High-purity quartz sand sorting method and device

Abstract

The invention provides a high-purity quartz sand sorting method, and relates to the technical field of quartz sand processing equipment. Capturing a quartz sand surface image in real time by using a high-definition camera; transmitting the quartz sand surface image to visual identification equipment; the visual identification equipment identifies impurities in the quartz sand surface image through a YoLov3 target detection model based on a preset impurity standard; when the impurities are detected, calculating the distance between the impurities and the high-definition camera by using a distance calculation model, and determining the specific positions of the impurities; according to the detected impurity position, a signal is sent to a high-pressure air pump, and impurities are removed through a sand suction pipe connected with the high-pressure air pump. Foreign matter impurities such as black spots are removed through visual identification equipment, manual selection is replaced, the automation degree is high, and misoperation caused by visual fatigue due to long-time manual operation is also avoided.

Description

Technical Field

[0001] The invention relates to the technical field of quartz sand processing equipment, and in particular to a high-purity quartz sand separation method and device. Background Art

[0002] High-purity quartz sand is an important raw material used in high-tech fields such as optics, electronic materials, and semiconductors. Its purity and quality directly affect the performance and reliability of downstream products. The traditional high-purity quartz sand beneficiation process generally includes ore crushing, sorting, roasting, water quenching, sand making, screening, magnetic separation, pickling, flotation, drying, drying, grading, strong magnetic separation, chlorination and other processes. In recent years, due to the increasing quality requirements of downstream customers for high-purity quartz sand raw materials, the quality of high-purity quartz sand processing needs to be strictly controlled in every link. However, since impurities (such as black spots, yellow spots, hair, etc.) are inevitably mixed into quartz sand during mining and processing, the presence of such impurities seriously affects the purity and quality of quartz sand. Therefore, in the process of high-purity quartz sand production, how to efficiently and accurately detect and remove these impurities has become a key technical problem.

[0003] Traditional methods for detecting and removing impurities from quartz sand usually rely on manual detection or simple screening equipment. However, these methods have the following major problems:

[0004] The detection efficiency is low, manual detection is time-consuming and labor-intensive and relies on the operator's experience and attention. It is easy to miss or misdetect, and it is difficult to meet the high-efficiency needs of modern production.

[0005] Low accuracy: Traditional detection methods have difficulty identifying tiny impurities, especially when lighting conditions change or the background is complex. Small impurities are often difficult to detect and remove accurately, affecting the final quality of the product.

[0006] Low degree of automation: The current impurity removal process mostly relies on manual or simple mechanical means, lacks intelligence and automation, and is difficult to adapt to the continuity and precision requirements of large-scale production lines. Summary of the invention

[0007] The purpose of the present invention is to solve the shortcomings of the prior art and provide a method for separating high-purity quartz sand.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A method for separating high-purity quartz sand comprises the following steps:

[0010] Use a high-definition camera to capture the surface image of quartz sand in real time;

[0011] transmitting the quartz sand surface image to a visual recognition device;

[0012] The visual recognition device identifies impurities in the surface image of quartz sand based on the preset impurity standards through the YoLov3 object detection model;

[0013] When impurities are detected, the distance calculation model is used to calculate the distance between the impurity and the high-definition camera to determine the specific location of the impurity;

[0014] According to the detected impurity position, a signal is sent to the high-pressure air pump, and the impurities are removed through the sand suction pipe connected to the high-pressure air pump.

[0015] Furthermore, the YoLov3 target detection model is optimized through the following steps:

[0016] Construct a dataset of quartz sand surface images for training, including images with impurities annotated;

[0017] Use data enhancement technology to generate impurity samples in a variety of different environments, including quartz sand surface images under different angles, lighting and noise conditions;

[0018] Through the K-means clustering algorithm, adjust the Anchor Box of the YoLov3 model and optimize the size of the Anchor Box;

[0019] Use a semi-supervised learning method to train the YoLov3 model using a combination of labeled and unlabeled data;

[0020] During the model training process, the weight of the loss function is adjusted to make it pay more attention to the detection accuracy and position positioning of small targets.

