A bulk coal pile density self-adaptive modeling and reserve metering method based on coal type clustering and multi-factor regression

Abstract

The application relates to a coal storage measurement and intelligent sensing field, and relates to a bulk coal pile density self-adaptive modeling and reserve measurement method based on coal type clustering and multi-factor regression. In order to solve the problems that the existing coal inventory measurement method can only obtain volume, a single fixed pile density conversion is adopted, precision is insufficient, and coal pile density is significantly differentiated due to different coal types, particle sizes and moisture contents, and it is difficult to describe the problems by using a unified coefficient. The application takes three-dimensional laser radar volume measurement data and a real weight of a measurement device calibrated as samples, clusters and groups the samples according to coal types and physical characteristics, establishes a pile density regression model in each group by taking volume, moisture content, layer height and ash content as factors, inverses the pile density by the real-time volume during normal measurement and calculates the inventory weight, simultaneously, the new samples are fed back to the model to realize parameter updating. The self-adaptive change of the pile density with the coal type and the physical property is realized, and the inventory weight measurement precision is improved. The application is used for automatic measurement of the inventory weight of a coal storage.

Description

Technical Field

[0001] This invention relates to the field of coal storage measurement and intelligent sensing technology, specifically to an adaptive modeling and storage measurement method for bulk coal based on coal type clustering and multi-factor regression. This method is applicable to the continuous, automated, and high-precision measurement of the weight of bulk coal in raw coal silos, refined coal silos, middlings and coal slime silos in coal washing plants, as well as in coal storage and distribution yards. Background Technology

[0002] Coal is a vital energy source for the national economy. Accurate measurement of coal inventory in warehouses or storage areas is crucial for production scheduling, sales settlement, and coal blending decisions, throughout the processes of washing, processing, storage, and distribution. How to quickly, continuously, and accurately obtain the weight of bulk coal in warehouses is a key technical issue in coal storage management.

[0003] Current inventory measurement methods mainly rely on manual stacking, cumulative calculation from inbound and outbound measurement ledgers, or single-point ultrasonic or single-point radar for material level measurement. These methods have sparse measurement points, making it difficult to reflect the overall shape and local undulations of the material surface; the ledger cumulative method gradually becomes distorted due to long-term error accumulation, spillage, and moisture changes, making it difficult to achieve continuous and automatic online measurement.

[0004] 3D lidar can acquire planar, high-coverage 3D point clouds in a short time, depicting the overall shape and local details of the coal surface, making the measurement of coal volume in the silo more reliable and stable. However, 3D lidar only directly acquires volume data; the volume must be converted to bulk density (mass per unit volume) to obtain the silo weight. The basic relationship is as follows:

[0005]

[0006] In the formula, V is the volume of coal in the storage silo measured by three-dimensional lidar, ρ is the bulk density, and M is the weight of the stored coal.

[0007] The bulk density of coal fluctuates significantly more than that of agricultural products such as grains. Its value primarily depends on physical properties such as coal type (raw coal, clean coal, middlings, coal slime), particle size distribution, moisture content, and ash content, and varies with the self-weight compaction caused by the height of the coal bed. Especially in coal washing plants, frequent coal type changes and the high and fluctuating moisture content of washed clean coal and coal slime can result in a bulk density difference of 5% to 15% for the same volume. Existing methods often use a single empirical bulk density or a fixed conversion factor for linear conversion, failing to capture the significant differences in bulk density with coal type and physical properties, leading to insufficient accuracy in inventory weight calculations. Summary of the Invention

[0008] The purpose of this invention is to address the problems of existing coal inventory measurement methods, which can only obtain volume, suffer from insufficient accuracy due to the use of a single fixed bulk density conversion, and are difficult to describe with a unified coefficient because coal bulk density varies significantly with different coal types, particle sizes, and moisture contents. Therefore, this invention provides an adaptive modeling and reserve measurement method for bulk coal bulk density based on coal type clustering and multi-factor regression.

