segments extracted from the near-infrared spectral time series data, and the corresponding segments extracted from the laser topography point cloud data at the same timestamp to form a dual-spectral raw data unit.

[0019] The multiple continuously acquired bispectral raw data units are arranged in chronological order to form the bispectral raw dataset.

[0020] As a further aspect of the present invention, the preprocessing of the X-ray fluorescence spectral data and the near-infrared spectral data to obtain the calibrated X-ray fluorescence spectral vector and the noise-reduced near-infrared spectral vector includes:

[0021] The X-ray fluorescence spectral data are subjected to background subtraction and scattering background correction to eliminate the influence of equipment background and continuous scattering background.

[0022] The X-ray fluorescence spectral data after background subtraction and scattering background correction were processed by decomposing the overlapping peaks of elemental spectral lines to separate the characteristic peak areas of individual elements.

[0023] The characteristic peak area of ​​the single element is intensity-corrected using standard samples to obtain a calibrated X-ray fluorescence spectral vector reflecting the absolute or relative content of the element.

[0024] The near-infrared spectral data are subjected to dark current subtraction and optical path length normalization to obtain a normalized near-infrared absorbance spectrum.

[0025] The standardized near-infrared absorbance spectrum is subjected to wavelet transform denoising to filter out high-frequency noise components and retain low-frequency characteristic signals related to the organic components of coal.

[0026] The denoised near-infrared absorbance spectrum is vector normalized to eliminate the light scattering effect caused by different sample particle size and packing density, and the denoised near-infrared spectral vector is obtained.

[0027] As a further aspect of the present invention, feature extraction is performed on the noise-reduced near-infrared spectral vector to obtain a feature set of coal organic components, including:

[0028] In the noise-reduced near-infrared spectral vector, the characteristic absorption wavelength range corresponding to the stretching vibration and combination vibration of water hydroxyl groups is located, and the integral value of the spectral absorbance within the characteristic absorption wavelength range corresponding to the stretching vibration and combination vibration of water hydroxyl groups is calculated as the water response characteristic value.

[0029] Locate the characteristic absorption wavelength range corresponding to the stretching vibration of aliphatic hydrocarbon groups in coal, and calculate the ratio of the peak height to the half-peak width of the spectral absorption peak within the characteristic absorption wavelength range corresponding to the stretching vibration of aliphatic hydrocarbon groups, which is used as the characteristic value of aliphatic hydrocarbon content.

[0030] Locate the characteristic absorption wavelength range corresponding to the stretching vibration of aromatic hydrocarbon groups in coal, and calculate the area of ​​the spectral absorption peak within the characteristic absorption wavelength range corresponding to the stretching vibration of aromatic hydrocarbon groups, which is used as the characteristic value of aromatic hydrocarbon content.

[0031] The characteristic absorption wavelength ranges related to hydrogen-containing functional groups of hydroxyl, amino, carboxyl, and phenolic hydroxyl groups in coal are located. The second derivative of the spectral curve within the characteristic absorption wavelength range is extracted, and the zero-crossing position and amplitude of the second derivative are used as characteristic values ​​of functional group type and content.

[0032] The moisture response characteristic value, the aliphatic hydrocarbon content characteristic value, the aromatic hydrocarbon content characteristic value, and the functional group type and content characteristic value are combined to form the characteristic set of coal organic components.

[0033] As a further aspect of the present invention, feature extraction is performed on the calibrated X-ray fluorescence spectral vector to obtain a set of coal elemental composition features, including:

[0034] In the calibrated X-ray fluorescence spectral vector, the characteristic X-ray peaks corresponding to silicon, aluminum, calcium and iron are identified, and the net intensity of each characteristic peak is recorded as the intensity value of the main ash element.

[0035] Identify the characteristic X-ray peaks corresponding to sulfur and phosphorus elements, calculate the ratio of their net intensity to the background intensity, and use it as the characteristic value of harmful elements.

[0036] Identify the characteristic peaks of low atomic number elements associated with organic components in coal, including the characteristic peaks of chlorine and potassium, and extract the peak position shift and full width at half maximum (FWHM) of these characteristic peaks as morphological characteristic values ​​of organic-related elements.

[0037] The intensity values ​​of the main ash elements, the characteristic values ​​of the harmful elements, and the morphological characteristic values ​​of the organically related elements are combined to form the coal elemental composition characteristic set.

[0038] As a further aspect of the present invention, feature-level fusion is performed on the coal organic component feature set and the coal elemental component feature set to generate a fused feature vector, including:

[0039] The feature values ​​in the set of organic components of coal are standardized and scaled to unify their value range to the range of zero to one, thus obtaining a standardized organic feature vector.

[0040] The characteristic values ​​in the coal elemental composition characteristic set are standardized and scaled to unify their value range to the range of zero to one, thus obtaining a standardized elemental feature vector.

[0041] Based on the historical contribution of each feature in the coal calorific value prediction model, an organic weight is assigned to each feature in the standardized organic feature vector, and an element weight is assigned to each feature in the standardized element feature vector.

[0042] The weighted standardized organic feature vector and the weighted standardized element feature vector are concatenated end to end to form an initial fusion vector;

[0043] Principal component analysis is performed on the initial fusion vector to extract the first few principal components whose cumulative variance contribution rate exceeds a preset threshold, thus forming the dimensionality-reduced fusion feature vector.

[0044] As a further aspect of the present invention, the improved partial least squares regression model adaptively adjusts weights based on laser morphology data of coal sample surfaces, and its working principle includes:

[0045] Acquire laser morphology point cloud data synchronized with the current coal sample to be tested;

[0046] The laser topography point cloud data is processed to calculate the topography feature parameters that reflect the surface morphology of the coal sample. The topography feature parameters include surface roughness, average particle projected area, and particle shape factor.

[0047] During the latent variable extraction process of the improved partial least squares regression model, the fused feature vector and the morphological feature parameters are input together.

[0048] When calculating each latent variable, the model dynamically adjusts the weight coefficients of each feature in the fused feature vector based on the morphological feature parameters. The principle for adjusting the weight coefficients is: for samples with large surface roughness or irregular particle shape, reduce the weight of features sensitive to spectral signal intensity and increase the weight of features sensitive to spectral shape and relative ratio.

[0049] The model recalculates the covariance matrix between samples based on the adjusted feature weights, and extracts latent variables and calculates regression coefficients for partial least squares regression based on this matrix, ultimately outputting a calorific value prediction that is insensitive to differences in the morphology of coal samples.

[0050] As a further aspect of the present invention, the laser morphology point cloud data is processed to calculate morphological feature parameters reflecting the surface morphology of the coal sample, including:

[0051] Plane fitting is performed on the laser topography point cloud data to remove the reference plane of the conveyor belt or the bottom of the sample container;

[0052] Calculate the standard deviation of the height values ​​of all points after removing the reference plane, and use it as the surface roughness.

[0053] Clustering and segmenting of the point cloud to identify point cloud clusters of individual coal particles, calculating the projected area of ​​each point cloud cluster on the horizontal plane, and taking the average of the projected areas of all particles as the average particle projected area.

[0054] For each identified coal particle point cloud cluster, calculate the aspect ratio of its minimum bounding rectangle, and calculate the ratio of the surface area of ​​the coal particle point cloud cluster to its volume. The product of the aspect ratio and the ratio is used as the shape factor of the coal particle corresponding to the coal particle in the coal particle point cloud cluster.

[0055] The average value of the shape factors for all coal particles is taken as the overall particle shape factor.

[0056] As a further aspect of the present invention, when calculating each latent variable, the model dynamically adjusts the weight coefficients of each feature in the fused feature vector based on the morphological feature parameters, including:

[0057] Obtain the morphological characteristic parameters corresponding to the current coal sample, including surface roughness, average particle projected area, and particle shape factor;

[0058] The morphological feature parameters are input into a preset weight mapping function, which is trained based on historical data and is used to output the initial weight adjustment factor corresponding to each original feature in the fused feature vector.

[0059] Identify all features in the fused feature vector that are sensitive to the absolute intensity of the spectral signal, including the net intensity of elemental feature peaks in the XRF feature set and features in the NIR feature set based on absorbance integral value and peak height.

