Method for recognizing coal gangue and sundries and automatic sorting system

Abstract

The application discloses a kind of coal gangue and sundry identification method and automatic sorting system, the method includes: obtaining the image information of coal flow, based on the appearance characteristics of coal, coal gangue and sundry determines gray threshold, the image of image information is separated, and the preprocessed image is obtained;The preprocessed image is distance conversion, is abstracted as three-dimensional image, the coal that is in contact with coal gangue and sundry is separated, and the coal gangue image is obtained;After the coal gangue image is preprocessed, the coal gangue image is segmented and extracted, and the coal gangue segmentation map is obtained;Based on the feature extraction model created in advance, the coal gangue and sundry in coal gangue segmentation map are extracted, and the identification of coal gangue and sundry is realized.The application realizes the sorting of coal gangue and sundry by image recognition, solves the occurrence of mis-sorting, missed detection and other situations, and improves the accuracy of coal gangue sorting.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coal gangue sorting, in particular to a coal gangue and sundry identification method and an automatic coal gangue and sundry sorting system. BACKGROUND

[0002] In the field of coal, coal gangue sorting is an important field of coal production. In the prior art, X-ray detection is usually used to identify coal gangue. However, in the process of coal gangue sorting, when the size of coal gangue is too large (for example, more than 300 mm), X-ray penetration is difficult, which leads to the failure to identify coal gangue smoothly in the sorting process, resulting in mis-sorting. In addition, in the existing coal gangue sorting technology, only coal gangue can be identified, and other sundries cannot be identified, resulting in incomplete sorting. SUMMARY

[0003] The technical problem to be solved by the present application is to provide a coal gangue and sundry identification method and an automatic coal gangue and sundry sorting system, so as to realize the sorting of coal gangue and sundry through image recognition, solve the occurrence of mis-sorting and missed detection, and improve the accuracy of coal gangue sorting.

[0004] To solve the above technical problems, the present application provides a coal gangue and sundry identification method, comprising:

[0005] Obtaining image information of coal flow, determining a gray threshold value based on the appearance characteristics of coal, coal gangue and sundry, separating the image foreground and background of the image information to obtain a pretreated image;

[0006] Distance transforming the pretreated image to abstract it into a three-dimensional image, separating the coal in contact with the coal gangue and sundry to obtain a coal gangue image;

[0007] After pretreating the coal gangue image, segmenting and extracting the coal gangue image to obtain a coal gangue segmentation image;

[0008] Based on a pre-created feature extraction model, extracting features of the coal gangue and sundry in the coal gangue segmentation image to realize the identification of the coal gangue and sundry.

[0009] In some embodiments, the feature extraction model adopts an encoder-decoder architecture, including an up-sampling layer, an intermediate layer and a down-sampling layer;

[0010] The down-sampling layer adopts a ResNet34 network structure to perform convolution, normalization and activation function down-sampling features on the coal gangue segmentation image to obtain a down-sampling feature segmentation image;

[0011] The up-sampling layer adopts a residual network and de-convolution, fuses path features of the encoder and the decoder, uses an activation function and convolution to fuse features, and outputs a predicted picture through convolution;

[0012] The intermediate layer includes a Transformer module and a composite attention multi-scale hollow convolution module.

[0013] In some embodiments, when the coal gangue image is segmented and extracted, the coal gangue image is segmented into a plurality of image blocks, and the plurality of image blocks are processed into sequences, and a class label is added to each sequence.

[0014] In some embodiments, before the down-sampling layer inputs the down-sampled feature segmentation image into the Transformer module, a position encoding is added to each image sequence, and the calculation formula is:

[0015] T = [W1X1 + W2X2 + L + W i X i ]+W pos

[0016] In the formula, T is the segmented feature block with added position encoding, W i is the weight of the i-th segmented feature block (i = 1, 2, …, I ∈ N+), X i is the i-th down-sampled segmented feature image (i = 1, 2, …, I ∈ N+), W pos ∈RD is a spatial position information encoding, where D is the linear space dimension.

[0017] In some embodiments, the Transformer module is formed by cascading a plurality of Transformer layers, and each Transformer layer includes a multi-head self-attention layer and a multi-layer perception layer.

