Bulk material detection method based on a granulometry-insensitive quantitative model
By constructing a particle size-insensitive quantitative model and filtering out particle size interference, the problem of decreased spectral analysis accuracy in wide-particle-size coal sample analysis was solved, achieving high-precision coal quality analysis and reducing hardware costs and failure rates.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING UNIV OF CHEM TECH
- Filing Date
- 2026-03-09
- Publication Date
- 2026-06-05
AI Technical Summary
When analyzing wide-particle-size coal samples in industrial settings, existing technologies exhibit drastic fluctuations in scattering coefficients due to variations in particle size and packing density. This leads to decreased accuracy in near-infrared spectroscopy analysis, failure of conventional partial least squares (PLS) models, and increased costs and failure rates due to the need for standardized hardware particle size distribution.
We adopt a granularity-insensitive quantitative model approach. By constructing a training dataset and using Gaussian process regression or support vector regression algorithms, we filter out granularity interference and establish a robust model that can adapt to the analysis of coal samples with wide granularity.
It enables improved coal quality analysis accuracy without crushing, reduces costs, adapts to coal type fluctuations and particle size changes, and improves analytical precision.
Smart Images
Figure CN122157902A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of chemometrics and spectroscopic analysis, and specifically to a method for detecting bulk materials based on a particle size-insensitive quantitative model. Background Technology
[0002] Near-infrared spectroscopy (NIRS) offers the advantages of being non-destructive and rapid in coal quality analysis. Based on diffuse reflectance theory (Kubelka-Munk theory), spectral absorbance (…) ) is not only related to the chemical absorption coefficient of the sample ( k It is related to ) and also to the scattering coefficient ( s Closely related to ).
[0003] In traditional laboratory analysis, coal samples are ground into uniform powders <0.2 mm, and the scattering coefficient... The absorbance is essentially constant, therefore The scattering coefficient is linearly correlated with chemical composition. However, in industrial settings (such as belt conveyors), coal samples are typically wide-particle-size (0-50 mm mixed). In this case, the scattering coefficient... With drastic fluctuations in particle size and packing density, this physical interference is often 1-2 orders of magnitude larger than the chemical absorption signal, causing the conventional partial least squares (PLS) model to completely fail.
[0004] Existing technologies typically attempt to unify particle size through complex hardware (such as fine crushers), but this increases system cost and failure rate. Summary of the Invention
[0005] To address the problems and needs existing in the background technology, the present invention aims to provide a method for detecting bulk materials based on a particle size-insensitive quantitative model. This method is a robust model construction approach that enables near-infrared spectrometers to directly adapt to coal samples with wide particle sizes. From the perspective of data processing and algorithm modeling, this method improves the accuracy of quantitative analysis by mathematically "filtering out" particle size interference, and is applicable to the analysis of other bulk materials with wide particle sizes, such as iron ore, tobacco, and grain.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: I. A method for detecting bulk materials based on a particle size-insensitive quantitative model S1: Obtain a sample set of bulk materials, as well as the standard properties and original near-infrared diffuse reflectance spectra of each sample, and perform spectral preprocessing to obtain the preprocessed spectra; S2: Construct a training dataset based on the sample types and the preprocessed spectra and standard properties of each sample in the bulk material sample set; construct a granularity-insensitive quantitative model based on this training dataset; S3: Obtain the original near-infrared diffuse reflectance spectrum of the sample to be tested and process it to obtain the corrected spectrum. Input the corrected spectrum into the particle size insensitive quantitative model, and the model outputs the prediction result.
[0007] S2 further includes: The classification model is trained based on the type of sample and the preprocessed spectra of each sample in the bulk material sample set to obtain a well-trained classification model.
[0008] In step S3, the original near-infrared diffuse reflectance spectrum of the sample to be tested is acquired and processed to obtain the corrected spectrum, including: The original near-infrared diffuse reflectance spectrum is corrected based on the average spectrum corresponding to the sample type of the sample to be tested, and the corresponding corrected spectrum is obtained.
[0009] The granularity-insensitive quantitative model includes Gaussian process regression or support vector regression.