[0060] Identify all features in the fused feature vector that are sensitive to spectral shape and relative ratios, including the feature peak intensity ratio, peak position shift and full width at half maximum (FWHM) in the XRF feature set, and features in the NIR feature set based on the second derivative of the absorption spectrum, peak area and peak width ratio.

[0061] For the feature that is sensitive to the absolute intensity of the spectral signal, its initial weight adjustment factor is multiplied by an intensity weight attenuation coefficient based on surface roughness and particle shape factor to obtain its final dynamic weight coefficient, wherein the intensity weight attenuation coefficient is negatively correlated with surface roughness and particle shape factor.

[0062] For the features that are sensitive to spectral shape and relative ratio, their initial weight adjustment factor is multiplied by a shape weight enhancement coefficient based on surface roughness and particle shape factor to obtain their final dynamic weight coefficient, wherein the shape weight enhancement coefficient is positively correlated with surface roughness and particle shape factor.

[0063] The calculated final dynamic weight coefficients are normalized to ensure that the sum of all weight coefficients is 1;

[0064] The normalized dynamic weight coefficients are multiplied one by one by each feature value in the fusion feature vector corresponding to the current latent variable calculation process to obtain a weighted feature vector, which is used to calculate the score of the current latent variable.

[0065] As a further aspect of the present invention, the method also includes dynamic compensation for the predicted calorific value:

[0066] Real-time monitoring of ambient temperature and humidity data during coal sample transportation;

[0067] Establish a compensation relationship model between ambient temperature and humidity and the volatility of surface moisture in coal;

[0068] After the improved partial least squares regression model outputs the predicted value of coal calorific value, the predicted value and the current environmental temperature and humidity data are input into the compensation relationship model.

[0069] The compensation relationship model outputs a heat compensation amount based on environmental conditions;

[0070] The predicted value of calorific value is added to the calorific value compensation amount to obtain the final coal calorific value test result after environmental dynamic compensation.

[0071] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0072] The coal organic component feature set, containing the characteristic band response values ​​of moisture and hydrogen-containing functional groups extracted from near-infrared spectral vectors, and the coal elemental component feature set, containing the characteristic peak intensity and position information of inorganic mineral elements and organic elements extracted from X-ray fluorescence spectral vectors, are fused at the feature level to generate a fused feature vector. This scheme can fully explore the core complementary information of the two types of spectra. The organic component features directly reflect the key material basis for the release of heat during coal combustion, while the elemental component features reflect the influence of inorganic minerals on calorific value. Feature-level fusion can eliminate redundant information, so that the fused feature vector simultaneously covers the core organic and inorganic features of coal, making up for the one-sidedness of single-spectral detection features, making the feature information more consistent with the actual influencing factors of calorific value, and improving the comprehensiveness and specificity of detection.

[0073] An improved partial least squares regression model was used to calculate the calorific value. This model incorporates laser morphology data of the coal sample surface, and uses this data to achieve adaptive weight adjustment of the model's internal parameters. Conventional partial least squares regression models have fixed weights and cannot adapt to coal samples with different surface morphologies. However, laser morphology data of the coal sample surface can reflect differences in sample particle size, surface smoothness, and other conditions. These differences can affect the accuracy of spectral signal acquisition. By using this data for adaptive weight adjustment, the model can dynamically optimize parameters according to the surface condition of different samples, adapt to the detection requirements of coal samples with different morphologies, reduce prediction bias caused by morphology differences, improve the stability and accuracy of detection results, and adapt to the scenario of coal sample conditions changing frequently in online detection. Attached Figure Description

[0074] Figure 1 This is a flowchart of the online detection method for coal calorific value based on XRF and NIR feature level fusion as described in this invention;

[0075] Figure 2 A flowchart for spectral data preprocessing;

[0076] Figure 3 A flowchart for obtaining the characteristic set of organic components in coal. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0078] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0079] See Figure 1This invention provides an online detection method for coal calorific value based on feature-level fusion of XRF and NIR spectra. The specific implementation process of this method is as follows: X-ray fluorescence (XRF) and near-infrared (NIR) spectral data of coal samples are simultaneously acquired to form a dual-spectral raw dataset. The XRF and NIR spectral data are preprocessed to obtain a calibrated XRF spectral vector and a denoised NIR spectral vector. Feature extraction is performed on the denoised NIR spectral vector to obtain a coal organic component feature set, which includes characteristic band response values ​​related to moisture and hydrogen-containing functional groups. Feature extraction is performed on the calibrated XRF spectral vector to obtain a coal elemental component feature set, which includes characteristic peak intensity and position information related to inorganic mineral elements and organic elements. Feature-level fusion is performed on the coal organic component feature set and the coal elemental component feature set to generate a fused feature vector. The fused feature vector is input into an improved partial least squares (PLS) regression model, which adaptively adjusts the weights based on the laser morphology data of the coal sample surface. The predicted value of coal calorific value is obtained by calculating using the improved partial least squares regression model.

[0080] In one embodiment of the present invention, the synchronous acquisition process of the dual-spectral raw dataset is described, wherein a coal sample is controlled to pass through an online detection area at a constant rate. As the coal sample passes through the online detection area, the coal sample is irradiated with an excitation source of an X-ray fluorescence spectrometer, and the X-ray fluorescence signal generated after the coal sample is excited is received, thus acquiring continuous X-ray fluorescence spectral time-series data. Simultaneously, the same coal sample is irradiated with a light source of a near-infrared spectrometer, and the near-infrared light signal reflected or transmitted by the coal sample is received, thus acquiring continuous near-infrared spectral time-series data. Simultaneously, a laser profilometer is used to scan the coal sample passing through the online detection area, synchronously acquiring laser morphology point cloud data of the coal sample surface. Corresponding segments extracted from the X-ray fluorescence spectral time-series data, the near-infrared spectral time-series data, and the laser morphology point cloud data at the same time stamp are aligned and bound to form a dual-spectral raw data unit. Multiple consecutively acquired dual-spectral raw data units are arranged in chronological order to constitute the dual-spectral raw dataset.

[0081] In practice, the coal sample transport device moves through a closed online detection area at a constant speed of 0.2 meters per second. This area integrates an X-ray fluorescence spectrometer, a near-infrared spectrometer, and a laser profilometer. As the coal sample passes through the online detection area, the rhodium-target X-ray tube of the X-ray fluorescence spectrometer acts as the excitation source, irradiating the surface of the coal sample on the conveyor belt at a 45-degree incident angle. The spectroscopic crystal and detector of the X-ray fluorescence spectrometer receive the characteristic X-ray fluorescence signal generated after the coal sample is excited, acquiring continuous X-ray fluorescence spectral time-series data at a frequency of 10 times per second, with each acquisition corresponding to a complete energy spectrum at a specific time point. Simultaneously, the light source of the near-infrared spectrometer emits a beam of near-infrared light with a wavelength range of 900 nm to 1700 nm, perpendicularly irradiating the same area of ​​the coal sample. The InGaAs array detector of the spectrometer receives the diffusely reflected near-infrared light signal from the coal sample, acquiring continuous near-infrared spectral time-series data at the same frequency of 10 times per second. At the same synchronized moment, the line laser generator of the laser profilometer projects a laser line onto the surface of the coal sample. The CCD camera captures the deformation profile of the laser line on the sample surface at a fixed frame rate. The three-dimensional laser topography point cloud data of the coal sample surface is generated through the triangulation principle. The scanning frequency of the laser profilometer is also set to 10 times per second to ensure time synchronization.

