A method and system for ore sorting

Abstract

Embodiments of the present application provide a kind of ore sorting method and system, comprising: based on preset line sweep frequency, using different energy X-ray to the ore material placed in feed belt is carried out synchronous dynamic data acquisition, obtain high-energy perspective image and low-energy perspective image, and adaptive registration, obtain the target low-energy image of pixel level alignment with high-energy perspective image;After that, extract high-energy basic features and low-energy basic features;In spatial differential domain, extract first-order gradient ratio feature and second-order gradient ratio feature;In the first sorting model, the confidence level of multilevel is calculated, the target ore of confidence level greater than the first confidence level is selected, and the fuzzy ore block with confidence level in the preset confidence level range is input into the second sorting model to separate the target ore of fuzzy ore block using the second sorting model;Finally, the target ore is physically separated using a gas valve spray array. The method improves the actual physical action hit rate.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically to an ore sorting method and system within the field of data processing technology. Background Technology

[0002] With the continuous depletion of mineral resources, the grade of raw ore being processed is decreasing day by day. Sensor-based intelligent ore pre-disposal (sorting) technology has become a key means to reduce mineral processing energy consumption and improve resource utilization.

[0003] In related technologies, material classification is achieved by collecting color and texture features from the surface. However, in actual industrial settings, ore surfaces are often covered by dust, mud, or oxidized, and optical surface features often fail to reflect the true internal mineral composition of the ore. This results in a very high misclassification rate for such methods under harsh working conditions.

[0004] In recent years, researchers have attempted to introduce deep convolutional neural networks (such as the lightweight network LOSN) to improve sorting accuracy. Although deep learning models have powerful feature extraction capabilities, in high-throughput, real-time industrial valve blowing scenarios (typically requiring millisecond-level response), even lightweight models deployed on edge devices still face severe computational bottlenecks and latency issues when performing complex deep forward inference on all target blocks, making it difficult to achieve a balance between computational power and throughput.

[0005] Dual-energy X-ray transmission (DE-XRT) based sorting technology has become the most promising sorting method due to its ability to penetrate the material surface and directly detect differences in effective atomic numbers within the material. DE-XRT can overcome the interference of the "thickness effect" in ores to some extent by utilizing the attenuation difference between high- and low-energy rays. However, most existing DE-XRT sorting algorithms rely on simple dual-energy ratio calculations and fixed static threshold segmentation. When faced with complex ore associations, extreme thickness variations, and material overlap and adhesion in actual belt conveyor transport, simple ratio features often fail to completely eliminate thickness interference. Furthermore, the static, single threshold decision logic lacks adaptive learning capabilities based on field data, exhibits poor generalization, and is difficult to meet the complex and ever-changing demands of efficient industrial sorting. Summary of the Invention

[0006] The purpose of this invention is to provide an ore sorting method and system, and the specific technical solution adopted is as follows: In a first aspect, embodiments of the present invention provide an ore sorting method, the method comprising: Based on a preset line scan frequency, X-rays of different energies are used to synchronously and dynamically acquire data on the ore material placed on the feed belt, resulting in high-energy and low-energy X-ray images. Adaptive registration is performed on the high-energy perspective image and the low-energy perspective image to obtain a target low-energy image that is pixel-level aligned with the high-energy perspective image; High-energy feature extraction and low-energy feature extraction are performed on the target low-energy image to obtain high-energy basic features and low-energy basic features; In the spatial differential domain of the high-energy fundamental features and the low-energy fundamental features, the first-order gradient ratio feature and the second-order gradient ratio feature between the high-energy fundamental features and the low-energy fundamental features are extracted. In the first sorting model, multi-level confidence calculations are performed on the first-order gradient ratio feature and the second-order gradient ratio feature. Target ores with confidence levels greater than the first confidence level are selected from the ore materials. Fuzzy ore blocks with confidence levels within a preset confidence range are input into the second sorting model to sort the target ores using the second sorting model. The second sorting model is obtained by incrementally fine-tuning the network parameters based on fuzzy ore block samples with confidence levels within the preset confidence range, and the network complexity of the second sorting model is greater than that of the first sorting model. The preset confidence range is greater than the second confidence level and less than the first confidence level, and the second confidence level is less than the first confidence level. Based on the region of interest of the target ore and the distance from the X-ray scan line to the gas valve jet array, the gas valve jet array is used to physically separate the target ore.

[0007] Secondly, an ore sorting system is provided, the system comprising: The acquisition module is used to synchronously and dynamically acquire data on the ore material placed on the feed belt using X-rays of different energies based on a preset line scan frequency, so as to obtain high-energy X-ray images and low-energy X-ray images. The registration module is used to adaptively register the high-energy perspective image and the low-energy perspective image to obtain a target low-energy image that is pixel-level aligned with the high-energy perspective image. The first extraction module is used to extract high-energy features and low-energy features from the target low-energy image to obtain high-energy basic features and low-energy basic features. The second extraction module is used to extract the first-order gradient ratio feature and the second-order gradient ratio feature between the high-energy basic feature and the low-energy basic feature in the spatial differential domain of the high-energy basic feature and the low-energy basic feature. The sorting module is used in the first sorting model to perform multi-level confidence calculations on the first-order gradient ratio feature and the second-order gradient ratio feature. It selects target ores with a confidence level greater than the first confidence level from the ore material, and inputs fuzzy ore blocks with confidence levels within a preset confidence range into the second sorting model. The second sorting model is then used to sort the fuzzy ore blocks for target ores. The second sorting model is obtained by incrementally fine-tuning the network parameters based on fuzzy ore block samples with confidence levels within the preset confidence range, and the network complexity of the second sorting model is greater than that of the first sorting model. The preset confidence range is greater than the second confidence level and less than the first confidence level, where the second confidence level is less than the first confidence level. The jetting module is used to physically separate the target ore using the gas valve jetting array based on the region of interest of the target ore and the distance from the X-ray scan line to the gas valve jetting array.

[0008] Thirdly, a computer program product is provided, comprising: computer program code, which, when run on a computer, causes the computer to perform the method described in the first aspect or any possible implementation thereof.

[0009] Fourthly, a computer-readable storage medium is provided that stores computer program code, which, when executed on a computer, causes the computer to perform the methods described in the first aspect or any possible implementation thereof.