[0010] S3 further includes: The original near-infrared diffuse reflectance spectrum of the sample to be tested is input into the trained classification model to obtain the sample category corresponding to the current sample to be tested.
[0011] In step S2, a training dataset is constructed based on the sample type and the preprocessed spectra and standard properties corresponding to each sample in the bulk material sample set, including: Based on the sample type and the preprocessed spectra corresponding to each sample in the bulk material sample set, an average spectrum corresponding to each sample type is generated; then, the preprocessed spectra of the same sample type are corrected using the average spectrum to obtain the corresponding corrected spectra; finally, a training dataset is constructed by combining the standard properties of each sample and the corrected spectra.
[0012] II. A bulk material detection device based on a particle size-insensitive quantitative model The data input unit is used to acquire a sample set of bulk materials and the standard properties and raw near-infrared diffuse reflectance spectra of each sample. The spectral preprocessing unit is used to preprocess the raw near-infrared diffuse reflectance spectrum. The average spectrum calculation unit is used to generate the average spectrum corresponding to each sample type based on the sample type and the preprocessed spectrum corresponding to each sample in the bulk material sample set. The spectral correction unit is used to correct the preprocessed spectra of the same sample category using the average spectrum of each sample category to obtain the corresponding corrected spectrum; The property prediction unit is used to construct a granularity-insensitive quantitative model based on the corrected spectrum and standard properties. A classification unit is used to store classification models and to predict sample categories using those models. The model training unit is used to train the classification model using the types of samples and the preprocessed spectra of each sample in the bulk material sample set.
[0013] III. A computer device The computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the bulk material detection method based on a particle size insensitive quantitative model.
[0014] IV. A computer-readable storage medium The medium stores a computer program, which, when executed by a processor, implements the steps of the bulk material detection method based on a particle size-insensitive quantitative model.
[0015] Compared with existing methods, the beneficial effects of the present invention are as follows: In the modeling stage, this invention artificially creates samples with "unchanged chemical values but drastic changes in particle size" to allow the model to "witness" the spectral changes of the same coal type under different physical states.
[0016] This invention establishes a classification and discrimination model based on the sample categories of materials, and at the same time adopts extended multivariate scattering correction according to the sample categories to filter out the variations in the spectrum that are not related to the chemical values (i.e. variations caused by particle size roughness) from the matrix. This invention uses a regression algorithm to establish a quantitative analysis model of properties, which more accurately captures the quantitative relationship between spectra and properties.
[0017] This invention achieves high-precision measurement without the need for fine crushing of coal samples, saving material and labor costs; it has strong anti-interference capabilities, adapts to fluctuations in coal type and changes in coal sample particle size, and improves the accuracy of coal quality analysis. Attached Figure Description
[0018] Figure 1 This is a flowchart of the method of the present invention.
[0019] Figure 2 The images show the spectra of all sub-samples of different coal types at different particle sizes.
[0020] Figure 3 To Figure 2 The projection of all samples onto a 2D plane after PCA.
[0021] Figure 4 This is the confusion matrix of the LDA classifier's classification predictions for the test set.
[0022] Figure 5 The projection of the training and test sets onto a 2D plane after modeling and predicting using an LDA classifier.
[0023] Figure 6 The projection of the training and test sets onto a 2D plane after EMSC granular correction and PCA.
[0024] Figure 7 The results of carbon fixation for the GPR prediction test set.
[0025] Figure 8 The results for predicting the volatile matter content of the GPR test set.
[0026] Figure 9 The result of predicting grayscale values for the GPR test set.
[0027] Figure 10 The results of GPR prediction for the moisture content of the test set. Detailed Implementation
[0028] To enable those skilled in the art to better understand the present invention and to more clearly define the scope of protection of the present invention, the present invention will be described in detail below with reference to certain specific embodiments and accompanying drawings. It should be noted that the following are only some specific embodiments of the present invention, and are merely a part of the embodiments of the present invention. The specific and direct descriptions of related structures are only for the convenience of understanding the present invention, and the specific features do not necessarily or directly limit the scope of the present invention. Conventional selections and substitutions made by those skilled in the art under the guidance of the present invention, as well as reasonable arrangements and combinations of several technical features under the guidance of the present invention, should all be considered within the scope of protection of the present invention.
