Method and system for automatically identifying fracture characteristics of mineral particles

Abstract

The invention relates to a mineral particle fracture characteristic automatic identification method and system, and the method comprises the steps: firstly scanning a large number of crushed and flatly arranged magnetite particles through an electron microscope to form a large number of electron microscope scanning pictures, and then manually calibrating each particle in the scanning pictures on a computer, the method comprises the following steps of: calibrating along-crystal fracture or cross-crystal fracture, counting the number of fracture types, importing most of calibrated pictures into a full convolutional neural network model based on an SSD (Solid State Disk) algorithm to learn and train the model, using the trained pictures as a picture recognition model, establishing recognition software on a computer, docking the recognition software with the model, and identifying the fractures. A scanning electron microscope is in butt joint with the computer to transmit data, so that the automatic mineral particle fracture characteristic identification system is integrally formed and used for identifying fracture types of magnetite particles formed after the magnetite crushing equipment crushes, counting the number and proportion of intergranular fractures and transgranular fractures, and determining the fracture characteristics of the magnetite particles according to the number and proportion of the intergranular fractures and the transgranular fractures. And basic data can be provided for researching and analyzing the relationship between the fracture type proportion and the crushing impact energy.

Description

Technical Field

[0001] The present invention relates to the technical field of mineral crushing and processing, and particularly to an automatic recognition method and system for the fracture characteristics of mineral particles. Background Art

[0002] The crushing and separation of metal ores are the prerequisites for subsequent smelting. The quality of the crushing effect directly affects the subsequent separation and smelting processes. Since metal ores are ores in which metal minerals are associated with some other minerals, and ores are formed by the admixture and combination of two or more different mineral crystals. The surface morphologies of the ore particles formed by different crushing energies are different. The surface morphology of the ore after crushing is mainly intergranular fracture, transgranular fracture, and intergranular-transgranular coupled fracture, which are all formed by the impact energy given to the ore particles by the crushing equipment during crushing. With the change of the crushing energy, the surface morphology of the crushed particles also changes. Whether there is a relationship between the two is not studied in the prior art.

[0003] To which fracture type does the surface morphology of the particles formed after ore crushing belong? How to quickly identify it and then perform statistical calculations can provide basic data for studying and analyzing the relationship between the fracture morphology and the crushing impact energy. Regarding the fracture type identification method, there is little mention in the prior art. The inventor believes that it is necessary to conduct research in this field and propose a feasible solution. Summary of the Invention

[0004] The purpose of the present invention is to solve the above problems and provide an automatic recognition method and system for the fracture characteristics of mineral particles.

[0005] The technical solution of the present invention is: an automatic recognition method for the fracture characteristics of mineral particles, including the following steps,

[0006] a. Crushing magnetite ore through a crushing device;

[0007] b. Collecting the ore particles after crushing, evenly dividing the particles into N groups, spreading each group flat and then performing electron microscopy scanning. The N scanned pictures after scanning are imported into a computer, where N≥1000; spreading and scanning enables multiple ore particles to be covered as much as possible in a single picture, so as to realize the recognition and statistics of multiple particles in one picture. After training the model in this way, the model also obtains the ability to recognize multiple particles simultaneously, which can improve the recognition efficiency of the model;

[0008] c. Manually calibrate the surface morphology of particles in each area of each scanned image. Different particle morphologies are respectively calibrated as one of the two fracture types: intergranular fracture or transgranular fracture, and summarize and calculate the number of intergranular fracture particles and the number of transgranular fracture particles in this image. Repeatedly operate this step to obtain N calibrated images. The manual calibration serves as the data sample for model training, which is a prerequisite for deep learning.

[0009] The manual calibration in step c is carried out based on the sample of the surface morphology image of ore particles with intergranular fracture. In the sample of this image, the fracture surface shows a sugar cube-like pattern, and the edges and corners of the fracture interface are clear. The manual calibration is also carried out based on certain criteria. Here, specific image samples of the two fracture morphologies are provided to solve the reference for calibration, improve the accuracy of calibration, make the calibrated images more reliable, and further ensure the reliability of the model obtained by deep learning.

[0010] The manual calibration in step c is carried out based on the sample of the surface morphology image of ore particles with transgranular fracture. In the sample of this image, the fracture surface shows a step pattern, or a river pattern, or a tearing pattern, or a secondary crack pattern.