[0082] In practical implementation, the data synchronization module uses each sampling moment as a timestamp, extracting the corresponding energy spectrum segment from the continuous X-ray fluorescence spectral time series data. This segment is an intensity vector containing 1024 channels. Similarly, it extracts the spectral segment corresponding to the same timestamp from the continuous near-infrared spectral time series data; this segment is an absorbance vector containing 256 wavelengths. From the synchronized laser topography point cloud data stream, it extracts a set of three-dimensional coordinate points reflecting the surface morphology of coal samples at the same physical location, corresponding to the current timestamp. The data synchronization module performs alignment and binding operations. The alignment operation precisely matches the X-ray fluorescence spectral segment, near-infrared spectral segment, and laser topography point cloud segment belonging to the same physical sampling moment on the time axis based on the precise time synchronization signal between devices. The binding operation associates and stores the three aligned data segments as a single dual-spectral raw data unit with a unified identifier. This data unit's data structure contains three data fields. In some embodiments, the continuous operation of the conveying device and the detection system allows the data synchronization module to generate and record 6,000 bispectral raw data units arranged in chronological order within a ten-minute sampling period. These 6,000 sequentially arranged data units together constitute a complete bispectral raw dataset for model training or real-time prediction.

[0083] In practical implementation, the laser topography point cloud data in each dual-spectral raw data unit is used not only for subsequent adaptive weight adjustment but also for preliminary quality assessment of the current data unit. The data processing system calculates the number of valid 3D data points in the current point cloud segment. If the number of valid points is lower than a preset threshold, it determines that the coal sample at that moment may have discontinuous stacking, dust obstruction, or material gaps. The dual-spectral raw data unit corresponding to this time point is marked as low-quality data and does not participate in subsequent feature extraction and model calculation processes, awaiting automatic removal by the system or manual review. In some embodiments, for data units marked as low-quality, the system records their timestamp and triggers an alarm, indicating potential equipment or sample flow abnormalities.

[0084] After generating the dual-spectral raw dataset, the system numbers the data units chronologically and creates an index to facilitate rapid retrieval and access to any data unit and its contained X-ray fluorescence spectral fragments, near-infrared spectral fragments, and laser morphology point cloud data in subsequent preprocessing, feature extraction, and model invocation steps. The bound storage of laser morphology point cloud data within the dual-spectral raw data units provides a direct data input source for the subsequent improved partial least squares regression model to adaptively adjust weights based on the coal sample surface morphology. The synchronous acquisition of morphology data avoids the mismatch between morphological and spectral information caused by sample flow.

[0085] In practical implementation, the accuracy of the timestamp is crucial to ensuring data synchronization. All devices' acquisition clocks are synchronized via a master clock signal, which is distributed by a high-precision time server to the controllers of the X-ray fluorescence spectrometer, near-infrared spectrometer, and laser profilometer via hardware trigger lines. The generation of each data unit strictly adheres to the principle of "time-driven, data-aligned." That is, whenever the master clock emits an acquisition pulse, all devices simultaneously perform a data acquisition action. The data synchronization module generates a timestamp based on this pulse moment and captures the latest acquired data segment from each device's buffer for alignment and binding. This mechanism fundamentally ensures that the X-ray fluorescence signal, near-infrared signal, and laser profilometer signal originate from the same tiny volume element or surface region of the coal sample. It is understandable that if there is a deviation in the time synchronization of the acquisition system—for example, if the near-infrared spectroscopy acquisition time lags slightly behind the X-ray fluorescence spectroscopy acquisition time—the physical positions of the two spectral acquisitions will shift due to the continuous movement of the coal sample on the conveyor belt. This will result in the subsequent fusion analysis not targeting the same sample, introducing unpredictable errors. Optionally, during the system initialization phase, a synchronization verification test is performed. This involves passing a test block with a specific marker through the detection area to check whether the marker appears in the same timestamp sequence in the three types of signal data, thus calibrating and confirming the synchronization performance of the entire acquisition chain. The precise synchronous acquisition and alignment of X-ray fluorescence spectral time-series data, near-infrared spectral time-series data, and laser topography point cloud data form the data foundation for all subsequent analysis steps.

[0086] In one embodiment of the present invention, the process of preprocessing X-ray fluorescence spectroscopy data and near-infrared spectroscopy data is described, see reference. Figure 2 The X-ray fluorescence (XRF) spectral data undergoes background subtraction and scattering background correction to eliminate the influence of equipment background and continuous scattering background. The XRF spectral data after background subtraction and scattering background correction is then processed to decompose overlapping peaks of elemental spectral lines, separating the characteristic peak areas of individual elements. The characteristic peak areas of these individual elements are then intensity-corrected using standard samples to obtain calibrated XRF spectral vectors reflecting the absolute or relative content of the elements. The near-infrared (NIIR) spectral data undergoes dark current subtraction and optical path length normalization to obtain standardized NIIR absorbance spectra. The standardized NIIR absorbance spectra are then subjected to wavelet transform denoising to filter out high-frequency noise components and retain low-frequency characteristic signals related to the organic components of coal. Finally, the denoised NIIR absorbance spectra are vector normalized to eliminate the light scattering effects caused by differences in sample particle size and packing density, resulting in denoised NIIR spectral vectors.

[0087] In practice, background subtraction and scattering background correction are performed on the X-ray fluorescence spectral data. Background subtraction involves subtracting the device background energy spectrum, measured beforehand under sample-free conditions, from the acquired raw X-ray fluorescence energy spectrum. The device background energy spectrum includes inherent signals such as stray radiation from the X-ray tube and detector electronic noise. Scattering background correction employs an iterative filtering method. For an X-ray fluorescence energy spectrum curve after background subtraction, a 15-channel moving average filter is first used for smoothing. The smoothed curve is used as the initial scattering background estimate. This initial background estimate is subtracted from the original spectrum. Peak identification is performed on the remaining spectral lines, and the identified peak regions are masked. The non-peak regions are smoothed again to update the background estimate. This process is repeated three times to obtain the final scattering background curve. This final scattering background curve is then subtracted from the original spectrum to complete the scattering background correction. In some embodiments, background subtraction and scattering background correction are performed independently on each channel of the energy spectrum. The processed X-ray fluorescence spectral data eliminates the influence of the constant background signal from the device itself and the continuous background signal generated by sample scattering, making subsequent characteristic peak identification and decomposition more accurate.

[0088] In practice, the X-ray fluorescence spectral data after background subtraction and scattering background correction are subjected to elemental spectral line overlap peak decomposition. For the spectral bands where the Kα peaks of iron and cobalt overlap, a mixing model containing two Gaussian peaks and a linear background is established. This mixing model can be expressed as:

[0089] in: Indicates energy X-ray fluorescence intensity count at the location, and These are the coefficients of the linear background function. and These are the amplitudes of the characteristic peaks of iron and cobalt, respectively. and These are the center energy positions of the characteristic peaks for iron and cobalt, respectively. and These are the Gaussian width parameters of the characteristic peaks of iron and cobalt, respectively. A nonlinear least-squares fitting algorithm is used to fit this mixture model to the data points in the overlapping spectral bands. The fitting process optimizes all model parameters until the sum of squared residuals between the calculated spectral lines and the actual observed spectral lines is minimized. After fitting, the area of ​​the iron characteristic peak is determined by the parameters. and Calculations show that the area of ​​the characteristic peak of cobalt is determined by parameters. and The characteristic peak areas of individual elements are thus separated through calculation. It can be understood that overlapping peak decomposition is applicable to all element pairs with overlapping spectral lines, such as manganese and chromium. Each decomposition process establishes a corresponding mixing model for a specific overlapping spectral band, and the number of Gaussian peaks in the model is consistent with the number of overlapping elements.

[0090] In practice, standard samples are used to perform intensity correction on the characteristic peak areas of a single element. The standard samples are a series of standard coal samples or artificially prepared standard samples with known elemental contents, whose elemental contents cover the expected content range of the coal sample to be tested. The standard samples are placed in the detection system, and X-ray fluorescence spectral data are acquired and preprocessed according to the same procedure, and the characteristic peak areas of each element in the standard samples are obtained. For each target element, such as iron, a linear correction curve is established between its characteristic peak area and its known absolute content. The mathematical expression of the correction curve is as follows: ,in Indicates the iron content. This represents the area of ​​the characteristic peak of iron obtained from the decomposition of the spectrum. and The correction coefficients and intercepts are determined through linear regression. For an unknown coal sample, after decomposing it to obtain the characteristic peak area of ​​iron, the absolute content of iron can be estimated by substituting it into this correction curve.

