Method for detecting coal content of sculpting screen based on binocular vision and dual-energy X-ray fusion

Abstract

The invention provides a method for detecting the coal content of a medium drainage screen based on binocular vision and dual-energy X-ray fusion, and relates to the technical field of visual detection, and the method comprises the following steps: synchronously collecting a binocular visible light image pair and a dual-energy X-ray image of a target material on the medium drainage screen; calculating and generating particle volume distribution, particle equivalent thickness distribution and particle surface feature distribution based on the binocular visible light image pair; based on the dual-energy X-ray image, the particle volume distribution and the particle equivalent thickness distribution, calculating the surface density and the equivalent density of each material particle, and generating particle material feature distribution and particle mass distribution; identifying each material particle in the target material by using a coal probability identification engine, and outputting predicted coal probability distribution; and calculating to obtain the total coal content of the target material, and evaluating the detection credibility to obtain a coal content detection result of the sculpting screen. The technical problems that in the prior art, the detection precision of the coal content of the medium drainage screen is low and is greatly influenced by working condition factors are solved.

Description

Technical Field

[0001] This invention relates to the field of visual inspection, and more particularly to a method for detecting coal content in a desliming screen based on the fusion of binocular vision and dual-energy X-ray. Background Technology

[0002] In the coal sorting and processing process, the desliming screen is a key piece of equipment. The detection of the coal content of the processed material is a core indicator for guiding production control, optimizing sorting efficiency, and ensuring product quality.

[0003] However, traditional methods for detecting coal content in desliming screens, such as manual sampling and analysis or single-sensor detection, generally suffer from low detection accuracy and are greatly affected by operating conditions such as material moisture and particle mixing, making it difficult to meet the demands of modern coal production for accurate, efficient, and continuous detection. Furthermore, single-sensor detection technologies have inherent limitations in acquiring material geometric parameters and identifying material composition, failing to fully cover the complex characteristics of materials used in desliming screens.

[0004] Therefore, in order to make up for the shortcomings of traditional detection methods and meet the demands of industrial production for accurate, real-time and reliable coal content detection, there is an urgent need for a coal content detection method based on binocular vision and dual-energy X-ray fusion for desliming screens, so as to achieve efficient and accurate detection of coal content in desliming screens. Summary of the Invention

[0005] This invention addresses the technical problems of low accuracy and high susceptibility to operating conditions in the detection of coal content in desliming screens in existing technologies by providing a method for detecting coal content in desliming screens based on the fusion of binocular vision and dual-energy X-ray.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0007] This invention provides a method for detecting coal content using a desliming screen based on binocular vision and dual-energy X-ray fusion, comprising: Triggered by a rotary encoder installed on the drive drum, the binocular visible light image pair and dual-energy X-ray image of the target material on the desliming screen are acquired simultaneously. Based on the binocular visible light image pair, the volume and equivalent thickness of each particle in the target material are calculated, and surface features are identified to generate particle volume distribution, particle equivalent thickness distribution, and particle surface feature distribution. Based on the dual-energy X-ray images, particle volume distribution, and particle equivalent thickness distribution, the material characteristics of each material particle are identified, the areal density and equivalent density of each material particle are calculated, and the particle mass is estimated to generate particle material characteristic distribution and particle mass distribution. Using a coal probability recognition engine, the particle size distribution, equivalent thickness distribution, surface feature distribution, material feature distribution, and mass distribution of the particles are used to identify each particle in the target material, and a predicted coal probability distribution is output. The total coal content of the target material is calculated by weighting the particle mass distribution and the predicted coal probability distribution, and the detection reliability is evaluated. The total coal content and the detection reliability are used as the coal content detection results of the desliming screen.

[0008] The beneficial effects of this invention are: Compared to existing technologies, this application first utilizes a rotary encoder installed on the drive drum to trigger and simultaneously acquire binocular visible light images and dual-energy X-ray images of the target material on the desliming screen. This simultaneous acquisition of multi-source information of the target material avoids problems such as timing misalignment or material region mismatch. Secondly, based on the binocular visible light images, the volume and equivalent thickness of each particle in the target material are calculated, and surface features are identified, generating particle volume distribution, particle equivalent thickness distribution, and particle surface feature distribution. This provides spatial and appearance dimension data support for subsequent calculations, effectively offsetting the interference of thickness effects on detection and overcoming the limitations of single detection technologies. Thirdly, based on the dual-energy X-ray images, particle volume distribution, and particle equivalent thickness distribution, the material characteristics of each particle are identified, and the areal density and equivalent density of each particle are calculated. Particle mass is estimated, generating particle material feature distribution and particle mass distribution. This allows for accurate identification of the material characteristics of each particle and the calculation of areal density, equivalent density, and particle mass estimation. Furthermore, utilizing a coal probability recognition engine, the system identifies each particle in the target material based on particle volume distribution, equivalent thickness distribution, surface feature distribution, material feature distribution, and mass distribution. It outputs a predicted coal probability distribution, accurately identifying the coal properties of each particle and stably outputting the predicted coal probability distribution. This approach balances prediction accuracy and real-time performance under different operating conditions, effectively reducing the impact of sample noise. Finally, the total coal content of the target material is calculated by weighting the particle mass distribution and the predicted coal probability distribution, and the detection reliability is evaluated. Using the total coal content and detection reliability as the coal content detection results for the desliming screen provides a direct reflection of the reliability of the prediction process, offering precise data for desliming screen production control and effectively overcoming the shortcomings of inaccurate and unreliable traditional detection data.

[0009] Through the above technical solution, this application effectively solves the problems of low accuracy, large interference from working conditions, and single-technology limitations of traditional detection methods: By synchronously acquiring binocular visible light image pairs and dual-energy X-ray images through a rotary encoder, the timeliness of data is ensured. Based on the binocular visible light image pairs and dual-energy X-ray images, particle volume, equivalent thickness, surface feature distribution, material characteristics, and mass distribution are obtained. Then, the coal probability recognition engine integrates multi-dimensional distribution data to output a reliable predicted coal probability distribution. Finally, the accurate total coal content is obtained through weighted calculation and a corresponding detection credibility assessment is provided, forming a complete detection link of real-time acquisition, multi-dimensional analysis, accurate identification, and reliable output. This provides coal content data support with accuracy, real-time performance, and reliability for the production control of desliming screens, and fully adapts to the needs of modern coal production. Attached Figure Description

[0010] Figure 1 A schematic flowchart of the coal content detection method based on binocular vision and dual-energy X-ray fusion for the present invention; Figure 2 This is a schematic diagram of the process for generating particle material feature distribution and particle mass distribution in the method for detecting coal content by desliming sieve based on binocular vision and dual-energy X-ray fusion provided by the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0013] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0014] Examples, such as Figure 1 As shown, this embodiment of the invention provides a method for detecting coal content using a desliming screen based on binocular vision and dual-energy X-ray fusion, including: S10: Triggered by a rotary encoder installed on the drive drum, it synchronously acquires binocular visible light image pairs and dual-energy X-ray images of the target material on the desliming screen.

[0015] When the desliming screen is working, the target material, which is a mixture of coal, gangue, and media, moves at a constant speed with the conveyor belt. After being washed by water spray, some of the gangue turns black due to immersion, making it difficult to distinguish from the appearance of coal. Single image detection cannot simultaneously obtain surface features and material information.

[0016] Meanwhile, the binocular visible light image pairs acquired by the binocular camera can extract surface features such as color, texture, and shape of material particles, and can also obtain depth information through stereo matching and three-dimensional reconstruction, thereby calculating particle volume and equivalent thickness; the dual-energy X-ray imaging device utilizes the difference in attenuation characteristics of two types of X-rays on different materials such as coal and gangue, and the acquired dual-energy X-ray images can directly reflect the differences in the internal structure and material composition of the materials.

[0017] To address the aforementioned issues, this application utilizes a rotary encoder installed on the drive drum to trigger and simultaneously acquire binocular visible light image pairs and dual-energy X-ray images of the target material on the desliming screen.

[0018] Specifically, step S10 in the method includes: Using the pulse signal emitted by the rotary encoder installed on the drive drum, the binocular camera and dual-energy X-ray imaging device are synchronously triggered to acquire binocular visible light image pairs and dual-energy X-ray images of the target material at the same sampling time. By using a pre-calibrated extrinsic parameter matrix, the binocular visible light image pair and the dual-energy X-ray image are uniformly registered to the belt coordinate system, establishing the spatial correspondence between the pixels of each image.

[0019] In this embodiment, the pulse signal emitted by the rotary encoder installed on the drive drum is first used to synchronously trigger the binocular camera and the dual-energy X-ray imaging device, acquiring binocular visible light image pairs and dual-energy X-ray images of the target material at the same sampling time. The drive drum is the power transmission component of the stripping screen, responsible for driving the belt to transport the target material. The rotary encoder is a sensor that converts mechanical rotational motion into electrical pulse signals; it is installed on the shaft of the drive drum, outputting a pulse signal every time the drive drum rotates a certain angle, with the pulse frequency proportional to the drum's rotational speed. Synchronous triggering avoids timing differences and ensures co-position imaging. The binocular camera refers to two visible light cameras with identical parameters, fixed installation positions, and maintained baseline distance, used to simultaneously photograph the target material on the stripping screen, acquiring a pair of binocular visible light image pairs with slightly different viewing angles. The dual-energy X-ray imaging device is a device capable of simultaneously emitting two different energy levels of X-rays and receiving the X-ray signals after penetrating the target material through a dedicated detector, forming low-energy X-ray images and high-energy X-ray images respectively.

[0020] Specifically, during the operation of the desliming screen, the target material, composed of a mixture of coal, gangue, and media, moves at a constant speed along the conveyor belt. After being washed by water spray, a large amount of gangue turns black due to wetting, making it difficult to distinguish from coal in visible light. A single image cannot simultaneously capture both surface features and material information. The pulse signals output by the rotary encoder are transmitted in real-time to the control unit of the binocular camera and dual-energy X-ray imaging device. When the pulse signal reaches a preset trigger threshold (e.g., triggering acquisition every 10 pulses), the control unit simultaneously sends a start command to both the binocular camera and the dual-energy X-ray imaging device. The binocular camera immediately captures a pair of binocular visible light images of the target material, while the dual-energy X-ray imaging device simultaneously emits low-energy rays (e.g., 30-60 keV) and high-energy rays (e.g., 80-120 keV), penetrating the target material to obtain a dual-energy X-ray image. This synchronous acquisition method ensures that multimodal data of the target material are acquired simultaneously.