[0021] Furthermore, the data enhancement technology specifically includes the following steps:

[0022] The quartz sand surface image is randomly rotated, scaled, translated, and flipped to simulate impurity detection scenarios at different angles;

[0023] Adjust image brightness, contrast, and saturation to suit different lighting conditions;

[0024] Add Gaussian noise and salt and pepper noise to simulate impurity detection in complex backgrounds;

[0025] Multiple extended data sets are generated through the above data augmentation operations, and the YoLov3 target detection model is iteratively trained.

[0026] Furthermore, the YoLov3 target detection model uses a non-maximum suppression algorithm to remove redundant detection boxes with high overlap and retain the prediction boxes with the highest confidence.

[0027] The process of the non-maximum suppression algorithm is as follows:

[0028] Sort all detection boxes from high to low according to their confidence;

[0029] Select the box with the highest confidence as the output box;

[0030] For the remaining boxes, the IoU with the output box is calculated. If the IoU value is higher than the set threshold, the box is considered as a duplicate detection and suppressed.

[0031] Furthermore, the semi-supervised learning method is used to train the YoLov3 model by combining labeled data and unlabeled data, including the following steps:

[0032] Extract some images with impurity annotations as labeled data and some unlabeled images as unlabeled data from the quartz sand surface image data set;

[0033] Use labeled data to perform preliminary supervised training on the YoLov3 model to obtain initial model parameters;

[0034] Use the preliminarily trained model to predict the unlabeled data, generate pseudo labels, and assign the corresponding impurity categories and positions to the unlabeled images;

[0035] Combine labeled data with pseudo-labeled data, retrain the YoLov3 model, and update the model weights and biases;

[0036] During the model training process, the loss functions of labeled data and unlabeled data are weighted, the prediction results of labeled data are optimized first, and the prediction accuracy of unlabeled data is improved through multiple iterations.

[0037] Furthermore, the steps of adjusting the Anchor Box of the YoLov3 model through the K-means clustering algorithm include:

[0038] Extract the width and height of all impurity targets from the quartz sand surface image dataset with impurities annotated;

[0039] Use K-means clustering algorithm to perform cluster analysis on the extracted impurity size data to determine the best AnchorBox size combination;

[0040] Adjust the size of the Anchor Box in the YoLov3 model according to the clustering results;

[0041] During model training, the updated Anchor Box is used for target detection.

[0042] Furthermore, the training method of the distance calculation model comprises the following steps:

[0043] The surface image data of quartz sand containing impurities is collected by a high-definition camera, and the actual distance between each impurity and the high-definition camera is recorded to form label data;

[0044] The surface image data of the quartz sand containing impurities and its corresponding actual distance label data are divided into a training set, a validation set and a test set;

[0045] Construct a distance calculation model, the input is the pixel data of the surface image data of quartz sand containing impurities, and the output is the distance between the impurities and the high-definition camera;

[0046] The distance calculation model consists of several hidden layers, each of which introduces a nonlinear relationship through an activation function;

[0047] The distance calculation model is supervised by using the training set. During the training process, the surface image of quartz sand containing impurities with known distance is input. The model gradually adjusts the distance calculation model parameters and fits the distance calculation function by learning the relationship between image features and actual distance.

[0048] The mean square error is used as the loss function to minimize the difference between the predicted value of the distance calculation model and the actual distance;

[0049] The Adam optimizer is used to dynamically adjust the weights and biases of the distance calculation model, and the initial learning rate is set to 0.001;

[0050] The validation set is used to evaluate the performance of the distance calculation model. By adjusting the number of neural network layers, the number of neurons, and the learning rate parameters, the prediction accuracy of the distance calculation model is optimized to obtain the final trained distance calculation model.

[0051] Furthermore, the distance calculation function is specifically expressed as:

[0052] d=f(W n ·σ(...(W2·σ(W1·I(x,y)+b1)+b2)...)+b n );

[0053] Where: d is the distance between the predicted impurity and the HD camera;

[0054] I(x,y) is the depth value or image feature value of the surface image of quartz sand at pixel point x,y);

[0055] W1,W2,...,W n It is the weight matrix of each layer of the neural network, which represents the parameters of the neural network model;

[0056] b1, b2, ..., bn are the bias terms of each layer;

[0057] σ is the activation function ReLU;

[0058] f is the linear activation function of the output layer, and the output is the regression value of the distance to the impurity.