[0018] The multi-head self-attention layer includes the following steps:

[0019] In the formula, Q is a query matrix, K is a key matrix, and V is a value matrix.

[0020] The weight of the segmented feature block is obtained by the following steps:

[0021] The query limit of input 1 is multiplied by the key vectors of input 1, input 2, and input 3, respectively, to obtain three vector values;

[0022] After softmax, the vector values are converted into respective importance degrees distributed between 0 and 1;

[0023] Each of the importance degrees is multiplied by the corresponding value vector and summed to obtain the weight of the segmented feature block.

[0024] In some embodiments, the composite attention multi-scale hollow convolution module comprises a composite channel spatial attention module and a multi-scale hollow convolution module.

[0025] The composite channel spatial attention module realizes feature extraction through one-dimensional convolution, performs normalization, obtains feature maps based on channel global max pooling and global average pooling, performs concat operation, combines into one channel, and finally performs point multiplication with the channel output feature to obtain the weight distribution of the spatial feature through Sigmoid.

[0026] The multi-scale hollow convolution module comprises three layers of hollow convolution, a 1x1 convolution layer, and a pooling layer for parallel sampling, and the hollow rates of the three layers of hollow convolution are 2, 4, and 8 respectively.

[0027] According to another aspect of the present application, an automatic sorting system for coal gangue and sundries is provided, comprising: a power supply, an identification device, a control device, and a grabbing device.

[0028] The power supply supplies power to the automatic sorting system.

[0029] The identification device comprises an image acquisition unit, a memory, and a processor, the image acquisition unit is used to acquire image information of the coal flow, the memory stores a computer program, and the program can realize the steps of the method of any one of the above embodiments when executed by the processor to identify the coal gangue and sundries in the coal flow.

[0030] The control device controls the grabbing device to grab the coal gangue and sundries based on the identified coal gangue and sundries.

[0031] In some embodiments, the power supply comprises a power monitoring unit to control the automatic sorting system to select a distributed charging potential arranged on a running track for charging when the power of the power supply reaches a preset value.

[0032] In some embodiments, the grabbing device comprises a mechanical arm and a clamping jaw, wherein the mechanical arm is a six-degree-of-freedom mechanical arm or a truss-type mechanical arm.

[0033] A connected coordinate system of the mechanical arm and the clamping jaw is established in the control device to control the movement trajectory of the mechanical arm and the clamping jaw according to the connected coordinate system.

[0034] In some embodiments, the automatic sorting system further comprises a host computer connected to the control module to monitor the operation of the automatic sorting system.

[0035] When there is an unidentified impurity in the image information obtained by the identification module, information of the unidentified impurity is obtained through a network, and the information is stored.

[0036] Compared with the prior art, the coal gangue and impurity identification method and the automatic sorting system have obvious advantages and beneficial effects.

[0037] The feature extraction model is used to extract and identify the features of the segmented coal gangue segmentation graph, which can effectively avoid the influence of environmental factors, effectively improve the identification accuracy of coal gangue and impurities, and further improve the operation efficiency of coal gangue sorting.

[0038] The above description is only a summary of the technical solutions of the present application, in order to more clearly understand the technical means of the present application, the content of the specification can be implemented, and in order to make the above and other purposes, characteristics and advantages of the present application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 A flowchart of a coal gangue and impurity identification method of an embodiment of the present application is shown;

[0040] Figure 2 A schematic block diagram of a feature extraction model of an embodiment of the present application is shown;

[0041] Figure 3 A schematic block diagram of a ResNet34 network structure of an embodiment of the present application is shown;

[0042] Figure 4 A schematic block diagram of a Transformer module of an embodiment of the present application is shown;

[0043] Figure 5 A flowchart of weights of a segmentation feature block of an embodiment of the present application is shown;

[0044] Figure 6 A schematic block diagram of a composite attention multi-scale hollow convolution module of an embodiment of the present application is shown;

[0045] Figure 7 A structural block diagram of an automatic sorting system of coal gangue and impurities of an embodiment of the present application is shown;

[0046] Figure 8 A structural block diagram of an automatic sorting system of coal gangue and impurities of another embodiment of the present application is shown. DETAILED DESCRIPTION

[0047] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined inventive objectives, and the specific implementation and effects of the coal gangue and sundry identification method and automatic sorting system according to the present application, a detailed description is given as follows in combination with the drawings and preferred embodiments.