[0021] For example, if the circumference of the drive roller is 1.2 meters, and the rotary encoder outputs 1000 pulse signals per revolution, and if the preset trigger threshold is set to trigger acquisition once every 10 pulse signals, during the acquisition process, a binocular camera positioned above the desliming screen and maintaining a baseline distance (dynamically determined according to actual conditions), and dual-energy X-ray imaging devices positioned on the upper and lower sides of the desliming screen, simultaneously acquire binocular visible light image pairs and dual-energy X-ray images of the target material at the same sampling moment. Furthermore, based on the correlation between pulse signals and belt movement, the belt movement distance corresponding to two adjacent sampling moments is 1.2 meters / 1000 × 10 = 0.012 meters, or 12 millimeters. This ensures continuous and complete acquisition of the target material on the belt, and that the binocular visible light image pairs and dual-energy X-ray images acquired at each sampling moment precisely correspond to the target material within the same 12-millimeter range on the belt.

[0022] Secondly, since the binocular camera and dual-energy X-ray imaging device acquire information from different angles, without a calibration criterion, it is impossible to determine whether the pixels of a certain particle in the binocular visible light image pair correspond to the same particle as the pixels in a certain attenuation region of the dual-energy X-ray image. Therefore, it is necessary to uniformly register the binocular visible light image pair and the dual-energy X-ray image to the belt coordinate system using a pre-calibrated extrinsic parameter matrix to establish the spatial correspondence between the pixels of each image. The extrinsic parameter matrix is ​​a parameter matrix describing the position and orientation relationship between the binocular camera, the dual-energy X-ray imaging device, and the belt coordinate system, which can be obtained through joint calibration using a calibration plate and a metal positioning block. Uniformly registering the binocular visible light image pair and the dual-energy X-ray image to the belt coordinate system ensures that the pixel positions of the same material particle correspond one-to-one in different images. The belt coordinate system is a three-dimensional Cartesian coordinate system established based on the desliming screen belt.

[0023] For example, firstly, a belt coordinate system is constructed based on the desliming screen belt. A standard calibration plate (such as a checkerboard calibration plate) with known size feature points and a metal positioning block are placed at different positions on the desliming screen belt. The binocular image of the calibration plate is captured by a binocular camera, and the dual-energy X-ray image of the calibration plate is captured by a dual-energy X-ray imaging device. Then, based on the known actual coordinates of the feature points of the calibration plate in the belt coordinate system, combined with the pixel coordinates of the feature points of the calibration plate in the binocular visible light image pair and the dual-energy X-ray image, the extrinsic parameter matrix of the binocular camera relative to the belt coordinate system and the extrinsic parameter matrix of the dual-energy X-ray imaging device relative to the belt coordinate system are calculated by perspective transformation algorithm to form a pre-calibrated extrinsic parameter matrix.

[0024] For example, a belt coordinate system is first constructed based on the desliming screen belt. For instance, the conveying direction is the X-axis, the width direction is the Y-axis, and the direction perpendicular to the belt surface upwards is the Z-axis. Then, a standard calibration plate (such as a checkerboard calibration plate) with known dimensional feature points and a metal positioning block are placed at different positions on the desliming screen belt. Subsequently, a pair of binocular visible light images of the calibration plate are captured by a binocular camera positioned above the belt, and a dual-energy X-ray image of the calibration plate is captured by a dual-energy X-ray imaging device positioned on the upper and lower sides of the belt. Based on the known actual three-dimensional coordinates of the feature points of the calibration plate in the belt coordinate system, combined with the two-dimensional pixel coordinates of the feature points of the calibration plate in the binocular visible light image pair and the dual-energy X-ray image, the extrinsic parameter matrices of the binocular cameras (left and right cameras) relative to the belt coordinate system and the extrinsic parameter matrix of the dual-energy X-ray imaging device relative to the belt coordinate system are calculated using a perspective transformation algorithm to form a pre-calibrated extrinsic parameter matrix.

[0025] For example, based on a pre-calibrated extrinsic matrix, if the pixel coordinates of a coal particle in the left camera image of a binocular visible light image are (320, 450), after transformation by the extrinsic matrix, its three-dimensional coordinates in the belt coordinate system are (650mm, 280mm, 30mm); if the pixel coordinates of the same coal particle in the right camera image of a binocular visible light image are (335, 450), after transformation by the extrinsic matrix, its three-dimensional coordinates in the belt coordinate system are (650mm, 280mm, 30mm); simultaneously, if the pixel coordinates of the same coal particle in the dual-energy X-ray image are (290, 420), after transformation by the extrinsic matrix, its three-dimensional coordinates in the belt coordinate system are also (650mm, 280mm, 30mm). In this way, the spatial correspondence between the pixels of the coal particle in each image is established, and the unique spatial correspondence of the coal particle's pixels in different images is clarified.

[0026] In summary, compared to existing technologies, this application utilizes a rotary encoder installed on the drive drum to trigger and simultaneously acquire binocular visible light image pairs and dual-energy X-ray images of the target material on the desliming screen. This simultaneous acquisition of multi-source information of the target material, followed by unified registration, avoids problems such as timing misalignment or material region mismatch, providing a reliable data foundation for subsequent multimodal feature fusion, coal gangue identification, and accurate calculation of coal content.

[0027] S20: Based on the binocular visible light image pair, calculate the volume and equivalent thickness of each material particle in the target material, identify surface features, and generate particle volume distribution, particle equivalent thickness distribution, and particle surface feature distribution.

[0028] The target material on the desliming screen is a mixture of coal, gangue, media and other components. Although some gangue turns black after being washed with water spray, making it difficult to distinguish from the appearance of coal, there are still inherent differences between coal, gangue and media in terms of surface properties and spatial morphology.

[0029] Specifically, coal is mostly dark black with a rough surface and irregular texture; dry / semi-dry gangue is often grayish-white or yellowish, with a dense texture and more distinct edges; medium (such as magnetite) is mostly black but has a relatively regular shape, and some have a metallic luster. These differences in color, texture and shape provide a basis for surface feature identification.

[0030] To address the aforementioned issues, this application calculates the volume and equivalent thickness of each particle in the target material based on the binocular visible light image pair, identifies surface features, and generates particle volume distribution, particle equivalent thickness distribution, and particle surface feature distribution.

[0031] Specifically, step S20 in the method includes: Based on the binocular visible light image pair, a disparity map is obtained through a stereo matching algorithm, and depth information is calculated based on the camera focal length and the baseline distance. Using the surface of the desliming screen belt as a reference plane, and combining the depth information, a three-dimensional model of each material particle is constructed through a surface mesh reconstruction algorithm. The volume of each material particle is calculated based on the three-dimensional model of each material particle, and the particle volume distribution is constructed. Based on the volume of each material particle and its projected area on the image plane, as well as the angle between the X-ray beam and the normal, the equivalent thickness of each material particle in the X-ray path direction is calculated, and the particle equivalent thickness distribution is constructed. A particle surface feature recognition model is constructed using a convolutional neural network. The model is then used to extract features from the binocular visible light image pairs to obtain the surface features of each particle in the target material. A particle surface feature distribution is then constructed, wherein the surface features include color features, texture features, and morphological features.

[0032] In this embodiment, firstly, a disparity map is obtained based on a pair of binocular visible light images using a stereo matching algorithm, and then depth information is calculated based on the camera focal length and baseline distance. The stereo matching algorithm refers to an algorithm that finds the corresponding pixel points of the same material particle in two images by comparing pixel grayscale values, texture features, and other information in the binocular visible light image pair; examples include semi-global matching algorithms and block matching algorithms. The disparity map is an image containing disparity information for each pixel. Disparity refers to the difference in horizontal pixel position of the same physical point on the imaging planes of the left and right cameras, serving as intermediate data for depth calculation. Depth information refers to the vertical distance from each point on the surface of the material particle to the lens plane of the binocular camera, directly reflecting the three-dimensional spatial position of the particle.

[0033] Specifically, because the two devices of the binocular camera have completely identical parameters and a fixed baseline distance, the imaging position of the same material particle in the binocular visible light image pair will experience a slight shift due to the difference in viewing angle, i.e., parallax d(x,y), where x and y are pixel coordinates. Furthermore, since the binocular camera is installed based on a horizontal baseline, the parallax calculation only focuses on the difference in the horizontal direction (i.e., the x-axis). According to the principle of triangulation, depth information can be calculated using the formula Z(x,y)=(f×B) / d(x,y), where Z(x,y) is the depth information corresponding to pixel (x,y), f is the camera focal length, and B is the baseline distance.

[0034] For example, if the camera focal length f = 10mm, the baseline distance between the binocular cameras B = 50mm, and the pixel coordinates of a point on a material particle in the binocular visible light image pair are (200, 300) and (185, 300), then the disparity of that point d = 200 - 185 = 15 pixels. Substituting into the formula, the depth information Z = (10 × 50) / 15 ≈ 33.3mm can be calculated, that is, the vertical distance from that point to the lens plane of the binocular camera is 33.3mm.

[0035] Secondly, using the surface of the desliming screen belt as the reference plane and combining depth information, a three-dimensional model of each material particle is constructed using a surface mesh reconstruction algorithm. The volume of each particle is then calculated based on its three-dimensional model, thus constructing the particle volume distribution. The surface mesh reconstruction algorithm refers to an algorithm that transforms discrete depth information into a continuous, smooth three-dimensional surface model, such as the Poisson reconstruction algorithm or the greedy projection triangulation algorithm. The particle volume distribution refers to calculating the volume of a single material particle through volume integration of the three-dimensional model, then dividing it into different intervals based on volume size, such as 0-5 cm³, 5-10 cm³, and above 10 cm³, and statistically analyzing the proportion of particles in each interval, reflecting the particle volume composition pattern of the target material.

[0036] Specifically, when the conveyor belt is empty, the depth information Zbelt of the upper surface of the desliming screen is collected; secondly, image segmentation algorithms (such as threshold segmentation and semantic segmentation) are used to separate individual material particles from the background from the binocular visible light image pair, determining the pixel set Ωi corresponding to each material particle; then, a surface mesh reconstruction algorithm is used to integrate the height information of all pixels into a continuous 3D particle model; finally, the volume integral of the 3D model is calculated, and the volume calculation formula for each material particle is: Where Z(x,y)-Z beltΔx and Δy represent the particle height corresponding to each pixel coordinate (x, y), i.e., the vertical distance from the pixel to the belt surface. Δx and Δy are the spatial resolution of the camera's imaging plane, i.e., the actual horizontal and vertical dimensions corresponding to one pixel. This volume calculation method based on a 3D model, compared to the traditional projected area extrapolation method, can more accurately reflect the true three-dimensional shape of material particles and effectively reduce volume estimation errors caused by particle stacking and irregular shapes.