[0009] The technical solution of this invention is: an adaptive modeling and reserve measurement method for bulk coal bulk density based on coal type clustering and multi-factor regression, comprising the following steps:

[0010] Step 1: Use the volume measurement data of the three-dimensional lidar and the actual weight calibrated by the measuring device as a sample, i.e., the "volume-weight" sample;

[0011] Step 2: Cluster and group the samples according to coal type and particle size distribution as physical property labels;

[0012] Step 3: Within each group, establish a bulk density regression model using volume, moisture content, bed height, and ash content as multiple factors;

[0013] Step 4: Invert the bulk density from the real-time volume and calculate the inventory weight.

[0014] Furthermore, the three-dimensional lidar in step one is deployed on the top of the coal bunker or above the material surface, with one or more units.

[0015] Furthermore, in step one, when the 3D lidar performs volume measurement, it first scans the material surface to obtain a 3D point cloud covering the entire material surface; after preprocessing such as coordinate registration, noise reduction and reference surface calibration, it then combines the silo geometric model to perform volume integration or mesh summation on the space between the material surface and the reference surface to obtain the volume V of the coal in the silo at the current moment.

[0016] Its volume integral is expressed as:

[0017]

[0018] In the formula, z s (x,y) represents the material surface elevation, z0(x,y) represents the datum elevation of the bottom of the silo, and Ω represents the horizontal projection area of ​​the silo.

[0019] Furthermore, the clustering and grouping process in step two is as follows:

[0020] For each "volume-weight" sample, its physical property feature vector is extracted, including coal type label, particle size distribution index, moisture content w, and ash content. Based on this, the sample is divided into several groups G1, G2, ..., GK, so that samples in the same group are similar in physical properties and the differences between groups are significant.

[0021] Furthermore, the clustering in step two can be performed manually by directly grouping coal types, or automatically using the K-measns algorithm, Gaussian mixture model, or unsupervised clustering method; for each sample, its measured bulk density is directly obtained by the following formula:

[0022] M is mass, V is volume, and density is calculated.

[0023] Furthermore, the process of establishing the multi-factor heap density regression model in step three is as follows:

[0024] Within each group Gj, using volume V, moisture content w, and bed height h as multiple factors as inputs, and measured bulk density ρ or stock weight M as the monitoring target, a bulk density regression model specific to that group is established:

[0025] .

[0026] Furthermore, the packing density regression model in step three adopts a multiple linear or low-order polynomial regression form:

[0027]

[0028] In the formula, β0 is the basic bulk density term, and β1, β2, and β3 are the influence coefficients of volume, moisture content, and bed height on bulk density, respectively. Among them, the moisture content coefficient β2 is negative, reflecting the physical law that the increase in moisture content leads to a decrease in the effective coal mass per unit volume.

[0029] Furthermore, the regression coefficients for each group were estimated using the least squares method:

[0030] .

[0031] Furthermore, the regression model employs spline regression, Gaussian process regression, support vector regression, or neural network learning models.

[0032] Furthermore, the process of real-time volumetric inversion of bulk density and calculation of inventory weight is as follows:

[0033] During the normal measurement phase, based on the current coal type and physical properties, its group Gj is determined, the corresponding regression model is called to retrieve the bulk density ρj, and the volume V measured in real time by three-dimensional lidar is combined to calculate the inventory weight using the following formula:

[0034]

[0035] Each time a stocktaking operation with the actual weight calibrated by the metering device is completed, a new sample is obtained, which is assigned to the corresponding group and the regression parameters of the group are incrementally updated. When a new coal type or significant drift in physical properties occurs, re-clustering is triggered, thus forming a closed loop of "measurement-grouping-regression-feedback-update", which continuously improves the bulk density model and measurement accuracy of each group as the operation progresses.

[0036] Compared with the prior art, the present invention has the following advantages:

[0037] 1. This invention proposes an adaptive modeling method for bulk coal bulk density that involves "first grouping by coal type and then performing multi-factor regression within the group". This method specifically addresses the problem of significant differentiation in coal bulk density due to differences in coal type and physical properties, and differs from the conventional approach of uniformly fitting a single model to all samples.

