Ore grade real-time identification and separation system and method
By identifying ore grade through image acquisition and deep learning models, and combining pneumatic sorting valves and robotic arms for automated sorting, the problems of low ore sorting efficiency and accuracy are solved, and costs are reduced.
Patent Information
- Application Number
- CN202511602589.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-04
- Publication Date
- 2026-03-03
AI Technical Summary
Existing ore sorting methods are inefficient and inaccurate, and costly. Manual sorting is highly subjective, and traditional physical sorting methods have strict requirements on ore properties, making them difficult to adapt to large-scale industrial production.
The system employs an image acquisition module, an image processing module, and an image recognition module, combined with a pneumatic sorting valve and a robotic arm, to sort ore. It utilizes a deep learning model to identify ore grade and combines it with ore size for automated sorting.
It improves the efficiency and accuracy of ore sorting, reduces sorting costs, and realizes automated identification and sorting of ore grades.
Smart Images
Figure CN121589049A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of ore processing technology, and in particular to a real-time ore grade identification and sorting system and method. Background Technology
[0002] In the development of the ore processing industry, traditional sorting methods have long been dominant, but as the industry's requirements for resource utilization efficiency have increased, their drawbacks have become increasingly prominent.
[0003] Manual sorting relies entirely on worker experience to judge ore grade, which has significant drawbacks. Firstly, human vision and experience are highly subjective; different workers may have vastly different grade judgments for the same ore, leading to inconsistent sorting results. Secondly, ore processing lines operate at high speeds and handle large volumes, making manual sorting extremely inefficient and unsuitable for large-scale industrial production. Furthermore, worker fatigue and emotional factors can easily cause high-grade ore to be misjudged and discarded, while low-grade ore is mixed into high-quality products, severely impacting resource recovery and enterprise profitability. Physical sorting, based on the physical properties of ore such as density, magnetism, and conductivity, is limited by the ore's inherent properties. For example, gravity sorting cannot effectively distinguish ores with small density differences (such as different grades of symbiotic ores); magnetic sorting is only effective for magnetic minerals and struggles to accurately screen out target minerals of different grades in complex polymetallic symbiotic ores. Moreover, physical sorting equipment has strict requirements for ore particle size and shape, resulting in high pre-treatment costs, further limiting its application scope and sorting accuracy.
[0004] Therefore, how to reduce the cost of ore sorting while ensuring efficiency and accuracy has become an urgent technical problem to be solved. Summary of the Invention
[0005] In view of this, it is necessary to provide a real-time ore grade identification and sorting system and method to solve the problems of low efficiency and accuracy and high cost of existing ore sorting methods.
[0006] To address the aforementioned problems, in a first aspect, the present invention provides a real-time ore grade identification and sorting system, comprising: The image acquisition module is used to acquire images of the ore surface in real time and transmit them to the image processing module; The image processing module is used to preprocess the ore surface image and transmit the preprocessed ore surface image to the image recognition module; The image recognition module is used to recognize the preprocessed ore surface image based on the image recognition model and output the ore grade to the ore sorting module. The image recognition model is trained on a residual network with added spatial attention mechanism based on a sample set of ore surface images with ore grade labels. The learning rate during the training process is determined based on the cosine annealing algorithm. The residual network includes multiple preset convolutional layers, which are composed of 1×1 convolutional kernels and 3×3 convolutional kernels. The ore sorting module is used to sort ore based on its grade using pneumatic sorting valves and robotic arms.
[0007] In one possible implementation, the preprocessing of the ore surface image includes: Image enhancement of ore surface images based on adaptive histogram equalization algorithm; The enhanced mineral surface image is adjusted to a preset size and pixel normalized based on the bicubic interpolation algorithm. The image of the ore surface after image normalization is expanded using at least one of the following image expansion methods: random flipping, rotation, cropping, and adding noise.
[0008] In one possible implementation, the preset convolutional layer is composed of a first convolutional kernel, a second convolutional kernel, and a third convolutional kernel connected by residuals in sequence, wherein the first and third convolutional kernels are 1×1 convolutional kernels, and the second convolutional kernel is a 3×3 convolutional kernel.
[0009] In one possible implementation, the ore sorting based on ore grade using pneumatic sorting valves and robotic arms includes: The ore is directed into the corresponding channel based on its grade. When the ore particle size is less than or equal to the particle size threshold, a pneumatic separation valve is used for ore separation. When the ore particle size is larger than the particle size threshold, a robotic arm is used for ore sorting.