[0037] For example, if the pixel set Ωi of a certain material particle contains 1000 pixels, the spatial resolution of the camera imaging plane is Δx=Δy=1mm, and the depth information Z of the belt surface is... belt To calculate the volume of a material particle, 1000 pixel coordinates are substituted into the formula for calculating the volume of each particle, resulting in a particle volume V = 10 cm³. The same method is used to calculate the volume of all particles in the target material, which are then divided into different ranges based on size, such as 0-5 cm³, 5-10 cm³, and above 10 cm³. The percentage of particles in each range is then calculated; for example, 30% of particles are in the 0-5 cm³ range, 50% in the 5-10 cm³ range, and 20% in the above 10 cm³ range. This results in a particle volume distribution that reflects the volume composition of the target material.

[0038] Next, based on the volume of each material particle, its projected area on the image plane, and the angle between the X-ray beam and the normal, the equivalent thickness of each material particle along the average penetration length in the X-ray path direction is calculated, thus constructing the particle equivalent thickness distribution. Here, the projected area refers to the area of ​​the two-dimensional projected region of the material particle on the binocular visible light image plane, i.e., the actual area corresponding to the pixel set Ωi, calculated using the following formula: , This refers to the projected area; the angle between the X-ray beam and the normal refers to the angle between the incident direction of the dual-energy X-ray and the normal direction of the surface of the desliming screen belt (i.e., the direction perpendicular to the belt surface), denoted as θ. θ is a fixed value that is pre-calibrated when installing the dual-energy X-ray imaging device; the equivalent thickness refers to the average penetration length of the material particles along the X-ray path direction, used to quantify the actual distance that X-rays penetrate the particles and to offset the influence of thickness differences caused by irregular particle shapes and different placement angles on the X-ray attenuation calculation; the particle equivalent thickness distribution refers to dividing all particles into intervals according to the size of the equivalent thickness and statistically analyzing the distribution ratio of particles in each interval.

[0039] Specifically, when dual-energy X-ray imaging devices perform X-ray detection, the degree of attenuation of the rays penetrating particles is directly related to the actual penetration length. However, due to the irregular shape of the particles, the penetration length varies at different locations and cannot be directly measured. Equivalent thickness, through the correlation calculation of volume, projected area, and angle, can achieve equivalent quantification of the penetration length. The formula for calculating equivalent thickness is: ,in, That is, the volume of each material particle The logic behind the equivalent thickness calculation formula is: particle volume Vi equals projected area. The average penetration length of the X-ray path is obtained by combining the average height perpendicular to the projection plane with the cosine of the angle θ between the X-ray beam and the normal.

[0040] For example, if the volume of a certain material particle Vi = 10 cm³ = 10000 mm³, and its projected area on the image plane is calculated from the image, =1000mm², the angle between the X-ray beam and the normal is θ=30°. Substituting these values ​​into the formula for calculating the equivalent thickness, we obtain the equivalent thickness of the material particle. =(Vi× ) / =(10000mm³× The average length of X-ray penetration through the material particles is approximately 8.66 mm (approximately 8.66 mm). This effectively compensates for the differences in actual penetration length caused by irregular particle shapes and varying placement angles, providing a consistent thickness reference for subsequent calculations. After calculating the equivalent thickness of all material particles using the same method, the particles are divided into intervals of 0-5 mm, 5-10 mm, and above 10 mm. The percentage of particles within each interval is then calculated to form an equivalent particle thickness distribution that reflects the material's thickness distribution pattern.

[0041] Finally, a particle surface feature recognition model was constructed using a convolutional neural network. This model was used to extract features from binocular visible light images, obtaining the surface features of each particle in the target material and constructing a particle surface feature distribution. These surface features include color, texture, and morphological features. Convolutional neural networks (CNNs), such as ResNet and MobileNet, are deep learning networks adept at processing image data. They can automatically extract low-level detail features and high-level semantic features from images, exhibiting strong feature representation capabilities. The particle surface feature recognition model, built upon a CNN, can automatically identify and output the surface features of individual material particles. The particle surface feature distribution refers to dividing all material particles according to surface feature type or parameter range and statistically analyzing the distribution ratio of each type / range. Surface features include color, texture, and morphological features: color features refer to the color information of the particle surface, reflecting the particle's color attributes; texture features refer to the texture pattern information of the particle surface, reflecting the roughness and texture density of the particle surface; and morphological features refer to the two-dimensional shape information of the particle, reflecting its geometric morphology.

[0042] For example, a particle surface feature recognition model can be constructed through the following technical path: 1. Data preparation: Collect binocular visible light image pairs from the actual working scene of the desliming screen, covering different lighting conditions, different material humidity and different particle sizes, and including mixtures of coal, gangue, media and other materials, to construct a sample image set containing at least 10,000 samples; for each sample image, label the RGB / HSV color space statistical values, gray-level co-occurrence matrix (GLCM) and local binary mode (LBP) texture parameters, aspect ratio and roundness and other parameters to form a sample feature set, and then randomly divide it into training set, validation set and test set in a ratio of 7:1.5:1.5.

[0043] 2. Model Construction: A particle surface feature recognition model can be constructed based on the lightweight MobileNetV2 convolutional neural network to balance detection accuracy and real-time performance. It mainly consists of an input layer, a feature extraction layer, a feature fusion layer, and an output layer. The input layer receives single-channel stereo images normalized to 224×224×3 dimensions, with pixel values ​​mapped to the [0,1] interval. The feature extraction layer contains 13 inverted residual modules, employing a combination of 1×1 convolution for dimensionality upscaling, 3×3 depthwise convolution for feature extraction, and 1×1 convolution for dimensionality reduction, combined with residual connections to preserve multi-scale features. The feature fusion layer upsamples feature maps from different sampling scales to the same size using transposed convolutions and then concatenates them, fusing low-level details with high-level semantic features. The output layer has three parallel fully connected branches, outputting 18-dimensional vectors related to color, texture, and morphological features, respectively. Dropout layers are embedded within the branches to suppress overfitting.

[0044] 3. Model Training: Using sample images from the training set as input features and corresponding sample features as supervision labels, an Adam optimizer with an initial learning rate of 1e-4 and weight decay of 1e-5 is employed. A weighted mean squared error loss function is used; for example, the weights for color, texture, and morphology features are 0.3, 0.4, and 0.3, respectively. During training, the ReduceLROnPlateau learning rate scheduling strategy is enabled. The learning rate decays by 50% when the validation set loss does not decrease for five consecutive epochs. A maximum training epoch of 100 and an early stopping strategy are also set. After each training epoch, the feature prediction MAE is evaluated using the validation set, and parameters are fine-tuned in real-time until the validation set loss converges and the test set MAE is below 0.05, resulting in a successfully trained particle surface feature recognition model. For example, a certain gangue particle has a grayish-white surface in a dry state. The particle surface feature recognition model extracts its color feature as the RGB mean value (200, 200, 190) from its binocular visible light image, its texture feature as a high GLCM energy value (indicating a uniform and dense surface texture), and its morphological feature as a roundness of 0.7 (indicating a relatively regular shape). Conversely, a certain coal particle has a dark black surface. The particle surface feature recognition model extracts its color feature as the RGB mean value (80, 70, 60) from its binocular visible light image, its texture feature as a high LBP value (indicating a rough texture), and its morphological feature as an aspect ratio of 2.5 (indicating an irregular shape). Thus, by extracting these different surface features through the particle surface feature recognition model, coal and impurity particles can be preliminarily distinguished.

[0045] In summary, compared to existing technologies, this application, based on binocular visible light image pairs, calculates the volume and equivalent thickness of each particle in the target material, identifies surface features, and generates particle volume distribution, particle equivalent thickness distribution, and particle surface feature distribution. In this way, it successfully extracts the volume, equivalent thickness, and surface features such as color, texture, and morphology of each particle in the target material, generating corresponding three types of distribution data. This provides spatial and appearance dimension data support for subsequent calculations, effectively offsetting the interference of thickness effects on detection and overcoming the limitations of single detection technologies.

[0046] S30: Based on the dual-energy X-ray image, particle volume distribution, and particle equivalent thickness distribution, identify the material characteristics of each material particle, calculate the areal density and equivalent density of each material particle, estimate the particle mass, and generate particle material characteristic distribution and particle mass distribution.

[0047] After being sprayed with water, the mixture of coal, gangue, and media on the desliming screen becomes partially confused in appearance. A single detection method cannot accurately distinguish the materials and calculate the particle mass at the same time. However, dual-energy X-rays have inherent differences in the attenuation characteristics of different materials such as carbonaceous materials, silicon aluminum oxides, and iron oxides. This can overcome the interference of surface state and achieve material identification. Combined with the particle volume and equivalent thickness obtained from previous binocular visible light images, the areal density, equivalent density, and particle mass can be further derived, making up for the limitation of single X-rays in obtaining three-dimensional and mass information of materials.

[0048] To address the aforementioned issues, this application identifies the material characteristics of each material particle based on the dual-energy X-ray image, particle volume distribution, and particle equivalent thickness distribution, calculates the areal density and equivalent density of each material particle, estimates the particle mass, and generates particle material characteristic distribution and particle mass distribution.

[0049] Specifically, such as Figure 2 As shown, step S30 in the method includes: Based on the dual-energy X-ray image, the attenuation value of each material particle in the target material under low-energy X-ray and high-energy X-ray is calculated, and according to the attenuation value, the material characteristics of each material particle are identified by the substrate decomposition model, and the particle material characteristic distribution is constructed. Based on the dual-energy X-ray image, the areal density of each particle in the target material is calculated using a dual-energy attenuation model. Based on the particle equivalent thickness distribution, the areal density of each material particle is divided by the corresponding equivalent thickness to obtain the equivalent density of each material particle, thus constructing the particle equivalent density distribution. Based on the particle volume distribution and particle equivalent density distribution, the mass of each material particle is estimated using a mass calculation formula to generate a particle mass distribution.