[0048] The present application provides a coal gangue and sundry identification method, as shown in the figure, the method comprises: Figure 1

[0049] Step S10, image information of coal flow is acquired, a gray threshold value is determined based on appearance characteristics of coal, coal gangue and sundry, image foreground and background of the image information are separated, and a pretreatment image is obtained.

[0050] The coal gangue and sundry identification method of the present application is applied to the automatic sorting system of coal gangue and sundry. During the operation of the automatic sorting system of coal gangue and sundry, the image information of coal flow is acquired in real time by the image acquisition unit (such as a camera) of the automatic sorting system. Then, based on the appearance characteristics such as luster and gray scale of coal, coal gangue and sundry, the gray threshold value of coal, coal gangue and sundry is determined by a gray histogram, the image foreground and background of the image information of coal flow are separated, and then a pretreatment image is obtained.

[0051] Step S20, distance transformation is performed on the pretreatment image, abstracted as a three-dimensional image, coal in contact with coal gangue and sundry is separated, and a coal gangue image is obtained.

[0052] In this step, distance transformation is performed on the pretreatment image, abstracted as a three-dimensional image, coal in contact with coal gangue and sundry is separated by a gradient descent algorithm, and then a coal gangue image is obtained. In this step, the gradient descent algorithm is a conventional technology in the field, and will not be described here.

[0053] Step S30, after pretreatment of the coal gangue image, the coal gangue image is segmented and extracted, and a coal gangue segmentation image is obtained.

[0054] In this step, after pretreatment such as enhancement, smoothing and denoising of the coal gangue image, a watershed algorithm based on distance transformation is used to segment and extract the pretreated coal gangue image, so as to obtain a coal gangue segmentation image. The watershed algorithm is a conventional technology in the field, and will not be described here.

[0055] Step S40, based on a pre-created feature extraction model, the coal gangue and sundry in the coal gangue segmentation image are subjected to feature extraction, so as to realize identification of the coal gangue and sundry.

[0056] ​The present application creates a feature extraction model in advance, which adopts an encoder-decoder structure, as shown in Figure 2 The encoder-decoder structure includes an up-sampling layer, an intermediate layer and a down-sampling layer.

[0057] The down-sampling layer adopts a ResNet34 network structure to perform convolution, normalization and activation function down-sampling features on the coal gangue segmentation graph, to obtain a down-sampling feature segmentation graph. Figure 3 The present application shows a schematic diagram of the ResNet34 network structure, which can connect the original input information directly to the output layer by adding a straight connection channel and a residual structure in the network, prevent network degradation, obtain enough image features even in the case of less input pictures, save network operation cost, and improve network running speed and precision. At the same time, the increased ResNet network structure can quickly make the network converge, and input the context feature information into the intermediate layer.

[0058] The channel numbers of the down-sampling X1 layer, X2 layer, X3 layer and X4 layer are 64, 128, 256 and 512 respectively, and the sizes of the feature maps after sampling are 64x64, 32x32, 16x16 and 8x8 respectively.

[0059] The up-sampling layer uses a residual network and deconvolution, and simultaneously uses Skip-connection to fuse the path features of the encoder and the decoder, uses a RelU activation function and a 3x3 convolution to perform feature fusion, and finally outputs a prediction picture through a 3x3 convolution of an EndConv layer.

[0060] The intermediate layer includes a Transformer module and a composite attention multi-scale dilated convolution module.

[0061] In an embodiment, there are many uncontrollable factors in the feature extraction process, such as the difficulty in identifying the overlapping and adhering parts of the picture, which brings great challenges to the segmentation of the overlapping part. By dividing the input image into multiple image blocks and processing them as sequences, adding a class token (Class Token) and adding a position encoding to each sequence, finally linearly mapping the added encoding as the input of the Transformer module to train the model in a supervised manner. Token will be a unit of sequence information together for feature extraction. The calculation formula is:

[0062] T=[W1X1+W2X2+L+W i X i ]+W pos

[0063] In the formula, T is the segmentation feature block with position encoding, Wi is the weight of the i-th split feature block (i = 1, 2, …, I ∈ N+), X i is the i-th down-sampled split feature map (i = 1, 2, …, I ∈ N+), W pos is the spatial position information code, where D is the linear spatial dimension.