[0010] In one possible implementation, the system further includes: The data management module is used to store ore image acquisition data, image recognition results, system operating parameters and equipment status data, and generate ore grade distribution reports and equipment operating status reports.
[0011] On the other hand, the present invention also provides a method for real-time ore grade identification and sorting based on the above-mentioned real-time ore grade identification and sorting system, comprising: Acquire real-time images of the ore surface and preprocess the images; The image recognition model is used to identify the preprocessed ore surface image and output the ore grade. The image recognition model is trained on a residual network with added spatial attention mechanism based on a sample set of ore surface images with ore grade labels. The learning rate during the training process is determined based on the cosine annealing algorithm. The residual network includes multiple preset convolutional layers, which are composed of 1×1 convolutional kernels and 3×3 convolutional kernels. Ore sorting is performed using pneumatic sorting valves and robotic arms based on ore grade.
[0012] In one possible implementation, the preprocessing of the ore surface image includes: Image enhancement of ore surface images based on adaptive histogram equalization algorithm; The enhanced mineral surface image is adjusted to a preset size and pixel normalized based on the bicubic interpolation algorithm. The image of the ore surface after image normalization is expanded using at least one of the following image expansion methods: random flipping, rotation, cropping, and adding noise.
[0013] In one possible implementation, the preset convolutional layer is composed of a first convolutional kernel, a second convolutional kernel, and a third convolutional kernel connected by residuals in sequence, wherein the first and third convolutional kernels are 1×1 convolutional kernels, and the second convolutional kernel is a 3×3 convolutional kernel.
[0014] In one possible implementation, the ore sorting based on ore grade using pneumatic sorting valves and robotic arms includes: The ore is directed into the corresponding channel based on its grade. When the ore particle size is less than or equal to the particle size threshold, a pneumatic separation valve is used for ore separation. When the ore particle size is larger than the particle size threshold, a robotic arm is used for ore sorting.
[0015] In one possible implementation, the method further includes: Ore grade distribution reports and equipment operation status reports are generated based on ore image acquisition data, image recognition results, system operating parameters, and equipment status data.
[0016] The beneficial effects of this invention are as follows: The real-time ore grade identification and sorting system and method provided by this invention obtains the ore grade level by acquiring and identifying images of the ore surface to ensure the accuracy of identification. Then, the ore is sorted according to the ore grade level. During the sorting process, pneumatic sorting valves and robotic arms can be used in combination with ore size to improve sorting efficiency. The ore grade and identification process can be automated, thereby reducing costs. This invention reduces sorting costs while ensuring the efficiency and accuracy of ore sorting. Attached Figure Description
[0017] Figure 1 A schematic diagram of an embodiment of the real-time ore grade identification and sorting system provided by the present invention; Figure 2 A schematic flowchart of an embodiment of the real-time ore grade identification and sorting method provided by the present invention; Figure 3 A schematic diagram of another embodiment of the real-time ore grade identification and sorting system provided by the present invention; Figure 4 A schematic flowchart of an embodiment of the ore grade identification and sorting process provided by the present invention; Figure 5 This is a schematic flowchart of an embodiment of the ore sorting process provided by the present invention. Detailed Implementation
[0018] 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 a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0019] In the description of the embodiments of the present invention, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.
[0020] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.
[0021] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.
[0022] This invention provides a real-time ore grade identification and sorting system and method, which are described below.
[0023] Figure 1 This is a schematic diagram of an embodiment of the real-time ore grade identification and sorting system provided by the present invention, as shown below. Figure 1 As shown, the real-time ore grade identification and sorting system 100 includes: Image acquisition module 101 is used to acquire images of the ore surface in real time and transmit them to image processing module 102; Image processing module 102 is used to preprocess the ore surface image and transmit the preprocessed ore surface image to image recognition module 103; Image recognition module 103 is used to recognize the preprocessed ore surface image based on the image recognition model and output the ore grade to the ore sorting module 104. The image recognition model is trained on a residual network with added spatial attention mechanism based on a sample set of ore surface images with ore grade labels. The learning rate during the training process is determined based on the cosine annealing algorithm. The residual network includes multiple preset convolutional layers, which are composed of 1×1 convolutional kernels and 3×3 convolutional kernels. The ore sorting module 104 is used to sort ore based on its grade using pneumatic sorting valves and robotic arms.
[0024] It should be noted that the real-time ore grade identification and sorting system provided by this invention can be applied to ore grade sorting scenarios. Specifically, the real-time ore grade identification and sorting system may include an image acquisition module, an image processing module, an image recognition module, and an ore sorting module.