[0091] The absolute or relative abundance estimates of all target elements (such as silicon, aluminum, calcium, iron, sulfur, phosphorus, chlorine, potassium, etc.) are arranged in elemental order to form a one-dimensional vector, which is the calibrated X-ray fluorescence spectral vector. Optionally, when only the relative abundance is needed, the relative intensity value can be obtained by comparing the characteristic peak area (after background and scattering correction) with the characteristic peak area of ​​the internal standard element to form the calibrated X-ray fluorescence spectral vector.

[0092] In practice, near-infrared spectral data undergoes dark current subtraction and optical path length standardization. Dark current subtraction involves turning off the light source of the near-infrared spectrometer before each spectral acquisition, collecting a set of dark spectra, and then subtracting the corresponding dark spectral intensity value from the original intensity value of each wavelength during subsequent sample spectral acquisitions. Optical path length standardization is performed on the reflectance spectrum by selecting a standard white plate with known reflectance and acquiring its near-infrared spectrum as a reference spectrum. The absorbance of the sample is then used as the reference spectrum. Through formula The calculation yielded, where It is the spectral intensity of the sample after dark current subtraction. The absorbance spectrum calculated after dark current subtraction is the reference spectral intensity, which is the standardized near-infrared absorbance spectrum. This processing eliminates the influence of light source intensity fluctuations and instrument response differences. In some embodiments, wavelet transform denoising is performed on the standardized near-infrared absorbance spectrum. The 'db4' wavelet is selected as the mother wavelet, and the absorbance spectral signal is decomposed into four levels to obtain wavelet coefficients at different scales. The high-frequency detail coefficients of the first and second levels are set to zero, and wavelet reconstruction is performed using the remaining low-frequency approximation coefficients and the non-zeroed detail coefficients. The reconstructed spectral signal filters out high-frequency noise components caused by electronic noise and random interference, while retaining low-frequency characteristic signals related to the vibration of organic components in coal.

[0093] In practice, the denoised near-infrared absorbance spectrum is vector normalized. Vector normalization is applied to an absorbance spectral vector containing m wavelengths. First, calculate the magnitude of the vector. Divide each absorbance value in the vector by this modulus to obtain the normalized vector. It is understandable that vector normalization alters the length of the spectral vector while preserving its direction, i.e., the shape characteristics of the spectrum. This effectively reduces the scattering effects on incident light caused by uneven particle size, varying surface roughness, and differences in packing density in coal samples, making the processed spectral vector more reflective of the sample's chemical composition. The spectral vector after vector normalization is the noise-reduced near-infrared spectral vector, which, along with the calibrated X-ray fluorescence spectral vector, will serve as input for subsequent feature extraction steps.

[0094] In one embodiment of the present invention, the process of extracting features from near-infrared spectral vectors to obtain a feature set of coal organic components, and extracting features from calibrated X-ray fluorescence spectral vectors to obtain a feature set of coal elemental components are described. (See reference...) Figure 3In the noise-reduced near-infrared spectral vector, the characteristic absorption wavelength range corresponding to the stretching vibration and combination vibration of hydroxyl groups in water is located, and the integral value of the spectral absorbance within this range is calculated as the moisture response characteristic value. The characteristic absorption wavelength range corresponding to the stretching vibration of aliphatic hydrocarbon groups in coal is located, and the ratio of the peak height to the full width at half maximum (FWHM) of the spectral absorption peak within this range is calculated as the aliphatic hydrocarbon content characteristic value. The characteristic absorption wavelength range corresponding to the stretching vibration of aromatic hydrocarbon groups in coal is located, and the area of ​​the spectral absorption peak within this range is calculated as the aromatic hydrocarbon content characteristic value. The characteristic absorption wavelength range related to the hydrogen-containing functional groups (hydroxyl, amino, carboxyl, and phenolic hydroxyl) in coal is located, and the second derivative of the spectral curve within this range is extracted. The zero-crossing position and amplitude of the second derivative are used as the functional group type and content characteristic values. The moisture response characteristic value, the aliphatic hydrocarbon content characteristic value, the aromatic hydrocarbon content characteristic value, and the functional group type and content characteristic value are combined to form the characteristic set of coal organic components.

[0095] In the calibrated X-ray fluorescence spectral vector, characteristic X-ray peaks corresponding to silicon, aluminum, calcium, and iron are identified, and the net intensity of each characteristic peak is recorded as the intensity value of the main ash elements. Characteristic X-ray peaks corresponding to sulfur and phosphorus are identified, and the ratio of their net intensity to background intensity is calculated as the characteristic value of harmful elements. Characteristic peaks of low atomic number elements related to the organic components in coal, including characteristic peaks of chlorine and potassium, are identified, and their peak position shift and full width at half maximum (FWHM) are extracted as the morphological characteristic value of organically related elements. The intensity values ​​of the main ash elements, the characteristic values ​​of harmful elements, and the morphological characteristic values ​​of organically related elements are combined to form the coal elemental composition characteristic set.

[0096] In practice, the characteristic absorption wavelength ranges corresponding to the stretching vibration and combination vibration of water hydroxyl groups are located in the noise-reduced near-infrared spectral vector. The characteristic absorption wavelength range of the first harmonic of the stretching vibration of water hydroxyl groups is located from about 1390 nm to 1420 nm, and the characteristic absorption wavelength range of the combination vibration of water hydroxyl groups is located from about 1890 nm to 1950 nm. The location operation is completed based on the wavelength index of the near-infrared spectrum, and each data point in the spectral vector corresponds to a specific wavelength. The integral value of the spectral absorbance within the characteristic absorption wavelength range corresponding to the stretching vibration and combination vibration of the water hydroxyl group is calculated as the water response characteristic value. The integral value is calculated using the trapezoidal rule in numerical integration. The absorbance data is integrated in the range of 1390 nm to 1420 nm, and the integral result is recorded as the water response characteristic value F_OH_1. The absorbance data is integrated in the range of 1890 nm to 1950 nm, and the integral result is recorded as the water response characteristic value F_OH_2. The two integral values ​​of the water hydroxyl group together constitute the final water response characteristic value. In some embodiments, the water response characteristic value is a weighted sum of the two integral values, and the weighting coefficients are determined by multiple linear regression of the standard sample set.

[0097] In practical implementation, the characteristic absorption wavelength range corresponding to the stretching vibrations of aliphatic hydrocarbon groups in coal is located. The characteristic absorption wavelength range of the first overtone of the CH stretching vibration of aliphatic hydrocarbon groups is located at approximately 1680 nm to 1780 nm. The ratio of the peak height to the half-peak width (WHM) of the spectral absorption peak within the characteristic absorption wavelength range corresponding to the stretching vibrations of aliphatic hydrocarbon groups is calculated as a characteristic value of aliphatic hydrocarbon content. The peak height is defined as the difference between the maximum absorbance and the baseline absorbance within this range. The baseline absorbance is determined by the straight line connecting the absorbance points corresponding to the start and end wavelengths of this range. The WHM is defined as the width of the absorption peak corresponding to half the peak height. The calculation formula is:

[0098] in: This indicates the peak height of the aliphatic CH absorption peak. This represents the full width at half maximum (FWHM) of the absorption peak. This ratio is correlated with the concentration of aliphatic CH groups. The formula is for extracting spectral shape features. To increase the absorption peak height, The half-peak width (WHM) of the absorption peak is used to characterize the relative content of aliphatic CH groups. This can be understood as locating the characteristic absorption wavelength range corresponding to the stretching vibrations of aromatic hydrocarbon groups in coal. The first-order overtone characteristic absorption wavelength range of the CH stretching vibrations of aromatic hydrocarbon groups is located between approximately 1610 nm and 1660 nm. The area of ​​the spectral absorption peak within this characteristic absorption wavelength range corresponding to the stretching vibrations of aromatic hydrocarbon groups is calculated as the characteristic value of aromatic hydrocarbon content. The peak area calculation also uses numerical integration, with the integration range being 1610 nm to 1660 nm. Baseline subtraction is required before integration, and the resulting integral is the characteristic value of aromatic hydrocarbon content, F_aromatic.