[0010] This invention offers the following advantages: Based on a preset line scan frequency, X-rays of different energies are used to synchronously and dynamically acquire data from ore materials placed on a feeding conveyor belt. This allows for the relatively accurate acquisition of high-energy and low-energy X-ray images, ensuring that the line scan frequencies of the high-energy and low-energy X-ray images are identical, facilitating subsequent image processing. Subsequently, adaptive registration is performed on the high-energy and low-energy X-ray images to obtain a target low-energy image that is pixel-level aligned with the high-energy X-ray image. This ensures that the target low-energy image, even in a low-energy system, still meets the pixel-level requirements of the high-energy X-ray image. By extracting high-energy and low-energy features from the target low-energy image, high-energy and low-energy basic features are obtained. Furthermore, in the spatial differential domain of these basic features, the first-order and second-order gradient ratio features are extracted. These features filter out low-frequency thickness interference caused by the smooth transition of the ore's outer contour, significantly amplifying the signal proportion of small associated minerals within the ore. Then, using the first sorting model, multi-level confidence calculations are performed on the first-order gradient ratio feature and the second-order gradient ratio feature. Target ores with a confidence level greater than the first confidence level are selected from the ore materials. The network parameters are incrementally fine-tuned using fuzzy ore block samples with confidence levels within a preset confidence range to obtain the trained sorting model, thus improving the classification accuracy of the trained model. Finally, the target ore is physically separated using the air valve jet array based on the distance from the region of interest of the target ore and the X-ray scan line to the air valve jet array. Thus, the algorithm inference delay is dynamically calculated based on the distance from the region of interest of the target ore and the X-ray scan line to the air valve jet array, and the opening pulse width and number of air valves are adaptively allocated according to the physical span of the ore. This hardware-software collaboration and dynamic compensation mechanism enables high-precision physical separation of the target ore, improving the actual physical action hit rate. Attached Figure Description

[0011] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram illustrating the implementation process of an ore sorting method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of another implementation process of an ore sorting method provided in an embodiment of the present invention; Figure 3This is a schematic diagram illustrating another implementation process of an ore sorting method provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of another implementation process of an ore sorting method provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the composition structure of an ore sorting system provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a computer block device provided in an embodiment of the present invention. Detailed Implementation

[0013] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following detailed description, in conjunction with the accompanying drawings and preferred embodiments, provides a method for ore sorting based on the present invention, including its specific implementation, structure, features, and effects. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments may be combined from any suitable form.

[0014] In the description of the embodiments of the present invention, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of the present invention, "multiple" means two or more.

[0015] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0016] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0017] In some embodiments, the mainstream detection and sorting technologies mainly include visible light / machine vision sorting, deep learning image sorting, and X-ray transmission sorting (XRT).

[0018] Real-time ore sorting technology based on visible light and surface features primarily relies on traditional machine vision. Its workflow involves acquiring real-time images of the ore surface using industrial red-green-blue (RGB) or multispectral cameras positioned above a high-speed conveyor belt. The algorithm then extracts color space features (such as RGB, hue, saturation, and brightness (HSV) components) and surface texture features (such as the gray-level co-occurrence matrix, GLCM) from the images. Finally, it classifies the extracted surface features using a support vector machine (SVM), the K-nearest neighbor algorithm (KNN), or a predefined color / grayscale threshold, thereby achieving the identification and separation of ore from waste rock. The core of this approach lies in utilizing the differences in optical reflectance on mineral surfaces for sorting.

[0019] To address the limitation of visible light in detecting the interior of materials, the industry has widely adopted basic X-ray transmission technology (XRT). This involves irradiating moving ore with a broadband X-ray source and using a linear array detector at the bottom to receive the residual X-ray intensity after penetration. Because minerals of different densities (such as high-density cassiterite and low-density gangue) have different absorption and attenuation coefficients for X-rays, the grayscale images generated by the detector can reflect the internal density distribution of the ore. The system performs median filtering and noise reduction on these grayscale images, segments connected components, and calculates the average grayscale value of the blocks. This average grayscale value is then compared to a preset single grayscale threshold. Blocks with values ​​below or above the threshold trigger a gas valve for separation.

[0020] To overcome the interference of ore thickness variations on single-energy X-ray attenuation, dual-energy X-ray transmission (DE-XRT) sorting technology was further developed. This technology simultaneously acquires two sets of X-ray transmission images: one high-energy and one low-energy. The core algorithm performs mathematical operations (usually direct division to obtain a ratio, or logarithmic subtraction) on the grayscale values ​​of the same pixel location or the same block region in both high- and low-energy images. Since the thickness variable is partially offset by the physical model during the ratio calculation, this ratio feature can more directly map the equivalent atomic number (Z-value) of the mineral. Subsequently, the system sets a classification threshold based on this dual-energy ratio feature to evaluate and separate target minerals such as sulfide minerals.

[0021] With the development of artificial intelligence, Convolutional Neural Network (CNN) technology has been introduced into ore image processing. To meet the extremely low latency requirements of industrial air valve injection, this type of solution typically designs a lightweight object detection or classification network for edge devices (such as embedded Graphics Processing Units (GPUs)). The implementation process includes: directly inputting the acquired ore image into the lightweight neural network; the network automatically extracts high-dimensional abstract features of the image through compression structures such as depthwise separable convolutions, and outputs a probability score indicating whether the object belongs to "ore" or "waste rock" through fully connected layers or a classification head. The system then drives the backend actuator to perform sorting based on the probability score.

[0022] However, sorting techniques based on visible light and traditional machine vision often fail to reflect the true internal mineral composition of ores, leading to extremely high misclassification rates under harsh working conditions. Sorting techniques based on deep learning and edge computing, while possessing powerful feature extraction capabilities, still face severe computational bottlenecks and latency issues in high-throughput, real-time industrial gas valve blowing scenarios (typically requiring millisecond-level response). Even lightweight models deployed on edge devices, performing complex deep forward inference on all target blocks, struggle to achieve a balance between computational power and throughput. Sorting techniques based on dual-energy X-ray transmission (DE-XRT), capable of penetrating the material surface and directly detecting differences in effective atomic numbers, have become the most promising sorting method. DE-XRT can overcome the interference of the "thickness effect" of ores to some extent through the attenuation difference between high- and low-energy rays. However, most existing DE-XRT sorting algorithms rely on simple dual-energy ratio calculations and fixed static threshold segmentation. When faced with complex ore associations, extreme thickness variations, and material overlap and adhesion in actual belt conveyor transportation, simple ratio features often cannot completely eliminate thickness interference. At the same time, static and single threshold decision logic lacks adaptive learning ability to field data, has poor generalization, and is difficult to meet the complex and ever-changing industrial efficient sorting needs.

[0023] In some embodiments, ore sorting based on dual-energy X-ray transmission (DE-XRT) technology mainly includes the following standard steps: 1. Data Acquisition and Dual-Energy Image Acquisition: The ore material enters the X-ray scanning area via a high-speed feed conveyor belt. A dual-energy X-ray linear array detector (typically a stacked structure with hardened filtering materials such as copper sheets sandwiched in between) is used to simultaneously acquire raw transmission images of both high-energy and low-energy X-ray levels after penetrating the ore. Dataset Preparation: Ore image data containing different grades, different types of associated vein minerals, and different particle size ranges are collected, and ground truth labeling is performed based on laboratory results.

[0024] 2. Image Preprocessing, Registration, Denoising, and Enhancement: Spatial filtering and morphological operations are used to remove detector background noise, dark current, and air scattering noise. High- and Low-Energy Spatial Registration: Due to the slight misalignment between the high- and low-energy detector crystals in physical space, and the high-frequency vibration of industrial conveyor belts, pixel-level registration of high-energy and low-energy images is required to ensure that the same coordinates accurately correspond to the same physical location of the same piece of ore during subsequent feature extraction.

[0025] 3. Region of Interest (ROI) Segmentation: Separating Ore from Background: Using threshold segmentation, edge detection, or connected component analysis algorithms, the conveyor belt background is removed. In scenarios with densely stacked materials, it is necessary to further separate individual ore pieces from adjacent or slightly adhered materials to form independent Regions of Interest (ROIs).