[0025] The image acquisition module utilizes a high-definition industrial camera paired with a customized light source. The industrial camera boasts high-speed acquisition capabilities, with a frame rate flexibly adjustable from 10-30 frames per second to adapt to production lines with varying ore conveying speeds. The light source system employs a highly uniform LED array, utilizing diffuse reflection lighting to effectively avoid issues such as glare and shadows on the ore surface, ensuring that the acquired images clearly reveal key features such as ore surface texture, color, and mineral particle distribution. The camera is mounted at a specific height and angle above the ore conveyor belt, guaranteeing that the shooting range covers the width of the belt, and that the ore imaging is proportionally accurate and features are complete.
[0026] The image processing module can perform image enhancement, image normalization, and data enhancement extension on the acquired ore surface images.
[0027] The image recognition module can identify the preprocessed ore surface image using an image recognition model, outputting the ore grade. The image recognition model is a residual network with an added attention mechanism. In the residual network, the residual block structure can be optimized, and the number and size of convolutional kernels can be adjusted to enhance the network's ability to extract subtle textures and color features of the ore. By introducing an attention mechanism (such as a spatial attention module), the network can focus on key feature regions related to ore grade, suppressing background noise interference and improving the targeting and effectiveness of feature extraction. During model training, a large-scale ore image training dataset can be constructed, covering image samples from different mineral types (such as iron ore, copper ore, coal ore, etc.), different grade levels (high, medium, low grade and boundary grade ore), and different collection environments (different lighting, different angles, different conveyor belt materials and backgrounds), ensuring the model has broad adaptability. A transfer learning strategy is adopted, first pre-training on the public image dataset ImageNet to obtain general feature extraction capabilities, and then fine-tuning on a ore-specific dataset to accelerate model convergence and improve training results. Set reasonable training hyperparameters, adopt cosine annealing learning rate, optimize batch size between 32 and 128 based on hardware resources and dataset size, control the number of training rounds between 100 and 300, combine early stopping mechanism and model checkpoint saving, and select the optimal model parameters.
[0028] The ore sorting module can sort ore by grade and can also use pneumatic sorting valves and robotic arms to sort ore according to size, thereby improving sorting efficiency.
[0029] In summary, the real-time ore grade identification and sorting system provided in this embodiment of the invention obtains the ore grade level by acquiring and identifying images of the ore surface to ensure identification accuracy. Then, the ore is sorted according to the ore grade level. During the sorting process, pneumatic sorting valves and robotic arms can be used in conjunction with ore size to improve sorting efficiency. The ore grade and identification process can be automated, thereby reducing costs. This invention reduces sorting costs while ensuring the efficiency and accuracy of ore sorting.
[0030] In some embodiments of the present invention, the preprocessing of the ore surface image includes: Image enhancement of ore surface images based on adaptive histogram equalization algorithm; The enhanced mineral surface image is adjusted to a preset size and pixel normalized based on the bicubic interpolation algorithm. The image of the ore surface after image normalization is expanded using at least one of the following image expansion methods: random flipping, rotation, cropping, and adding noise.
[0031] It should be noted that image enhancement can employ an adaptive histogram equalization algorithm. This algorithm dynamically adjusts pixel distribution based on brightness and contrast differences in different regions of the ore image, improving overall image contrast and making ore surface features (such as mineral texture boundaries and color differences) clearer. It can also be combined with guided filtering algorithms to preserve image edge details while removing noise, avoiding feature blurring problems caused by traditional filtering algorithms and effectively removing noise introduced during image acquisition due to dust, light fluctuations, etc.
[0032] When performing image normalization, the image can be scaled to a fixed size (such as 224×224 pixels, 256×256 pixels, etc., which can be flexibly set according to the needs of model training and deployment) according to the input requirements of the deep learning model. The bicubic interpolation algorithm is used to ensure the quality of the scaled image. Then, the pixel values are normalized. By calculating the image mean and standard deviation, the pixel values are mapped to the range of [0, 1] or [-1, 1], which accelerates the training and inference process of the deep learning model and improves the model stability and convergence speed.
[0033] Furthermore, operations such as random flipping, rotation, cropping, and adding noise can be used to expand the training dataset and improve the model's generalization ability. During the real-time sorting stage, appropriate enhancement algorithms can be applied in real-time to adaptively adjust for special conditions in the acquired images (such as local overexposure or underexposure).
[0034] In some embodiments of the present invention, the preset convolutional layer is composed of a first convolutional kernel, a second convolutional kernel, and a third convolutional kernel connected by residuals in sequence, wherein the first convolutional kernel and the third convolutional kernel are 1×1 convolutional kernels, and the second convolutional kernel is a 3×3 convolutional kernel.