[0038] 2. This invention incorporates physical property factors such as moisture content and bed height into the regression along with volume, thereby achieving a multi-factor quantitative characterization of bulk density as a function of physical properties. It is particularly suitable for the working conditions of large fluctuations in the moisture content of clean coal and coal slime in coal washing plants.

[0039] 3. This invention uses a three-dimensional lidar with explosion-proof certification to acquire point clouds of high-coverage material surfaces and perform volume integration. The bulk density is then inverted using a grouped regression model to calculate the inventory weight, achieving continuous, automated, and high-precision measurement.

[0040] 4. This invention constructs a closed loop of "measurement-grouping-regression-feedback-update", supports incremental updates of parameters within groups and re-clustering triggered by new coal types, and enables the model to be continuously optimized as data accumulates.

[0041] 5. This invention establishes a coal yard-specific bulk density model through clustering and grouping followed by multi-factor regression within each group. This invention enables continuous, automated, and high-precision measurement of coal inventory weight within the storage facility. The main advantages are as follows:

[0042] (1) By replacing a single fixed coefficient with coal type clustering, the significant differences in bulk density of different coal types can be fundamentally adapted to, and the measurement accuracy can be significantly improved.

[0043] (2) The group of multifactor regression explicitly incorporates moisture content and bed height, which has good adaptability to high moisture and variable moisture conditions in washing plants;

[0044] (3) Preferably, a three-dimensional lidar with explosion-proof certification is selected to meet the explosion-proof safety requirements of the coal yard and support redundancy and cross-verification of multiple devices;

[0045] (4) The clustering method and regression form are not limited and can be flexibly selected according to the field data, with strong engineering applicability and scalability;

[0046] (5) It supports incremental updates and re-clustering of new coal types, and can be connected with existing metering systems such as truck scales and belt scales to form a closed loop, with accuracy continuously improving as it runs.

[0047] 6. The main performance indicators that this invention can achieve are as follows:

[0048] The mean absolute error (MAE), mean absolute percentage error (MAPE), maximum error ratio, and coefficient of determination (R²) of inventory weight measurement can be used for evaluation, with the actual weight measured by truck scales, belt scales, or inbound / outbound measurement systems as reference values. The coefficient of determination is defined as follows:

[0049]

[0050] In the formula, n is the sample size, and yi and Let be the reference value and the predicted value of the i-th sample, respectively. The average value is used as a reference. In a specific embodiment of the present invention, the mean absolute percentage error (MAPE) of the predicted inventory weight of each coal type group after grouping can be controlled within 1.0%, and the coefficient of determination R² is higher than 0.998, which meets the accuracy requirements of on-site inventory, inventory monitoring and coal blending scheduling in coal yards (the specific value depends on the equipment and on-site data). Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the system composition and three-dimensional lidar volume measurement of the present invention. Figure 2 This is a schematic diagram of the overall method flow of the present invention. Figure 3 This is a schematic diagram of coal type clustering and intragroup multifactor regression of the present invention. Figure 4 This is a graph showing the inventory weight measurement results of this invention. Detailed Implementation

[0052] It is important to note that in the application scenarios targeted by this invention, inventory measurement primarily serves static or quasi-static inventory estimation for inventory checks and coal blending scheduling, and does not require differentiation of dynamic operations such as warehousing and outbound processes. Therefore, this invention does not introduce correction mechanisms related to the stacking and reclaiming path, but rather focuses on resolving the core contradiction of "significant differentiation in bulk density due to different coal types and physical properties." The key to improving measurement accuracy lies in how to automatically identify and group different coal types and establish a bulk density model dominated by physical property factors within each group.