[0026] 4. Basic Feature Extraction: Within the segmented ROI, calculate the average gray value and gray extreme values ​​of the high- and low-energy images respectively. Dual-energy ratio feature: Based on the X-ray physical attenuation model, traditional methods usually directly calculate the ratio or logarithmic subtraction of the high- and low-energy gray values ​​(or transmittance), attempting to use this ratio to map the effective atomic number (Z value) of the ore and use it as the core feature to distinguish the target ore from waste rock.

[0027] 5. Model Decision and Classification: Static Threshold Classification: Set one or more fixed ratio thresholds, and directly compare the extracted ore features with the thresholds to achieve binary or multi-class classification. Basic Machine Learning Classification: Extract multiple features to construct feature vectors, and input them into pre-trained offline Support Vector Machines (SVM), Random Forests (RF), or lightweight neural networks for online forward inference calculation to obtain classification probabilities.

[0028] 6. Real-time control and valve execution system integration for high-speed sorting: After the decision model outputs the classification results, the host computer combines the centroid physical coordinates of the ore block with the belt speed to calculate the precise delay time and sends a trigger signal to the underlying programmable logic controller (PLC) or FPGA. Within milliseconds, the high-pressure electromagnetic valve array of the corresponding channel is controlled to perform short-term high-pressure airflow, changing the trajectory of the target ore and achieving physical separation. Based on this, this embodiment of the invention provides an ore sorting method, constructing a multi-level confidence ore sorting algorithm (MLC-OSA). By introducing high- and low-energy multi-dimensional derived features (such as second-order gradient ratio) to completely eliminate physical interference caused by changes in ore thickness and shape; at the same time, a nonlinear feature activation mechanism, dynamic weight fusion, and cascaded hybrid decision logic based on confidence boundaries (integrating lightweight threshold decision and complex machine learning model) are introduced to establish a decision closed loop that balances extremely low latency and high robustness. Ultimately, by combining closed-loop feedback with online parameter optimization, an efficient and precise industrial-grade ore sorting system is formed that can dynamically adapt to fluctuations in the properties of raw ore in different mining areas.

[0029] The specific scheme of the ore sorting method provided by the present invention will be described in detail below with reference to the accompanying drawings. For example... Figure 1 The diagram shown illustrates the implementation flow of an ore sorting method according to an embodiment of the present invention. The method includes: 101. Based on a preset line scan frequency, X-rays of different energies are used to synchronously and dynamically acquire data on the ore material placed on the feed belt, resulting in high-energy and low-energy X-ray images.

[0030] Here, a physical structure of top-mounted X-ray source and top-mounted dual-energy linear array detector is adopted. After the ore material is uniformly fed into the high-speed conveyor belt by the vibrating feeder (the belt speed is usually 2-3 meters per second (m / s)), the material enters the X-ray scanning dark chamber.

[0031] When X-rays penetrate moving ore, they attenuate to varying degrees due to differences in the equivalent atomic number (Z value) and physical thickness (t) within different ores. This attenuation process follows the Beer-Lambert Law.

[0032] The detector array consists of an upper high-energy response crystal (for acquiring high-energy images), a middle metal-hardened filter (such as a copper sheet, used to absorb high-energy photons), and a lower low-energy response crystal (for acquiring low-energy images). The system synchronously outputs high-energy raw transmission images in real time at a set linear scan frequency. and low-energy raw transmission images ,in These are the pixel coordinates of the image.

[0033] Real-time data stream monitoring is introduced during the acquisition phase. The system statistically analyzes the background grayscale of the image according to time windows. If a drift in the background grayscale is detected (caused by detector temperature drift or X-ray tube aging), the system automatically triggers the underlying background correction algorithm to update the air reference value, ensuring the stability of the absolute numerical reference for subsequent feature extraction. In addition to using a dual-energy linear array detector with high / low physical crystal layers, a photon counting detector (PCD) can also be used. Multi-energy transmission images are acquired by setting multiple energy thresholds, and the gradient ratio feature extraction method described in this invention is applied in the multi-energy spectrum dimension.

[0034] 102. Adaptive registration is performed on the high-energy perspective image and the low-energy perspective image to obtain a target low-energy image that is pixel-level aligned with the high-energy perspective image.

[0035] Here, due to the minute longitudinal spacing between the high-energy and low-energy detection crystals in physical space, and the high-frequency vibration and belt misalignment in the industrial environment, the directly acquired... and The alignment is not perfect at the pixel level. If misaligned, subsequent calculations of the "high-low energy gradient ratio" will produce significant edge artifacts, severely interfering with sorting decisions. Therefore, this step introduces a fast, flexible registration method based on edge contour constraints, including: First, adaptive noise reduction and background removal are performed: First, the original image is processed... and A guided filter is applied to smooth speckle noise from the detector while preserving the internal texture and edge gradient information of the ore. Subsequently, a global thresholding method (combined with the air background reference value) is used to remove the belt background, generating an initial ore foreground mask.

[0036] Secondly, coarse registration is achieved based on centroid translation: Let the foreground region of the same ore in the high-energy image be... The foreground region in the low-energy image is Calculate the centroid coordinates of both. and .

[0037] The initial translation vector is calculated as shown in formula (1): (1); use A coarse translation is performed on the low-energy image to obtain a preliminarily aligned low-energy image. .

[0038] Secondly, based on local optimization of the edge contour, precise registration is achieved: Because the ore may undergo minute tumbling during high-speed movement, simple translation cannot meet the requirements for precise alignment. This system uses high-energy images... Using first-order gradient operators (such as the Sobel operator) as a benchmark, edge response maps of high- and low-energy images are extracted. and .

[0039] The objective optimization function based on normalized cross-correlation (NCC) is constructed as shown in Equation (2): (2); in, For local fine-tuning of parameters, and This is used to calculate the local gradient mean within the window.

[0040] In coarse translation Searching within the neighborhood makes Maximize the fine-tuning parameters By applying this transformation, a low-energy image that is pixel-precisely aligned with the high-energy image is finally obtained. .

[0041] Finally, a registration quality feedback evaluation is conducted: The system calculates the sum of absolute differences in the edge regions of the final aligned image. If this sum exceeds a preset safety threshold, it indicates that the block has undergone violent tumbling or severe overlap during scanning. The system marks this sample as a "high uncertainty sample" and assigns it an extremely low confidence level in subsequent cascaded decisions to prevent accidental gas valve activation that could lead to the mixing of pure waste rock.

[0042] 103. High-energy feature extraction and low-energy feature extraction are performed on the target low-energy image to obtain high-energy basic features and low-energy basic features.

[0043] Here, after completing the high- and low-energy image registration, independent blocks are extracted for complex overlapping conditions, and multi-dimensional feature vectors with strong robustness to changes in physical thickness are extracted to obtain high-energy basic features and low-energy basic features.

[0044] In some possible implementations, step 103 above can be achieved by... Figure 2 The steps shown are to be implemented as follows: 201. Perform independent block segmentation on the target low-energy image to obtain a set of region masks for each independent ore block.

[0045] Here, the ore on the conveyor belt often exhibits slight adhesion, and traditional single-threshold connected component algorithms easily misclassify multiple pieces of ore as one. This system proposes a segmentation mechanism that integrates high- and low-energy edge information: First, calculate the logarithmic decay rate map of the aligned high and low energy images: as well as ,in This is the baseline value for air quality.