[0035] It should be noted that using small-sized convolutional kernels for residual connections can increase network depth, thereby improving the accuracy of the model in grade recognition.
[0036] In some embodiments of the present invention, the ore sorting based on ore grade using pneumatic sorting valves and robotic arms includes: The ore is directed into the corresponding channel based on its grade. When the ore particle size is less than or equal to the particle size threshold, a pneumatic separation valve is used for ore separation. When the ore particle size is larger than the particle size threshold, a robotic arm is used for ore sorting.
[0037] It should be noted that when using pneumatic sorting valves and robotic arms for ore sorting based on ore grade, the appropriate valve and arm can be selected based on the ore particle size. For sorting small-particle ore (particle size ≤ 20 mm), a high-speed response pneumatic sorting valve is selected. For sorting large-particle ore (particle size > 20 mm), a six-degree-of-freedom multi-joint robot is used as the mechanical sorting arm.
[0038] In some embodiments of the present invention, the system further includes: The data management module is used to store ore image acquisition data, image recognition results, system operating parameters and equipment status data, and generate ore grade distribution reports and equipment operating status reports.
[0039] It should be noted that the real-time ore grade identification and sorting system provided by this invention also includes a data management module, which can store ore image acquisition data, image recognition results, system operating parameters and equipment status data, and generate ore grade distribution reports and equipment operating status reports. This provides a foundation for exploring the potential relationship between ore grade and image features, acquisition environment and equipment parameters, and provides data support for optimizing image acquisition strategies, adjusting model parameters and improving sorting processes.
[0040] The present invention also provides a method for real-time ore grade identification and sorting based on the above-mentioned real-time ore grade identification and sorting system. Figure 2 This is a schematic flowchart of an embodiment of the real-time ore grade identification and sorting method provided by the present invention, as shown below. Figure 2 As shown, the real-time ore grade identification and sorting method includes: S201. Acquire real-time images of the ore surface and preprocess the images.
[0041] It should be noted that: images of the ore surface can be obtained using a high-definition industrial camera with a customized light source. Preprocessing of the ore surface image can include image enhancement, image normalization, and data augmentation. Image preprocessing can improve the efficiency of subsequent image recognition, thereby improving sorting efficiency.
[0042] S202. Based on the image recognition model, the preprocessed ore surface image is identified and the ore grade is output. The image recognition model is trained on the residual network with added spatial attention mechanism based on the ore surface image sample set with ore grade label. The learning rate during the training process is determined based on the cosine annealing algorithm. The residual network includes multiple preset convolutional layers, which are composed of 1×1 convolutional kernels and 3×3 convolutional kernels.
[0043] It should be noted that the image recognition model is a residual network with an added attention mechanism. In a residual network, the residual block structure can be optimized, and the number and size of convolutional kernels can be adjusted to enhance the network's ability to extract subtle textures and color features of the ore. By introducing attention mechanisms (such as channel attention and spatial attention modules), the network can focus on key feature regions related to ore grade, suppress background noise interference, and improve the targeting and effectiveness of feature extraction.
[0044] S203. Based on the ore grade, pneumatic sorting valves and robotic arms are used for ore sorting.
[0045] It should be noted that: when sorting ore, the ore can be sorted by grade, and pneumatic sorting valves and robotic arms can be used to sort the ore according to its size to improve sorting efficiency.
[0046] In summary, the real-time ore grade identification and sorting method provided by this invention obtains the ore grade level by acquiring and identifying images of the ore surface to ensure identification accuracy. Then, the ore is sorted according to the ore grade level. During the sorting process, pneumatic sorting valves and robotic arms can be used in conjunction with ore size to improve sorting efficiency. The ore grade and identification process can be automated, thereby reducing costs. This invention reduces sorting costs while ensuring the efficiency and accuracy of ore sorting.
[0047] Combination Figure 3 To overcome the challenges of ore sorting, break through the limitations of manual and traditional physical sorting, and meet the real-time sorting requirements of production lines, this invention proposes a real-time ore grade identification and sorting system based on deep learning. Combined with... Figure 4 The specific operation process of the deep learning-based real-time ore grade identification and sorting system includes the following steps: 1. During ore transportation, the camera continuously acquires images of the ore surface according to a set triggering mechanism (which can be triggered by the belt speed sensor signal or by a timed trigger), and transmits the raw image data to the data preprocessing module in real time. Simultaneously, the camera's built-in temperature and humidity sensors provide feedback on environmental parameters, helping to determine whether the image acquisition quality is affected by environmental factors.