[0011] d. Establish a fully convolutional neural network model based on the SSD algorithm for learning and training the calibrated images after step c in a computer.

[0012] The fully convolutional neural network model based on the SSD algorithm described in step d

[0013] Extract image features using the first few layers of the classic VGG16 network as the backbone network and remove the fully convolutional layer; after the VGG16 network, add 8 convolutional layers, and then select the feature maps generated by 6 convolutional layers in the entire network for particle detection, namely Conv4_3, Conv7, Conv8_2, Conv9_2, Conv10_2, Conv11_2. The feature maps of these 6 different convolutional layers predict bounding boxes with different sizes and aspect ratios; in this way, without changing the size of the feature map, a wider field of view of the image can be obtained, which is beneficial to the classification and localization of the target.

[0014] The SSD predefines prior boxes of different sizes and scales on the above 6 prediction layers. Among them, the feature maps output by Conv4_3, Conv10_2, and Conv11_2 predict 4 types of prior boxes at each position, with aspect ratios of 1:1, 1:2, and 2:1; while the feature maps output by Conv7, Conv8_2, and Conv9_2 predefine 6 types of prior boxes, with aspect ratios of 1:1, 1:2, 2:1, 1:3, and 3:1, and there are two different scales for the 1:1 prediction boxes; the scales of the bounding boxes predicted by the output feature maps of Conv4_3, Conv7, Conv8_2, Conv9_2, Conv10_2, and Conv11_2 are 38x38, 19x19, 10x10, 5x5, 3x3, and 1x1 respectively. The total number of predicted bounding boxes is the sum of the 6 detection layers, a total of 8732;

[0015] The loss function of the SSD object detection network includes two parts: class loss and location loss. The formula is as follows:

[0016]

[0017] Among them, x represents the Jaccard coefficient of the predicted box matching the ground truth box, c represents the classification confidence, l represents the four values of the center point coordinates, width, and height of the predicted box, and g represents the four values of the center point coordinates, width, and height of the ground truth box; α represents the weight, N represents the number of boundaries that match the ground truth box with a threshold greater than 0.5, and Lconf and Loc are the classification confidence loss function and the localization loss function respectively; among them, the class confidence loss and the localization loss respectively select Logloss and SmoothL1 as the loss functions.

[0018] e. Import most of the calibrated pictures completed in step c into the fully convolutional neural network model based on the SSD algorithm on the computer for deep learning training. After deep learning training, a picture recognition model is formed. Use the remaining calibrated pictures to test the picture recognition model. If the test accuracy meets the requirements, stop the deep learning training and retain the trained model; if the test accuracy does not meet the requirements, continue the deep learning training and testing until the test accuracy meets the requirements, and retain the trained picture recognition model; the test accuracy is a manually set training condition, which determines the number of iterations of the learning training. The higher the requirement, the more times of learning training, and the longer the time required. After meeting the requirement, stop the iteration and retain the trained picture recognition model that meets the standard for use.

[0019] In step e, 70% of the calibrated images are used for deep learning training, and the remaining 30% of the calibrated images are used for testing. Only by performing deep learning training on the model with a large number of images can the recognition accuracy be improved and the recognition ability of the model be ensured.

[0020] During the learning and training process of the model for the calibrated images in step e, the model extracts the shape, texture, and color features of the magnetite ore particles in the images and combines them with manual calibration for learning and training. When performing manual calibration, these features will be considered in combination with the sample images. Of course, when the model is trained, such features in the images will also be extracted and added to the training and learning. In this way, when the model obtained through training performs image recognition, it will also make a comprehensive judgment based on these features to ensure the accuracy of the judgment.

[0021] The required test accuracy in step e is not less than 94%.

[0022] f. Establish a connection between the software in the computer and the image recognition model obtained in step e above.

[0023] g. Collect the magnetite ore particles that have completed the crushing work and the ore particles that need to be identified for particle morphology. After spreading them out evenly, take pictures by electron microscope scanning to form pictures to be identified. The pictures to be identified also use electron microscope scanning after spreading out, which is the same as the process of obtaining the training sample pictures. What ability the model learns and what problem it is aimed at solving correspond to each other before and after.