[0046] Subsequently, the fused edge map is extracted: Using the watershed algorithm combined with The local maximum ridge line forcibly severs the soft connection region of the bonded ore, and outputs the region mask set of each independent ore block. .

[0047] 202. High-energy feature extraction and low-energy feature extraction are performed on the region mask set of each independent ore block to obtain high-energy basic features and low-energy basic features.

[0048] Here, we are referring to the extracted individual ore areas. Calculate the basic characteristics of its high-energy and low-energy decay maps. This includes: the mean high-energy decay value. Low-energy decay mean High and low energy attenuation variance and Thus, based on the dual-energy fusion gradient for independent block segmentation, high-energy and low-energy basic features can be accurately extracted for each independent ore block.

[0049] 104. In the spatial differential domain of the high-energy fundamental feature and the low-energy fundamental feature, extract the first-order gradient ratio feature and the second-order gradient ratio feature between the high-energy fundamental feature and the low-energy fundamental feature.

[0050] Here, in the spatial differential domain, by calculating the high-energy and low-energy attenuation coefficients, and combining them with the first and second-order partial derivatives, the first-order gradient ratio feature and the second-order gradient ratio feature between the high-energy fundamental feature and the low-energy fundamental feature are calculated. When extracting spatially derived features, in addition to using classic first- / second-order differential operators such as Sobel and Laplacian, Gabor filter banks, wavelet transforms, or fractional differential operators can also be used to extract thickness-independent features at specific spatial frequencies, achieving the same anti-thickness interference effect.

[0051] In some possible implementations, firstly, the attenuation logarithms of X-rays under the influence of the high-energy fundamental feature and the low-energy fundamental feature are determined respectively to obtain the high-energy attenuation coefficient and the low-energy attenuation coefficient; Here, according to Beer-Lambert's law, the logarithm of X-ray attenuation... ,in The attenuation coefficient of X-rays is affected by the atomic number of the material. For thickness. Traditional dual-energy ratio. . The attenuation coefficient of X-rays is influenced by the atomic number of the material, a fundamental characteristic of high-energy radiation. This is the attenuation coefficient of X-rays due to the influence of the atomic number of the material, which is a fundamental characteristic of low energy.

[0052] Secondly, in the spatial differential domain, the first-order spatial partial derivatives of the high-energy attenuation coefficient and the low-energy attenuation coefficient are calculated respectively. Here, we introduce the ratio characteristic of the spatial differential domain. We assume the ore has a homogeneous material in local regions ( (The thickness is constant), but the thickness gradually changes in a wedge or spherical shape. In this case, for Find the first-order spatial partial derivative: ; Furthermore, the first-order gradient ratio characteristic is obtained based on the ratio between the first-order spatial partial derivatives of the high-energy decay coefficient and the low-energy decay coefficient. Here, the first-order gradient ratio feature is calculated: ; The first-order gradient ratio characteristic It cleverly utilizes the rate of change of adjacent pixels in the image space, due to the thickness gradient. The high and low energy images at the same physical point are completely coplanar, so this ratio can more strictly cancel out the influence of the absolute value of the overall thickness, and only retain the difference in the absorption sensitivity of the material to different energy level rays.

[0053] Finally, the second-order gradient ratio characteristic is obtained based on the ratio between the second-order spatial partial derivatives of the high-energy decay coefficient and the low-energy decay coefficient.

[0054] Here, the ratio of the high-energy and low-energy second-order gradients (Laplace operator) is calculated: ; The second-order gradient characterizes the change in curvature. The interior of an ore often contains high-density target mineral grains (such as sulfide phenocrysts), and at the grain boundaries... It exhibits a violent impulse response. This feature, acting like a high-pass filter, filters out low-frequency thickness interference caused by the smooth transition of the ore's outer contour, greatly amplifying the signal proportion of tiny associated minerals within. Thus, spatial calculus-derived features such as the ratio of high and low energy first / second-order gradients are extracted. This feature effectively removes low-frequency thickness interference caused by the smooth transition of the ore's outer contour, maximally amplifying the attenuated signal of the target minerals within, and significantly improving the sorting accuracy for extremely thick or thin ore samples.

[0055] 105. In the first sorting model, the confidence levels of the first-order gradient ratio feature and the second-order gradient ratio feature are calculated at multiple levels. Target ore with a confidence level greater than the first confidence level is selected from the ore material. Fuzzy ore blocks with confidence levels within a preset confidence range are input into the second sorting model so that the target ore is sorted from the fuzzy ore blocks using the second sorting model.

[0056] The second sorting model is obtained by incrementally fine-tuning the network parameters based on fuzzy ore block samples with confidence levels within a preset confidence range. The network complexity of the second sorting model is greater than that of the first sorting model. The preset confidence range is greater than a second confidence level but less than a first confidence level, with the second confidence level being less than the first confidence level. The subsequent complex machine learning model (i.e., the second sorting model) is not limited to random forests or support vector machines (SVMs). Where edge computing power allows, it can be replaced with lightweight convolutional neural networks (such as MobileNet or YOLO series detection heads) or lightweight variants of the Visual Transformer (ViT) to handle difficult samples in fuzzy regions.

[0057] In some possible implementations, step 105 above can be achieved through... Figure 3 The steps shown are to be implemented as follows: 301. The first-order gradient ratio feature and the second-order gradient ratio feature are concatenated to obtain the initial feature vector of each independent ore block.

[0058] Here, the aforementioned basic features are combined with calculus-derived features to form each ore block. Output a Initial eigenvectors of dimension .

[0059] 302. Using the first sorting model, multi-level feature denoising and feature weighted fusion are performed on the initial feature vector to obtain the confidence level of each independent ore block.

[0060] Here, in the first sorting model, the initial feature vector is subjected to feature denoising and weighted fusion through the activation layer, feature fusion layer, and cascaded decision layer in the model, thereby obtaining the confidence level of each independent ore block.

[0061] In some possible implementations, firstly, a preset feature threshold matrix of the activation layer in the first sorting model is obtained; secondly, the difference between the initial feature vector and the preset feature threshold matrix is ​​determined; thirdly, based on the ratio of the difference to the preset feature threshold matrix, activation weights are obtained, and the activation weights are used to sparsify the noise features in the initial feature vector that are lower than the preset feature threshold matrix to obtain activated features; finally, the activated features are linearly weighted and fused using a preset feature contribution coefficient matrix to obtain the confidence level of each independent ore block.

[0062] Here, the initial feature vector Input activation layer. To eliminate background noise and invalid feature terms, the system sets a feature threshold matrix. The activation weights are calculated using a specific nonlinear truncation function. As shown in formula (3): (3); This formula forces a value below the physically defined threshold. The noise features are set to zero (sparsed) and the effective features are adaptively amplified by relative gain.

[0063] Using a pre-set feature contribution coefficient matrix The activated features are then linearly weighted and fused to obtain the overall confidence score of the ore. As shown in formula (4): (4); In this embodiment of the invention, by constructing a multi-level confidence ore sorting algorithm (MLC-OSA), the real-time requirements of millisecond-level pulse jetting in industrial belt conveyors are met while ensuring accuracy, and it also has online self-learning capabilities.