[0050] In this embodiment, firstly, based on dual-energy X-ray images, the attenuation values ​​of each material particle in the target material under low-energy and high-energy X-rays are calculated. Then, based on the attenuation values, the material characteristics of each particle are identified using a substrate decomposition model, and a particle material characteristic distribution is constructed. The attenuation value refers to the degree of intensity reduction of X-rays after penetrating the material particles, following the Lambert-Beer law, and is calculated using the following formula: In the formula, This represents the logarithmic decay value corresponding to energy E. The background radiation intensity when there is no material. The intensity of the X-rays after penetrating the material is considered. The substrate decomposition model is a mathematical model that infers the material composition of particles based on the differences in the mass attenuation coefficients of different materials to dual-energy X-rays. The core logic assumes that the target material is mainly composed of two types of substrates, such as the carbonaceous substrate of coal and the silicon-aluminum oxide substrate of gangue. Material characteristics refer to the identification results of the material type, and particle material characteristic distribution refers to the quantity or mass proportion of different material characteristics in the target material.

[0051] Specifically, firstly, particle segmentation is performed on the dual-energy X-ray image to determine the pixel region corresponding to each material particle; then, the logarithmic attenuation value of this region at low and high energies is calculated using the aforementioned attenuation value calculation formula. , Then , Substitute into the substrate decomposition model: =μ1(L)×t1+μ2(L)×t2、 =μ1(H)×t1+μ2(H)×t2, where μ1(E) and μ2(E) are the mass attenuation coefficients of the two types of substrates at energy E, which can be obtained from the X-ray mass attenuation coefficient table, and t1 and t2 are the thicknesses of the two types of substrates; after obtaining μ1(E) and μ2(E) from the table, the logarithmic attenuation value is... , Substituting μ1(E) and μ2(E) into the substrate decomposition model, the thicknesses t1 and t2 of the two types of substrates can be obtained. Then, the proportion of the equivalent thickness of the two types of substrates is calculated as the material characteristics of each material particle. Finally, the material types of all particles are counted to form the particle material characteristic distribution.

[0052] It should be noted that the X-ray mass attenuation coefficient table is a standardized data table compiled by an international authoritative organization based on experimental measurements and theoretical calculations. It can be used to look up the mass attenuation coefficient of different substances (elements or compounds) at a specific X-ray energy. It is an existing mature technology that can be directly used without the need for additional experimental determination.

[0053] For example, the intensity of low-energy rays when there is no material =1000, after penetrating a certain particle =600, then the low-energy decay value =ln(1000 / 600)≈0.511; High-energy ray intensity =1000, after penetrating a certain particle =700, then the high-energy decay value =ln(1000 / 700)≈0.357. For example, if we look up the X-ray mass attenuation coefficient table and find that the carbonaceous substrate μ1(L) = 0.15cm² / g and μ1(H) = 0.10cm² / g, and the silicon aluminate oxide substrate μ2(L) = 0.30cm² / g and μ2(H) = 0.18cm² / g, and substitute them into the substrate decomposition model: =0.15cm² / g×t1+0.30cm² / g×t2、 =0.10cm² / g×t1+0.18cm² / g×t2, solving for t1≈2.8g / cm² and t2≈0.3g / cm², we get the carbon content = 2.8 / (2.8+0.3)≈90% and the silicon-aluminum oxide content = 0.3 / (2.8+0.3)≈10%. Therefore, the material characteristics of the particle can be determined to be coal. Furthermore, the material characteristics of each particle in the target material are calculated using the same method, and the proportion of particles of different material types by quantity or mass is statistically analyzed to construct a particle material characteristic distribution that reflects the material composition law of the target material.

[0054] Secondly, based on dual-energy X-ray images, the areal density of each particle in the target material is calculated using a dual-energy attenuation model. Areal density refers to the mass per unit projected area of ​​a particle, measured in kg / m² or g / cm², and is directly related to particle thickness and density, but is not affected by irregular particle shape. The dual-energy attenuation model utilizes the attenuation laws of low-energy and high-energy X-rays, solving simultaneous equations to eliminate thickness variables and directly calculating areal density. Its core advantage lies in its ability to counteract the interference of uneven particle thickness and stacking on detection, solving the problem that single-energy X-rays cannot separate the effects of thickness and density.

[0055] Specifically, the attenuation value of a single-energy X-ray =μ×σ, where μ is the mass decay coefficient of the material, σ is the areal density, and μ is related to the material. Therefore, for the segmented particle region, the logarithmic decay values ​​at low and high energies are calculated first. , Furthermore, based on the dual-energy decay model: =μ(L)×σ、 =μ(H)×σ, solve and calculate the mean to get the final σ, where μ(L) and μ(H) are the average mass decay coefficients of the particle material at low and high energies, respectively, which are determined by the material characteristics and can be obtained directly by querying.

[0056] For example, if the low energy decay value of a certain coal particle =0.511, high-energy decay value =0.357, and the values ​​of μL=0.15cm² / g and μH=0.10cm² / g for coal are obtained from the query. Substituting these values ​​into the above dual-energy decay model, the surface densities are calculated to be 0.511 / 0.15≈3.41g / cm² and 0.357 / 0.10≈3.57g / cm², respectively. The average of the two values ​​is then calculated to obtain the surface density of the particle σ=(3.41+3.57) / 2=3.49g / cm².

[0057] Secondly, based on the particle equivalent thickness distribution, the areal density of each material particle is divided by its corresponding equivalent thickness to obtain the equivalent density of each material particle, thus constructing the particle equivalent density distribution. Here, equivalent density refers to the ratio of the areal density of a material particle to its equivalent thickness, expressed in g / cm³ or kg / m³. It reflects the density of the particles and is an important auxiliary indicator for distinguishing materials. For example, the equivalent density of coal is typically 1.3-1.8 g / cm³, gangue is 2.5-3.0 g / cm³, and media is 4.5-5.0 g / cm³. The particle equivalent density distribution involves dividing all particles into different intervals according to their equivalent density, such as 1.0-2.0 g / cm³, 2.0-3.0 g / cm³, and above 3.0 g / cm³. The distribution percentage of particles in each interval is statistically analyzed to further verify the accuracy of the material identification results.

[0058] Specifically, the physical essence of surface density σ is density × thickness, therefore the equivalent density can be obtained through the formula ρ. i =σ i / h i It is derived that, where σ i Let h be the areal density of the i-th particle. i This corresponds to the equivalent thickness. The core logic of this formula is: to obtain the equivalent thickness h from binocular visible light images. i The average length of the particles penetrated by X-rays was quantified, and then compared with the areal density σ calculated from dual-energy X-ray images. i By combining these methods, density estimation after thickness offset can be achieved, solving the problem that traditional X-rays cannot obtain the true density of particles.

[0059] For example, if the surface density σ of a certain particle i =3.49 g / cm², equivalent thickness hi = 0.866 cm, substituting these values ​​into the above formula, we can obtain the equivalent density ρ. i =3.49g / cm² ÷ 0.866cm ≈ 4.03g / cm³. The equivalent density of each material particle was calculated using the same method. Then, all particles were divided into preset intervals based on their equivalent density, such as 1.0-2.0g / cm³, 2.0-3.0g / cm³, and above 3.0g / cm³. The percentage of particles by quantity or mass within each interval was statistically analyzed to construct a particle equivalent density distribution that reflects the density composition of the target material.

[0060] Finally, based on the particle volume distribution and particle equivalent density distribution, the mass of each material particle is estimated using the mass calculation formula to generate the particle mass distribution.

[0061] The formula for calculating mass is: m i = ×ρ i In the formula, m i Let be the mass of the i-th particle. Let ρ be the volume of the i-th particle. i Let be the equivalent density of the i-th particle. Particle mass distribution refers to dividing all material particles into different ranges based on their mass size, such as 0-5g, 5-10g, 10-20g, and over 20g, and statistically analyzing the proportion of particles in each range, reflecting the mass composition pattern of the target material.

[0062] Specifically, the key to mass estimation is to accurately fuse the particle volume calculated based on binocular visible light images with the equivalent particle density calculated based on dual-energy X-ray images: firstly, the volume of each particle is extracted from the particle volume distribution in step S20. Extract the corresponding ρ from the particle equivalent density distribution in this step. i Then, substitute the values ​​into the mass calculation formula to calculate the mass m of a single particle. i .

[0063] For example, if the volume of a certain particle =10cm³, equivalent density ρ i =4.03 g / cm³, substituting into the mass calculation formula, we get the mass m. i =10cm³×4.03g / cm³=40.3g. The mass of each material particle is calculated using the same method, and then divided into intervals of 0-10g, 10-20g, and above 20g to generate a particle mass distribution reflecting the material's mass composition.

[0064] In summary, compared to existing technologies, this application, based on the aforementioned dual-energy X-ray images, particle volume distribution, and particle equivalent thickness distribution, identifies the material characteristics of each particle, calculates the areal density and equivalent density of each particle, and estimates the particle mass, generating particle material characteristic distribution and particle mass distribution. This allows for accurate identification of the material characteristics of each particle, calculation of areal density and equivalent density, and estimation of particle mass, providing reliable data support for the subsequent accurate calculation of coal content in terms of material and mass dimensions, effectively overcoming the limitations of single detection technologies.

[0065] S40: Using a coal probability recognition engine, identify each particle in the target material based on the particle volume distribution, particle equivalent thickness distribution, particle surface feature distribution, particle material feature distribution, and particle mass distribution, and output a predicted coal probability distribution.

[0066] After the desliming screen mixture is sprayed with water, the appearance of some coal gangue is easily confused. In actual working conditions, the volume, thickness, surface characteristics, material and mass distribution of the materials vary in complexity. A single recognition model is difficult to balance the prediction accuracy and real-time performance under different working conditions.

[0067] To address the aforementioned issues, this application utilizes a coal probability recognition engine to identify each particle in the target material based on the particle volume distribution, particle equivalent thickness distribution, particle surface feature distribution, particle material feature distribution, and particle mass distribution, and outputs a predicted coal probability distribution.