[0059] Furthermore, the loss function is defined as:

[0060]

[0061] Among them, d i is the actual distance, d i ' is the distance predicted by the model, and N is the number of samples.

[0062] Furthermore, according to the detected impurity position, a signal is sent to a high-pressure air pump, and impurities are removed through a sand suction pipe connected to the high-pressure air pump, which specifically includes:

[0063] The output results of the distance calculation model are combined with the impurity position prediction results of the YoLov3 target detection model to control the movement path and suction strength of the high-pressure air pump and sand suction pipe to achieve accurate positioning and removal of impurities;

[0064] Among them, the distance between the impurities and the high-definition camera output by the distance calculation model is used to determine the vertical position of the sand suction pipe;

[0065] The impurity position output by the YoLov3 model is used to determine the horizontal movement position of the sand suction pipe;

[0066] The suction strength of the sand suction pipe is dynamically adjusted according to the size and type of impurities.

[0067] Compared with the prior art, the advantages and positive effects of the present invention are:

[0068] The present invention provides a high-purity quartz sand sorting method based on a high-definition camera, a YoLov3 target detection model and a distance calculation model. The method can not only efficiently and accurately identify impurities on the surface of quartz sand, but also achieve accurate removal of impurities through an automated control system, effectively detect small impurities in a variety of complex environments, and optimize the impurity removal process in combination with a distance calculation model, thereby improving the purity and quality of quartz sand and meeting the needs of modern production lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 The present invention provides a schematic flow chart of a method for separating high-purity quartz sand. DETAILED DESCRIPTION

[0070] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0071] like Figure 1 As shown, a high-purity quartz sand separation method provided by the present invention comprises the following steps:

[0072] S1, using a high-definition camera to capture the surface image of quartz sand in real time;

[0073] S2, transmitting the quartz sand surface image to a visual recognition device;

[0074] S3, the visual recognition device identifies impurities in the quartz sand surface image through the YoLov3 target detection model based on the preset impurity standard;

[0075] S4. When impurities are detected, the distance between the impurities and the high-definition camera is calculated using a distance calculation model to determine the specific location of the impurities;

[0076] S5. According to the detected impurity position, a signal is sent to the high-pressure air pump, and the impurities are removed through the sand suction pipe connected to the high-pressure air pump.

[0077] Specifically, a high-definition camera is used to capture the surface image of quartz sand in real time. The high-definition camera is set above the quartz sand sorting equipment, which can monitor the surface of quartz sand in real time and capture the details of its surface. The camera should have high resolution to ensure that tiny impurities on the quartz sand, such as black spots, yellow spots, hair, etc., can be clearly captured.

[0078] The surface image of quartz sand captured by a high-definition camera is transmitted in real time to the visual recognition device, which contains a preset image processing algorithm. First, the image is subjected to basic filtering, grayscale processing and edge detection to enhance the impurity features in the image. Afterwards, the visual recognition device analyzes the image using a preset impurity standard to determine whether there are impurities.

[0079] The YoLov3 target detection model is used to detect impurities in the processed quartz sand surface image. The model is optimized based on the quartz sand surface image data set during training, and different types of impurities are identified through the preset Anchor Box. The impurities in the quartz sand surface image are classified and detected to identify the location and type of impurities.

[0080] Among them, the specific steps and mechanisms of the YoLov3 target detection model in identifying impurities on the surface of quartz sand are as follows:

[0081] The YoLov3 object detection model uses a convolutional neural network (CNN) to extract multi-scale features of images. The YoLov3 object detection model extracts features at multiple scales of quartz sand images and can handle both small and large impurity targets. In the impurity detection of quartz sand, many impurities (such as black spots, yellow spots, hair, etc.) are very small, so the model needs to be able to extract subtle features at a smaller scale.

[0082] The network structure of the YoLov3 object detection model generates multiple feature maps through downsampling operations. These feature maps capture information at different levels. The bottom feature map focuses on capturing details, such as the shape and texture of impurities, while the high-level feature map focuses on overall semantic information, such as large-scale surface changes. For small impurities in quartz sand, the bottom feature map is more important.