[0064] In some possible implementations, in the first sorting model, to address the long-term drift of the properties of the raw ore and the fluctuations in the computing power at the system's underlying level, a multi-dimensional parameter online feedback correction mechanism is introduced to achieve dynamic optimization of the preset feature contribution coefficient matrix, the first confidence level, and the second confidence level. This can be achieved through the following process: First, the error value between the predicted classification result and the actual classification result of each independent ore block output by the first sorting model is obtained. Second, based on the error value and the activation weight, the preset feature contribution coefficient matrix is ​​dynamically updated. Third, based on the resource occupancy rate of the first sorting model in the device and the inference queue length, the load coefficient of the device is calculated. If the load coefficient indicates that the computing power of the device is less than the preset computing power threshold, the preset confidence range is narrowed; if the load coefficient indicates that the computing power of the device is greater than the preset computing power threshold, the preset confidence range is expanded. In this way, an MLC-OSA multi-level confidence cascaded sorting architecture is constructed. The vast majority of ordinary samples are intercepted through a "lightweight high-speed channel," and only a very small number of "fuzzy interval samples" are routed to the complex machine learning model (i.e., the second sorting model). At the same time, an "elastic breathing regulation" mechanism based on edge computing load and error feedback is introduced to dynamically scale the decision boundary, effectively preventing computing power downtime and valve delays caused by high throughput.

[0065] Here, the error feedback of the chemical grade of the sorted concentrate and tailings is obtained periodically. An online gradient descent algorithm is used to dynamically update the preset feature contribution coefficient matrix of the fusion layer. If formula (5) is as shown: (5); in The learning rate is used. This mechanism enables the model to adapt to feature preferences over a long period of time.

[0066] Introducing an edge device computing power monitoring module (real-time monitoring of CPU / GPU utilization and inference queue length, comprehensively evaluating the load factor). ) and downstream grade error trend feedback (error coefficient) The system uses these two indicators to determine the first confidence level and... Second confidence level Perform periodic "flexible breathing" adjustments: When edge computing power approaches its limit ( When this happens, the system executes the "shrinkage strategy":

[0067]

[0068] By narrowing the ambiguity range, more samples are forced to use the lightweight channel, sacrificing a small margin of precision for system stability and preventing misoperation of the air valve due to sorting delay. When computing power is abundant and sorting errors increase, the system executes an "expansion strategy":

[0069]

[0070] By expanding the fuzzy range, more edge samples are routed to complex machine learning models, making full use of remaining computing power to explore the ultimate sorting accuracy. and To adjust the step size, and It is a nonlinear mapping function. For example... Figure 4 As shown, the initial multidimensional feature vector V is input into the sorting model, and then enters the first layer: the feature activation layer, where the activation weight Wi is calculated through a nonlinear activation and truncation unit. The activation weight Wi is output to the second layer: the feature fusion layer, where the confidence score S is calculated through a weighted fusion scoring unit and input to the third layer: the three-part cascaded decision layer. The confidence score is judged by combining the first confidence score and the second confidence score. If S is less than or equal to the second confidence score, it is determined to be waste rock; if S is greater than or equal to the first confidence score, it is determined to be target ore. For the data between the first confidence score and the second confidence score, it is interpreted as a boundary model and triggers secondary precise inference, which is input into a complex machine learning model (such as SVM or random forest). The machine learning model further determines whether the material ore is waste rock or target ore. Finally, after classification, multi-source feedback data (e.g., calculating the test error E and edge computing load L) is input into the dynamic optimization and closed-loop control module. This data then passes through the online parameter update unit to achieve gradient descent and elastic breathing control, dynamically adjusting the first and second confidence levels (i.e., the physical threshold Ti) and the contribution coefficient Ci (i.e., the preset feature contribution coefficient matrix). For the online dynamic update of the preset feature contribution coefficient matrix, the first confidence level, and the second confidence level, in addition to gradient descent, particle swarm optimization, genetic algorithms, or a reward mechanism based on reinforcement learning (using the actual test quality as the reward) can also be used for global optimization of multi-dimensional parameters.

[0071] 303, among the individual ore blocks of the ore material, select the target ore with a confidence level greater than the first confidence level.

[0072] In the application phase, this invention introduces a high-value sample mining and semi-automated iterative annotation mechanism based on cascading boundary triggering to achieve data closure.

[0073] First, in cascaded decision-making, most samples are processed through the lightweight high-speed channel (score). or The system processes images quickly. It does not save these "simple samples" locally. Instead, it only targets and saves images falling within the blurred boundary regions. The original dual-energy X-ray image and feature vector of the "difficult sample".

[0074] Then, for the captured difficult samples, the backend complex machine learning model (i.e., the second sorting model) outputs a predicted probability distribution. Simultaneously calculate its predicted information entropy. The variance of energy attenuation in the bound material. Construct an improved sample uncertainty index The system will Samples exceeding the set threshold are pushed to the "queue for review" in the host computer visualization client.

[0075] Finally, through the client interface, human operators only need to visually verify and confirm the true category of the very small portion of the most deceptive "high-entropy samples" selected by the system. The data after confirmation is stored in the incremental database, and the system uses its idle computing power to perform mini-batch incremental fine-tuning on the backend machine learning model, thereby enabling the entire sorting system to have a targeted self-evolution capability that "gets smarter with use".

[0076] In steps 301 to 303 above, by concatenating the feature vectors and then performing multi-level feature denoising and feature weighted fusion on the initial feature vectors, the confidence level of each independent ore block can be accurately calculated, thereby improving the accuracy of classifying each independent ore block.

[0077] 106. Based on the region of interest of the target ore and the distance from the X-ray scan line to the gas valve jet array, the gas valve jet array is used to physically separate the target ore.

[0078] Here, after obtaining the region of interest of the target ore, the dynamic delay time of the target ore reaching the gas valve blowing array is calculated by combining the distance from the X-ray scan line to the gas valve blowing array. The gas valve blowing array is then controlled to physically separate the target ore according to the dynamic delay time.

[0079] In some possible implementations, step 106 above can be achieved through the following steps 161 and 162 (not shown in the figure): 161. Based on the region of interest of the target ore and the distance from the X-ray scan line to the gas valve blowing array, calculate the dynamic delay time of the target ore reaching the gas valve blowing array.

[0080] Here, the physical centroid coordinates of the region of interest of the target ore, the maximum span of the region of interest in the belt running direction, and the maximum lateral width are first determined. Then, the difference between the physical centroid coordinates and the distance from the X-ray scan line to the air valve jet array is determined. Next, the ratio between the difference and the real-time belt speed is calculated. Finally, based on the algorithm inference time corresponding to the target ore, the ratio is adjusted to obtain the dynamic delay time of the target ore reaching the air valve jet array.

[0081] The system adopts a collaborative working mode between a high-performance PC client (host computer) and a programmable logic controller (PLC) / field-programmable gate array (FPGA) (slave computer). The host computer focuses on running high-intensity image preprocessing, MLC-OSA multi-level feature extraction, algorithm inference, and visual interactive interface; the slave computer focuses on handling the I / O timing control of the high real-time multi-channel solenoid valve array.

[0082] After completing the classification decision for a single piece of ore, the host computer extracts the region of interest (ROI) mask for that target ore and calculates its physical centroid coordinates. and its maximum physical span in the belt running direction (Y-axis). and the maximum width in the horizontal direction (X-axis) .