[0064] In an embodiment, as shown in FIG. 2, the Transformer module is composed of a plurality of Transformer layers in cascade, each of which includes a multi-head self-attention layer and a multi-layer perception layer (BN). Figure 4

[0065] The multi-head self-attention layer includes:

[0066] In the formula, Q is a query matrix, K is a key matrix, and V is a value matrix.

[0067] The weight of the split feature block is obtained by the following steps:

[0068] As shown in FIG. 3, the query limit of input 1 is multiplied by the key vectors of input 1, input 2 and input 3 respectively to obtain three vector values. Figure 5 After softmax, the vector values are converted into respective importance degrees distributed between [0, 1].

[0069] Each importance degree is multiplied by the corresponding value vector and summed to obtain the weight of the split feature block.

[0070] In an embodiment, to strengthen the edge feature extraction of the network, a multi-scale dilated convolution module with attention mechanism is proposed. After the Transformer module extracts high-level semantic features, a parallel dilated convolution and pooling with attention mechanism are used to further collect features. The attention mechanism increases the weight of the edge and filters out some unnecessary features. The dilated convolution module expands the receptive field on the basis of obtaining finer features to obtain more comprehensive information.

[0071] As shown in FIG. 4, the composite attention multi-scale dilated convolution module includes a composite channel spatial attention module and a multi-scale dilated convolution module.

[0072] Figure 6

[0073] ​​​The composite channel and space attention module realizes feature extraction through one-dimensional convolution and performs normalization, and the features obtained based on the global max pooling and the global average pooling of the channel are concatenated, combined into one channel, and finally subjected to Sigmoid and dot product with the channel output feature to obtain the weight distribution of the spatial feature.

[0074] The core of the attention mechanism of computer vision segmentation is to hope that the model can "focus on global features and focus on key features" in the input picture, and the Transformer can well save the features focused by the self-attention. However, since the Transformer layer splits the input layer into multiple sequences for direct processing, long-distance structures may cause information to be missed, so a composite channel and space attention mechanism is added after the Transformer to capture context information.

[0075] In the traditional attention model, the amount of information used by the input picture is limited, resulting in an output that is still not comprehensive enough. To solve this problem, in order to make full use of effective information, a composite channel and space attention module is introduced. This module is a lightweight module that can achieve good results with very few parameters. The channel attention module of this module uses channel attention to distribute weights, so as to enhance useful information and suppress useless information. One-dimensional convolution is used to realize feature extraction without dimension reduction, which can avoid the influence of dimension reduction on learning channels. Then, Sigmoid is used to normalize the information. The spatial attention first performs a global max pooling and a global average pooling based on the channel, concatenates the feature maps obtained by the two, combines them into one channel, and finally performs Sigmoid and dot product with the channel output feature to obtain the weight distribution of the spatial feature.

[0076] As shown in Figure 6 , the global average pooling is an average pooling layer, the global max pooling is a max pooling layer, InChnnel, OutChannel and Conv are 1x1 convolution, and the last Sigmoid is an activation function that plays a normalizing role.

[0077] As shown in Figure 6 , the multi-scale dilated convolution module includes three dilated convolution layers, a 1x1 convolution layer, and a pooling layer for parallel sampling, and the dilated rates of the three dilated convolution layers are 2, 4 and 8 respectively.

[0078] The empty hole convolution can expand the receptive field of the feature map without losing the image resolution, the high resolution can accurately locate the target, and the different receptive fields can perceive different scale information, and through parallel combination of different sampling rates, multiple different size receptive fields can be obtained, which to some extent reduces the problem of too much detail loss caused by too much resolution reduction due to ordinary convolution. The present application adopts three layers of empty hole convolution, one 1*1 convolution layer and one pooling layer for parallel sampling, the empty hole rates are 2, 4 and 8 respectively, and the channel number is reduced by four times, which is to reduce the amount of calculation. Then the pooling layer is connected to the up-sampling, and finally the four different scale features are fused to make the output image contain different sizes of receptive fields.