[0024] h. Import the pictures to be identified into the above software through the software, and then apply the image recognition model to perform image recognition, output the number of transgranular fractures and intergranular fractures of the ore particles in the picture, calculate the intergranular fracture ratio, and output the recognized and calculated data through the software.

[0025] An automatic recognition system for the fracture characteristics of mineral particles established according to the method shown above. The system at least includes a scanning electron microscope, a computer, and software for identifying the fracture types of magnetite ore particles installed in the computer.

[0026] The scanning electron microscope is used to scan and take pictures of the sampled magnetite ore particles to obtain scanning pictures.

[0027] The computer is used to receive and store the scanning pictures of the scanning electron microscope and store the identified pictures.

[0028] The identification software includes

[0029] A basic operation module for pictures, which at least includes sub-modules for reading, displaying, saving, and reloading pictures.

[0030] An image recognition and classification module, which calls the image recognition model in step e above to classify and locate the particles in the picture.

[0031] A human-computer interaction module for manually selecting pictures for recognition and classification;

[0032] A data display module for displaying the recognition and classification results of pictures, including the positioning and classification of each magnetite ore particle target in the picture, as well as the quantity statistics of intergranular fractures and transgranular fracture zones;

[0033] This software is the software in step f of the described method.

[0034] The beneficial effects of the present invention are as follows: An automatic recognition method and system for the fracture characteristics of mineral particles of the present invention. The recognition method first scans a large number of flattened magnetite ore particles after crushing through an electron microscope to form a large number of electron microscope scanning pictures. Then, each particle in these scanning pictures is calibrated manually on a computer, calibrating intergranular fractures or transgranular fractures, and counting the quantity of fracture types. Most of the calibrated pictures are imported into a fully convolutional neural network model based on the SSD algorithm for learning and training of the model. After the training reaches the standard, it is used as a picture recognition model. Then, an identification software is established on the computer to dock with this model, and the scanning electron microscope is docked with this computer to transmit data, forming an automatic recognition system for the fracture characteristics of mineral particles as a whole, which is used to identify the fracture types of magnetite ore particles formed after the crushing of magnetite ore crushing equipment, and count the quantity and proportion of intergranular fractures and transgranular fractures, and can provide basic data for studying and analyzing the relationship between the proportion of fracture types and the crushing impact energy. Description of the Drawings

[0035] Figure 1 It is a picture of an intergranular fracture sample of magnetite ore particles to be detected in the present invention;

[0036] Figure 2 It is one of the pictures of transgranular fracture samples (step pattern) of magnetite ore particles to be detected in the present invention;

[0037] Figure 3 It is the second picture of transgranular fracture samples (river pattern) of magnetite ore particles to be detected in the present invention;

[0038] Figure 4 It is the third picture of transgranular fracture samples (tear pattern) of magnetite ore particles to be detected in the present invention;

[0039] Figure 5 It is the fourth picture of transgranular fracture samples (secondary crack) of magnetite ore particles to be detected in the present invention;

[0040] Figure 6 It is an electron microscope scanning picture of magnetite ore particles identified by the method of the present invention;

[0041] Figure 7 Schematic diagram of the fully convolutional neural network structure of the SSD algorithm in the method of the present invention;

[0042] Figure 8 Graph showing the relationship between the number of model training times and the detection accuracy in the method of the present invention;

[0043] Figure 9 Flow chart of the software operation in the recognition system of the present invention. Specific implementation mode

[0044] Example 1: Refer to Figures 1-8 , an automatic recognition method for the fracture characteristics of mineral particles, comprising the following steps,

[0045] a. Crushing magnetite ore through a crushing device;

[0046] b. Collecting the ore particles after crushing, evenly dividing the particles into N groups, spreading each group flat and then performing electron microscopy scanning, importing the N scanned pictures after scanning into a computer, N≥1000; spreading flat for scanning, so that as many ore particles as possible are covered in a single picture, so as to realize the recognition and statistics of multiple particles in one picture. After training the model in this way, the model also obtains the ability to recognize multiple particles simultaneously, which can improve the recognition efficiency of the model.