[0048] Image acquisition utilizes a high-definition industrial camera with a resolution of at least 2000×1500 pixels, paired with a customized light source system. The industrial camera boasts high-speed acquisition capabilities, with a frame rate flexibly adjustable from 10-30 frames per second to adapt to production lines with varying ore conveying speeds. The light source system employs a highly uniform LED array, employing diffuse reflection lighting design to effectively avoid issues such as glare and shadows on the ore surface, ensuring that the acquired images clearly present key features such as ore surface texture, color, and mineral particle distribution. The camera is installed at a specific height and angle above the ore conveyor belt (determined through simulation and on-site debugging; typically 0.8-1.2 meters high, with an angle of 30-60 degrees to the belt conveyor direction), ensuring the shooting range covers the width of the belt and that the ore imaging is proportionally accurate and features are complete.
[0049] 2. Construct a multi-level image preprocessing algorithm to optimize the original image in all aspects.
[0050] Multi-level image preprocessing algorithms include image enhancement, image normalization, and data augmentation extension.
[0051] Image enhancement employs an adaptive histogram equalization algorithm, which dynamically adjusts pixel distribution to address brightness and contrast differences in different regions of the ore image, thereby improving the overall image contrast and making ore surface features (such as mineral texture boundaries and color differences) clearer. Combined with a guided filtering algorithm, it preserves image edge details while removing noise, avoiding feature blurring caused by traditional filtering algorithms, and effectively removing noise introduced during image acquisition due to dust, light fluctuations, etc.
[0052] Image normalization refers to first uniformly adjusting the image size. Based on the input requirements of deep learning models, the image is scaled to a fixed size (such as 224×224 pixels, 256×256 pixels, etc., which can be flexibly set according to the model training and deployment requirements). A bicubic interpolation algorithm is used to ensure the quality of the scaled image. Then, the pixel values are normalized by calculating the image mean and standard deviation, mapping the pixel values to the range of [0, 1] or [-1, 1], which accelerates the training and inference process of deep learning models and improves the stability and convergence speed of the models.
[0053] Data augmentation refers to the fact that during the model training phase, the preprocessing module can also perform offline data augmentation, such as random flipping, rotation, cropping, and adding noise, to expand the size of the training dataset and improve the model's generalization ability. During the real-time sorting phase, for special cases of the acquired images (such as local overexposure or underexposure), the corresponding augmentation algorithm is called in real time for adaptive adjustment.
[0054] 3. Receive standardized image data transmitted from the data preprocessing module, call the optimized deep learning model for feature extraction and classification recognition, and transmit the recognition results (grade level, confidence level) to the sorting execution module and data storage and management module in real time. At the same time, record key parameters in the model inference process (such as inference time and activated feature layers) for model performance analysis and iterative optimization.
[0055] An optimized deep learning model refers to a deep learning model that has been optimized based on the selection and improvement of the basic network, the model training strategy, and the model inference.
[0056] The basic network selection and improvement are based on the classic ResNet network, with improvements made to suit the characteristics of ore images. In the ResNet network, the residual block structure is optimized, and the number and size of convolutional kernels are adjusted to enhance the network's ability to extract subtle textures and color features of the ore. Attention mechanisms (such as channel attention and spatial attention modules) are introduced to focus the network on key feature regions related to ore grade, suppressing background noise interference and improving the targeting and effectiveness of feature extraction.
[0057] The model training strategy involves constructing a large-scale ore image training dataset, covering image samples from different mineral types (such as iron ore, copper ore, and coal ore), different grade levels (high, medium, low, and boundary grade ores), and different collection environments (different lighting, different angles, and different conveyor belt materials and backgrounds), ensuring the model has broad adaptability. A transfer learning strategy is employed, first pre-training on the public image dataset ImageNet to acquire general feature extraction capabilities, and then fine-tuning on a specific ore dataset to accelerate model convergence and improve training performance. Reasonable training hyperparameters are set, with a cosine annealing learning rate, and the batch size optimized between 32-128 based on hardware resources and dataset size. The number of training epochs is controlled between 100-300 epochs, combined with an early stopping mechanism and model checkpoint saving, to select the optimal model parameters.
[0058] Model inference optimization involves quantizing and pruning the trained model to reduce computational and storage requirements and improve inference speed. The TensorRT inference acceleration engine is employed and deployed and optimized on industrial control computer hardware to ensure that the inference time is kept below 100 milliseconds in real-time sorting scenarios, meeting the high-speed operation requirements of the production line. The model outputs ore grade levels (high grade, medium grade, low grade, waste rock) and corresponding confidence scores. The confidence threshold can be dynamically adjusted (set from 0.7 to 0.9, adapted to sorting accuracy and production line requirements) as the basis for sorting execution decisions.