[0079] According to another embodiment of the present application, an automatic sorting system for coal gangue and sundries is provided, as shown in the figure, the automatic sorting system comprises a power supply 10, an identification device 20, a control device 30 and a grabbing device 40. Figure 7

[0080] The power supply 10 supplies power to the automatic sorting system; the identification device 20 comprises an image acquisition unit, a memory and a processor, the image acquisition unit is used to acquire image information of the coal flow, the memory stores a computer program, and the program can realize the identification method of the coal gangue and sundries in any of the above embodiments when executed by the processor to identify the coal gangue and sundries in the coal flow; the control device 30 controls the grabbing device 40 to grab the coal gangue and sundries based on the identified coal gangue and sundries.

[0081] Specifically, the power supply 10 is an explosion-proof and intrinsic safety power supply box, which provides a 12V intrinsic safety circuit and ensures the safety of the site through a direct current contact. In addition, the power supply also comprises a power monitoring unit, which controls the automatic sorting system to select a distributed charging potential arranged on the running track for charging when the power of the power supply reaches a preset value. A distributed charging point structure is arranged on the track of the automatic sorting system, and a charging device is scientifically configured on the running track to realize autonomous charging through the power supply self-contained power monitoring system.

[0082] The image acquisition unit of the identification device 20 is a camera, which is used to acquire image information of the coal flow in real time, and the memory stores a computer program, which can realize the identification method of the coal gangue and sundries in the above embodiments when executed, and further identify the coal gangue and sundries in the coal flow.

[0083] The control device 30 controls the grabbing device 40 to grab the identified coal gangue and sundries according to the coal gangue and sundries identified by the identification device 20. The control device 30 can be a PLC controller.

[0084] ​The grabbing device 40 comprises a mechanical arm and a gripper, wherein the mechanical arm is a six-degree-of-freedom mechanical arm or a truss-type mechanical arm; a connected coordinate system of the mechanical arm and the gripper is established in the control device, so as to control the moving track of the mechanical arm and the gripper according to the connected coordinate system.

[0085] Specifically, in an embodiment, the mechanical arm is a six-degree-of-freedom mechanical arm, and the mechanical structure and the control system matched with the gripper are designed to have the grabbing capability for the coal gangue and the sundries in shape, size and weight.

[0086] Taking the six-degree-of-freedom mechanical arm as the research object, the basic structure of the mechanical arm is simplified, the connected coordinate system of the mechanical arm is established based on the D-H method, the mathematical model of the mechanical arm is established based on the Robotics Toolbox toolbox of MATLAB, the joint space of the mechanical arm is planned by using the quintic polynomial method, and simulation is performed, the moving track and the gripper design suitable for the accurate and rapid grabbing and sorting of the coal gangue are researched, and the grabbing of the coal gangue and the sundries in shape change, different size and weight is realized.

[0087] In another embodiment, the mechanical arm is a truss-type mechanical arm, and the mechanical structure and the control system matched with the gripper are designed to have the grabbing capability for the coal gangue and the sundries in shape, size and weight.

[0088] Taking the non-standard truss 3-degree-of-freedom and 4-degree-of-freedom mechanical arm as the research object, the basic structure of the mechanical arm is simplified, the connected coordinate system of the mechanical arm is established based on the D-H method, the mathematical model of the mechanical arm is established based on the Robotics Toolbox toolbox of MATLAB, the joint space of the mechanical arm is planned by using the quintic polynomial method, and simulation is performed, the moving track and the gripper design suitable for the accurate and rapid grabbing and sorting of the coal gangue are researched, and the grabbing of the coal gangue and the sundries in shape change, different size and weight is realized. In order to ensure that the truss-type mechanical arm can reliably move along the truss following the sundries.