[0099] In the specific implementation, the characteristic absorption wavelength ranges related to the hydrogen-containing functional groups of hydroxyl, amino, carboxyl, and phenolic hydroxyl groups in coal are located. These characteristic absorption wavelength ranges include hydroxyl (OH) at approximately 1400-1450 nm, amino (NH) at approximately 1490-1550 nm, carboxyl (OH) at approximately 1850-1900 nm, and phenolic hydroxyl (OH) at approximately 1400-1450 nm and approximately 1900-1950 nm. The second derivative of the spectral curves within the characteristic absorption wavelength range is extracted. The second derivative is calculated using a Savitzky-Golay filter with a polynomial order of 2 and a window width of 11 data points, resulting in the second derivative spectrum. The zero-crossing position and amplitude of the second derivative are used as characteristic values ​​of functional group type and content. The zero-crossing position corresponds to the center wavelength of the absorption peak in the original spectrum, and the amplitude reflects the sharpness and intensity of the absorption peak. The characteristic values ​​of moisture response, aliphatic hydrocarbon content, and aromatic hydrocarbon content, along with the zero-crossing position and amplitude characteristics extracted from the second derivative spectrum, are combined to form a feature set of coal organic components. This feature set is a feature vector containing multiple numerical elements. Optionally, the feature set of coal organic components can be listed in tabular form. The table clearly shows the characteristics extracted from the near-infrared spectrum and their sources. Refer to Table 1, which illustrates the composition of the coal organic component feature set.

[0100] Table 1: Characteristic Set of Organic Components in Coal

[0101] In practice, characteristic X-ray peaks corresponding to silicon, aluminum, calcium, and iron are identified in the calibrated X-ray fluorescence spectral vector. Silicon corresponds to the SiKα peak, aluminum to the AlKα peak, calcium to the CaKα peak, and iron to the FeKα peak. The channel position of each characteristic peak in the energy spectrum is determined by energy calibration, and the net intensity of each characteristic peak is recorded as the intensity value of the main ash element. The net intensity of the characteristic peak is obtained by subtracting the count of continuous background from the total count in the channel region where the peak is located. Characteristic X-ray peaks corresponding to sulfur and phosphorus are identified. Sulfur corresponds to the SKα peak, and phosphorus corresponds to the PKα peak. The ratio of the net intensity of the characteristic peak of sulfur and phosphorus to the background intensity is calculated as the characteristic value of the harmful element. The background intensity is taken as the average count of the interference-free regions on both sides of the characteristic peak.

[0102] In some embodiments, characteristic peaks of low atomic number elements associated with organic components in coal are identified, including characteristic peaks of chlorine and potassium. Chlorine corresponds to the ClKα peak, and potassium corresponds to the KKα peak. The peak position shift and half-width at half-maximum (HWHM) of the characteristic peaks of chlorine and potassium are extracted as morphological feature values ​​of organically related elements. The peak position shift is defined as the difference between the energy channel of the actual measured characteristic peak center and the standard energy channel, and the HWHM is defined as the full width at half the height of the characteristic peak. It can be understood that the intensity values ​​of major ash elements, the characteristic values ​​of harmful elements, and the morphological feature values ​​of organically related elements are combined to form a coal elemental component feature set. This coal elemental component feature set is also a numerical feature vector, which, together with the coal organic component feature set, serves as the input for subsequent feature-level fusion.

[0103] In one embodiment of the present invention, the working principle of feature-level fusion to generate a fused feature vector and the improved partial least squares regression model adjusting weights based on laser morphology data are described. Feature-level fusion of the coal organic component feature set and the coal elemental component feature set is performed to generate a fused feature vector. Specifically, this includes: standardizing and scaling each feature value in the coal organic component feature set to unify its value range to zero to one, obtaining a standardized organic feature vector; standardizing and scaling each feature value in the coal elemental component feature set to unify its value range to zero to one, obtaining a standardized elemental feature vector; assigning an organic weight to each feature in the standardized organic feature vector and an elemental weight to each feature in the standardized elemental feature vector based on the historical contribution of each feature in the coal calorific value prediction model; concatenating the weighted standardized organic feature vector and the weighted standardized elemental feature vector to form an initial fused vector; and performing principal component analysis on the initial fused vector to extract the top principal components whose cumulative variance contribution rate exceeds a preset threshold, constituting the dimensionality-reduced fused feature vector.

[0104] The improved partial least squares regression model adaptively adjusts weights based on laser morphology data of the coal sample surface. Its working principle includes: acquiring laser morphology point cloud data synchronized with the current coal sample. During the latent variable extraction process of the improved partial least squares regression model, the fused feature vector and the morphology feature parameters calculated from the laser morphology point cloud data are input together. When calculating each latent variable, the model dynamically adjusts the weight coefficients of each feature in the fused feature vector according to the morphology feature parameters. The adjustment principle is: for samples with large surface roughness or irregular particle shapes, the weights of features sensitive to spectral signal intensity are reduced, while the weights of features sensitive to spectral shape and relative ratios are increased. The model recalculates the covariance matrix between samples based on the adjusted feature weights, and performs latent variable extraction and regression coefficient calculation for partial least squares regression based on this matrix, ultimately outputting a predicted calorific value that is insensitive to differences in coal sample morphology.

[0105] In practice, feature-level fusion is performed on the coal organic component feature set and the coal elemental component feature set to generate a fused feature vector. Each feature value in the coal organic component feature set is then standardized and scaled using a min-max normalization method. Its value after standardization and scaling Calculated using the formula:

[0106] in: This represents the set of all values ​​for that feature in the training set samples. It is the minimum value of this feature in the training set. This is the maximum value of the feature in the training set. After this calculation, the value range of each feature in the coal organic component feature set is unified to the range of zero to one, resulting in a standardized organic feature vector. The formula is a data standardization formula. Through min-max normalization, the original feature values ​​are mapped to the [0,1] interval, eliminating dimensional differences and facilitating model calculation. It can be understood that the same standardization and scaling process is also applied to each feature value in the coal elemental component feature set, mapping each feature value to the range of zero to one, resulting in a standardized elemental feature vector. The standardization and scaling process eliminates the influence of differences in dimensions and numerical ranges between different features.

[0107] In practical implementation, based on the historical contribution of each feature in the coal calorific value prediction model, an organic weight is assigned to each feature in the standardized organic feature vector, and an element weight is assigned to each feature in the standardized element feature vector. The historical contribution is obtained by analyzing an established, unweighted benchmark partial least squares regression model based on the same features. In this benchmark model, the average absolute value of the loadings of each feature on the model's latent variables is calculated as the initial historical contribution score for that feature. The initial historical contribution scores of each feature in the standardized organic feature vector are normalized so that the sum of the weights of all organic features is 1. The normalized value is the organic weight assigned to each feature. Element weights are assigned to each feature in the standardized element feature vector using the same method. The weighted standardized organic feature vector and the weighted standardized element feature vector are then concatenated end-to-end. The weighting operation involves multiplying each feature value in the standardized organic feature vector by its corresponding organic weight, and multiplying each feature value in the standardized element feature vector by its corresponding element weight. The two weighted vectors are then connected sequentially to form a higher-dimensional initial fusion vector. In some embodiments, the feature dimension of the initial fusion vector is the sum of the normalized organic feature vector dimension and the normalized element feature vector dimension.

[0108] In the specific implementation, principal component analysis (PCA) is performed on the initial fusion vector. PCA takes the initial fusion vector as input, calculates its covariance matrix, and performs eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. The eigenvectors are the principal component directions. The principal components are arranged in descending order of eigenvalues, and the variance contribution rate and cumulative variance contribution rate of each principal component are calculated. The top few principal components with a cumulative variance contribution rate exceeding a preset threshold (set to 95%) are extracted, forming the dimensionality-reduced fusion feature vector. The dimension of the fusion feature vector is equal to the number of extracted principal components, much lower than the dimension of the initial fusion vector. PCA removes potential multicollinearity in the initial fusion vector and extracts new features that represent most of the original feature information. Optionally, the preset threshold can be adjusted according to the actual application's requirements for model complexity and accuracy, for example, set to 90% or 98%. Table 2 illustrates a feature-level fusion process, where the coal organic component feature set contains 4 features, and the coal elemental component feature set contains 5 features.