[0047] c. Manually calibrating the surface morphology of the particles in each area of each picture after scanning, calibrating different particle morphologies as one of the two fracture types of intergranular fracture or transgranular fracture respectively, and summarizing and calculating the number of intergranular fracture particles and the number of transgranular fracture particles in this picture. Repeatedly operating this step to obtain N calibrated pictures. Manual calibration establishes data samples for model training, which is a prerequisite for deep learning.

[0048] d. Establishing a fully convolutional neural network model based on the SSD algorithm for learning and training the calibrated pictures completed in step c in a computer,

[0049] e. Import most of the calibrated pictures after step c into the fully convolutional neural network model based on the SSD algorithm on the computer for deep learning training. After the deep learning training, a picture recognition model is formed. Use the remaining calibrated pictures to test the picture recognition model. If the test accuracy meets the requirements, stop the deep learning training and retain the trained model. If the test accuracy does not meet the requirements, continue the deep learning training and testing until the test accuracy meets the requirements, and retain the trained picture recognition model. The test accuracy is a manually set training condition, which determines the number of iterations of the learning training. The higher the requirement, the more times of learning training, and the longer the time required. In this embodiment, 94% is used as the passing requirement for the test accuracy. After meeting it, stop the iteration and retain the trained and qualified picture recognition model for use.

[0050] f. Establish a connection between the software in the computer and the picture recognition model obtained in step e above.

[0051] g. Collect the ore particles that have completed the crushing work and need to be identified for particle morphology. After spreading them out flat, take pictures by electron microscope scanning to form pictures to be identified. The pictures to be identified are also scanned by electron microscope after spreading out flat, which is the same as the acquisition process of the training sample pictures. What capabilities the model learns and what problems it solves correspond to each other before and after.

[0052] h. Import the pictures to be identified into the above software through the software, and then apply the picture recognition model to identify the pictures, output the number of transgranular fractures and intergranular fractures of the ore particles in the pictures, calculate the intergranular fracture ratio, and output the identified and calculated data through the software.

[0053] In step e above, 70% of the calibrated pictures are used for deep learning training, and the remaining 30% of the calibrated pictures are used for testing. Only by performing deep learning training on the model with a large number of pictures can its recognition accuracy be improved and the recognition ability of the model be guaranteed.

[0054] In step c above, the manual calibration is carried out according to the sample of the surface morphology pictures of the ore particles with intergranular fractures. In the sample of this picture, the fracture surface shows a sugar cube pattern, and the edges and corners of the fracture interface are clear.

[0055] In step c above, the manual calibration is carried out according to the sample of the surface morphology pictures of the ore particles with transgranular fractures. In the sample of this picture, the fracture surface shows a step pattern, or a river pattern, or a tearing pattern, or a secondary crack pattern.

[0056] The manual calibration is also carried out based on certain conditions. Here, specific picture samples are provided for the two fracture morphologies to solve the reference for calibration, improve the accuracy of calibration, make the calibrated pictures more reliable, and further ensure the reliability of the model obtained by deep learning.

[0057] In the learning and training process of the model for the calibration pictures in step e, the model extracts the shape, texture, and color features of magnetite ore particles in the pictures and combines them with manual calibration for learning and training. When performing manual calibration, these features will be considered in combination with sample pictures. Of course, when the model is trained, such features in the pictures will also be extracted and added to the training and learning. The model obtained through such training will also make a comprehensive judgment of multiple features through these features when performing picture recognition, ensuring the accuracy of the judgment.

[0058] The required test accuracy rate in step e is not less than 94%.

[0059] In the fully convolutional neural network model based on the SSD algorithm described in step d, the first few layers of the classic VGG16 network are used as the backbone network to extract image features, and the fully convolutional layer is removed; in this way, without changing the size of the feature map, a wider field of view of the image can be obtained, which is beneficial to the classification and positioning of the target; after the VGG16 network, 8 convolutional layers are added, and then the feature maps generated by 6 convolutional layers in the entire network are selected for particle detection, namely Conv4_3, Conv7, Conv8_2, Conv9_2, Conv10_2, Conv11_2. The predicted sizes and aspect ratios of the bounding boxes of the feature maps of 6 different convolutional layers are different;