[0053] Specific Implementation Method 1: This implementation method, based on coal type clustering and multi-factor regression, includes the following steps:

[0054] Step 1: Use the volume measurement data of the three-dimensional lidar and the actual weight calibrated by the measuring device as a sample, i.e., the "volume-weight" sample;

[0055] Step 2: Cluster and group the samples according to coal type and particle size distribution as physical property labels;

[0056] Step 3: Within each group, establish a bulk density regression model using volume, moisture content, bed height, and ash content as multiple factors;

[0057] Step 4: Invert the bulk density from the real-time volume and calculate the inventory weight.

[0058] Specific implementation method two: In step one of this implementation method, the three-dimensional lidar is arranged on the top of the coal bunker or above the material surface, and the number is one or more.

[0059] As an optional solution, two or more three-dimensional radars of different models or systems can be deployed in parallel to collect volume data in order to perform cross-verification and select the main metrology equipment, thereby improving the reliability and redundancy of volume measurement.

[0060] The purpose of using one or more radars in this implementation is that multiple devices can simultaneously collect data for cross-validation, and the device with higher accuracy or more stable data can be selected as the main metering device. At the same time, when one device fails, the other devices can continue to work, avoiding measurement interruption, thereby improving the reliability and robustness of the entire metering system.

[0061] Specific implementation method three: In step one of this implementation method, when the three-dimensional lidar performs volume measurement, it first scans the material surface to obtain a three-dimensional point cloud covering the entire material surface; after preprocessing such as coordinate registration, noise reduction and reference surface calibration, it then combines the silo geometric model to perform volume integration or grid summation on the space between the material surface and the reference surface to obtain the volume V of the coal in the silo at the current moment.

[0062] Its volume integral is expressed as:

[0063]

[0064] In the formula, z s (x,y) represents the material surface elevation, z0(x,y) represents the datum elevation of the silo bottom (ground level), and Ω represents the horizontal projection area of ​​the silo body.

[0065] This implementation method concretizes the general "volume measurement data" in step one into an operable technical means, providing accurate volume input for subsequent clustering, regression modeling, and weight inversion. Specifically, by acquiring a full-coverage 3D point cloud, the undulations of the coal pile surface can be completely recorded, avoiding the problem that single-point measurements cannot reflect the overall shape of the material surface; coordinate registration is used to unify the coordinate system of multi-source point clouds, denoising is used to remove outliers caused by dust or equipment errors, and reference surface calibration is used to determine the zero potential surface elevation of the silo bottom or ground level. These three aspects together ensure the reliability of the basic data for volume calculation. The volume integral formula obtains the true volume by double integral of the difference between the material surface elevation and the reference surface elevation over the horizontal projection area of ​​the silo body, accumulating point by point. It does not require simplified assumptions about the shape of the coal pile and can adapt to arbitrary undulating material surface shapes. This limitation ensures the accuracy and feasibility of the volume data source in the entire method.

[0066] Specific Implementation Method Four: The clustering and grouping process in step two of this implementation method is as follows:

[0067] For each "volume-weight" sample, its physical property feature vector is extracted, including coal type label, particle size distribution index, moisture content w, and ash content. Based on this, the sample is divided into several groups G1, G2, ..., GK, so that samples in the same group are similar in physical properties and the differences between groups are significant.

[0068] The bulk density of coal in this embodiment varies significantly due to differences in coal type, particle size, moisture content, and ash content. If all samples are modeled uniformly without clustering, samples with different physical properties will interfere with each other, causing the regression model to fail to accurately depict the density variation pattern of any particular coal type. By clustering, samples with similar physical properties are grouped into the same group, making the coal type and physical properties relatively uniform within the group. This distinguishes coals of different types and physical properties, avoiding interference from density differences between different coal types in the modeling. This step, together with the subsequent intra-group multi-factor regression modeling, constitutes the core scheme of this application: "clustering first, then intra-group multi-factor regression," which is one of the key aspects that distinguishes this application from existing single fixed bulk density conversion methods.

[0069] Specific Implementation Method Five: In step two of this implementation method, clustering is performed manually by directly grouping according to coal type, or automatically using the K-measns algorithm, Gaussian mixture model, or unsupervised clustering method; for each sample, its measured bulk density is directly obtained by the following formula:

[0070] M is mass, V is volume, and density is calculated.