[0083] Anchor Box is a series of predefined rectangular boxes of fixed size, which are used to generate candidate regions during the detection process. For impurities on the surface of quartz sand, the model uses Anchor Box to locate potential impurity areas. The size of Anchor Box can be optimized according to the impurity size in the training data through K-means clustering algorithm to ensure that it can accurately match impurities of different sizes.

[0084] During the detection process, Anchor Boxes are applied to different locations on the feature map to form multiple candidate regions. Each Anchor Box will be matched with the actual impurity target in the image, and the model will measure the overlap between the Anchor Box and the actual impurity through IoU (Intersection over Union). Based on the overlap, the model will select the best Anchor Box as the prediction box.

[0085] After the Anchor Box matches the actual impurities, the YoLov3 object detection model predicts the category (i.e., impurities or background) for each candidate region. For the surface of quartz sand, common impurities may include black spots, yellow spots, hair, etc. The model outputs the probability of each candidate region belonging to different impurity categories through the Softmax function.

[0086] During model training, the classification loss function is optimized based on the error between the actual category of the target (such as black dots, yellow dots, etc.) and the predicted category. The YoLov3 object detection model adjusts weights and biases to make the predicted category as close to the actual category label as possible.

[0087] The YoLov3 object detection model not only needs to identify the category of impurities, but also needs to accurately predict the bounding box of the impurities. In the process of position regression, the model calculates the position correction of each Anchor Box through regression. The correction values ​​output by the model include the center point coordinates, width and height correction values ​​of the candidate box, which are used to accurately adjust the Anchor Box to make it close to the actual impurity target.

[0088] Bounding box regression formula:

[0089] For each candidate box, the YoLov3 target detection model outputs 4 regression values ​​(t x ,t y ,t w ,t h ), which are used to adjust the center position and size of the Anchor Box respectively. The regression formula is as follows:

[0090] b x =σ(t x )+c x ; b y =σ(t y )+c y ; b w =p w ·e tw ; b h =p h ·e th .

[0091] in:

[0092] b x ,b y is the center coordinate of the prediction box;

[0093] c x ,c y It is the coordinate of the upper left corner of the Anchor Box;

[0094] b w ,b h is the width and height of the prediction box;

[0095] p w ,p h Is the width and height of the Anchor Box;

[0096] t x ,t y ,t w ,t h is the predicted value of the model;

[0097] σ is the Sigmoid activation function, which is used to compress the predicted value to between 0 and 1.

[0098] In target detection, an impurity target may be recognized by multiple Anchor Boxes at the same time, resulting in multiple duplicate prediction boxes. To solve this problem, the YoLov3 target detection model uses the non-maximum suppression (NMS) algorithm to remove redundant detection boxes with high overlap and retain the prediction boxes with the highest confidence. The NMS process is as follows: sort all detection boxes from high to low according to confidence; select the box with the highest confidence as the output box; for the remaining boxes, calculate the IoU with the output box. If the IoU value is higher than the set threshold (for example, 0.5), the box is considered a duplicate detection and suppressed.

[0099] For impurity detection on the surface of quartz sand, the IoU threshold of NMS can be adjusted according to the actual situation. Generally speaking, the distance between small impurities is small, and setting a lower IoU threshold (such as 0.3) can reduce missed detection.

[0100] Finally, the YoLov3 object detection model will output the detected impurities and their corresponding bounding boxes and confidence values. The confidence represents the model's credibility in the detection results. Usually a confidence threshold (such as 0.5) is set, and detection results below this threshold will be filtered out.

[0101] In this specific implementation, regarding the impurity removal signal transmission: for the detected impurities, the model transmits the location and category information of the impurities to the control system of the high-pressure air pump, and the control system accurately adjusts the position and suction force of the sand suction pipe to automatically remove the impurities detected on the surface of the quartz sand.

[0102] In some preferred embodiments, the YoLov3 target detection model is optimized by the following steps:

[0103] Construct a dataset of quartz sand surface images for training, including images with impurities annotated;

[0104] Use data enhancement technology to generate impurity samples in a variety of different environments, including quartz sand surface images under different angles, lighting and noise conditions;

[0105] Through the K-means clustering algorithm, adjust the Anchor Box of the YoLov3 model and optimize the size of the Anchor Box;

[0106] Use a semi-supervised learning method to train the YoLov3 model using a combination of labeled and unlabeled data;

[0107] During the model training process, the weight of the loss function is adjusted to make it pay more attention to the detection accuracy and position positioning of small targets.