[0029] like Figure 1 As shown, the bulk material detection method based on a particle size-insensitive quantitative model proposed in this invention includes the following steps: S1: Obtain a sample set of bulk materials, along with the standard properties and raw near-infrared diffuse reflectance spectra for each sample. After spectral preprocessing, obtain the preprocessed spectra. The raw near-infrared diffuse reflectance spectra are used to characterize chemical and strong physical scattering information. Standard properties are those obtained in the laboratory using national standards for sample analysis. For example, for coal, standard properties include industrial analytical indicators (moisture, volatile matter, ash, fixed carbon), calorific value, etc.
[0030] In bulk material sample sets, each sample contains multiple particle size distributions. The specific particle size distributions must be set according to the actual situation, that is, depending on the scenario in which the method is applied. For example, for any coal sample with a specific chemical composition (moisture, ash, etc.), it is physically sieved into at least three different particle size states: fine powder (<1mm), fine-grained (<3mm), medium-grained (3-13mm), coarse-grained (13-50mm), in-situ rough state, and in-situ smooth state. When constructing the mixed state, the particle size is recombined according to proportion and mass. All coal samples in the single-size state and the recombined particle size state form a subsample set; all of the above subsample sets are assigned the same property values and included in the model training, forcing the model to learn the spectral commonalities of the same chemical composition under different physical forms.
[0031] Optionally, the bulk material sample set may include one or more sample types.
[0032] In this embodiment of the invention, 12 types of coal were collected from different mining areas, namely 0.3 coking coal, anthracite coal grade 1, anthracite coal grade 2, coking coal, fat coal, gas coal, gas-fat coal, lean coal, lean-thin coal, lignite coal grade 1, lignite coal grade 2, and thin coal. Eleven samples of different particle sizes were prepared for each type of coal, including nine particle sizes and two roughnesses. The nine particle sizes were 0.045 mm, 0.074 mm, 0.1 mm, 0.21 mm, 0.5 mm, 1 mm, 2.5 mm, 4.75 mm, and 8 mm. The two roughnesses were a rough block with in-situ fractures and a smooth block with a polished surface roughness of less than 0.3 mm. The fixed carbon (%), volatile matter (%), ash (%), and moisture (%) of each coal were analyzed using standard methods on an air-dried basis, for a total of four properties.
[0033] In one feasible implementation, spectral preprocessing includes: Abnormal spectra caused by gaps between coal sample particles, shadows, or specular reflections are removed from the original near-infrared diffuse reflectance spectrum.
[0034] Specifically, an absorbance energy threshold range is set, and the average absorbance of each spectrum is calculated. If the average absorbance of a spectrum is lower than the lower limit of the absorbance energy threshold, it is identified as a "gap / shadow spectrum" and discarded. If the average absorbance of a spectrum is higher than the upper limit of the absorbance energy threshold, it is identified as a "supersaturated / specular reflection spectrum" and discarded.
[0035] Near-infrared diffuse reflectance spectra were collected for all 12 types of coal and 11 particle sizes in the embodiments of the present invention, and the effectiveness was verified. Figure 2 The spectrum of all 132 samples of 12 types of coal at 11 different particle sizes is shown, which shows that different particle sizes have a great influence on the spectral morphology. Figure 3 The projections of 132 samples after principal component analysis (PCA) on the feature plane formed by the first and second principal component axes, labeled by coal type, are shown. It can be seen that the physical morphology of the samples seriously affects the identification of spectrochemical information, and the coal types overlap with each other.
[0036] S2: Construct a training dataset based on the sample types and the preprocessed spectra and standard properties of each sample in the bulk material sample set; construct a granularity-insensitive quantitative model based on this training dataset; In one feasible implementation, a training dataset is constructed based on the sample type and the preprocessed spectra and standard properties corresponding to each sample in the bulk material sample set, including: Based on the sample type and the preprocessed spectra corresponding to each sample in the bulk material sample set, an average spectrum is generated for each sample type. Then, the preprocessed spectra of the same sample type are corrected using the average spectrum to obtain the corresponding corrected spectra, so as to filter out physical scattering information unrelated to chemical composition in the spectral matrix. Finally, a training dataset is constructed by combining the standard properties of each sample and the corrected spectrum.