[0089] In an embodiment, the automatic sorting system further comprises a host computer connected to the control module to monitor the operation of the automatic sorting system; and

[0090] When there is sundry that cannot be identified in the image information obtained by the identification module, information of the sundry that cannot be identified is obtained through a network, and the information is stored.

[0091] In a specific embodiment, as shown in Figure 8 The automatic sorting system of the coal gangue and the sundries comprises a power supply, a host computer, a coal gangue automatic sorting workbench, a rectifier, a PLC controller, an electromagnetic valve, a magnetic switch, an air compressor and a sensor.

[0092] The power supply is an explosion-proof and intrinsic safety power supply box, which provides a 12V intrinsic safety circuit and a direct current contact to ensure safety in the field. The multi-parameter mobile inspection device is powered by a 5Ah lithium battery to ensure sufficient endurance. A distributed charging point structure is used to scientifically configure charging devices along the running track, and the device automatically charges through the built-in power monitoring system. The system provides the required voltage. The actuator is an electromagnetic valve, which is driven by the output signal of the PLC to drive the cylinder to act. The driving mechanism is an air compressor or a robot motor. The control mechanism is a PLC, which is the core of the control system and performs all operations according to the PLC program. The detection mechanism is a magnetic switch and a sensor, which provides an input signal to the PLC. The upper computer is a PC (front-end server, which uses a combination of C / S, B / S, and mobile client APP, has a clear modular architecture, complete external interfaces, and can share data externally through the WEB server. The dedicated user terminal software can remotely control the mobile inspection device, configure tasks, access videos, and access data, etc.). The communication transmission uses a WiFi wireless signal network communication system that meets TCP / IP standards, which can be connected to the PLC through RS232 cable (RS485 interface or USB 4.0 optical fiber interface), and carries a monitoring system to work in the process to achieve remote mobile monitoring and real-time data transmission to the centralized management control center (the cloud server of the management control center uses a combination of C / S, B / S, and mobile client APP, has a clear modular architecture, complete external interfaces, and can comprehensively manage, coordinate, and coordinate various auxiliary equipment in the machine room. The device collects data through the WEB server and shares it externally through the external interface. The dedicated user terminal software can remotely control the mobile inspection device, configure tasks, access videos, and access data, etc.). The queued coal and gangue mixture is transmitted to the computer system for physical property judgment through image recognition and sensor data collection during transmission, and the coal and gangue are distinguished and corresponding instructions are issued. The front camera of the robot and the front-end server (the front-end server uses a combination of C / S, B / S, and mobile client APP, has a clear modular architecture, complete external interfaces, and can share data externally through the WEB server. The dedicated user terminal software can remotely control the mobile inspection device, configure tasks, access videos, and access data, etc.) are used to track the line body (according to the situation of the roadway, the robot track design and construction scheme is determined, and 10# I-beam is used as the inspection device track, which is treated with special anti-rust treatment, is solid, durable, and not easy to erode. The track is 6 meters long, covering the entire inspection area, and is the main carrier of the inspection device) to achieve the grabbing of coal and gangue.

[0093] The above merely describes preferred embodiments of the present application, and is not intended to limit the present application in any form. Although the present application has been disclosed with the preferred embodiments as above, it is not intended to limit the present application. Any person skilled in the art, without departing from the technical solution of the present application, can make some changes or modifications to the above disclosed technical content to obtain equivalent embodiments with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application, as long as it does not deviate from the technical solution of the present application, shall still fall within the scope of the technical solution of the present application.