[0068] Specifically, step S40 in the method includes: Based on historical coal content detection records from desliming screens, we collected sample particle volume distribution sets, sample particle equivalent thickness distribution sets, sample particle surface feature distribution sets, sample particle material feature distribution sets, and sample particle mass distribution sets. We then used the historical particle coal content ratio distributions corresponding to different sample particle volume distributions, sample particle equivalent thickness distributions, sample particle surface feature distributions, sample particle material feature distributions, and sample particle mass distributions as sample coal probability distributions to obtain sample coal probability distribution sets. The sample particle volume distribution set, sample particle equivalent thickness distribution set, sample particle surface feature distribution set, sample particle material feature distribution set, sample particle mass distribution set, and sample coal probability distribution set are used as sample training data. The sample training data is divided into P equal parts and randomly iterated with replacement to obtain P sample training sets, where P is an integer greater than or equal to 20. Using the P sample training sets, deep learning models are trained until convergence, resulting in P coal probability identification plugins. The P coal probability identification plugins are then integrated and fused according to the mean fusion strategy to generate a coal probability identification engine.

[0069] In this embodiment, firstly, based on historical coal content detection records of desliming screens, a set of sample particle volume distribution, a set of sample particle equivalent thickness distribution, a set of sample particle surface feature distribution, a set of sample particle material feature distribution, and a set of sample particle mass distribution are collected. Then, the historical particle coal content distribution corresponding to different sample particle volume distribution, sample particle equivalent thickness distribution, sample particle surface feature distribution, sample particle material feature distribution, and sample particle mass distribution is used as the sample coal probability distribution to obtain the sample coal probability distribution set. Here, the historical particle coal content distribution refers to the actual coal content probability distribution extracted from historical detection data corresponding to a certain set of sample particle volume distribution, sample particle equivalent thickness distribution, sample particle surface feature distribution, sample particle material feature distribution, and sample particle mass distribution. In this way, a precise mapping sample of five types of material distribution combinations and actual coal content probability distributions is constructed, providing reliable supervised training data for the coal probability recognition engine and ensuring that the model training direction is highly consistent with the actual working conditions of the desliming screen.

[0070] Secondly, the sample training data is obtained using the sample particle volume distribution set, sample particle equivalent thickness distribution set, sample particle surface feature distribution set, sample particle material feature distribution set, sample particle mass distribution set, and sample coal probability distribution set. This sample training data is divided into P equal parts, and random iterations with replacement are performed to obtain P sample training sets, where P is an integer greater than or equal to 20. For example, the complete sample training data is divided into P equal parts. Assuming there are N sets of sample training data, each part contains N / P sets of data. Then, P independent sampling processes are initiated. Each sampling randomly selects one sample from the P parts and includes it in the current training set. After sampling, the sample part is returned to the original sample pool to ensure equal probability in each sampling. When the sample size reaches N sets, sampling stops, forming one sample training set. This process is repeated P times to obtain P mutually independent sample training sets with some sample overlap.

[0071] Finally, using a training set of P samples, deep learning models are trained until convergence, resulting in P coal probability identification plugins. These P coal probability identification plugins are then integrated and fused according to the mean fusion strategy to generate a coal probability identification engine. For example, the coal probability recognition engine can be constructed through the following technical path: 1. Model construction: P coal probability recognition plugins with completely identical structures can be constructed using a lightweight convolutional neural network (CNN). Each coal probability recognition plugin consists of an input layer, a feature encoding layer, a multi-dimensional feature fusion layer, and an output layer. The input layer converts the 5 types of distribution data into a 64-dimensional feature vector and normalizes it through a LayerNorm layer. The feature encoding layer extracts deep correlation features layer by layer through 3 convolution and pooling modules. The multi-dimensional feature fusion layer uses an attention mechanism to weight and fuse multi-scale features, compresses the dimensions through a fully connected layer, and embeds a Dropout layer to suppress overfitting. The output layer outputs the probability of a single coal particle and the proportion of coal in the interval through parallel fully connected branches, and maps it to the [0,1] interval through a Sigmoid activation function to form a coal probability distribution output vector.

[0072] 2. Model Training: The training process for all P coal probability identification plugins is consistent. Taking the training process of any coal probability identification plugin as an example: Randomly select one sample training set from the P sample training sets and divide it into training set, validation set, and test set in a ratio of 7:1.5:1.5. Use the sample particle volume distribution, sample particle equivalent thickness distribution, sample particle surface feature distribution, sample particle material feature distribution, and sample particle mass distribution in the training set as input features, and the corresponding sample coal probability distribution set as supervision label. Use the AdamW optimizer with an initial learning rate of 1e-4 and a weight decay of 1e-5, and weighted mean square error (MSE) loss for training. Set the maximum number of training rounds to 100 and enable the early stopping strategy (Patience=8). When the validation set loss does not decrease for 8 consecutive rounds and the test set prediction MSE is lower than 0.03 and the accuracy is higher than 92%, it is considered converged. Save the parameters to obtain one trained coal probability identification plugin, and train P coal probability identification plugins in the same way.

[0073] 3. Model Integration: The P coal probability identification plugins are integrated and fused according to the mean fusion strategy to generate a coal probability identification engine. The structure and parameters of all coal probability identification plugins are solidified to form a unified system. When new five types of distribution data are input, the coal probability identification engine can simultaneously call the P coal probability identification plugins, and each coal probability identification plugin independently outputs the coal probability distribution prediction result. Then, the arithmetic mean of the single particle coal probability, the proportion of coal in the interval, and other corresponding dimensions of the P prediction results are calculated one by one and integrated into a stable coal probability distribution result. In this way, by offsetting the random prediction error of individual coal probability identification plugins and reducing the generalization variance, the prediction accuracy and stability of the coal probability identification engine under different working conditions are ensured.

[0074] Thus, the coal probability identification engine, through the integration of P coal probability identification plugins, is more resistant to sample noise interference than the prediction of a single coal probability identification plugin, effectively improving the prediction accuracy, stability, and generalization ability of the coal probability distribution of the target material and the complex working conditions of desliming screen.

[0075] Furthermore, the phrase "using a coal probability recognition engine to identify each particle in the target material based on the particle volume distribution, particle equivalent thickness distribution, particle surface feature distribution, particle material feature distribution, and particle mass distribution, and outputting a predicted coal probability distribution" includes: The uniformity of the particle volume distribution, particle equivalent thickness distribution, particle surface feature distribution, particle material feature distribution, and particle mass distribution is calculated respectively, and the uniformity coefficients of volume distribution, thickness distribution, surface feature distribution, material feature distribution, and mass distribution are output. The overall particle distribution uniformity coefficient is obtained by weighting the volume distribution uniformity coefficient, thickness distribution uniformity coefficient, surface feature distribution uniformity coefficient, material feature distribution uniformity coefficient, and mass distribution uniformity coefficient. The ratio of the preset standard particle overall distribution uniformity coefficient to the particle overall distribution uniformity coefficient is multiplied by the initial number of selected plug-ins and rounded to obtain the number of selected plug-ins K, where the initial number of selected plug-ins is 8, and K is greater than or equal to 3 and less than or equal to P. K coal probability identification plugins are randomly selected from the P coal probability identification plugins of the coal probability identification engine. Based on the particle volume distribution, particle equivalent thickness distribution, particle surface feature distribution, particle material feature distribution and particle mass distribution, each material particle in the target material is identified. The mean of the K identification results is fitted to obtain the predicted coal probability distribution.

[0076] In this embodiment, the uniformity of particle volume distribution, particle equivalent thickness distribution, particle surface feature distribution, particle material feature distribution, and particle mass distribution is first calculated, and the uniformity coefficients of volume distribution, thickness distribution, surface feature distribution, material feature distribution, and mass distribution are output. Here, uniformity refers to the degree of concentration of a certain type of distribution data; the more concentrated the distribution, the higher the uniformity, and the more dispersed the distribution, the lower the uniformity.

[0077] For example, the variances of particle volume distribution, particle equivalent thickness distribution, particle surface feature distribution, particle material feature distribution, and particle mass distribution can be calculated separately. Then, the reciprocal of the variance is calculated and normalized to the [0,1] interval to obtain the uniformity coefficients for volume distribution, thickness distribution, surface feature distribution, material feature distribution, and mass distribution. For instance, if the particle volume distribution is 20% for 0-5cm³, 30% for 5-10cm³, 35% for 10-15cm³, and 15% for over 15cm³, the variance of the particle volume distribution is first calculated to be 62.5%². Then, the reciprocal of the variance is calculated and normalized to obtain the uniformity coefficient for volume distribution, such as 0.64. The normalization logic can be based on the reciprocal of the largest variance of similar distributions to set a benchmark value. The uniformity coefficients for thickness distribution, surface feature distribution, material feature distribution, and mass distribution can be calculated using the same logic and method.

[0078] Secondly, the overall particle distribution uniformity coefficient is calculated by weighting the volume distribution uniformity coefficient, thickness distribution uniformity coefficient, surface feature distribution uniformity coefficient, material feature distribution uniformity coefficient, and mass distribution uniformity coefficient. The sum of the weights of these four coefficients is 1.

[0079] For example, based on actual needs and the degree of influence of each distribution on coal probability identification, the weights of the volume distribution uniformity coefficient, thickness distribution uniformity coefficient, surface feature distribution uniformity coefficient, material feature distribution uniformity coefficient, and mass distribution uniformity coefficient can be preset to 0.15, 0.15, 0.2, 0.3, and 0.2, respectively. Then, a weighted sum is performed to obtain the overall particle distribution uniformity coefficient. The overall particle distribution uniformity coefficient reflects the concentration and mixing degree of the target material composition. The larger the overall particle distribution uniformity coefficient, the more concentrated the volume, thickness, surface features, material, and mass distribution of the material, the stronger the consistency of particles with different attributes, and the simpler the working conditions corresponding to the desliming screen. Conversely, the smaller the coefficient, the more mixed the material composition, the more complex the working conditions, and the more coal probability identification plugins are needed for prediction to improve prediction accuracy.

[0080] For example, if the uniformity coefficients of volume distribution, thickness distribution, surface feature distribution, material feature distribution, and mass distribution are 0.64, 0.75, 0.7, 0.85, and 0.78 respectively, and then the preset weights are substituted into the above values, the overall particle distribution uniformity coefficient is obtained by weighted summation: 0.15×0.64+0.15×0.75+0.2×0.7+0.3×0.85+0.2×0.78=0.7595. That is, the current overall particle distribution uniformity coefficient is approximately 0.7595, indicating that the overall working condition of the material is relatively simple.