[0037] S2 also includes: The classification model is trained based on the sample type and the preprocessed spectra corresponding to each sample in the bulk material sample set to obtain a trained classification model. Optionally, the classification model is a linear discriminant classifier (LDA). LDA is a classic supervised learning algorithm whose core objective is to "maximize inter-class distance and minimize intra-class distance". LDA utilizes two matrices when finding the optimal projection direction: the intra-class scatter matrix (...). The inter-class scatter matrix measures how "scattered" the data is within each category; the smaller the better. The larger the better. The mathematical goal of LDA is to find a projection direction such that... The ratio is the largest.
[0038] S3: Obtain and process the original near-infrared diffuse reflectance spectrum of the sample to be tested to obtain the corrected spectrum. Input the corrected spectrum into the particle size-insensitive quantitative model, and the model outputs the prediction result. If the original near-infrared diffuse reflectance spectrum of the sample to be tested is an anomalous spectrum, the spectral information is reacquired.
[0039] In one feasible implementation, the original near-infrared diffuse reflectance spectrum of the sample to be tested is obtained and processed to obtain a corrected spectrum, including: Based on the average spectrum corresponding to the sample type of the sample to be tested, the original near-infrared diffuse reflectance spectrum is corrected using extended multivariate scattering correction (EMSC) to obtain the corresponding corrected spectrum.
[0040] Granularity-insensitive quantitative models include Gaussian process regression (GPR) or support vector regression (SVR).
[0041] Granularity-insensitive quantitative models are obtained based on one or more standard properties, meaning that a granularity-insensitive quantitative model can be used to predict one or more standard properties.
[0042] S3 also includes: The original near-infrared diffuse reflectance spectrum of the sample to be tested is input into the trained classification model to obtain the sample category corresponding to the current sample to be tested.
[0043] In this embodiment, 132 samples were randomly divided into a training set and a test set in a 7:3 ratio; a classification model was built using LDA for 92 training set samples according to coal type, and classification and discrimination were performed on 40 test samples. Figure 4 The confusion matrix for the predicted test set is shown. Figure 5 The training set (marked with circles) and the test set (marked with asterisks) are projected onto a 2D plane after being modeled and predicted using the LDA classifier. It can be seen that all test samples are classified into the correct coal type category.
[0044] In one feasible implementation, particle size effects are removed using Extended Multiplicative Scatter Correction (EMSC) based on coal type. EMSC is an advanced preprocessing algorithm for spectral data, an improved version of the classic multiplicative scattering correction (MSC). It primarily eliminates baseline drift and nonlinear effects caused by physical scattering (optical path difference, uneven particle size distribution) in the spectrum, thus preserving pure chemical absorption information. In spectral analysis (especially of solid powders, turbid liquids, or biological tissues), the spectral signal typically contains two parts: chemical absorption (desired) and physical scattering (to be removed). Chemical absorption consists of characteristic peaks determined by the sample composition; physical scattering includes additive effects—overall baseline shifts, multiplicative effects—amplification or reduction of overall spectral intensity (optical path change), and wavelength dependence—baseline tilting or curvature (which is difficult for ordinary MSC to handle). By introducing more parameters, EMSC can handle tilted and curved baselines and even remove known interferences. The core idea of EMSC is to fit a mathematical model to each sample spectrum, assuming the spectrum to be corrected is... EMSC established the following regression model: In the formula, It is the reference spectrum (usually the average spectrum of all samples). It is a multiplicative factor, representing the proportional difference in optical path or concentration. It is an additive constant (baseline translation). It is a linear term (baseline tilt / slope). It is a quadratic term (baseline curvature); Perform multiple linear regression on each sample and calculate the coefficients. Then, the physical effects are subtracted using the formula, and the chemical information is normalized:
[0045] The average spectrum of the training set samples in this embodiment of the invention is calculated for each of the 12 different coal types. The average spectrum is used as the reference spectrum for each coal type, and then EMSC correction is performed. Figure 6 The diagram shows the distribution of samples after EMSC correction of the training set (marked with circles) and PCA, and after LDA identification of the coal type of the test set (marked with asterisks), EMSC correction is performed using the average spectrum of the class, and then the samples are projected onto the feature plane of the training set PCA. It can be seen that samples of the same coal type are closely clustered together, while samples of different coal types are separated.