[0108] Among them, the data enhancement technology specifically includes the following steps:

[0109] The quartz sand surface image is randomly rotated, scaled, translated, and flipped to simulate impurity detection scenarios at different angles;

[0110] Adjust image brightness, contrast, and saturation to suit different lighting conditions;

[0111] Add Gaussian noise and salt and pepper noise to simulate impurity detection in complex backgrounds;

[0112] Multiple extended data sets are generated through the above data augmentation operations, and the YoLov3 target detection model is iteratively trained.

[0113] The steps to adjust the Anchor Box of the YoLov3 model through the K-means clustering algorithm include:

[0114] Extract the width and height of all impurity targets from the quartz sand surface image dataset with impurities annotated;

[0115] Use K-means clustering algorithm to perform cluster analysis on the extracted impurity size data and determine the best AnchorBox size combination to match impurities of different sizes;

[0116] According to the clustering results, the size of the Anchor Box in the YoLov3 model is adjusted to make it more suitable for the diversity and distribution of impurities on the surface of quartz sand;

[0117] During the model training process, the updated Anchor Box is used for target detection to improve the accuracy and recall of impurity detection.

[0118] Among them, the semi-supervised learning method is used to combine labeled data and unlabeled data to train the YoLov3 model, which specifically includes the following steps:

[0119] Extract some images with impurity annotations as labeled data and some unlabeled images as unlabeled data from the quartz sand surface image data set;

[0120] Use labeled data to perform preliminary supervised training on the YoLov3 model to obtain initial model parameters;

[0121] Use the preliminarily trained model to predict the unlabeled data, generate pseudo labels, and assign the corresponding impurity categories and positions to the unlabeled images;

[0122] Combine labeled data with pseudo-labeled data, retrain the YoLov3 model, and update the model weights and biases;

[0123] During the model training process, the loss functions of labeled data and unlabeled data are weighted, the prediction results of labeled data are optimized first, and the prediction accuracy of unlabeled data is improved through multiple iterations.

[0124] The training method of the distance calculation model provided in this specific implementation includes the following steps:

[0125] The surface image data of quartz sand containing impurities is collected by a high-definition camera, and the actual distance between each impurity and the high-definition camera is recorded to form label data;

[0126] The surface image data of the quartz sand containing impurities and its corresponding actual distance label data are divided into a training set, a validation set and a test set;

[0127] Construct a distance calculation model, the input is the pixel data of the surface image data of quartz sand containing impurities, and the output is the distance between the impurities and the high-definition camera;

[0128] The distance calculation model consists of several hidden layers, each of which introduces a nonlinear relationship through an activation function;

[0129] The distance calculation model is supervised by using the training set. During the training process, the surface image of quartz sand containing impurities with known distance is input. The model gradually adjusts the distance calculation model parameters and fits the distance calculation function by learning the relationship between image features and actual distance.

[0130] The mean square error is used as the loss function to minimize the difference between the predicted value of the distance calculation model and the actual distance;

[0131] The Adam optimizer is used to dynamically adjust the weights and biases of the distance calculation model, and the initial learning rate is set to 0.001;

[0132] The validation set is used to evaluate the performance of the distance calculation model. By adjusting the number of neural network layers, the number of neurons, and the learning rate parameters, the prediction accuracy of the distance calculation model is optimized to obtain the final trained distance calculation model.

[0133] In some preferred embodiments, the distance calculation function is specifically expressed as:

[0134] d=f(W n ·σ(...(W2·σ(W1·I(x,y)+b1)+b2)...)+b n );

[0135] Where: d is the distance between the predicted impurity and the HD camera;

[0136] I(x,y) is the depth value or image feature value of the surface image of quartz sand at pixel point x,y);

[0137] W1,W2,...,W n It is the weight matrix of each layer of the neural network, which represents the parameters of the neural network model;

[0138] b1, b2, ..., bn are the bias terms of each layer;

[0139] σ is the activation function ReLU;

[0140] f is the linear activation function of the output layer, and the output is the regression value of the distance to the impurity.