[0059] 4. Receive the ore grade identification results transmitted by the deep learning recognition module, and combine them with pre-set sorting rules (e.g., high-grade ore enters channel A, medium-grade ore enters channel B, low-grade ore enters channel C, and waste rock enters channel D) to generate control commands in real time. Transmit the commands to the controller of the sorting actuator via the industrial control bus to drive the actuator's actions. Simultaneously, configure photoelectric sensors to monitor the actuator's status in real time, providing feedback on whether the sorting action is in place and whether there are any abnormalities such as jamming, ensuring a stable and reliable sorting process. Combination Figure 5 Based on the physical characteristics of the ore (hardness, particle size, etc.) and the production line layout, the hardware components of the sorting actuator include pneumatic sorting valves and robotic arms. Pneumatic Sorting Valve: For sorting small-particle ore (particle size ≤ 20 mm), a high-speed response pneumatic sorting valve is selected. This valve uses a direct-acting solenoid valve structure with a response time ≤ 5 milliseconds. By controlling the on / off state of the solenoid valve, the direction and force of compressed air injection are precisely controlled, blowing small-particle ore into the corresponding grade collection channel. The internal channel of the sorting valve adopts a streamlined design to reduce ore flow resistance and avoid blockage, while also featuring an anti-wear coating to extend its service life. Mechanical Sorting Arm: For sorting large-particle ore (particle size > 20 mm), a six-degree-of-freedom multi-joint robot is used as the mechanical sorting arm, with a repeatability accuracy ≤ 10 mm. The mechanical sorting arm is equipped with a high-precision vision guidance system. It acquires the ore's position and posture information in real time through an industrial camera, combines this with the grade results output by a deep learning recognition module, plans the optimal grasping path, accurately grasps the ore, and places it into the designated collection box. The mechanical sorting arm is equipped with an adaptive gripper that automatically adjusts the gripping force according to the shape and material of the ore, ensuring a stable and reliable gripping process and preventing ore damage or drop. The sorting actuator is installed at the end of the ore conveyor belt or at a specific sorting station, matching the speed of the conveyor belt to ensure a stable residence time of the ore in the sorting area and improve sorting accuracy.
[0060] 5. Build a distributed database system to implement data management functions, storing multi-dimensional data such as ore image acquisition data (raw images, pre-processed images), deep learning model recognition results (grade, confidence level, inference time), system operating parameters (camera parameters, pre-processing algorithm parameters, model hyperparameters), and equipment status data (camera temperature, sorting actuator working time). A combination of cloud and local storage is adopted. Local storage meets the need for fast real-time data access, while cloud storage enables long-term data backup and off-site sharing, ensuring data security and scalability.
[0061] The data management functions also include query and statistics functions, providing a flexible data query interface that supports queries based on multiple conditions such as time, ore type, and grade, generating ore grade distribution reports, equipment operation status reports, etc., to intuitively display the production line operation status and ore quality distribution; data analysis and mining functions, using data analysis algorithms (including cluster analysis and association rule mining) to perform in-depth processing on stored data, mining the potential relationship between ore grade and image features, acquisition environment, and equipment parameters, providing data support for optimizing image acquisition strategies, adjusting model parameters, and improving sorting processes; and model iteration support functions, periodically adding new stored data (especially ore image data that the model misidentifies) to the training dataset after labeling, providing a data source for the iterative training of deep learning models and achieving continuous optimization of model performance.
[0062] The following is a specific example of ore sorting, and the specific sorting steps include: 1. Hardware deployment.
[0063] In the ore crushing and screening workshop of a large iron ore mine, a high-definition industrial camera (model: HDC-X300, resolution 2448×2048 pixels, frame rate 30 frames / second) was installed 1.5 meters above the ore conveyor belt, with the camera's optical axis at a 45° angle to the belt, ensuring that the field of view covered a 1.2-meter width of the belt. A ring-shaped shadowless light source (power: 6.7W, color temperature: 6000-6500K) was installed around the camera, and the light was evenly distributed through a diffuser to eliminate reflections and shadows on the ore surface.
[0064] Through-beam photoelectric sensors (model: GTB10-N1212, detection accuracy: 1 mm) are installed at 2-meter intervals on both sides of the conveyor belt to trigger the camera to acquire images. The industrial control computer is placed in a control room 5 meters away from the conveyor belt and is connected to the industrial camera and photoelectric sensors via gigabit network cable.