[0046] In one feasible implementation, Gaussian Process Regression (GPR) is used to correlate the spectral properties of coal samples after removing particle size effects with their analytical properties. Gaussian Process Regression (GPR) is a non-parametric regression model based on Bayesian statistics, a powerful tool for small-sample learning and uncertainty estimation in machine learning. GPR comprises three core elements: Gaussian distribution, kernel function, and posterior distribution. GPR assumes that the joint distribution of any finite number of data points follows a multivariate Gaussian distribution; the kernel function... The "similarity" between two points is defined. and If they are very close together and the kernel function value is very large, then the model considers them to be... The values should also be very close. RBF (Gaussian kernel) is the most commonly used kernel function, which produces a smooth and infinitely differentiable function. There are also Matern kernel and Periodic kernel. Before looking at the data, countless lines could be regression lines. The prior distribution usually assumes that the regression line has a mean of 0 and a large variance. The observed data points are like "nails". All lines in the posterior distribution must pass through (or be very close to) these nails, which narrows down the range of possible functions and forms the final prediction band.
[0047] In this embodiment of the invention, the training set after removing the granularity effect is trained using GPR, and then the test set after removing the granularity effect is predicted. The kernel function used is the RBF Gaussian kernel, and the four properties of fixed carbon, volatile matter, ash content and dry basis moisture are modeled and predicted. Figure 7 , Figure 8 , Figure 9 and Figure 10 The prediction results for each property are shown separately. The training and test sets were randomly split five times, and the average error of GPR predictions for the four properties is presented. The values are fixed carbon 0.08%, volatile matter 0.06%, ash 0.03%, and moisture 0.003%, all close to 0, with a coefficient of determination for each property. The values are all close to 1, which means that the data in this embodiment has basically achieved the most ideal result. Although each type of coal has different particle sizes and different spectra, its properties are the same. Therefore, under ideal conditions, the test set error should be 0. and The calculation formula is as follows:
[0048] in, It is the sample size. It is the true property value of the i-th coal sample. It is the predicted value of the i-th coal sample.
[0049]
[0050] in, It is the average of all true values. These are fluctuations that the model cannot explain (i.e., prediction errors). It is the original fluctuation of the data itself (a manifestation of variance). The closer the model is to 1, the more accurate it is.
[0051] For comparison, Table 1 lists the results using other spectral preprocessing methods and regression methods. The training and test sets were randomly divided five times, and the table shows the average of the five divisions. Preprocessing methods included Extended Multivariate Scattering Correction (EMSC) and Savitzky-Golay Second-Order Differentiation (SGD2). Regression methods included Gaussian Process Regression (GPR), Support Vector Regression (SVR), Partial Least Squares (PLS), Extreme Learning Machine (ELM), and XGBoost. Each regression method used cross-validation to select optimal parameter values. It can be seen that EMSC+GPR has the highest prediction accuracy, followed by EMSC+SVR. However, the SVR algorithm requires sequential single-property regression and searching for the optimal parameters for the sample set within a grid space of hundreds of parameter combinations, which can be time-consuming. Next is the ELM algorithm, which requires determining the appropriate number and type of neurons. Other algorithms have larger errors.
[0052] Table 1. RMSE of test sets for different preprocessing functions + regression algorithms
[0053] The present invention proposes a bulk material detection device based on a particle size-insensitive quantitative model, comprising: The data input unit is used to acquire a sample set of bulk materials and the standard properties and raw near-infrared diffuse reflectance spectra of each sample. The spectral preprocessing unit is used to preprocess the raw near-infrared diffuse reflectance spectrum. The average spectrum calculation unit is used to generate the average spectrum corresponding to each sample type based on the sample type and the preprocessed spectrum corresponding to each sample in the bulk material sample set. The spectral correction unit is used to correct the preprocessed spectra of the same sample category using the average spectrum of each sample category to obtain the corresponding corrected spectrum; The property prediction unit is used to construct a granularity-insensitive quantitative model based on the corrected spectrum and standard properties. The bulk material detection device based on a particle size-insensitive quantitative model proposed in this invention also includes: A classification unit is used to store classification models and to predict sample categories using those models. The model training unit is used to train the classification model using the types of samples and the preprocessed spectra of each sample in the bulk material sample set.