[0141] In some preferred embodiments, the loss function is defined as:

[0142]

[0143] Among them, d i is the actual distance, d i ' is the distance predicted by the model, and N is the number of samples.

[0144] Specifically, the specific steps and mechanisms of the distance calculation model in predicting impurity distances are as follows:

[0145] The input of the distance calculation model is the pixel data in the quartz sand surface image, and each pixel corresponds to the depth value or image feature value in the image. By capturing the image of the quartz sand surface with a high-definition camera, the model can obtain the specific location of the impurity target in the image.

[0146] The input layer of the distance calculation model receives the depth information of the pixel points. The structure of the neural network allows the model to process this information through multiple layers of the network and gradually extract features related to distance prediction. These features are used for subsequent distance regression calculations.

[0147] The distance calculation model adopts a multi-layer perceptron structure, which gradually converts the data of image pixels into distance prediction values ​​through nonlinear transformation of multiple layers of neurons. The distance calculation model contains several hidden layers, and each hidden layer introduces nonlinearity through an activation function (such as ReLU), thereby enhancing the model's ability to capture complex data patterns.

[0148] Different hidden layers are responsible for feature extraction at different levels. The lower hidden layers focus on capturing local features, such as the size and shape of impurities, while the neurons in higher layers combine more abstract global features to further optimize the accuracy of distance prediction.

[0149] At the output layer of the neural network, the distance calculation model predicts the distance between the impurity and the camera based on the input feature value. After each input data point passes through the network, the output is the predicted distance value d. The output layer of the neural network is usually a linear activation function f(x) = x, so that the distance can be directly regressed.

[0150] By using the Adam optimizer, the distance calculation model dynamically adjusts weights and biases in each training iteration to minimize the loss function. The optimization process will gradually improve the prediction accuracy of the distance calculation model, enabling it to more accurately fit the distance between impurities and the high-definition camera.

[0151] During the training process of the distance calculation model, the validation set is used to evaluate the generalization ability of the model. By adjusting the hyperparameters of the neural network, such as the number of layers, number of neurons, and learning rate, the performance of the distance calculation model on different data sets is optimized.

[0152] Based on the feedback from the validation set, the distance calculation model can be further fine-tuned to ensure that it can accurately predict the distance of impurities under different quartz sand samples and environmental conditions. This process can be optimized through multiple iterations to ultimately obtain a high-precision distance calculation model.

[0153] Ultimately, the distance calculation model will output the predicted distance between the impurity and the high-definition camera. This distance value is combined with the impurity detection results to provide accurate location information for the high-pressure air pump and sand suction pipe.

[0154] The distance calculation model provided in this specific embodiment inputs pixel data in the quartz sand image and uses the multi-level feature extraction and regression calculation mechanism of the neural network to accurately predict the distance between impurities and the high-definition camera. Through dynamic adjustment of the mean square error loss function and the Adam optimizer, the model can improve its prediction accuracy in multiple iterative training. Combined with the distance output and impurity detection results, the system can achieve efficient and automatic impurity removal.

[0155] In this specific implementation, according to the detected impurity position, a signal is sent to a high-pressure air pump, and impurities are removed through a sand suction pipe connected to the high-pressure air pump, specifically including:

[0156] The output results of the distance calculation model are combined with the impurity position prediction results of the YoLov3 target detection model to control the movement path and suction strength of the high-pressure air pump and sand suction pipe to achieve accurate positioning and removal of impurities;

[0157] Among them, the distance between the impurities and the high-definition camera output by the distance calculation model is used to determine the vertical position of the sand suction pipe;

[0158] The impurity position output by the YoLov3 model is used to determine the horizontal movement position of the sand suction pipe;

[0159] The suction strength of the sand suction pipe is dynamically adjusted according to the size and type of impurities.

[0160] Specifically, they include:

[0161] After completing impurity detection and distance calculation, the impurity position prediction result of the YoLov3 target detection model is combined with the output result of the distance calculation model. The output of the YoLov3 target detection model includes the horizontal coordinate position of the impurity in the image (such as x and y coordinates), which is used to determine the position of the impurity in the image; the distance calculation model outputs the vertical distance between the impurity and the high-definition camera, which is used to determine the depth of the impurity in the actual space (i.e., the vertical position of the sand suction tube). By fusing these two parts of data, the precise coordinate position of the impurity in three-dimensional space can be calculated; the fused three-dimensional coordinates are passed to the control system of the high-pressure air pump and the drive system of the sand suction tube to determine the movement path of the sand suction tube.