[0065] For the sorting actuators, a pneumatic sorting valve (model: Matrix 850, response time 10 ms) and a six-degree-of-freedom mechanical sorting arm (model: CR-15iA, repeatability 0.5 mm) are installed at the end of the belt. The pneumatic sorting valve is connected to a compressed air tank via an air pipe, and an adaptive gripper and a vision-guided camera are installed at the end of the mechanical sorting arm. Both are controlled by a programmable logic controller (PLC), which is connected to an industrial computer via an industrial Ethernet network to achieve data interaction and command transmission.
[0066] 2. Dataset construction.
[0067] Iron ore samples were collected from the iron mine and three surrounding mining areas, covering ores of different grades (TFe content from 30% to 65%) and different forms (massive, granular, and powdery). In a laboratory environment, using the same image acquisition equipment, a total of 100,000 ore images were acquired under different lighting conditions (strong light at noon on a sunny day, weak light on a cloudy day, and supplemental lighting at night).
[0068] Five mining engineers and laboratory technicians with over 10 years of experience were organized to annotate the acquired images based on the iron ore chemical analysis results (grade determined using X-ray fluorescence spectrometry and chemical titration). The annotations included ore grade levels (high grade: TFe ≥ 60%, medium grade: 50% ≤ TFe < 60%, low grade: 25% ≤ TFe < 50%, waste rock: TFe < 25%) and the distribution areas of high-grade minerals (delineated on the image using a polygon tool). After annotation, senior experts reviewed the data to ensure an accuracy rate of over 98%, thus constructing the initial training dataset.
[0069] 3. Model training.
[0070] A training environment based on the PyTorch deep learning framework was set up on an industrial control computer. An improved ResNet-50 model was used as the basic architecture, and channel attention and spatial attention modules were introduced. First, pre-training was performed on the ImageNet dataset to learn general image features; the pre-training time was 36 hours.
[0071] Then, the constructed iron ore image dataset was divided into training, validation, and test sets in an 8:1:1 ratio. During training, the initial learning rate was set to 0.001, a cosine annealing learning rate scheduler was used, the batch size was 64, and the number of training epochs was 200. The Adam optimizer and cross-entropy loss function were used to train the model. The model was evaluated on the validation set every 5 epochs, and training stopped when the validation set accuracy no longer improved after 10 consecutive epochs. Ultimately, the model achieved a recognition accuracy of 96.5%, a recall of 95.8%, and an F1 score of 96.1% on the test set, meeting the requirements of actual production.
[0072] 4. System integration and testing.
[0073] The image acquisition module, data preprocessing module, deep learning recognition module, sorting execution module, and data storage and management module were activated sequentially. By manually placing iron ore samples of different grades on the conveyor belt, the image acquisition quality and triggering accuracy of the image acquisition module were tested to ensure that the camera could accurately capture clear and complete images as the ore passed by.
[0074] Examine the data preprocessing module's effect on the acquired images, observing whether the results after image enhancement, denoising, and normalization meet expectations. In the deep learning recognition module, input the preprocessed image and test the model's inference speed and recognition accuracy, ensuring that the inference time for a single image is within 80 milliseconds and the recognition results are accurate.
[0075] The control and sorting execution module was tested to assess the action response of the pneumatic sorting valve and the mechanical sorting arm, verifying their ability to accurately sort the ore to the corresponding collection channel based on the identification results. Simultaneously, the data storage and management module was checked to ensure data integrity and accuracy. Through joint debugging, any issues identified were promptly addressed, such as optimizing image acquisition parameters, adjusting model inference thresholds, and calibrating the action parameters of the sorting execution mechanism.
[0076] 5. Iterative model updates.
[0077] During system operation, representative data (including misclassified images and images of ore from newly acquired mining areas) are automatically selected weekly from newly acquired ore image data. These images are then labeled by professionals and added to the training dataset. An incremental learning algorithm is used to fine-tune the existing trained model, adjusting its parameters to adapt to the new data features.
[0078] A comprehensive evaluation of the model is conducted quarterly, using a new test dataset (containing ore images under various operating conditions that newly emerged during system operation) to test model performance. If the model's recognition accuracy drops by more than 95%, the model is retrained, and its structure and parameters are optimized to ensure that the model maintains high recognition accuracy at all times.