Claims

1. A method of identifying coal gangue and trash, characterized by, The method comprises the following steps: Obtain image information of the coal flow, determine a gray threshold value based on the appearance characteristics of coal, coal gangue and sundries, separate the image foreground and background of the image information, and obtain a pretreated image; Perform distance transformation on the pretreated image, abstract it into a three-dimensional image, separate the coal in contact with the coal gangue and sundries, and obtain a coal gangue image; After pretreating the coal gangue image, perform segmentation and extraction on the coal gangue image to obtain a coal gangue segmentation image; Perform feature extraction on the coal gangue and sundries in the coal gangue segmentation image based on a pre-created feature extraction model to realize recognition of the coal gangue and sundries; The feature extraction model adopts an encoder-decoder architecture, which comprises an up-sampling layer, an intermediate layer and a down-sampling layer; The down-sampling layer adopts a ResNet34 network structure to perform convolution, normalization and activation function down-sampling features on the coal gangue segmentation image to obtain a down-sampling feature segmentation image; The up-sampling layer adopts a residual network and deconvolution, fuses the path features of the encoder and the decoder, uses an activation function and convolution to perform feature fusion, and outputs a prediction picture through convolution; The intermediate layer comprises a Transformer module and a composite attention multi-scale dilated convolution module; The composite attention multi-scale dilated convolution module comprises a composite channel spatial attention module and a multi-scale dilated convolution module; The composite channel spatial attention module realizes feature extraction through one-dimensional convolution, performs normalization, obtains feature maps based on channel global max pooling and global average pooling, performs concat operation, combines into one channel, and finally performs point multiplication on the spatial feature weight distribution through Sigmoid and channel output features; The multi-scale dilated convolution module comprises three dilated convolution layers, a 1x1 convolution layer and a pooling layer for parallel sampling, and the hole rates of the three dilated convolution layers are 2, 4 and 8 respectively.

2. The coal waste and foreign matter identification method according to claim 1, characterized by, Before inputting the down-sampling feature segmentation image into the Transformer module, the down-sampling segmentation feature image is mapped into a linear space and labeled and position encoded, and the calculation formula is: ； wherein: T is a segmented feature block with position coding, W i is the weight of the i-th segmented feature block (i = 1, 2, …, I ∈ N+), X i is the i-th down-sampled segmented feature map (i = 1, 2, …, I ∈ N+), W pos ∈ RD is the spatial position information coding, wherein D is the linear spatial dimension.

3. The coal waste and foreign matter identification method according to claim 1, characterized by, When the coal gangue image is segmented and extracted, the coal gangue image is segmented into multiple image blocks, and the multiple image blocks are processed into sequences, and position encoding is added to each sequence.

4. The coal waste and foreign matter identification method according to claim 3, characterized by, The Transformer module is formed by cascading multiple Transformer layers, and each Transformer layer comprises a multi-head self-attention layer and a multi-layer perception layer; wherein the multi-head self-attention layer ; In the formula, Q is a query matrix, K is a key matrix, and V is a value matrix; The weight of the segmentation feature block is obtained through the following steps: Input 1 query limit, multiply by the key vectors of input 1, input 2 and input 3 respectively to obtain three vector values; After softmax, the vector values are converted into respective importance degrees distributed between 0 and 1; Each importance degree is multiplied by the corresponding value vector and summed to obtain the weight of the segmentation feature block.

5. An automatic sorting system of coal gangue and sundries, characterized in that, The method comprises the following steps: The power supply, the identification device, the control device and the grabbing device; The power supply supplies power for the automatic sorting system; The identification device comprises an image acquisition unit, a memory and a processor, the image acquisition unit is used for acquiring image information of the coal flow, the memory stores a computer program, and the program can realize the steps of the method in any one of claims 1 to 4 when executed by the processor to identify coal gangue and sundries in the coal flow; The control device controls the grabbing device to grab the coal gangue and sundries based on the identified coal gangue and sundries.

6. The automatic sorting system of coal refuse and trash according to claim 5, wherein, The power supply comprises an electric quantity monitoring unit to control the automatic sorting system to select a distributed charging potential arranged on a running track to charge when the electric quantity of the power supply reaches a preset value.

7. The automatic sorting system of coal refuse and trash according to claim 5, wherein, The grabbing device comprises a mechanical arm and a clamping jaw, wherein the mechanical arm is a six-degree-of-freedom mechanical arm or a truss type mechanical arm; A connected coordinate system of the mechanical arm and the clamping jaw is established in the control device to control the moving track of the mechanical arm and the clamping jaw according to the connected coordinate system.

8. The automatic sorting system of coal refuse and trash according to claim 5, wherein, The automatic sorting system further comprises a host computer connected to the control module to monitor the operation of the automatic sorting system, and when there is sundries that cannot be identified in the image information acquired by the identification module, the information of the sundries that cannot be identified is acquired through a network and stored.

Images