[0162] Vertical direction: According to the output of the distance calculation model, the movement of the sand suction pipe on the Z axis is determined, that is, the vertical height of the sand suction pipe is adjusted to accurately approach the depth where the impurities are located.

[0163] Horizontal direction: The XY plane coordinates output by the YoLov3 target detection model are used to control the movement of the sand suction tube on the X-axis and Y-axis to achieve precise positioning of the sand suction tube in the horizontal position.

[0164] Through the control of three-dimensional coordinates, the sand suction pipe can be moved to the impurity position with high precision in complex environments.

[0165] The suction strength of the sand suction pipe is dynamically adjusted according to the type and size of impurities detected:

[0166] The YoLov3 target detection model not only outputs the coordinate position of the impurities, but also can identify the types of impurities (such as black spots, yellow spots, hair, etc.). The suction strength can be adjusted according to the different types of impurities.

[0167] Based on the size of the bounding box of the detected impurities, the volume and surface area of ​​the impurities can be judged, and the suction strength can be further determined. For larger impurities, the suction tube will increase the suction to ensure that the impurities can be completely adsorbed, while for smaller impurities, the suction will be reduced to avoid sucking in too many quartz sand particles.

[0168] After the sand suction tube starts to move, real-time detection information is continuously obtained to dynamically adjust the position and suction of the sand suction tube; if the impurities are not completely removed during the suction process, the suction strength is readjusted or the sand suction tube is repositioned according to the residual detection results to ensure that the impurities are completely removed.

[0169] Since the size and position of impurities vary, the trajectory of the sand suction tube is continuously corrected by updating the position and distance information in real time to ensure that the sand suction tube can effectively remove each impurity in turn in the case of complex background and multiple impurities.

[0170] The above description is only a preferred embodiment of the present invention and does not limit the present invention in other forms. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A method for separating high-purity quartz sand, characterized in that: The steps include: Use a high-definition camera to capture the surface image of quartz sand in real time; transmitting the quartz sand surface image to a visual recognition device; The visual recognition device identifies impurities in the surface image of quartz sand based on the preset impurity standards through the YoLov3 object detection model; When impurities are detected, the distance calculation model is used to calculate the distance between the impurity and the high-definition camera to determine the specific location of the impurity; According to the detected impurity position, a signal is sent to the high-pressure air pump, and the impurities are removed through the sand suction pipe connected to the high-pressure air pump.

2. A method for separating high-purity quartz sand according to claim 1, characterized in that: The YoLov3 target detection model is optimized by the following steps: Construct a dataset of quartz sand surface images for training, including images with impurities annotated; Use data enhancement technology to generate impurity samples in a variety of different environments, including quartz sand surface images under different angles, lighting and noise conditions; Through the K-means clustering algorithm, adjust the Anchor Box of the YoLov3 model and optimize the size of the Anchor Box; Use a semi-supervised learning method to train the YoLov3 model using a combination of labeled and unlabeled data; During the model training process, the weight of the loss function is adjusted to make it pay more attention to the detection accuracy and position positioning of small targets.

3. A method for separating high-purity quartz sand according to claim 2, characterized in that: The data enhancement technology specifically includes the following steps: The quartz sand surface image is randomly rotated, scaled, translated, and flipped to simulate impurity detection scenarios at different angles; Adjust image brightness, contrast, and saturation to suit different lighting conditions; Add Gaussian noise and salt and pepper noise to simulate impurity detection in complex backgrounds; Multiple extended data sets are generated through the above data augmentation operations, and the YoLov3 target detection model is iteratively trained.

4. A method for separating high-purity quartz sand according to claim 1, characterized in that: The YoLov3 target detection model uses a non-maximum suppression algorithm to remove redundant detection boxes with high overlap and retain the prediction box with the highest confidence. The process of the non-maximum suppression algorithm is as follows: Sort all detection boxes from high to low according to their confidence; Select the box with the highest confidence as the output box; For the remaining boxes, the IoU with the output box is calculated. If the IoU value is higher than the set threshold, the box is considered as a duplicate detection and suppressed.