[0060] SSD predefines prior boxes of different sizes and scales on the above 6 prediction layers. Among them, the feature maps output by Conv4_3, Conv10_2, and Conv11_2 predict 4 types of prior boxes at each position, and their aspect ratios are 1:1, 1:2, and 2:1; while the feature maps output by Conv7, Conv8_2, and Conv9_2 predefine 6 types of prior boxes, and the aspect ratios are 1:1, 1:2, 2:1, 1:3, and 3:1 respectively. Among them, the prediction boxes with an aspect ratio of 1:1 have two different scales; the scales of the bounding boxes predicted by the output feature maps of Conv4_3, Conv7, Conv8_2, Conv9_2, Conv10_2, and Conv11_2 are 38x38, 19x19, 10x10, 5x5, 3x3, and 1x1 respectively. The total number of predicted bounding boxes is the sum of the 6 detection layers, a total of 8732;

[0061] The loss function of the SSD object detection network includes two parts: class loss and location loss. The formula is as follows:

[0062]

[0063] Wherein, x represents the Jaccard coefficient of the prediction box matching the ground truth box, c represents the classification confidence, l represents the four values of the center point coordinates, width, and height of the prediction box, and g represents the four values of the center point coordinates, width, and height of the ground truth box; α represents the weight, N represents the number of boundaries matching the ground truth box with a threshold greater than 0.5, and L conf and L oc are the classification confidence loss function and the localization loss function respectively; among them, the categorical confidence loss and the localization loss respectively select Logloss and SmoothL1 as the loss functions.

[0064] Example 2, see Figure 9 , an automatic recognition system for mineral particle fracture characteristics established according to the method described in Example 1, which at least includes a scanning electron microscope, a computer, and software for identifying the fracture types of magnetite ore particles installed in the computer;

[0065] The scanning electron microscope is used to scan and photograph the sampled magnetite ore particles to obtain scanning pictures;

[0066] The computer is used to receive and store the scanning pictures of the scanning electron microscope and store the identified pictures;

[0067] The described recognition software includes

[0068] a basic operation module for pictures, which at least includes sub-modules for reading, displaying, saving, and reloading pictures;

[0069] a picture recognition and classification module, which calls the picture recognition model in step e of Example 1 to classify and locate the particles in the picture;

[0070] a human-computer interaction module for manually selecting pictures for recognition and classification;

[0071] a data display module for displaying the recognition and classification results of pictures, including the location and classification of each magnetite ore particle target in the picture, as well as the statistics of the number of intergranular fracture and transgranular fracture zones;

[0072] This software is the software in step f of Example 1 in the described method.

[0073] This system scans and photographs the symmetric mineral sample through a scanning electron microscope to form pictures, and then uses the trained software for identifying mineral particle fracture characteristics to identify the mineral sample pictures, obtaining the mineral particle fracture characteristic data in the mineral sample. This data can provide basic data for studying and analyzing the relationship between the proportion of fracture types and the crushing impact energy, and further provide a comparison basis for studying the selection of crushing impact energy in mineral grinding.

Claims

1. A method for automatically identifying fracture characteristics of mineral particles, comprising the following steps: a. Crushing the magnetite ore through crushing equipment; b. Collect the ore particles after crushing, divide the particles into N groups, spread each group flat and scan them with an electron microscope, and import the N scanned pictures into the computer, N ≥ 1000; c. Manually calibrate the surface morphology of particles in each area of ​​each scanned image. Different particle morphologies are calibrated as one of the two fracture types, intergranular fracture and transgranular fracture. The number of intergranular fracture particles and transgranular fracture particles in the image are calculated. Repeat this step to obtain N calibration images. d. Establish a fully convolutional neural network model based on the SSD algorithm in the computer for learning and training the calibrated images after step c is completed. e. Import most of the calibration images completed in step c into the fully convolutional neural network model based on the SSD algorithm on the computer for deep learning training. After the deep learning training, an image recognition model is formed. The image recognition model is tested with the remaining calibration images. If the test accuracy meets the requirements, the deep learning training is stopped and the trained model is retained; if the test accuracy does not meet the requirements, the deep learning training and testing are continued until the test accuracy meets the requirements, and the trained image recognition model is retained; f. Establish software in the computer to connect with the image recognition model obtained in step e above, g. Collect the magnetite ore after crushing and the ore particles that need to be identified, spread them out and scan them with an electron microscope to form photos to be identified. h. Import the photo to be identified into the above software through software, and then apply the image recognition model to perform image recognition, output the number of transgranular fractures and intergranular fractures of the ore particles in the image, calculate the proportion of intergranular fractures, and output the identified and calculated data through the software.