[0083] Combined with real-time feedback of belt speed from the belt side shaft encoder The host computer calculates the dynamic delay time of the ore reaching the air valve injection array. The expression is: ; in, This refers to the fixed physical distance from the X-ray scan line to the valve array. The algorithm inference time for this sample in the MLC-OSA architecture (dynamic variable).

[0084] 162, The target ore is physically separated using the air valve jet array based on the dynamic delay time.

[0085] In some possible implementations, a preset mapping table characterizing the correspondence between the maximum lateral width and the number of air valve channels is used to determine the target number of air valve channels matching the maximum lateral width; then, pulse width mapping calculation is performed on the maximum span and the estimated quality of the target ore to obtain the pulse width corresponding to the target ore; and an adaptive blowing control signal is generated using the target number of air valve channels, the dynamic delay time, and the pulse width; in response to the adaptive blowing control signal, the air valve blowing array corresponding to the target number of air valve channels is controlled to physically separate the target ore.

[0086] Here, the preset mapping table can be a custom correspondence table between the maximum horizontal width and the number of air valve channels. The target number of air valve channels, dynamic delay time, and pulse width are combined to form a control signal to control the air valves in the air valve injection array that meet the target number of air valve channels, and to physically separate the target ore according to the dynamic delay time and pulse width.

[0087] The host computer not only outputs a Boolean signal indicating whether to spray, but also... Map the number of air valve channels that need to be opened, and according to Its quality is estimated, and the opening and holding time (pulse width) of the air valve is dynamically calculated. Larger ore pieces are assigned longer pulse widths for blowing, while smaller pieces are assigned shorter pulse widths to avoid "airflow crosstalk" blowing away surrounding waste rock. Finally, a high-frequency command frame containing [valve channel number, delay time, pulse width] is sent to the lower-level computer for execution, completing the high-precision physical separation of the target ore. In this way, the upper-level computer, while outputting its decision, dynamically calculates the algorithm's inference delay and adaptively allocates the valve opening pulse width and channel number based on the physical span of the ore. This hardware-software collaboration and dynamic compensation mechanism completely eliminates "airflow crosstalk" and "blowing lag," ensuring an extremely high hit rate in actual physical actions.

[0088] This invention creates a multi-level confidence cascade architecture (MLC-OSA) ore sorting system, comprising a complete workflow including image acquisition, flexible registration, multi-dimensional derived feature extraction, nonlinear feature activation, feature weighted fusion, and classification routing based on dynamic confidence boundaries (transmitting samples whose scores fall into the fuzzy range to a subsequent complex machine learning model). For the feature engineering of extracting the "ratio of high- and low-energy first-order gradient magnitudes" and the "ratio of high- and low-energy second-order gradients (Laplacian operator)" from log-decayed images, this feature can effectively remove low-frequency thickness interference caused by the smooth transition of the ore's outer contour, like a high-pass filter, and maximizes the attenuation signal of the internal target minerals, significantly improving the sorting accuracy of extremely thick or thin ore samples. A multi-level confidence cascaded sorting architecture of MLC-OSA is constructed, which intercepts the vast majority of ordinary samples through a "lightweight high-speed channel" and routes only a very small number of "fuzzy interval samples" to complex machine learning models. At the same time, an "elastic breathing regulation" mechanism based on edge computing load and error feedback is introduced to dynamically scale the decision boundary, effectively preventing computing power downtime and valve delays caused by high throughput. By monitoring the real-time computing load status (CPU / GPU utilization) of edge devices and the downstream classification error trend, the system dynamically "shrinks" or "expands" the high / low confidence thresholds to achieve an adaptive control strategy that balances computing power and accuracy in real time. Based on the actual grade testing error of the sorted concentrate / tailings, a closed-loop control system is used to dynamically update the feature layer contribution coefficient and nonlinear activation threshold using online optimization algorithms (such as gradient descent). This introduces a parameter closed-loop dynamic optimization and semi-automatic iterative labeling mechanism, enabling online reverse updates of feature fusion weights and activation thresholds based on the actual downstream testing grade error. Simultaneously, it targets and captures "difficult samples" with high information entropy for expert fine-tuning. This allows the sorting system to adapt to the "drift in raw ore properties" caused by changes in mining depth, becoming more accurate with use and significantly reducing the costs of subsequent equipment recalibration and maintenance.

[0089] To address the challenges of severe conveyor belt vibration and dense material distribution, this system introduces a fast elastic registration algorithm based on edge contour constraints, an adaptive segmentation mechanism (ROI) that integrates high and low energy gradients, a registration scheme that uses local normalized cross-correlation (NCC) combined with first-order gradient extraction for fine-tuning, and a segmentation mechanism that uses high and low energy maximum fusion edge extraction combined with a watershed algorithm to forcibly cut off areas of adherent ore. This effectively solves the edge artifact problem caused by misalignment of high and low energy images and achieves accurate stripping of adherent ore, ensuring the purity of feature extraction from the source. The dynamic inference time of the host computer algorithm is incorporated into the formula architecture for calculating the air valve injection delay time, and the air valve opening pulse width and channel number are dynamically mapped according to the physical span of the ore. This decoupling of hardware and software and adaptive execution improves the final physical separation yield.

[0090] This invention provides an ore sorting system, such as... Figure 5 As shown, the system 500 includes: The acquisition module 501 is used to synchronously and dynamically acquire data on the ore material placed on the feed belt using X-rays of different energies based on a preset line scan frequency, so as to obtain high-energy X-ray images and low-energy X-ray images. The registration module 502 is used to perform adaptive registration of the high-energy perspective image and the low-energy perspective image to obtain a target low-energy image that is pixel-level aligned with the high-energy perspective image. The first extraction module 503 is used to extract high-energy features and low-energy features from the target low-energy image to obtain high-energy basic features and low-energy basic features. The second extraction module 504 is used to extract the first-order gradient ratio feature and the second-order gradient ratio feature between the high-energy basic feature and the low-energy basic feature in the spatial differential domain of the high-energy basic feature and the low-energy basic feature. The sorting module 505 is used to perform multi-level confidence calculations on the first-order gradient ratio feature and the second-order gradient ratio feature in the first sorting model, select target ores with a confidence level greater than the first confidence level from the ore material, and input fuzzy ore blocks with confidence levels within a preset confidence range into the second sorting model, so as to sort the target ores using the second sorting model; wherein, the second sorting model is obtained by incrementally fine-tuning the network parameters based on fuzzy ore block samples with confidence levels within the preset confidence range, and the network complexity of the second sorting model is greater than the network complexity of the first sorting model; the preset confidence range is greater than the second confidence level and less than the first confidence level, and the second confidence level is less than the first confidence level; The blowing module 506 is used to physically separate the target ore using the air valve blowing array based on the region of interest of the target ore and the distance from the X-ray scan line to the air valve blowing array.

[0091] In some possible implementations, the blowing module 506 is further configured to calculate the dynamic delay time of the target ore reaching the gas valve blowing array based on the distance from the region of interest of the target ore and the X-ray scan line to the gas valve blowing array; and to use the gas valve blowing array to physically separate the target ore based on the dynamic delay time.

[0092] In some possible implementations, the blowing module 506 is further configured to determine the physical centroid coordinates of the region of interest of the target ore, the maximum span and maximum lateral width of the region of interest in the belt running direction; determine the difference between the physical centroid coordinates and the distance from the X-ray scan line to the air valve blowing array; determine the ratio between the difference and the real-time belt speed of the belt; and adjust the ratio based on the algorithm inference time corresponding to the target ore to obtain the dynamic delay time of the target ore reaching the air valve blowing array.