[0079] In this invention, the texture, color, and mineral distribution of the ore surface can be converted into digital information through high-definition image acquisition. A deep learning model can then extract subtle differences in features from massive image data, enabling accurate identification of ores of different grades and overcoming the limitations of manual and traditional physical sorting. After optimization and deployment, the deep learning model can complete image recognition and decision output in a short time, meeting the real-time sorting needs of the production line and laying the foundation for the construction of an intelligent sorting system.
[0080] The above provides a detailed description of the real-time ore grade identification and sorting system and method provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A real-time ore grade identification and sorting system, characterized in that, include: The image acquisition module is used to acquire images of the ore surface in real time and transmit them to the image processing module; The image processing module is used to preprocess the ore surface image and transmit the preprocessed ore surface image to the image recognition module; The image recognition module is used to recognize the preprocessed ore surface image based on the image recognition model and output the ore grade to the ore sorting module. The image recognition model is trained on a residual network with added spatial attention mechanism based on a sample set of ore surface images with ore grade labels. The learning rate during the training process is determined based on the cosine annealing algorithm. The residual network includes multiple preset convolutional layers, which are composed of 1×1 convolutional kernels and 3×3 convolutional kernels. The ore sorting module is used to sort ore based on its grade using pneumatic sorting valves and robotic arms.
2. The real-time ore grade identification and sorting system according to claim 1, characterized in that, The preprocessing of the ore surface image includes: Image enhancement of ore surface images based on adaptive histogram equalization algorithm; The enhanced mineral surface image is adjusted to a preset size and pixel normalized based on the bicubic interpolation algorithm. The image of the ore surface after image normalization is expanded using at least one of the following image expansion methods: random flipping, rotation, cropping, and adding noise.
3. The real-time ore grade identification and sorting system according to claim 1, characterized in that, The preset convolutional layer is composed of a first convolutional kernel, a second convolutional kernel, and a third convolutional kernel connected by residuals in sequence. The first and third convolutional kernels are 1×1 convolutional kernels, and the second convolutional kernel is a 3×3 convolutional kernel.
4. The real-time ore grade identification and sorting system according to claim 1, characterized in that, The ore sorting based on ore grade using pneumatic sorting valves and robotic arms includes: The ore is directed into the corresponding channel based on its grade. When the ore particle size is less than or equal to the particle size threshold, a pneumatic separation valve is used for ore separation. When the ore particle size is larger than the particle size threshold, a robotic arm is used for ore sorting.
5. The real-time ore grade identification and sorting system according to claim 1, characterized in that, The system also includes: The data management module is used to store ore image acquisition data, image recognition results, system operating parameters and equipment status data, and generate ore grade distribution reports and equipment operating status reports.
6. A method for real-time identification and sorting of ore grade based on the real-time ore grade identification and sorting system according to any one of claims 1 to 5, characterized in that, include: Acquire real-time images of the ore surface and preprocess the images; The image recognition model is used to identify the preprocessed ore surface image and output the ore grade. The image recognition model is trained on a residual network with added spatial attention mechanism based on a sample set of ore surface images with ore grade labels. The learning rate during the training process is determined based on the cosine annealing algorithm. The residual network includes multiple preset convolutional layers, which are composed of 1×1 convolutional kernels and 3×3 convolutional kernels. Ore sorting is performed using pneumatic sorting valves and robotic arms based on ore grade.
7. The method for real-time identification and sorting of ore grade according to claim 6, characterized in that, The preprocessing of the ore surface image includes: Image enhancement of ore surface images based on adaptive histogram equalization algorithm; The enhanced mineral surface image is adjusted to a preset size and pixel normalized based on the bicubic interpolation algorithm. The image of the ore surface after image normalization is expanded using at least one of the following image expansion methods: random flipping, rotation, cropping, and adding noise.
8. The method for real-time identification and sorting of ore grade according to claim 6, characterized in that, The preset convolutional layer is composed of a first convolutional kernel, a second convolutional kernel, and a third convolutional kernel connected by residuals in sequence. The first and third convolutional kernels are 1×1 convolutional kernels, and the second convolutional kernel is a 3×3 convolutional kernel.
9. The method for real-time identification and sorting of ore grade according to claim 6, characterized in that, The ore sorting based on ore grade using pneumatic sorting valves and robotic arms includes: The ore is directed into the corresponding channel based on its grade. When the ore particle size is less than or equal to the particle size threshold, a pneumatic separation valve is used for ore separation. When the ore particle size is larger than the particle size threshold, a robotic arm is used for ore sorting.
10. The method for real-time identification and sorting of ore grade according to claim 6, characterized in that, The method further includes: Ore grade distribution reports and equipment operation status reports are generated based on ore image acquisition data, image recognition results, system operating parameters, and equipment status data.