2. The method for automatically identifying fracture characteristics of mineral particles according to claim 1, characterized in that: In the step e, 70% of the calibration images are used for deep learning training, and the remaining 30% of the calibration images are used for testing.

3. The method for automatically identifying fracture characteristics of mineral particles according to claim 1, characterized in that: The manual calibration in step c is performed based on a sample of a surface morphology picture of ore particles fractured along the crystal, wherein the fracture in the sample of the picture is in the shape of rock candy, and the edges and corners of the fracture interface are clear.

4. The method for automatically identifying fracture characteristics of mineral particles according to claim 1 is characterized by: The manual calibration in step c is performed based on a sample of a surface morphology image of a ore particle with transgranular fracture, wherein the fracture in the sample of the image is a step pattern, a river pattern, a tear pattern, or a secondary crack pattern.

5. The method for automatically identifying fracture characteristics of mineral particles according to claim 1, characterized in that: In the learning and training process of the model for the calibration image in step e, the model extracts the shape, texture, and color features of the magnetite particles in the image, and performs learning and training in combination with manual calibration.

6. The method for automatically identifying fracture characteristics of mineral particles according to claim 1, characterized in that: The accuracy of the test that meets the requirements in step e is not less than 94%.

7. The method for automatically identifying fracture characteristics of mineral particles according to claim 1, characterized in that: The fully convolutional neural network model based on the SSD algorithm described in step d, The first few layers of the classic VGG16 network are used as the backbone network to extract image features, and the full convolution layer is removed. After the VGG16 network, 8 convolution layers are added, and then the feature maps generated by 6 convolution layers in the entire network are selected for particle detection, namely Conv4_3, Conv7, Conv8_2, Conv9_2, Conv10_2, and Conv11_2. The feature maps of 6 different convolution layers predict bounding boxes of different sizes and aspect ratios. SSD predefines prior frames of different sizes and scales on the above 6 prediction layers. The feature maps output by Conv4_3, Conv10_2 and Conv11_2 predict 4 types of prior frames at each position, with aspect ratios of 1:1, 1:2 and 2:1; while the feature maps output by Conv7, Conv8_2 and Conv9_2 predefine 6 types of prior frames, with aspect ratios of 1:1, 1:2, 2:1, 1:3 and 3:1, respectively, of which the 1:1 prediction frame has two different scales; the scales of the bounding boxes predicted by the output feature maps of Conv4_3, Conv7, Conv8_2, Conv9_2, Conv10 2, Conv11_2 are 38x38, 19x19, 10x10, 5x5, 3x3 and 1x1 respectively, and the total number of predicted bounding boxes is the sum of the 6 detection layers, a total of 8732; The loss function of the SSD target detection network includes two parts: category loss and position loss. The formula is as follows: Among them, x represents the Jaccard coefficient of the predicted box matching the true box, c represents the confidence of the classification, l represents the four values ​​of the center point coordinates, width and height of the predicted box, g represents the four values ​​of the center point coordinates, width and height of the true value box; α represents the weight, N represents the number of boundaries that match the threshold of the true box greater than 0.5, L conf and L oc They are classification confidence loss function and positioning loss function respectively; among them, Log loss and SmoothL1 are used as loss functions for category confidence loss and positioning loss respectively.

8. An automatic identification system for mineral particle fracture characteristics established according to any one of claims 1 to 6, the system comprising at least a scanning electron microscope, a computer, and magnetite ore particle fracture type identification software installed in the computer; The scanning electron microscope is used to scan and photograph the sampled magnetite particles to obtain scanned images; The computer is used to receive and store the scanned images of the scanning electron microscope, and is used to store the recognized images; The identification software includes The basic operation module of the picture includes at least the submodules of reading, displaying, saving and reloading the picture; An image recognition and classification module, which calls the image recognition model in step e to classify and locate particles in the image; Human-computer interaction module, used to manually select pictures for identification and classification; The data display module is used to display the recognition and classification results of the image, including the location and classification of each magnetite particle target in the image, as well as the number statistics of intergranular fractures and transgranular fracture zones; The software is the software in step f of the method described above.

Images