[0093] In some possible implementations, the injection module 506 is further configured to: determine the number of target air valve channels matching the maximum lateral width based on a preset mapping table characterizing the correspondence between the maximum lateral width and the number of air valve channels; perform pulse width mapping calculation on the maximum span and the estimated quality of the target ore to obtain the pulse width corresponding to the target ore; generate an adaptive injection control signal based on the number of target air valve channels, the dynamic delay time, and the pulse width; and, in response to the adaptive injection control signal, control the air valve injection array corresponding to the number of target air valve channels to physically separate the target ore.

[0094] In some possible implementations, the first extraction module 503 is further configured to perform independent block segmentation on the target low-energy image to obtain a set of region masks for each independent ore block; and to perform high-energy feature extraction and low-energy feature extraction on the set of region masks for each independent ore block to obtain high-energy basic features and low-energy basic features.

[0095] In some possible implementations, the sorting module 505 is further configured to concatenate the first-order gradient ratio feature and the second-order gradient ratio feature to obtain the initial feature vector of each independent ore block; using the first sorting model, perform multi-level feature denoising and feature weighted fusion on the initial feature vector to obtain the confidence level of each independent ore block; and select target ore with a confidence level greater than the first confidence level from each independent ore block of the ore material.

[0096] In some possible implementations, the sorting module 505 is further configured to: obtain a preset feature threshold matrix of the activation layer in the first sorting model; determine the difference between the initial feature vector and the preset feature threshold matrix; obtain activation weights based on the ratio of the difference to the preset feature threshold matrix; use the activation weights to sparsify the noise features in the initial feature vector that are lower than the preset feature threshold matrix to obtain activated features; and use a preset feature contribution coefficient matrix to perform linear weighted fusion on the activated features to obtain the confidence level of each independent ore block.

[0097] In some possible implementations, the sorting module 505 is further configured to: obtain the error value between the predicted classification result and the actual classification result of each independent ore block output by the first sorting model; dynamically update the preset feature contribution coefficient matrix based on the error value and the activation weight; calculate the load coefficient of the device based on the resource occupancy rate of the first sorting model in the device and the inference queue length; if the load coefficient indicates that the computing power of the device is less than a preset computing power threshold, shrink the preset confidence range; if the load coefficient indicates that the computing power of the device is greater than the preset computing power threshold, expand the preset confidence range.

[0098] In some possible implementations, the second extraction module 504 is further configured to determine the attenuation logarithm of X-rays under the influence of the high-energy fundamental feature and the low-energy fundamental feature, respectively, to obtain the high-energy attenuation coefficient and the low-energy attenuation coefficient; calculate the first-order spatial partial derivatives of the high-energy attenuation coefficient and the low-energy attenuation coefficient in the spatial differential domain, respectively; obtain the first-order gradient ratio feature based on the ratio between the first-order spatial partial derivatives of the high-energy attenuation coefficient and the low-energy attenuation coefficient; and obtain the second-order gradient ratio feature based on the ratio between the second-order spatial partial derivatives of the high-energy attenuation coefficient and the low-energy attenuation coefficient.

[0099] Optionally, the transmission medium can be a wired link (e.g., but not limited to, coaxial cable, optical fiber, and Digital Subscriber Line (DSL)) or a wireless link (e.g., but not limited to, Wireless Fidelity (WIFI), Bluetooth, and mobile block device networks). It should be noted that the control block device provided in the above embodiments is only an example illustrating the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer block device can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the method embodiments provided in the above embodiments belong to the same concept, and their specific implementation processes are detailed in the method embodiments, and will not be repeated here.

[0100] Figure 6 This is a schematic diagram of the structure of a computer block device provided in an embodiment of the present invention. For example, as shown... Figure 6 As shown, the computer block device 600 includes: a memory 601, a processor 602, and a computer program 603 stored in the memory 601 and running on the processor 602, wherein when the processor 602 executes the computer program 603, the computer block device can perform any of the ore sorting methods described above.

[0101] Furthermore, embodiments of the present invention also protect a control block device, which may include a memory and a processor. The memory stores executable program code, and the processor is used to call and execute the executable program code to perform an ore sorting method provided by the embodiments of the present invention. Embodiments of the present invention can divide the control block device into functional modules based on the above method examples. For example, each module may correspond to a specific function, or two or more functions may be integrated into a processing module. The integrated module can be implemented in hardware. It should be noted that the module division in the embodiments of the present invention is illustrative and only represents a logical functional division; in actual implementation, there may be other division methods. It should be noted that all relevant content of each step involved in the above method embodiments can be referenced to the functional description of the corresponding functional module, and will not be repeated here. It should be understood that the control block device provided by the embodiments of the present invention is used to execute the above-described ore sorting method, and therefore can achieve the same effect as the above-described implementation method. When using integrated units, the control block device may include a processing module and a storage module. When the control block device is applied to a block device, the processing module can be used to control and manage the actions of the block device. The storage module can be used to support block devices in executing mutual program code, etc. The processing module can be a processor or controller, which can implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. The processor can also be a combination of functions that implement computing capabilities, such as a combination of one or more microprocessors, a combination of digital signal processing (DSP) and microprocessors, etc., and the storage module can be a memory.

[0102] Furthermore, the control block device provided in the embodiments of the present invention may specifically be a chip, component, or module. The chip may include a connected processor and a memory; wherein, the memory is used to store instructions, and when the processor calls and executes the instructions, the chip can execute the ore sorting method provided in the above embodiments. The embodiments of the present invention also provide a computer-readable storage medium storing computer program code. When the computer program code is run on a computer, the computer executes the aforementioned method steps to implement the ore sorting method provided in the above embodiments.

[0103] This invention also provides a computer program product. When the computer program product is run on a computer, it causes the computer to execute the aforementioned related steps to achieve the ore sorting method provided in the above embodiments. The control block device, computer-readable storage medium, computer program product, or chip provided in this invention are all used to execute the corresponding methods provided above. Therefore, the beneficial effects they achieve can be referred to in the beneficial effects of the corresponding methods provided above, and will not be repeated here. Through the description of the above embodiments, those skilled in the art can understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the control block device can be divided into different functional modules to complete all or part of the functions described above. In the embodiments provided by this invention, it should be understood that the disclosed control block device and method can be implemented in other ways. For example, the control block device embodiments described above are merely illustrative. For example, the division of modules or units is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another control block device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, control block device or unit, and can be electrical, mechanical or other forms.