[0071] When the coal composition is relatively simple and the incoming material is stable, clustering can be quickly completed by manually grouping by coal type, which is simple to operate and low in cost. When coal type labels are missing or the coal composition is complex, unsupervised clustering methods such as K-measns algorithm and Gaussian mixture model can be used. Taking physical property feature vectors such as moisture content, ash content, and particle size distribution as input, the algorithm automatically discovers sample groups with similar physical properties and completes the grouping, without the need for pre-manual coal type division. This is especially suitable for coal storage and distribution yards with mixed coal types and large fluctuations in incoming material. The two methods can be flexibly selected according to the actual site conditions, giving the method stronger engineering adaptability and scalability. At the same time, for each sample, the measured bulk density is directly calculated by ρ=M / V. The purpose of this calculation is to transform the collected "volume-weight" paired samples into a "volume-physical property-bulk density" sample subset, providing a supervised target value for subsequent intra-group regression modeling, so that model training can be implemented.

[0072] Specific Implementation Method Six: Combination Figure 1 and Figure 2 This embodiment describes the process of establishing a multi-factor heap density regression model in step three of this embodiment as follows:

[0073] Within each group Gj, using volume V, moisture content w, and bed height h as multiple factors as inputs, and measured bulk density ρ or stock weight M as the monitoring target, a bulk density regression model specific to that group is established:

[0074] .

[0075] In this implementation method, within the same group, the coal type and physical properties are relatively homogeneous. At this point, the main influencing factors of bulk density converge to three quantifiable factors: volume (reflecting the degree of compaction), moisture content (reflecting the effect of moisture on the effective coal mass per unit volume), and bed height (reflecting the effect of self-weight compaction). By using these three factors as inputs and measured bulk density as the monitoring target for regression modeling, a quantitative mapping relationship between bulk density and multi-factor changes can be established within the group. This makes the bulk density no longer a fixed constant, but dynamically changes with volume, moisture content, and bed height. This step, together with the clustering grouping in step two, constitutes the core scheme of this invention: "clustering grouping first, then multi-factor regression within the group." Clustering grouping addresses the significant differences in bulk density between different coal types, while multi-factor regression within the group addresses the continuous fluctuations in bulk density within the same coal type due to changes in moisture content and compaction state. If multi-factor regression within the group is not performed and only a fixed density value is used after grouping, the density differences within the same coal type caused by different batch moisture contents and different storage locations cannot be reflected, and the measurement accuracy remains limited. This step enables the bulk density model to adapt to the actual working conditions of large fluctuations in the moisture content of clean coal and coal slime in coal washing plants, providing accurate density input for subsequent inventory weight calculations.

[0076] Specific Implementation Method Seven: In step three of this implementation method, the bulk density regression model group adopts a multiple linear or low-order polynomial regression form:

[0077]

[0078] In the formula, β0 is the basic bulk density term, and β1, β2, and β3 are the influence coefficients of volume, moisture content, and bed height on bulk density, respectively. Among them, the moisture content coefficient β2 is negative, reflecting the physical law that the increase in moisture content leads to a decrease in the effective coal mass per unit volume.

[0079] Specific Implementation Method Eight: The regression coefficients of each group in this implementation method are estimated using the least squares method:

[0080] .

[0081] Based on the known sample data, the optimal estimates of β0, β1, β2, and β3 are obtained by minimizing the sum of squared residuals between the measured bulk density and the model prediction, so that the regression model achieves the best fit on the training samples and ensures the accuracy and reliability of the model parameters.

[0082] Specific implementation method nine: The regression model in this implementation method adopts spline regression, Gaussian process regression, support vector regression or neural network learning model.