[0081] Next, the ratio of the preset standard particle overall distribution uniformity coefficient to the particle overall distribution uniformity coefficient is multiplied by the initial number of selected plug-ins and rounded to obtain the number of compatible plug-ins selected, K. The initial number of selected plug-ins is 8, and K is greater than or equal to 3 and less than or equal to P. The preset standard particle overall distribution uniformity coefficient refers to a pre-set overall uniformity coefficient corresponding to a uniform material composition and ideal operating conditions. For example, setting the preset standard particle overall distribution uniformity coefficient to 0.8 serves as a benchmark for measuring the complexity of the current operating conditions. Those skilled in the art can dynamically set this coefficient according to the actual operating conditions of the desliming screen, material characteristics, and other requirements.

[0082] The initial number of plugins selected is preset to 8. This value is a reasonable value determined after comprehensively balancing prediction accuracy and detection real-time performance. K must satisfy the constraint 3≤K≤P. The lower limit of 3 ensures prediction accuracy and avoids excessive random errors due to too few coal probability identification plugins. The upper limit P is the total number of all coal probability identification plugins in the coal probability identification engine. The ratio of the preset standard particle overall distribution uniformity coefficient to the particle overall distribution uniformity coefficient is calculated because a smaller particle overall distribution uniformity coefficient indicates a more mixed material composition and more complex working conditions. A larger number of compatible plugins K is selected to cover complex distributions through the diversity of coal probability identification plugins and ensure prediction accuracy. Conversely, a larger particle overall distribution uniformity coefficient indicates a more concentrated material composition and simpler working conditions. A smaller number of compatible plugins K is selected to reduce the computational load of the coal probability identification plugins and ensure detection efficiency.

[0083] For example, if the preset standard particle overall distribution uniformity coefficient is 0.8 and the current particle overall distribution uniformity coefficient is 0.7595, then the number of adapter plugins selected, K = 0.8 / 0.7595×8 =9, which is within the constraint range of 3≤K≤P. This means that the collaborative prediction of 9 coal probability identification plug-ins is required to cover the distribution characteristics of the current working condition and ensure the accuracy of coal probability prediction. At the same time, it does not increase the unnecessary computational load due to the excessive number of plug-ins, and takes into account the real-time requirements.

[0084] Finally, K coal probability identification plugins are randomly selected from the P coal probability identification plugins in the coal probability identification engine. Based on particle volume distribution, particle equivalent thickness distribution, particle surface feature distribution, particle material feature distribution, and particle mass distribution, each particle in the target material is identified. The K identification results are then subjected to mean fitting to obtain the predicted coal probability distribution. Mean fitting involves taking the arithmetic mean of the identification results output by the K coal probability identification plugins to obtain the final predicted coal probability distribution. This reduces the random error of individual coal probability identification plugins and improves the stability and accuracy of the prediction.

[0085] For example, if K=9, 9 are randomly selected from 20 coal probability recognition plugins. The particle volume distribution, particle equivalent thickness distribution, particle surface feature distribution, particle material feature distribution, and particle mass distribution are input. The 9 plugins predict the coal probability of a certain particle as 92%, 90%, 93%, ..., 89%, respectively. The average of these values ​​is taken: (92%+90%+93%+...+89%) / 8≈91%, and the predicted coal probability distribution of the particle is 91%.

[0086] In summary, compared to existing technologies, this application utilizes a coal probability recognition engine to identify each particle in the target material based on its particle volume distribution, equivalent thickness distribution, surface feature distribution, material properties distribution, and mass distribution, and outputs a predicted coal probability distribution. Thus, by fusing five types of data—particle volume distribution, equivalent thickness distribution, surface feature distribution, material properties distribution, and mass distribution—the coal probability recognition engine accurately identifies the coal properties of each particle, stably outputs a predicted coal probability distribution, balances prediction accuracy and real-time performance under different operating conditions, effectively reduces the impact of sample noise, and provides reliable data support for subsequent accurate calculation of coal content.

[0087] S50: The total coal content of the target material is calculated by weighting the particle mass distribution and the predicted coal probability distribution, and the detection reliability is evaluated. The total coal content and the detection reliability are used as the coal content detection results of the desliming screen.

[0088] The core requirement for coal content detection in desliming screens is to provide accurate and reliable coal content data for production control. However, traditional detection methods are prone to systematic bias in particle mass estimation and lack credibility assessment criteria for single coal content results. The aforementioned steps have already obtained particle mass distribution and predicted coal probability distribution, and the total coal content can be calculated by weighting, while the credibility of the detection can be evaluated simultaneously to quantify the reliability of the results.

[0089] To address the aforementioned issues, this application calculates the total coal content of the target material using a weighted average of the particle mass distribution and the predicted coal probability distribution, assesses the reliability of the detection, and uses the total coal content and the detection reliability as the coal content detection results of the desliming screen.

[0090] Specifically, step S50 in the method includes: A mass deviation analysis is performed based on the historical total mass sequence of particles and the historical total mass sequence of detected particles within a preset historical time zone. A mass correction factor is calculated based on the mass estimation deviation. The historical total mass of detected particles is the actual particle mass obtained by weighing equipment. The particle mass distribution is compensated using the aforementioned mass correction factor to generate a compensated particle mass distribution; The total coal content of the target material is calculated by weighting the compensated particle mass distribution and the predicted coal probability distribution.

[0091] In this embodiment, a quality deviation analysis is first performed based on the historical total particle mass sequence and the historical total particle mass sequence detected within a preset historical time zone. A quality correction factor is calculated based on the quality estimation deviation. The historical total particle mass is the actual particle mass obtained through weighing equipment. The preset historical time zone refers to a pre-defined, statistically significant historical detection period, such as the most recent 24 hours or the time period corresponding to the most recent 100 sets of detection data, ensuring a sufficient sample size to reflect common deviation patterns. The historical total particle mass sequence refers to the data sequence composed of the total mass of each group of materials calculated using the particle mass distribution method within the preset historical time zone. The historical total particle mass sequence also refers to the sequence of the actual total mass of each group of materials directly measured using weighing equipment, such as electronic belt scales or high-precision platform scales, within the preset historical time zone.

[0092] For example, the ratio of estimated mass to actual mass can be calculated by aligning the historical total mass sequence of particles and the historical total mass sequence of detected particles within a preset historical time zone. The mean of all ratios is then calculated as the mass deviation analysis result. The reciprocal of this result is then calculated to obtain the mass correction factor. The mass deviation analysis result represents the average deviation factor between the estimated and actual masses, reflecting the overall degree of deviation. A mass deviation analysis result closer to 1 indicates a smaller deviation, a closer mass correction factor to 1, and a smaller compensation magnitude. Conversely, a larger difference between the mass deviation analysis result and 1 indicates a more significant deviation, and a more prominent compensation effect of the mass correction factor.

[0093] For example, if the historical total particle mass sequence is [102kg, 98kg, 101kg, ..., 103kg], and the historical total detected particle mass sequence is [100kg, 100kg, 99kg, ..., 100kg], first align and calculate the ratio of estimated mass to actual mass one by one, such as 102kg / 100kg=1.02, 98kg / 100kg=0.98, 101kg / 99kg≈1.0202, etc., then calculate the mean of all ratios to obtain the mass deviation analysis result of 1.00804. Then calculate the reciprocal of the mass deviation analysis result to obtain the mass correction factor = 1 / 1.00804 = 0.992.

[0094] Secondly, a mass correction factor is used to compensate for the particle mass distribution, generating a compensated particle mass distribution. Specifically, the compensation process follows the principle of adjusting the mass of each particle individually. The estimated mass of each material particle is extracted from the particle mass distribution, and then each estimated mass is multiplied by the mass correction factor to obtain the compensated particle mass. Finally, the compensated particle masses of all particles are recalculated according to their original mass ranges to form the compensated particle mass distribution. This corrects the systematic deviation between the model's estimated mass and the actual mass, ensuring more accurate mass data for each particle.

[0095] For example, if the estimated mass of a certain particle in the particle mass distribution is 10g and the mass correction factor is 0.992, then the compensated particle mass = 10g × 0.992 ≈ 9.92g, and the particle mass distribution is compensated in the same way to generate the compensated particle mass distribution.

[0096] Finally, the total coal content of the target material is calculated by weighting the compensated particle mass distribution and the predicted coal probability distribution. Specifically, the compensated particle mass and predicted coal probability of the corresponding particles are first extracted from the compensated particle mass distribution and the predicted coal probability distribution. The product of the two is calculated to obtain the coal mass contribution value of a single particle, representing the estimated mass of the particle that is identified as coal. Then, the contribution values ​​of all particles are summed to obtain the estimated total coal mass in the target material. At the same time, the sum of the compensated masses of all particles is calculated to obtain the total compensated mass of the target material. Finally, the estimated total coal mass is divided by the total compensated mass and multiplied by 100% to obtain the total coal content of the target material.

[0097] For example, suppose the target material contains three representative particles: Particle 1, Particle 2, and Particle 3, with compensated particle masses of 9.92g, 19.84g, and 4.96g, respectively, and corresponding predicted coal probabilities of 91%, 34%, and 89%, respectively. Then, the coal mass contribution value of a single particle is calculated as follows: Coal mass contribution value of Particle 1 = 9.92g × 91% ≈ 9.027g, Coal mass contribution value of Particle 2 = 19.84g × 34% ≈ 6.746g, Coal mass contribution value of Particle 3 = 4.96g × 89% ≈ 4.414g. Further calculations yielded: the estimated total mass of coal = 9.027g + 6.746g + 4.414g ≈ 20.187g, and the total compensation mass of the target material was calculated as = 9.92g + 19.84g + 4.96g ≈ 34.72g. Finally, the total coal content was calculated as = (20.187g / 34.72g) × 100% ≈ 58.1%. This weighted calculation ensures that the total coal content accurately reflects the proportion of coal by mass, aligning with actual testing requirements.

[0098] Furthermore, the step of "assessing the reliability of the detection" includes: Calculate the mean values ​​of the particle volume distribution, particle equivalent thickness distribution, and particle mass distribution to obtain the mean particle volume, mean particle equivalent thickness, and mean particle mass. The deviations of the average particle volume, average particle equivalent thickness, and average particle mass from the preset particle standard values ​​are calculated respectively to obtain the particle volume deviation, particle equivalent thickness deviation, and particle mass deviation. The detection reliability is obtained by comprehensively evaluating the particle volume deviation range, particle equivalent thickness deviation range, particle mass deviation range, particle overall distribution uniformity coefficient, and mass estimation deviation. The detection reliability is negatively correlated with the particle volume deviation range, particle equivalent thickness deviation range, particle mass deviation range, and mass estimation deviation, and positively correlated with the particle overall distribution uniformity coefficient.