[0109] Table 2: Feature-level fusion process table

[0110] In its implementation, the improved partial least squares regression model adaptively adjusts weights based on laser morphology data of the coal sample surface, acquiring laser morphology point cloud data synchronized with the current coal sample. The laser morphology point cloud data and spectral data are aligned and bound at the same timestamp. During latent variable extraction in the improved partial least squares regression model, the fused feature vector and morphological feature parameters calculated from the laser morphology point cloud data are input together. These morphological feature parameters include surface roughness, average particle projected area, and particle shape factor. When calculating each latent variable, the improved partial least squares regression model dynamically adjusts the weight coefficients of each feature in the fused feature vector based on the morphological feature parameters. The adjustment principle is to reduce the weight of features sensitive to the absolute intensity of the spectral signal and increase the weight of features sensitive to spectral shape and relative ratio for samples with high surface roughness or irregular particle shapes. The model recalculates the covariance matrix between samples based on the adjusted feature weights and performs latent variable extraction and regression coefficient calculation for partial least squares regression based on this covariance matrix, ultimately outputting a predicted calorific value that is insensitive to differences in coal sample morphology.

[0111] It is understandable that, since the features in the fused feature vector have already undergone standardization and principal component analysis, the improved partial least squares regression model at this stage adjusts the importance of the fused feature vector of each sample in the covariance calculation. In some embodiments, the morphological feature parameters are quantized into an adjustment vector, which is then multiplied by the fused feature vector before participating in the calculation of latent variables. This adaptive weight adjustment mechanism enables the model to dynamically adjust the interpretation of spectral information according to different physical states of the sample surface, thereby reducing the impact of sample morphology changes on the accuracy of calorific value prediction.

[0112] In one embodiment of the present invention, the specific process of calculating morphological feature parameters and dynamically adjusting the weight coefficients of the model from laser morphological point cloud data, as well as the dynamic compensation of the calorific value prediction results, is described. The laser morphological point cloud data is processed to calculate morphological feature parameters reflecting the surface morphology of the coal sample, including: performing planar fitting on the laser morphological point cloud data to remove the reference plane of the conveyor belt or the bottom of the sample container; calculating the standard deviation of the height values ​​of all points after removing the reference plane, as the surface roughness; clustering the point cloud to identify point cloud clusters of individual coal particles; calculating the projected area of ​​each point cloud cluster on the horizontal plane; and calculating the average of the projected areas of all particles, as the average particle projected area; for each identified coal particle point cloud cluster, calculating the aspect ratio of its minimum bounding rectangle, and calculating the ratio of the surface area to the volume of the coal particle point cloud cluster; using the product of the aspect ratio and the ratio as the shape factor of the coal particle corresponding to the coal particle point cloud cluster; and calculating the average of the shape factors of all coal particles, as the overall particle shape factor.

[0113] When calculating each latent variable, the model dynamically adjusts the weight coefficients of each feature in the fused feature vector based on the morphological feature parameters. This includes: obtaining the morphological feature parameters corresponding to the current coal sample, including surface roughness, average particle projected area, and particle shape factor; inputting the morphological feature parameters into a preset weight mapping function, which is trained based on historical data and used to output the initial weight adjustment factor corresponding to each original feature in the fused feature vector; identifying all features in the fused feature vector that are sensitive to the absolute intensity of the spectral signal, including the net intensity of elemental feature peaks in the XRF feature set and features based on absorbance integral value and peak height in the NIR feature set; and identifying all features in the fused feature vector that are sensitive to spectral shape and relative ratios, including the feature peak intensity ratio, peak position offset, and full width at half maximum (FWHM) in the XRF feature set, and features based on the second derivative of the absorption spectrum, peak area, and peak width ratio in the NIR feature set.

[0114] For the feature sensitive to the absolute intensity of the spectral signal, its initial weight adjustment factor is multiplied by an intensity weight attenuation coefficient based on surface roughness and particle shape factors to obtain its final dynamic weight coefficient, wherein the intensity weight attenuation coefficient is negatively correlated with surface roughness and particle shape factors. For the feature sensitive to spectral shape and relative ratio, its initial weight adjustment factor is multiplied by a shape weight enhancement coefficient based on surface roughness and particle shape factors to obtain its final dynamic weight coefficient, wherein the shape weight enhancement coefficient is positively correlated with surface roughness and particle shape factors. The calculated final dynamic weight coefficients are normalized to ensure that the sum of all weight coefficients is 1. The normalized dynamic weight coefficients are multiplied one by one by each feature value in the fused feature vector corresponding to the current latent variable calculation process to obtain a weighted feature vector, which is used to calculate the score of the current latent variable.

[0115] The method also includes dynamic compensation for the predicted calorific value: real-time monitoring of ambient temperature and humidity data during coal sample transportation. A compensation relationship model is established between ambient temperature and humidity and the volatile matter content of coal surface moisture. After the improved partial least squares regression model outputs the predicted value of coal calorific value, the predicted value and the current ambient temperature and humidity data are input into the compensation relationship model. The compensation relationship model outputs a calorific value compensation amount based on the environmental conditions. The predicted calorific value and the calorific value compensation amount are added to obtain the final coal calorific value detection result after dynamic environmental compensation.

[0116] In practice, the laser morphology point cloud data is processed to calculate morphological characteristic parameters reflecting the surface morphology of the coal sample. Plane fitting is performed on the laser morphology point cloud data, and a reference plane is estimated from the data using a random sampling consensus algorithm. Points in the laser morphology point cloud data whose distance to this estimated plane is less than a set threshold are considered in-plane points. Through multiple iterations, a plane model containing the most in-plane points is found. This plane model represents the bottom surface of the conveyor belt or sample container. The height of this reference plane is then subtracted from the original laser morphology point cloud data to complete the reference plane removal operation. The standard deviation of the height values ​​of all points after removing the reference plane is calculated as the surface roughness. Surface roughness The calculation formula is based on the height values ​​of all points. Its mean The deviation is addressed. In some embodiments, the point cloud is clustered to identify point cloud clusters of individual coal particles. The clustering and segmentation employs the DBSCAN algorithm based on Euclidean distance. The maximum bounding rectangle of each point cloud cluster on the horizontal plane is used as the projected contour of the particle. The projected area of ​​each point cloud cluster is calculated, and the average of the projected areas of all particles is taken as the average particle projected area. .

[0117] It is understandable that, for each identified coal particle point cloud cluster, the aspect ratio of its minimum bounding rectangle is calculated. The minimum bounding rectangle is obtained using the rotating caliper algorithm, and the aspect ratio is defined as the length of the longer side of the rectangle. Divide by the length of the shorter side Simultaneously calculate the surface area of ​​coal particle point cloud clusters. The ratio of its volume V to its surface area The volume is obtained by summing the areas of all triangles after Delaunay triangulation. The aspect ratio and the product of the 3D convex hull volume of the point cloud cluster are used as the shape factor of the coal particles corresponding to the point cloud cluster. The formula for calculating the shape factor is:

[0118] in: The shape factor representing a single coal particle. This represents the length of the longest side of the smallest bounding rectangle of the particle point cloud cluster. This represents the length of the shorter side of the smallest bounding rectangle of the particle point cloud cluster. This represents the surface area of ​​the particle point cloud cluster. This represents the volume of the particle point cloud cluster. The shape factor for all coal particles. The arithmetic mean is calculated and used as the overall particle shape factor reflecting the overall morphology of the current sample. Surface roughness Average particle projection area and overall particle shape factor Together, they constitute the morphological feature parameters, which are used for subsequent dynamic adjustment of model weights.