[0083] Multiple linear or low-order polynomial regression is suitable for situations with limited sample size and relatively simple relationships between bulk density and factors. However, when sufficient sample accumulation is available and complex nonlinear relationships exist between bulk density and volume, moisture content, and bed height, more flexible learning models such as spline regression, Gaussian process regression, support vector regression, or neural networks can be used to automatically capture the nonlinear mapping relationships between variables without pre-setting specific function forms. By providing multiple optional models, the most suitable modeling method can be flexibly selected on-site based on the data foundation and accuracy requirements, without being bound to a specific algorithm, thus enhancing the engineering applicability and scalability of the method.

[0084] Specific Implementation Method Ten: The real-time volume inversion of bulk density and calculation of inventory weight in step four of this implementation method is as follows:

[0085] During the normal measurement phase, based on the current coal type and physical properties, its group Gj is determined, the corresponding regression model is called to retrieve the bulk density ρj, and the volume V measured in real time by three-dimensional lidar is combined to calculate the inventory weight using the following formula:

[0086]

[0087] Each time a stocktaking operation with the actual weight calibrated by the metering device is completed, a new sample is obtained, which is assigned to the corresponding group and the regression parameters of the group are incrementally updated. When a new coal type or significant drift in physical properties occurs, re-clustering is triggered, thus forming a closed loop of "measurement-grouping-regression-feedback-update", which continuously improves the bulk density model and measurement accuracy of each group as the operation progresses.

[0088] This implementation method applies the bulk density regression model established in steps one through three to a practical application, enabling continuous and automated estimation of inventory weight based on real-time volume data. Simultaneously, by using the actual weight calibrated by the metering device as a new sample to feed back to the model for incremental updates, the model parameters remain synchronized with actual operating conditions. When a new coal type or significant shifts in physical properties occur, re-clustering is triggered, allowing the grouping structure to adaptively adjust. These mechanisms together form a closed loop of "measurement—grouping—regression—feedback—update," ensuring that the bulk density models and metering accuracy of each group continuously improve with operation, eliminating the need for repeated manual calibration.

[0089] Combination Figures 1 to 4 Explanation of the working principle of this invention:

[0090] As attached Figure 2 As shown, this invention provides an adaptive modeling and reserve measurement method for bulk coal bulk density based on coal type clustering and multi-factor regression, specifically including the following steps:

[0091] T1: One or more 3D LiDARs developed by Harbin Zhiwu Technology Co., Ltd. are installed on the top of the coal bunker or above the material surface, in a full-coverage scanning position, and the coordinates of the radar and the geometric model of the bunker are calibrated to determine structural parameters such as the inner diameter of the bunker wall and the elevation of the reference surface of the bottom of the bunker.

[0092] T2: The material surface is scanned using 3D LiDAR to obtain a three-dimensional point cloud covering the entire material surface. After preprocessing such as coordinate registration, noise reduction, and reference surface calibration, the volume of coal in the silo is obtained by volume integration of the space between the material surface and the reference surface.

[0093] T3: Synchronously collect the actual inventory weight M provided by truck scales, belt scales or inbound / outbound metering systems, and record physical property information such as coal type label, particle size distribution, moisture content w, and material layer height h to form a "volume-physical property-weight" sample set.

[0094] T4: Based on the coal type and physical properties, the samples are clustered and grouped to obtain several groups G1,…,GK.

[0095] T5: Within each group, using multiple factors such as volume, moisture content, and material layer height as inputs, and the measured bulk density (or inventory weight) as the objective, a group-specific bulk density regression model ρj=fj(V,w,h) is established using methods such as least squares.

[0096] T6: Combine the various models into a grouped bulk density model specific to this coal yard, and verify the goodness of fit using indicators such as R².

[0097] T7: During the normal measurement phase, the group to which the product belongs is determined based on the current physical properties, the bulk density is inverted, and the inventory weight is calculated using M=V·ρj(V,w,h).

[0098] T8 and T9: After each inventory check operation that can be calibrated by the metering device is completed, the new samples are assigned to the corresponding group and the regression parameters are updated incrementally; when a new coal type or significant drift in physical properties occurs, re-clustering is triggered, the inventory weight is output and the accuracy is continuously improved.