5. A method for separating high-purity quartz sand according to claim 1, characterized in that: in, The semi-supervised learning method is used to train the YoLov3 model by combining labeled data and unlabeled data, including the following steps: Extract some images with impurity annotations as labeled data and some unlabeled images as unlabeled data from the quartz sand surface image data set; Use labeled data to perform preliminary supervised training on the YoLov3 model to obtain initial model parameters; Use the preliminarily trained model to predict the unlabeled data, generate pseudo labels, and assign the corresponding impurity categories and positions to the unlabeled images; Combine labeled data with pseudo-labeled data, retrain the YoLov3 model, and update the model weights and biases; During the model training process, the loss functions of labeled data and unlabeled data are weighted, the prediction results of labeled data are optimized first, and the prediction accuracy of unlabeled data is improved through multiple iterations.

6. A method for separating high-purity quartz sand according to claim 1, characterized in that: The steps to adjust the Anchor Box of the YoLov3 model through the K-means clustering algorithm include: Extract the width and height of all impurity targets from the quartz sand surface image dataset with impurities annotated; Use K-means clustering algorithm to perform cluster analysis on the extracted impurity size data to determine the best anchor box size combination; Adjust the size of the Anchor Box in the YoLov3 model according to the clustering results; During model training, the updated Anchor Box is used for target detection.

7. A method for separating high-purity quartz sand according to claim 1, characterized in that: The training method of the distance calculation model comprises the following steps: The surface image data of quartz sand containing impurities is collected by a high-definition camera, and the actual distance between each impurity and the high-definition camera is recorded to form label data; The surface image data of the quartz sand containing impurities and its corresponding actual distance label data are divided into a training set, a validation set and a test set; Construct a distance calculation model, the input is the pixel data of the surface image data of quartz sand containing impurities, and the output is the distance between the impurities and the high-definition camera; The distance calculation model consists of several hidden layers, each of which introduces a nonlinear relationship through an activation function; The distance calculation model is supervised by using the training set. During the training process, the surface image of quartz sand containing impurities with known distance is input. The model gradually adjusts the distance calculation model parameters and fits the distance calculation function by learning the relationship between image features and actual distance. The mean square error is used as the loss function to minimize the difference between the predicted value of the distance calculation model and the actual distance; The Adam optimizer is used to dynamically adjust the weights and biases of the distance calculation model, and the initial learning rate is set to 0.001; The validation set is used to evaluate the performance of the distance calculation model. By adjusting the number of neural network layers, the number of neurons, and the learning rate parameters, the prediction accuracy of the distance calculation model is optimized to obtain the final trained distance calculation model.

8. A method for separating high-purity quartz sand according to claim 7, characterized in that: The distance calculation function is specifically expressed as: d=f(W n ·σ(…(W2·σ(W1·I(x,y)+b1)+b2)…)+b n ); Where: d is the distance between the predicted impurity and the HD camera; I(x,y) is the depth value or image feature value of the surface image of quartz sand at pixel point x,y); W1,W2,...,W n It is the weight matrix of each layer of the neural network, which represents the parameters of the neural network model; b1, b2, ..., bn are the bias terms of each layer; σ is the activation function ReLU; f is the linear activation function of the output layer, and the output is the regression value of the distance to the impurity.

9. A method for separating high-purity quartz sand according to claim 8, characterized in that: The loss function is defined as: Among them, d i is the actual distance, d i ' is the distance predicted by the model, and N is the number of samples.

10. A method for separating high-purity quartz sand according to claim 1, characterized in that: The method of sending a signal to a high-pressure air pump according to the detected impurity position, and removing impurities through a sand suction pipe connected to the high-pressure air pump, specifically includes: The output results of the distance calculation model are combined with the impurity position prediction results of the YoLov3 target detection model to control the movement path and suction strength of the high-pressure air pump and sand suction pipe to achieve accurate positioning and removal of impurities; Among them, the distance between the impurities and the high-definition camera output by the distance calculation model is used to determine the vertical position of the sand suction pipe; The impurity position output by the YoLov3 model is used to determine the horizontal movement position of the sand suction pipe; The suction strength of the sand suction pipe is dynamically adjusted according to the size and type of impurities.

Images