[0104] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous. The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. The above content is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for ore sorting, characterized in that, The method includes: Based on a preset line scan frequency, X-rays of different energies are used to synchronously and dynamically acquire data on the ore material placed on the feed belt, resulting in high-energy and low-energy X-ray images. Adaptive registration is performed on the high-energy perspective image and the low-energy perspective image to obtain a target low-energy image that is pixel-level aligned with the high-energy perspective image; High-energy feature extraction and low-energy feature extraction are performed on the target low-energy image to obtain high-energy basic features and low-energy basic features; In the spatial differential domain of the high-energy fundamental features and the low-energy fundamental features, the first-order gradient ratio feature and the second-order gradient ratio feature between the high-energy fundamental features and the low-energy fundamental features are extracted. In the first sorting model, multi-level confidence calculations are performed on the first-order gradient ratio feature and the second-order gradient ratio feature. Target ores with confidence levels greater than the first confidence level are selected from the ore materials. Fuzzy ore blocks with confidence levels within a preset confidence range are input into the second sorting model to sort the target ores using the second sorting model. The second sorting model is obtained by incrementally fine-tuning the network parameters based on fuzzy ore block samples with confidence levels within the preset confidence range, and the network complexity of the second sorting model is greater than that of the first sorting model. The preset confidence range is greater than the second confidence level and less than the first confidence level, and the second confidence level is less than the first confidence level. Based on the region of interest of the target ore and the distance from the X-ray scan line to the gas valve jet array, the gas valve jet array is used to physically separate the target ore.

2. The method according to claim 1, characterized in that, The method of physically separating the target ore using the gas valve jet array based on the distance from the region of interest of the target ore and the X-ray scan line to the gas valve jet array includes: Based on the region of interest of the target ore and the distance from the X-ray scan line to the gas valve blowing array, the dynamic delay time of the target ore reaching the gas valve blowing array is calculated. The target ore is physically separated using the air valve jet array based on the dynamic delay time.

3. The method according to claim 2, characterized in that, The calculation of the dynamic delay time of the target ore reaching the gas valve blowing array based on the distance from the region of interest of the target ore and the X-ray scan line to the gas valve blowing array includes: Determine the physical centroid coordinates of the region of interest of the target ore, the maximum span of the region of interest in the direction of belt conveyor operation, and the maximum lateral width of the region of interest. Determine the difference between the physical centroid coordinates and the distance from the X-ray scan line to the gas valve jet array; Determine the ratio between the difference and the real-time belt speed of the belt; Based on the algorithm inference time corresponding to the target ore, the ratio is adjusted to obtain the dynamic delay time of the target ore reaching the gas valve injection array.

4. The method according to claim 3, characterized in that, The physical separation of the target ore using the air valve injection array based on the dynamic delay time includes: Based on a preset mapping table representing the correspondence between the maximum lateral width and the number of valve channels, the target number of valve channels matching the maximum lateral width is determined. A pulse width mapping calculation is performed on the maximum span and the estimated mass of the target ore to obtain the pulse width corresponding to the target ore; An adaptive jet control signal is generated based on the number of target air valve channels, the dynamic delay time, and the pulse width. In response to the adaptive injection control signal, the air valve injection array corresponding to the number of target air valve channels is controlled to physically separate the target ore.

5. The method according to claim 1, characterized in that, The step of extracting high-energy features and low-energy features from the target low-energy image to obtain high-energy basic features and low-energy basic features includes: The target low-energy image is segmented into individual blocks to obtain a set of region masks for each individual mineral block; High-energy feature extraction and low-energy feature extraction are performed on the region mask set of each independent ore block to obtain high-energy basic features and low-energy basic features.

6. The method according to claim 1, characterized in that, In the first sorting model, multi-level confidence calculations are performed on the first-order gradient ratio feature and the second-order gradient ratio feature. Target ores with a confidence level greater than the first confidence level are selected from the ore materials, including: The first-order gradient ratio feature and the second-order gradient ratio feature are concatenated to obtain the initial feature vector of each independent ore block. Using the first sorting model, multi-level feature denoising and feature weighted fusion are performed on the initial feature vector to obtain the confidence level of each independent ore block; Among the individual ore blocks of the ore material, the target ore with a confidence level greater than the first confidence level is selected.

7. The method according to claim 6, characterized in that, The first sorting model is used to perform multi-level feature denoising and feature weighting fusion on the initial feature vector to obtain the confidence level of each independent ore block, including: Obtain the preset feature threshold matrix of the activation layer in the first sorting model; Determine the difference between the initial feature vector and the preset feature threshold matrix; The activation weight is obtained based on the ratio of the difference to the preset feature threshold matrix; Using the activation weights, the noise features in the initial feature vector that are below the preset feature threshold matrix are sparsified to obtain activated features; The activated features are linearly weighted and fused using a preset feature contribution coefficient matrix to obtain the confidence level of each independent ore block.

8. The method according to claim 7, characterized in that, The method further includes: Obtain the error value between the predicted classification result and the actual classification result of each independent ore block output by the first sorting model; Based on the error value and the activation weight, the preset feature contribution coefficient matrix is ​​dynamically updated; Based on the resource occupancy rate of the first sorting model in the device and the inference queue length, the load factor of the device is calculated. If the load coefficient indicates that the computing power of the device is less than a preset computing power threshold, the preset confidence range is narrowed. If the load coefficient indicates that the computing power of the device is greater than the preset computing power threshold, the preset confidence range is expanded.

9. The method according to claim 1, characterized in that, The extraction of the first-order gradient ratio feature and the second-order gradient ratio feature between the high-energy fundamental feature and the low-energy fundamental feature within the spatial differential domain of the high-energy fundamental feature and the low-energy fundamental feature includes: The attenuation logarithms of X-rays under the influence of the high-energy fundamental feature and the low-energy fundamental feature are determined respectively to obtain the high-energy attenuation coefficient and the low-energy attenuation coefficient. In the spatial differential domain, calculate the first-order spatial partial derivatives of the high-energy attenuation coefficient and the low-energy attenuation coefficient, respectively. The first-order gradient ratio characteristic is obtained based on the ratio between the first-order spatial partial derivatives of the high-energy decay coefficient and the low-energy decay coefficient. The second-order gradient ratio characteristic is obtained based on the ratio between the second-order spatial partial derivatives of the high-energy decay coefficient and the low-energy decay coefficient.

10. An ore sorting system, characterized in that, The system includes: The acquisition module is used to synchronously and dynamically acquire data on the ore material placed on the feed belt using X-rays of different energies based on a preset line scan frequency, so as to obtain high-energy X-ray images and low-energy X-ray images. The registration module is used to adaptively register the high-energy perspective image and the low-energy perspective image to obtain a target low-energy image that is pixel-level aligned with the high-energy perspective image. The first extraction module is used to extract high-energy features and low-energy features from the target low-energy image to obtain high-energy basic features and low-energy basic features. The second extraction module is used to extract the first-order gradient ratio feature and the second-order gradient ratio feature between the high-energy basic feature and the low-energy basic feature in the spatial differential domain of the high-energy basic feature and the low-energy basic feature. The sorting module is used in the first sorting model to perform multi-level confidence calculations on the first-order gradient ratio feature and the second-order gradient ratio feature. It selects target ores with a confidence level greater than the first confidence level from the ore material, and inputs fuzzy ore blocks with confidence levels within a preset confidence range into the second sorting model. The second sorting model is then used to sort the fuzzy ore blocks for target ores. The second sorting model is obtained by incrementally fine-tuning the network parameters based on fuzzy ore block samples with confidence levels within the preset confidence range, and the network complexity of the second sorting model is greater than that of the first sorting model. The preset confidence range is greater than the second confidence level and less than the first confidence level, where the second confidence level is less than the first confidence level. The jetting module is used to physically separate the target ore using the gas valve jetting array based on the region of interest of the target ore and the distance from the X-ray scan line to the gas valve jetting array.

Images