[0099] Example 1: Grouping of three coal types and multiple linear regression (as attached) Figure 3 , Figure 4 (As shown)

[0100] This embodiment uses a coal storage facility in a coal washing plant as an example, employing a 3D LiDAR developed by Harbin Zhiwu Technology Co., Ltd. to scan the material surface. The coal types on site include clean coal, raw coal, and coal slime; the samples are clustered into three groups according to the coal type labels. The scatter plots of the measured bulk density for each group are shown below. Figure 3 As shown in (a), there are significant differences in bulk density levels among the groups: approximately 0.88–0.93 t / m³ for the clean coal group, approximately 0.80–0.85 t / m³ for the raw coal group, and approximately 0.66–0.74 t / m³ for the coal slime group.

[0101] Multiple linear regression was established within each group with moisture content as the main factor. The exemplary regression models for the three groups are as follows (in the following formulas, ρ1, ρ2, and ρ3 correspond to the clean coal group, raw coal group, and coal slime group, respectively; the coefficients are schematic values ​​after desensitization treatment, used only to illustrate the model form, and the actual values ​​will depend on the field data):

[0102]

[0103]

[0104]

[0105] like Figure 3 As shown in (b), the bulk density of all three groups decreased with increasing moisture content, and the rate of decrease varied for different coal types, verifying the necessity of group modeling.

[0106] Inventory weights are uniformly converted using M=V·ρj. Backtesting results ( Figure 4The results show that the predicted values ​​are highly consistent with the reference values. The average absolute percentage error (MAPE) of the inventory weight prediction for the clean coal group, raw coal group, and coal slime group is approximately 0.62%, 0.78%, and 0.95%, respectively, with a comprehensive MAPE of approximately 0.81%. The coefficient of determination R² is higher than 0.998, and the maximum error ratio does not exceed approximately 1.8%, which meets the accuracy requirements for on-site inventory, inventory monitoring, and coal blending scheduling (the specific values ​​depend on the equipment and on-site data).

[0107] Example 2: Automatic Grouping Based on Unsupervised Clustering

[0108] As another implementation method, when coal type labels are missing or the coal composition is complex, manual coal type labels can be avoided. Instead, physical property characteristic vectors such as moisture content, ash content, and particle size distribution can be used as inputs. The K-measns algorithm or Gaussian mixture model is used to automatically cluster the samples, adaptively determining the number of groups K and the boundaries of each group. Then, a bulk density regression model is established within each group. This method does not require pre-manual classification of coal types and can automatically discover sample groups with similar physical properties. It is especially suitable for coal storage and distribution yards with multiple coal types and large fluctuations in incoming materials.

[0109] Example 3: Nonlinear Upgrading of Within-Group Regression Model

[0110] As samples accumulate, the multiple linear regression within groups in Example 1 can be upgraded to nonlinear forms such as low-order polynomial regression, spline regression, or Gaussian process regression to more precisely characterize the nonlinear changes in bulk density with volume and moisture content; alternatively, the optimal regression form can be selected for each group. When there are significant changes in the coal yard structure, coal type composition, or incoming material characteristics, the clustering and modeling process is restarted, and the grouping and regression parameters are updated based on the new samples. Therefore, the overall scheme of this invention is not limited to a specific clustering method or regression form.

[0111] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make other changes within the spirit of the invention and apply it to fields not mentioned in the invention. Of course, all such changes made in accordance with the spirit of the invention should be included within the scope of protection claimed by the invention.

Claims

1. A method for adaptive modeling and reserve measurement of bulk coal bulk density based on coal type clustering and multi-factor regression, characterized in that: Includes the following steps: Step 1: Use the volume measurement data of the three-dimensional lidar and the actual weight calibrated by the measuring device as a sample, i.e., the "volume-weight" sample; Step 2: Cluster and group the samples according to coal type and particle size distribution as physical property labels; Step 3: Within each group, establish a bulk density regression model using volume, moisture content, bed height, and ash content as multiple factors; Step 4: Invert the bulk density from the real-time volume and calculate the inventory weight.