[0119] In practical implementation, the improved partial least squares regression model dynamically adjusts the weight coefficients of each feature in the fused feature vector based on the morphological feature parameters when calculating each latent variable, thereby obtaining the morphological feature parameters corresponding to the current coal sample, including surface roughness. Average particle projection area and particle shape factor Surface roughness Average particle projection area and particle shape factor These three morphological feature parameters are input to a preset weight mapping function, which is a feedforward neural network trained based on historical data. The neural network takes the morphological feature parameters as input and uses the initial weight adjustment factor corresponding to each original feature in the fused feature vector. For output. Identify all features in the fused feature vector that are sensitive to the absolute intensity of the spectral signal, including the net intensity of silicon, aluminum, calcium, and iron characteristic peaks in the XRF feature set, and the moisture response feature value based on absorbance integral and the feature value based on peak height in the NIR feature set. Identify all features in the fused feature vector that are sensitive to spectral shape and relative ratios, including the ratio of the net intensity of sulfur characteristic peak to background intensity, the ratio of the net intensity of phosphorus characteristic peak to background intensity, the peak position shift and full width at half maximum (FWHM) of chlorine characteristic peak, the peak position shift and FWHM of potassium characteristic peak in the XRF feature set, and the zero-crossing position and amplitude features extracted based on the second derivative of the absorption spectrum, the aromatic hydrocarbon content feature value based on peak area integral, and the aliphatic hydrocarbon content feature value based on the ratio of peak height to FWHM in the NIR feature set.

[0120] In practical implementation, for features sensitive to the absolute intensity of spectral signals, their initial weighting adjustment factors are... With a surface roughness-based and particle shape factor Intensity weight attenuation coefficient Multiplying them yields the final dynamic weighting coefficients. In which the intensity weight attenuation coefficient With surface roughness and particle shape factor There is a negative correlation, and the intensity weight attenuation coefficient is... The calculation method is as follows ,in and The attenuation coefficient is determined based on training with historical data. For features sensitive to spectral shape and relative ratios, their initial weights are adjusted by... With a surface roughness-based and particle shape factor Shape weight enhancement factor Multiplying them yields the final dynamic weighting coefficients. The shape weight enhancement coefficient With surface roughness and particle shape factor There is a positive correlation, shape weight enhancement coefficient The calculation method is as follows ,in and The enhancement coefficients are determined based on training with historical data.

[0121] It is understandable that the intensity weight attenuation coefficient is... and shape weight enhancement coefficient The design principle is that when the surface roughness Increase or particle shape factor As the sample surface becomes more uneven or the particle shape more irregular, the absolute intensity of the spectral signal will be weakened or distorted due to effects such as scattering and shadowing. Therefore, the weight of intensity-sensitive features should be reduced, while the shape features and relative ratios between features are less affected by physical morphology, thus the weight of these features should be increased. The final dynamic weight coefficients of all calculated features are then used. Normalization is performed to ensure that the sum of all weight coefficients is 1. The normalization operation is to normalize each... Divide by the sum of all weight coefficients. Multiply the normalized dynamic weight coefficients one by one with each feature value in the fusion feature vector corresponding to the current latent variable calculation process to obtain the weighted feature vector, which is used to calculate the score of the current latent variable. Optionally, the weight mapping function can also use a multiple linear regression model or a support vector machine model to establish the mapping relationship between the morphological feature parameters and the initial weight adjustment factor.

[0122] In practice, the predicted calorific value is dynamically compensated, and the ambient temperature and humidity data during coal sample transportation are monitored in real time. These data are collected by environmental sensors installed above the conveyor belt or near the sample testing area, with a sampling frequency of once per second. A compensation relationship model is established between ambient temperature and humidity and the volatility of surface moisture in coal. This model, based on extensive historical experimental data, describes the relationship under different ambient temperatures. and ambient humidity The difference between the evaporation loss of surface moisture in coal samples during the detection process and the moisture loss under standard conditions affects the predicted calorific value. The improved partial least squares regression model outputs the predicted calorific value of coal. Then, the predicted value Compared with current ambient temperature data Current ambient humidity data The input is the compensation relationship model, which outputs a heat compensation amount based on the environmental conditions. In some embodiments, the compensation relationship model is a temperature-based model. and humidity For input, with compensation amount The output is a bivariate polynomial function, whose coefficients are determined by fitting calibration experimental data. The predicted value of the calorific value is then used. With heat compensation The results are added together to obtain the final coal calorific value after environmental dynamic compensation. , It is understandable that the application of real-time monitoring and compensation models of environmental temperature and humidity data can correct the systematic biases in calorific value prediction models caused by differences in the surface moisture state of coal due to changes in environmental conditions.

[0123] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for online detection of coal calorific value based on the fusion of XRF and NIR characteristic levels, characterized in that, include: Simultaneously collect X-ray fluorescence spectral data and near-infrared spectral data of coal samples to form a dual-spectral raw dataset; The X-ray fluorescence spectral data and the near-infrared spectral data are preprocessed respectively to obtain the calibrated X-ray fluorescence spectral vector and the noise-reduced near-infrared spectral vector. Feature extraction is performed on the noise-reduced near-infrared spectral vector to obtain a coal organic component feature set, which includes characteristic band response values ​​related to moisture and hydrogen-containing functional groups. Feature extraction is performed on the calibrated X-ray fluorescence spectral vector to obtain a coal elemental composition feature set, which includes characteristic peak intensity and position information related to inorganic mineral elements and organic elements. The feature-level fusion of the coal organic component feature set and the coal elemental component feature set is performed to generate a fused feature vector; The fused feature vector is input into the improved partial least squares regression model, which adaptively adjusts the weights based on the laser morphology data of the coal sample surface. The predicted value of coal calorific value is calculated and output using the improved partial least squares regression model.

2. The online detection method for coal calorific value based on XRF and NIR characteristic-level fusion according to claim 1, characterized in that, The X-ray fluorescence spectral data and near-infrared spectral data of the simultaneously acquired coal samples constitute a dual-spectral raw dataset, including: The coal sample is controlled to pass through the online detection area at a constant rate; When the coal sample passes through the online detection area, the coal sample is irradiated with the excitation source of an X-ray fluorescence spectrometer, and the X-ray fluorescence signal generated after the coal sample is excited is received, thereby acquiring continuous X-ray fluorescence spectral time-series data. Simultaneously, the same coal sample is irradiated with the light source of a near-infrared spectrometer, and the near-infrared light signals reflected or transmitted by the coal sample are received to obtain continuous near-infrared spectral time-series data. A laser profilometer is used to scan the coal sample passing through the online detection area, and laser morphology point cloud data of the coal sample surface is acquired simultaneously. Align and bind the corresponding segments extracted from the X-ray fluorescence spectral time series data, the corresponding segments extracted from the near-infrared spectral time series data, and the corresponding segments extracted from the laser topography point cloud data at the same timestamp to form a dual-spectral raw data unit. The multiple continuously acquired bispectral raw data units are arranged in chronological order to form the bispectral raw dataset.

3. The online detection method for coal calorific value based on XRF and NIR characteristic-level fusion according to claim 1, characterized in that, The preprocessing of the X-ray fluorescence spectral data and the near-infrared spectral data to obtain the calibrated X-ray fluorescence spectral vector and the denoised near-infrared spectral vector includes: The X-ray fluorescence spectral data are subjected to background subtraction and scattering background correction to eliminate the influence of equipment background and continuous scattering background. The X-ray fluorescence spectral data after background subtraction and scattering background correction were processed by decomposing the overlapping peaks of elemental spectral lines to separate the characteristic peak areas of individual elements. The characteristic peak area of ​​the single element is intensity-corrected using standard samples to obtain a calibrated X-ray fluorescence spectral vector reflecting the absolute or relative content of the element. The near-infrared spectral data are subjected to dark current subtraction and optical path length normalization to obtain a normalized near-infrared absorbance spectrum. The standardized near-infrared absorbance spectrum is subjected to wavelet transform denoising to filter out high-frequency noise components and retain low-frequency characteristic signals related to the organic components of coal. The denoised near-infrared absorbance spectrum is vector normalized to eliminate the light scattering effect caused by different sample particle size and packing density, and the denoised near-infrared spectral vector is obtained.