[0099] In this embodiment, the mean values ​​of particle volume distribution, particle equivalent thickness distribution, and particle mass distribution are first calculated to obtain the mean particle volume, mean particle equivalent thickness, and mean particle mass. The mean particle volume reflects the overall volume of the material particles, the mean particle equivalent thickness reflects the overall average equivalent thickness of the particles, and the mean particle mass reflects the overall mass level of the particles.

[0100] Specifically, the mean particle volume, mean particle equivalent thickness, and mean particle mass can be calculated using the interval midpoint weighted average method. That is, based on the interval division of particle volume distribution, particle equivalent thickness distribution, and particle mass distribution, the midpoint value of each interval is taken as the representative value of the particles in that interval, and then multiplied by the proportion of that interval. After summing, the corresponding mean particle volume, mean particle equivalent thickness, and mean particle mass are obtained.

[0101] For example, taking particle volume distribution as an example, if the particle volume distribution is 20% for 0-5cm³, 30% for 5-10cm³, 35% for 10-15cm³, and 15% for above 15cm³, with the midpoints of the intervals being 2.5cm³, 7.5cm³, 12.5cm³, and 17.5cm³ respectively, then using the interval midpoint weighted average method, the average particle volume is calculated as 2.5×20%+7.5×30%+12.5×35%+17.5×15%=9.75cm³. Similarly, the average equivalent particle thickness and average particle mass can be obtained using the same calculation.

[0102] Secondly, the deviations of the average particle volume, average particle equivalent thickness, and average particle mass from the preset particle standard values ​​are calculated to obtain the particle volume deviation, particle equivalent thickness deviation, and particle mass deviation. The preset particle standard values ​​refer to the typical particle size, thickness, and mass of materials conventionally processed by the desliming screen, such as coal and gangue. These preset particle volume standard values, equivalent thickness standard values, and particle mass standard values ​​can be dynamically adjusted and determined by those skilled in the art based on the characteristics of the actual coal type, gangue properties, and production conditions to ensure a high degree of compatibility between the standard values ​​and the actual processing scenario.

[0103] For example, the relative deviations of the average particle volume, average particle equivalent thickness, and average particle mass from the preset particle standard values ​​can be calculated separately, and these deviations can be used as the particle volume deviation, particle equivalent thickness deviation, and particle mass deviation. The relative deviation is calculated as follows: Relative deviation = |(Average - Preset particle standard value) / Preset particle standard value| × 100%.

[0104] For example, considering the characteristics of bituminous coal and sandstone gangue processed daily by the desliming screen, and determining the preset particle standard values ​​as 10 cm³ for particle volume, 1.2 cm for equivalent thickness, and 12 g for particle mass, when the average particle volume is 9.75 cm³, the particle volume deviation range = |(10-9.75) / 10|×100%≈2.5%. Similarly, the particle equivalent thickness deviation range, such as 8.33%, and the particle mass deviation range, such as 2.08%, can be calculated using the same logic and method. The particle volume deviation range, particle equivalent thickness deviation range, and particle mass deviation range quantify the degree of deviation of the overall volume, equivalent thickness, and mass of the current material from the properties of conventional materials.

[0105] Finally, the detection reliability is obtained by comprehensively evaluating the particle volume deviation amplitude, particle equivalent thickness deviation amplitude, particle mass deviation amplitude, particle overall distribution uniformity coefficient, and mass estimation deviation. Among them, the detection reliability is negatively correlated with the particle volume deviation amplitude, particle equivalent thickness deviation amplitude, particle mass deviation amplitude, and mass estimation deviation, and positively correlated with the particle overall distribution uniformity coefficient.

[0106] The detection reliability refers to the degree of reliability of the coal content detection result, with a value range of [0,1], where 0 represents completely unreliable and 1 represents completely reliable. The detection reliability is negatively correlated with particle volume deviation, particle equivalent thickness deviation, particle mass deviation, and mass estimation deviation because the larger these deviations are, the more significant the difference between the current material properties and conventional standard materials, or the more obvious the deviation between the mass estimation data and the actual value. This increases the prediction error of the coal probability recognition engine based on conventional patterns, naturally reducing the reliability of the detection result. Conversely, the detection reliability is positively correlated with the overall particle distribution uniformity coefficient because a larger coefficient indicates a more concentrated distribution of material attributes such as volume, thickness, and material properties, resulting in simpler and more stable operating conditions. This makes it easier for the coal probability recognition engine to accurately capture material patterns, reduce prediction errors, and thus increase the reliability of the detection result.

[0107] For example, if the particle volume deviation, particle equivalent thickness deviation, particle mass deviation, particle overall distribution uniformity coefficient, and mass estimation deviation are 2.5%, 8.33%, 2.08%, 0.7595, and 1.00804 respectively, and their respective weights are 0.2, 0.2, 0.2, 0.3, and 0.1, then the detection reliability = , and the detection reliability comprehensively reflects .

[0108] For example, a comprehensive evaluation can be conducted using a method of reverse processing of negatively correlated indicators and weighted score normalization. The steps are as follows: First, assign weights to the particle volume deviation amplitude, particle equivalent thickness deviation amplitude, particle mass deviation amplitude, particle overall distribution uniformity coefficient, and mass estimation deviation, such as 0.2, 0.2, 0.2, 0.3, and 0.1, with the sum of all weights being 1. The weight allocation can be adjusted according to the actual detection priority. Second, indicator preprocessing: For the particle volume deviation amplitude, particle equivalent thickness deviation amplitude, particle mass deviation amplitude, and mass estimation deviation, first normalize them to [0,1] according to the actual value / preset maximum reasonable value of the indicator, and then pass the normalization process. The 1-normalized value is converted into a positive score. The higher the positive score, the greater its contribution to the credibility. The particle overall distribution uniformity coefficient is directly used as the positive score based on the original value. The third step is to calculate the weighted total score: Detection credibility = w1×S1+w2×S2+w3×S3+w4×S4+w5×S5, where w1, w2, w3, w4, and w5 are the weights of particle volume deviation, particle equivalent thickness deviation, particle mass deviation, particle overall distribution uniformity coefficient, and mass estimation deviation, respectively. S1, S2, S3, S4, and S5 are the positive scores of the five indicators. The final result is normalized to the [0,1] interval.

[0109] For example, based on the data above, the particle volume deviation, particle equivalent thickness deviation, particle mass deviation, particle overall distribution uniformity coefficient, and mass estimation deviation are 2.5%, 8.33%, 2.08%, 0.7595, and 1.00804, respectively, with their respective weights of 0.2, 0.2, 0.2, 0.3, and 0.1. The preset maximum reasonable values ​​for each negatively correlated indicator are: deviation amplitude 20% and relative deviation of mass estimation 10%. Based on the above calculation logic, the specific calculations are as follows: 1. Positive score calculation: S1=1-(2.5% / 20%)=0.875, S2=1-(8.33% / 20%)≈0.5835, S3=1-(2.08% / 20%)=0.896, S4=0.7595 (positively correlated, directly used), S5=1-(0.804% / 10%)≈0.9196; ② Weighted total score calculation: Detection reliability = 0.2×0.875+0.2×0.5835+0.2×0.896+0.3×0.7595+0.1×0.9196≈0.7908. The detection reliability comprehensively reflects the impact of the stability of current material properties and the accuracy of data estimation on the coal content detection results, and intuitively reflects the reliability of this detection result.

[0110] Furthermore, the total coal content (e.g., 58.1%) and the detection confidence level (e.g., 0.7908) are used as the coal content detection results for the desliming screen. The total coal content directly reflects the mass proportion of coal in the target material, while the detection confidence level quantifies the reliability of the coal content result. This combined result can provide a tiered basis for the operation control of the desliming screen and the optimization of separation system parameters. For example, when the detection confidence level is high (e.g., ≥0.8), the total coal content data can be directly used to guide production adjustments; when the confidence level is low (e.g., <0.6), a data verification process is triggered (e.g., recalibrating the quality correction factor and supplementing sample training) to avoid production decision deviations due to unreliable data and ensure a deep fit between the detection results and actual production needs.

[0111] In summary, compared to existing technologies, this application calculates the total coal content of the target material based on a weighted average of the particle mass distribution and the predicted coal probability distribution, and evaluates the reliability of the detection. The total coal content and the detection reliability are then used as the coal content detection result for the desliming screen. Thus, obtaining a coal content detection result for the desliming screen that includes both the total coal content and the detection reliability can directly reflect the reliability of the prediction process, providing a precise basis for the production control of the desliming screen and effectively overcoming the shortcomings of inaccurate and unreliable traditional detection data.

[0112] In summary, the embodiments of this application have at least the following technical effects: Compared to existing technologies, this application first utilizes a rotary encoder installed on the drive drum to trigger and simultaneously acquire binocular visible light image pairs and dual-energy X-ray images of the target material on the desliming screen. In this way, multi-source information of the target material is acquired simultaneously and uniformly registered, avoiding problems such as timing misalignment or material region mismatch, providing a reliable data foundation for subsequent multimodal feature fusion, coal gangue identification, and accurate calculation of coal content.

[0113] Secondly, this application, based on binocular visible light image pairs, calculates the volume and equivalent thickness of each particle in the target material, identifies surface features, and generates particle volume distribution, particle equivalent thickness distribution, and particle surface feature distribution. In this way, the volume, equivalent thickness, and surface features such as color, texture, and morphology of each particle in the target material are successfully extracted, and corresponding three types of distribution data are generated. This provides spatial and appearance dimension data support for subsequent calculations, effectively offsetting the interference of thickness effects on detection and overcoming the limitations of single detection technologies.

[0114] Furthermore, based on the aforementioned dual-energy X-ray images, particle volume distribution, and particle equivalent thickness distribution, this application identifies the material characteristics of each particle, calculates the areal density and equivalent density of each particle, and estimates the particle mass, generating particle material characteristic distribution and particle mass distribution. In this way, the material characteristics of each particle can be accurately identified, areal density and equivalent density calculations and particle mass estimations can be completed, providing reliable data support for the subsequent accurate calculation of coal content in terms of material and mass dimensions, effectively overcoming the limitations of single detection technologies.