2. The adaptive modeling and reserve measurement method for bulk coal bulk density based on coal type clustering and multi-factor regression according to claim 1, characterized in that: In step one, the three-dimensional lidar is deployed on the top of the coal bunker or above the material surface, with one or more units.

3. The adaptive modeling and reserve measurement method for bulk coal bulk density based on coal type clustering and multi-factor regression according to claim 1, characterized in that: In step one, when the 3D lidar performs volume measurement, it first scans the material surface to obtain a 3D point cloud covering the entire material surface. After preprocessing such as coordinate registration, noise reduction and reference surface calibration, it combines the silo geometric model to perform volume integration or mesh summation on the space between the material surface and the reference surface to obtain the volume V of the coal in the silo at the current moment. Its volume integral is expressed as: In the formula, z s (x,y) represents the material surface elevation, z0(x,y) represents the datum elevation of the bottom of the silo, and Ω represents the horizontal projection area of ​​the silo.

4. The adaptive modeling and reserve measurement method for bulk coal bulk density based on coal type clustering and multi-factor regression according to claim 1, characterized in that: The clustering and grouping process in step two is as follows: For each "volume-weight" sample, its physical property feature vector is extracted, including coal type label, particle size distribution index, moisture content w, and ash content. Based on this, the sample is divided into several groups G1, G2, ..., GK, so that samples in the same group are similar in physical properties and the differences between groups are significant.

5. The adaptive modeling and reserve measurement method for bulk coal bulk density based on coal type clustering and multi-factor regression according to claim 4, characterized in that: The clustering in step two can be done manually by directly grouping coal types, or automatically using the K-measns algorithm, Gaussian mixture model, or unsupervised clustering methods; for each sample, the measured bulk density is directly obtained by the following formula: M is mass, V is volume, and density is calculated.

6. The adaptive modeling and reserve measurement method for bulk coal bulk density based on coal type clustering and multi-factor regression according to claim 1, characterized in that: The process of establishing the multi-factor heap density regression model in step three is as follows: Within each group Gj, using volume V, moisture content w, and bed height h as multiple factors as inputs, and measured bulk density ρ or stock weight M as the monitoring target, a bulk density regression model specific to that group is established: 。 7. The adaptive modeling and reserve measurement method for bulk coal bulk density based on coal type clustering and multi-factor regression according to claim 6, characterized in that: The bulk density regression model in step three uses multiple linear or low-order polynomial regression: In the formula, β0 is the basic bulk density term, and β1, β2, and β3 are the influence coefficients of volume, moisture content, and bed height on bulk density, respectively. Among them, the moisture content coefficient β2 is negative, reflecting the physical law that the increase in moisture content leads to a decrease in the effective coal mass per unit volume.

8. The adaptive modeling and reserve measurement method for bulk coal bulk density based on coal type clustering and multi-factor regression according to claim 7, characterized in that: The regression coefficients for each group were estimated using the least squares method: 。 9. The adaptive modeling and reserve measurement method for bulk coal bulk density based on coal type clustering and multi-factor regression according to claim 6, characterized in that: The regression model employs spline regression, Gaussian process regression, support vector regression, or neural network learning models.

10. The adaptive modeling and reserve measurement method for bulk coal bulk density based on coal type clustering and multi-factor regression according to claim 1, characterized in that: The process of real-time volumetric inversion of bulk density and calculation of inventory weight in step four is as follows: During the normal measurement phase, based on the current coal type and physical properties, its group Gj is determined, the corresponding regression model is called to retrieve the bulk density ρj, and the volume V measured in real time by three-dimensional lidar is combined to calculate the inventory weight using the following formula: Each time a stocktaking operation with the actual weight calibrated by the metering device is completed, a new sample is obtained, which is then assigned to the corresponding group and the regression parameters of that group are incrementally updated. When a new coal type or significant drift in physical properties occurs, re-clustering is triggered, thus forming a closed loop of "measurement-grouping-regression-feedback-update", which continuously improves the bulk density model and metering accuracy of each group as the operation progresses.

Images