4. The online detection method for coal calorific value based on XRF and NIR characteristic-level fusion according to claim 1, characterized in that, Feature extraction is performed on the denoised near-infrared spectral vector to obtain a feature set of coal organic components, including: In the noise-reduced near-infrared spectral vector, the characteristic absorption wavelength range corresponding to the stretching vibration and combination vibration of water hydroxyl groups is located, and the integral value of the spectral absorbance within the characteristic absorption wavelength range corresponding to the stretching vibration and combination vibration of water hydroxyl groups is calculated as the water response characteristic value. Locate the characteristic absorption wavelength range corresponding to the stretching vibration of aliphatic hydrocarbon groups in coal, and calculate the ratio of the peak height to the half-peak width of the spectral absorption peak within the characteristic absorption wavelength range corresponding to the stretching vibration of aliphatic hydrocarbon groups, which is used as the characteristic value of aliphatic hydrocarbon content. Locate the characteristic absorption wavelength range corresponding to the stretching vibration of aromatic hydrocarbon groups in coal, and calculate the area of ​​the spectral absorption peak within the characteristic absorption wavelength range corresponding to the stretching vibration of aromatic hydrocarbon groups, which is used as the characteristic value of aromatic hydrocarbon content. The characteristic absorption wavelength ranges related to hydrogen-containing functional groups of hydroxyl, amino, carboxyl, and phenolic hydroxyl groups in coal are located. The second derivative of the spectral curve within the characteristic absorption wavelength range is extracted, and the zero-crossing position and amplitude of the second derivative are used as characteristic values ​​of functional group type and content. The moisture response characteristic value, the aliphatic hydrocarbon content characteristic value, the aromatic hydrocarbon content characteristic value, and the functional group type and content characteristic value are combined to form the characteristic set of coal organic components.

5. The online detection method for coal calorific value based on XRF and NIR characteristic-level fusion according to claim 1, characterized in that, Feature extraction is performed on the calibrated X-ray fluorescence spectral vector to obtain a set of coal elemental composition features, including: In the calibrated X-ray fluorescence spectral vector, the characteristic X-ray peaks corresponding to silicon, aluminum, calcium and iron are identified, and the net intensity of each characteristic peak is recorded as the intensity value of the main ash element. Identify the characteristic X-ray peaks corresponding to sulfur and phosphorus elements, calculate the ratio of their net intensity to the background intensity, and use it as the characteristic value of harmful elements. Identify the characteristic peaks of low atomic number elements associated with organic components in coal, including the characteristic peaks of chlorine and potassium, and extract the peak position shift and full width at half maximum (FWHM) of these characteristic peaks as morphological characteristic values ​​of organic-related elements. The intensity values ​​of the main ash elements, the characteristic values ​​of the harmful elements, and the morphological characteristic values ​​of the organically related elements are combined to form the coal elemental composition characteristic set.

6. The online detection method for coal calorific value based on XRF and NIR characteristic-level fusion according to claim 1, characterized in that, The feature-level fusion of the coal organic component feature set and the coal elemental component feature set is performed to generate a fused feature vector, including: The feature values ​​in the set of organic components of coal are standardized and scaled to unify their value range to the range of zero to one, thus obtaining a standardized organic feature vector. The characteristic values ​​in the coal elemental composition characteristic set are standardized and scaled to unify their value range to the range of zero to one, thus obtaining a standardized elemental feature vector. Based on the historical contribution of each feature in the coal calorific value prediction model, an organic weight is assigned to each feature in the standardized organic feature vector, and an element weight is assigned to each feature in the standardized element feature vector. The weighted standardized organic feature vector and the weighted standardized element feature vector are concatenated end to end to form an initial fusion vector; Principal component analysis is performed on the initial fusion vector to extract the first few principal components whose cumulative variance contribution rate exceeds a preset threshold, thus forming the dimensionality-reduced fusion feature vector.

7. The online detection method for coal calorific value based on XRF and NIR characteristic-level fusion according to claim 1, characterized in that, The improved partial least squares regression model adaptively adjusts weights based on laser morphology data of coal sample surfaces. Its working principle includes: Acquire laser morphology point cloud data synchronized with the current coal sample to be tested; The laser topography point cloud data is processed to calculate the topography feature parameters that reflect the surface morphology of the coal sample. The topography feature parameters include surface roughness, average particle projected area, and particle shape factor. During the latent variable extraction process of the improved partial least squares regression model, the fused feature vector and the morphological feature parameters are input together. When calculating each latent variable, the model dynamically adjusts the weight coefficients of each feature in the fused feature vector based on the morphological feature parameters. The principle for adjusting the weight coefficients is: for samples with large surface roughness or irregular particle shape, reduce the weight of features sensitive to spectral signal intensity and increase the weight of features sensitive to spectral shape and relative ratio. The model recalculates the covariance matrix between samples based on the adjusted feature weights, and extracts latent variables and calculates regression coefficients for partial least squares regression based on this matrix, ultimately outputting a calorific value prediction that is insensitive to differences in the morphology of coal samples.

8. The online detection method for coal calorific value based on XRF and NIR characteristic-level fusion according to claim 7, characterized in that, The laser-generated point cloud data is processed to calculate morphological characteristic parameters reflecting the surface morphology of the coal sample, including: Plane fitting is performed on the laser topography point cloud data to remove the reference plane of the conveyor belt or the bottom of the sample container; Calculate the standard deviation of the height values ​​of all points after removing the reference plane, and use it as the surface roughness. Clustering and segmenting of the point cloud to identify point cloud clusters of individual coal particles, calculating the projected area of ​​each point cloud cluster on the horizontal plane, and taking the average of the projected areas of all particles as the average particle projected area. For each identified coal particle point cloud cluster, calculate the aspect ratio of its minimum bounding rectangle, and calculate the ratio of the surface area of ​​the coal particle point cloud cluster to its volume. The product of the aspect ratio and the ratio is used as the shape factor of the coal particle corresponding to the coal particle in the coal particle point cloud cluster. The average value of the shape factors for all coal particles is taken as the overall particle shape factor.

9. The online detection method for coal calorific value based on XRF and NIR characteristic-level fusion according to claim 7, characterized in that, When calculating each latent variable, the model dynamically adjusts the weight coefficients of each feature in the fused feature vector based on the morphological feature parameters, including: Obtain the morphological characteristic parameters corresponding to the current coal sample, including surface roughness, average particle projected area, and particle shape factor; The morphological feature parameters are input into a preset weight mapping function, which is trained based on historical data and is used to output the initial weight adjustment factor corresponding to each original feature in the fused feature vector. Identify all features in the fused feature vector that are sensitive to the absolute intensity of the spectral signal, including the net intensity of elemental feature peaks in the XRF feature set and features in the NIR feature set based on absorbance integral value and peak height. Identify all features in the fused feature vector that are sensitive to spectral shape and relative ratios, including the feature peak intensity ratio, peak position shift and full width at half maximum (FWHM) in the XRF feature set, and features in the NIR feature set based on the second derivative of the absorption spectrum, peak area and peak width ratio. For the feature that is sensitive to the absolute intensity of the spectral signal, its initial weight adjustment factor is multiplied by an intensity weight attenuation coefficient based on surface roughness and particle shape factor to obtain its final dynamic weight coefficient, wherein the intensity weight attenuation coefficient is negatively correlated with surface roughness and particle shape factor. For the features that are sensitive to spectral shape and relative ratio, their initial weight adjustment factor is multiplied by a shape weight enhancement coefficient based on surface roughness and particle shape factor to obtain their final dynamic weight coefficient, wherein the shape weight enhancement coefficient is positively correlated with surface roughness and particle shape factor. The calculated final dynamic weight coefficients are normalized to ensure that the sum of all weight coefficients is 1; The normalized dynamic weight coefficients are multiplied one by one by each feature value in the fusion feature vector corresponding to the current latent variable calculation process to obtain a weighted feature vector, which is used to calculate the score of the current latent variable.

10. The online detection method for coal calorific value based on XRF and NIR characteristic-level fusion according to claim 1, characterized in that, The method also includes dynamic compensation for the predicted calorific value: Real-time monitoring of ambient temperature and humidity data during coal sample transportation; Establish a compensation relationship model between ambient temperature and humidity and the volatility of surface moisture in coal; After the improved partial least squares regression model outputs the predicted value of coal calorific value, the predicted value and the current environmental temperature and humidity data are input into the compensation relationship model. The compensation relationship model outputs a heat compensation amount based on environmental conditions; The predicted value of calorific value is added to the calorific value compensation amount to obtain the final coal calorific value test result after environmental dynamic compensation.

Images