[0054] This invention proposes a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of a bulk material detection method based on a particle size insensitive quantitative model.
[0055] This invention proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a bulk material detection method based on a particle size-insensitive quantitative model.
Claims
1. A method for detecting bulk materials based on a particle size-insensitive quantitative model, characterized in that, Includes the following steps: S1: Obtain a sample set of bulk materials, as well as the standard properties and original near-infrared diffuse reflectance spectra of each sample, and perform spectral preprocessing to obtain the preprocessed spectra; S2: Construct a training dataset based on the sample types and the preprocessed spectra and standard properties of each sample in the bulk material sample set; construct a granularity-insensitive quantitative model based on this training dataset; S3: Obtain the original near-infrared diffuse reflectance spectrum of the sample to be tested and process it to obtain the corrected spectrum. Input the corrected spectrum into the particle size insensitive quantitative model, and the model outputs the prediction result.
2. The method for detecting bulk materials based on a particle size-insensitive quantitative model according to claim 1, characterized in that, S2 further includes: The classification model is trained based on the type of sample and the preprocessed spectra of each sample in the bulk material sample set to obtain a well-trained classification model.
3. The method for detecting bulk materials based on a particle size-insensitive quantitative model according to claim 1, characterized in that, In step S3, the original near-infrared diffuse reflectance spectrum of the sample to be tested is acquired and processed to obtain the corrected spectrum, including: The original near-infrared diffuse reflectance spectrum is corrected based on the average spectrum corresponding to the sample type of the sample to be tested, and the corresponding corrected spectrum is obtained.
4. The method for detecting bulk materials based on a particle size-insensitive quantitative model according to claim 1, characterized in that, The granularity-insensitive quantitative model includes Gaussian process regression or support vector regression.
5. The method for detecting bulk materials based on a particle size-insensitive quantitative model according to claim 2, characterized in that, S3 further includes: The original near-infrared diffuse reflectance spectrum of the sample to be tested is input into the trained classification model to obtain the sample category corresponding to the current sample to be tested.
6. The method for detecting bulk materials based on a particle size-insensitive quantitative model according to claim 1, characterized in that, In step S2, a training dataset is constructed based on the sample type and the preprocessed spectra and standard properties corresponding to each sample in the bulk material sample set, including: Based on the sample type and the preprocessed spectra corresponding to each sample in the bulk material sample set, an average spectrum corresponding to each sample type is generated; then, the preprocessed spectra of the same sample type are corrected using the average spectrum to obtain the corresponding corrected spectra; finally, a training dataset is constructed by combining the standard properties of each sample and the corrected spectra.
7. A bulk material detection device based on a particle size-insensitive quantitative model, characterized in that, include: The data input unit is used to acquire a sample set of bulk materials and the standard properties and raw near-infrared diffuse reflectance spectra of each sample. The spectral preprocessing unit is used to preprocess the raw near-infrared diffuse reflectance spectrum. The average spectrum calculation unit is used to generate the average spectrum corresponding to each sample type based on the sample type and the preprocessed spectrum corresponding to each sample in the bulk material sample set. The spectral correction unit is used to correct the preprocessed spectra of the same sample category using the average spectrum of each sample category to obtain the corresponding corrected spectrum; The property prediction unit is used to construct a granularity-insensitive quantitative model based on the corrected spectrum and standard properties.
8. The bulk material detection device based on a particle size-insensitive quantitative model according to claim 7, characterized in that, Also includes: A classification unit is used to store classification models and to predict sample categories using those models. The model training unit is used to train the classification model using the types of samples and the preprocessed spectra of each sample in the bulk material sample set.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the bulk material detection method based on a particle size insensitive quantitative model as described in any one of claims 1 to 6.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the bulk material detection method based on a particle size insensitive quantitative model as described in any one of claims 1 to 6.