[0115] Furthermore, this application utilizes a coal probability recognition engine to identify each particle in the target material based on the particle volume distribution, particle equivalent thickness distribution, particle surface feature distribution, particle material feature distribution, and particle mass distribution, and outputs a predicted coal probability distribution. Thus, by fusing five types of data—particle volume distribution, particle equivalent thickness distribution, particle surface feature distribution, particle material feature distribution, and particle mass distribution—the coal probability recognition engine accurately identifies the coal properties of each particle, stably outputs a predicted coal probability distribution, balances prediction accuracy and real-time performance under different operating conditions, effectively reduces the impact of sample noise, and provides reliable data support for subsequent accurate calculation of coal content.

[0116] Finally, this application calculates the total coal content of the target material based on the weighted average of the particle mass distribution and the predicted coal probability distribution, and evaluates the detection reliability. The total coal content and the detection reliability are then used as the coal content detection result for the desliming screen. In this way, obtaining the coal content detection result for the desliming screen, which includes both the total coal content and the detection reliability, can intuitively reflect the reliability of the prediction process, providing a precise basis for the production control of the desliming screen and effectively overcoming the shortcomings of inaccurate and unreliable traditional detection data.

[0117] Through the above technical solution, this application effectively solves the problems of low accuracy, large interference from working conditions, and single-technology limitations of traditional detection methods: By synchronously acquiring binocular visible light image pairs and dual-energy X-ray images through a rotary encoder, the timeliness of data is ensured. Based on the binocular visible light image pairs and dual-energy X-ray images, particle volume, equivalent thickness, surface feature distribution, material characteristics, and mass distribution are obtained. Then, the coal probability recognition engine integrates multi-dimensional distribution data to output a reliable predicted coal probability distribution. Finally, the accurate total coal content is obtained through weighted calculation and a corresponding detection credibility assessment is provided, forming a complete detection link of real-time acquisition, multi-dimensional analysis, accurate identification, and reliable output. This provides coal content data support with accuracy, real-time performance, and reliability for the production control of desliming screens, and fully adapts to the needs of modern coal production.

[0118] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0119] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0121] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0122] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0123] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0124] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for detecting coal content using a desliming screen based on binocular vision and dual-energy X-ray fusion, characterized in that the method... include: Triggered by a rotary encoder installed on the drive drum, the binocular visible light image pair and dual-energy X-ray image of the target material on the desliming screen are acquired simultaneously. Based on the binocular visible light image pair, the volume and equivalent thickness of each particle in the target material are calculated, and surface features are identified to generate particle volume distribution, particle equivalent thickness distribution, and particle surface feature distribution. Based on the dual-energy X-ray images, particle volume distribution, and particle equivalent thickness distribution, the material characteristics of each material particle are identified, the areal density and equivalent density of each material particle are calculated, and the particle mass is estimated to generate particle material characteristic distribution and particle mass distribution. Using a coal probability recognition engine, the particle size distribution, equivalent thickness distribution, surface feature distribution, material feature distribution, and mass distribution of the particles are used to identify each particle in the target material, and a predicted coal probability distribution is output. The total coal content of the target material is calculated by weighting the particle mass distribution and the predicted coal probability distribution, and the detection reliability is evaluated. The total coal content and the detection reliability are used as the coal content detection results of the desliming screen.

2. The method for detecting coal content using a desliming screen based on binocular vision and dual-energy X-ray fusion as described in claim 1, characterized in that, Triggered by a rotary encoder mounted on the drive drum, binocular visible light images and dual-energy X-ray images of the target material on the desliming screen are simultaneously acquired, including: Using the pulse signal emitted by the rotary encoder installed on the drive drum, the binocular camera and dual-energy X-ray imaging device are synchronously triggered to acquire binocular visible light image pairs and dual-energy X-ray images of the target material at the same sampling time. By using a pre-calibrated extrinsic parameter matrix, the binocular visible light image pair and the dual-energy X-ray image are uniformly registered to the belt coordinate system, establishing the spatial correspondence between the pixels of each image.

3. The method for detecting coal content in a desliming screen based on binocular vision and dual-energy X-ray fusion as described in claim 1, characterized in that, Based on the binocular visible light image pair, the volume and equivalent thickness of each particle in the target material are calculated, and surface features are identified to generate particle volume distribution, particle equivalent thickness distribution, and particle surface feature distribution, including: Based on the binocular visible light image pair, a disparity map is obtained through a stereo matching algorithm, and depth information is calculated based on the camera focal length and the baseline distance. Using the surface of the desliming screen belt as a reference plane, and combining the depth information, a three-dimensional model of each material particle is constructed through a surface mesh reconstruction algorithm. The volume of each material particle is calculated based on the three-dimensional model of each material particle, and the particle volume distribution is constructed. Based on the volume of each material particle and its projected area on the image plane, as well as the angle between the X-ray beam and the normal, the equivalent thickness of each material particle in the X-ray path direction is calculated, and the particle equivalent thickness distribution is constructed. A particle surface feature recognition model is constructed using a convolutional neural network. The model is then used to extract features from the binocular visible light image pairs to obtain the surface features of each particle in the target material. A particle surface feature distribution is then constructed, wherein the surface features include color features, texture features, and morphological features.

4. The method for detecting coal content using a desliming screen based on binocular vision and dual-energy X-ray fusion as described in claim 1, characterized in that, Based on the dual-energy X-ray images, particle volume distribution, and particle equivalent thickness distribution, the material characteristics of each particle are identified, and the areal density and equivalent density of each particle are calculated. Particle mass is estimated, and particle material characteristic distribution and particle mass distribution are generated, including: Based on the dual-energy X-ray image, the attenuation value of each material particle in the target material under low-energy X-ray and high-energy X-ray is calculated, and according to the attenuation value, the material characteristics of each material particle are identified by the substrate decomposition model, and the particle material characteristic distribution is constructed. Based on the dual-energy X-ray image, the areal density of each particle in the target material is calculated using a dual-energy attenuation model. Based on the particle equivalent thickness distribution, the areal density of each material particle is divided by the corresponding equivalent thickness to obtain the equivalent density of each material particle, thus constructing the particle equivalent density distribution. Based on the particle volume distribution and particle equivalent density distribution, the mass of each material particle is estimated using a mass calculation formula to generate a particle mass distribution.

5. The method for detecting coal content using a desliming screen based on binocular vision and dual-energy X-ray fusion as described in claim 1, characterized in that, The method for constructing the coal probability identification engine includes: Based on historical coal content detection records from desliming screens, we collected sample particle volume distribution sets, sample particle equivalent thickness distribution sets, sample particle surface feature distribution sets, sample particle material feature distribution sets, and sample particle mass distribution sets. We then used the historical particle coal content ratio distributions corresponding to different sample particle volume distributions, sample particle equivalent thickness distributions, sample particle surface feature distributions, sample particle material feature distributions, and sample particle mass distributions as sample coal probability distributions to obtain sample coal probability distribution sets. The sample particle volume distribution set, sample particle equivalent thickness distribution set, sample particle surface feature distribution set, sample particle material feature distribution set, sample particle mass distribution set, and sample coal probability distribution set are used as sample training data. The sample training data is divided into P equal parts and randomly iterated with replacement to obtain P sample training sets, where P is an integer greater than or equal to 20. Using the P sample training sets, deep learning models are trained until convergence, resulting in P coal probability identification plugins. The P coal probability identification plugins are then integrated and fused according to the mean fusion strategy to generate a coal probability identification engine.

6. The method for detecting coal content using a desliming screen based on binocular vision and dual-energy X-ray fusion as described in claim 5, characterized in that, Using a coal probability recognition engine, the particle size distribution, equivalent thickness distribution, surface feature distribution, material feature distribution, and mass distribution of the target material are used to identify each particle, and a predicted coal probability distribution is output, including: The uniformity of the particle volume distribution, particle equivalent thickness distribution, particle surface feature distribution, particle material feature distribution, and particle mass distribution is calculated respectively, and the uniformity coefficients of volume distribution, thickness distribution, surface feature distribution, material feature distribution, and mass distribution are output. The overall particle distribution uniformity coefficient is obtained by weighting the volume distribution uniformity coefficient, thickness distribution uniformity coefficient, surface feature distribution uniformity coefficient, material feature distribution uniformity coefficient, and mass distribution uniformity coefficient. The ratio of the preset standard particle overall distribution uniformity coefficient to the particle overall distribution uniformity coefficient is multiplied by the initial number of selected plug-ins and rounded to obtain the number of selected plug-ins K, where the initial number of selected plug-ins is 8, and K is greater than or equal to 3 and less than or equal to P. K coal probability identification plugins are randomly selected from the P coal probability identification plugins of the coal probability identification engine. Based on the particle volume distribution, particle equivalent thickness distribution, particle surface feature distribution, particle material feature distribution and particle mass distribution, each material particle in the target material is identified. The mean of the K identification results is fitted to obtain the predicted coal probability distribution.

7. The method for detecting coal content using a desliming screen based on binocular vision and dual-energy X-ray fusion as described in claim 6, characterized in that, The total coal content of the target material is calculated by weighting the particle mass distribution and the predicted coal probability distribution, including: A mass deviation analysis is performed based on the historical total mass sequence of particles and the historical total mass sequence of detected particles within a preset historical time zone. A mass correction factor is calculated based on the mass estimation deviation. The historical total mass of detected particles is the actual particle mass obtained by weighing equipment. The particle mass distribution is compensated using the aforementioned mass correction factor to generate a compensated particle mass distribution; The total coal content of the target material is calculated by weighting the compensated particle mass distribution and the predicted coal probability distribution.

8. The method for detecting coal content in a desliming screen based on binocular vision and dual-energy X-ray fusion according to claim 7, characterized in that, The steps for assessing the reliability of a test include: Calculate the mean values ​​of the particle volume distribution, particle equivalent thickness distribution, and particle mass distribution to obtain the mean particle volume, mean particle equivalent thickness, and mean particle mass. The deviations of the average particle volume, average particle equivalent thickness, and average particle mass from the preset particle standard values ​​are calculated respectively to obtain the particle volume deviation, particle equivalent thickness deviation, and particle mass deviation. The detection reliability is obtained by comprehensively evaluating the particle volume deviation range, particle equivalent thickness deviation range, particle mass deviation range, particle overall distribution uniformity coefficient, and mass estimation deviation. The detection reliability is negatively correlated with the particle volume deviation range, particle equivalent thickness deviation range, particle mass deviation range, and mass estimation deviation, and positively correlated with the particle overall distribution uniformity coefficient.

Images