A portable flake cake pressing sampling device and sampling method

Abstract

The application discloses a portable shaving cake pressing and sampling device and a sampling method, and belongs to the technical field of scrap steel recycling equipment.The portable shaving cake pressing and sampling device comprises a vehicle frame base, a telescopic box body is arranged on one side of the top of the vehicle frame base, a vehicle frame main body is arranged above the telescopic box body, an electric lifting mechanism is arranged on the vehicle frame main body, a telescopic mechanical arm is installed on the electric lifting mechanism, an electromagnetic chuck is arranged at the end of the telescopic mechanical arm, a transverse driving device is arranged above the vehicle frame main body, a multi-stage telescopic arm is arranged at the bottom of the transverse driving device, a hydraulic crushing pick head is installed at the bottom end of the multi-stage telescopic arm, and an image acquisition device is further arranged at the bottom of the transverse driving device.The application solves the problem that the existing scrap steel cake pressing and block moving, crushing, sampling and detection work is not convenient and fast enough.Through three-level technologies of visual analysis, defect positioning and intelligent crushing, the application realizes automatic sampling.

Description

Technical Field

[0001] This invention relates to the field of scrap steel recycling and processing equipment technology, specifically to a convenient wood shavings cake sampling device and sampling method. Background Technology

[0002] A portable shavings briquette sampler is a crusher used to process scrap steel briquettes and granulated steel blocks. Scrap steel recycling is a crucial step in the steel industry. Scrap steel is typically stored and transported in briquettes or blocks for subsequent reuse. Common forms of scrap steel include shavings briquettes and granulated steel blocks, which are made through the following methods:

[0003] Wood shavings patties: Wood shavings are thin, flaky waste produced during steel processing. They are typically pressed into patties using a hydraulic press. During pressing, the shavings are placed in a mold and high pressure is applied to bind them tightly together, forming a flat, patty structure. Wood shaving patties are typically 90-300mm in diameter and 30-80mm thick (cube side length ≤ 300mm), weighing approximately 0.5-100kg, depending on the patty's density and material composition.

[0004] Particle steel briquettes: Particle steel is made from fine scrap steel particles produced during steel production. It is typically pressed into briquettes using a hydraulic press or mechanical press. During pressing, the particle steel is placed in a mold and subjected to high pressure to bind it tightly, forming a regular block structure. Common dimensions for particle steel briquettes are 300-600mm in side length (cold pressing) or 260mm in diameter × 240mm in height (hot pressing), with a weight between 25-80kg, depending on the briquettes' density and material composition.

[0005] In the manufacturing, distribution, and application of shavings briquettes and granulated steel briquettes, multiple sampling, crushing, and quality inspections are required. Currently, the sampling, crushing, and sampling processes are all done manually by operators who handle the samples, crush them with small hammers, and take samples by hand. However, due to the large weight and tight bonding of shavings briquettes, manual operation is very inconvenient, laborious, inefficient, and poses safety hazards, making it difficult to meet the needs of the modern scrap steel industry. Summary of the Invention

[0006] The automated sampling achieved through a three-stage technology of visual analysis, defect location, and intelligent crushing solves the problems mentioned in the background.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A convenient shavings cake sampling device includes a frame base. A retractable box is provided on one side of the top of the frame base. A frame body is provided above the retractable box. A linear guide rail is provided on the side wall of the frame body facing the frame base. An electric lifting mechanism is provided on the linear guide rail. A retractable robotic arm is installed on the electric lifting mechanism. An electromagnetic chuck is provided at the end of the retractable robotic arm. A transverse slide rail fixing frame is provided above the frame body. A transverse drive device is installed on the bottom surface of the transverse slide rail fixing frame via a slide rail. A multi-stage retractable arm is provided at the bottom of the transverse drive device. A hydraulic crushing pick is installed at the bottom end of the multi-stage retractable arm. An image acquisition device for real-time image acquisition of the scrap steel cake block on the electromagnetic chuck is also provided at the bottom of the transverse drive device. The image acquisition device is located on one side of the multi-stage retractable arm.

[0009] The image acquisition device is used to acquire the surface cracks and iron oxide distribution characteristics of scrap steel in real time, and transmit the images to the edge computing terminal.

[0010] The edge computing terminal is equipped with a deep learning model based on a convolutional neural network for target detection and has deployed an attention-based metal defect detection algorithm to analyze the morphological characteristics of scrap steel transmitted by the image acquisition device in real time and generate three-dimensional crushing path planning instructions.

[0011] Preferably, the telescopic housing includes a fixed housing, a movable housing, a folding linkage assembly, and a hydraulic cylinder. The fixed housing is fixed to the upper surface of the vehicle frame base by bolts. The folding linkage assembly consists of two sets of V-shaped hydraulic rods, one end of which is hinged to the inner wall of the fixed housing, and the other end of which is hinged to the top of the movable housing. The hydraulic cylinder is horizontally arranged between the fixed housing and the vehicle frame body, and the end of the piston rod is hinged to the intersection point of the folding linkage assembly.

[0012] Preferably, a battery is connected to the side of the frame body away from the linear guide rail, and a handrail assembly is installed on the side of the battery away from the frame body. A control handle and an edge computing terminal are provided on the top of the handrail assembly, and braked swivel wheels are installed at the bottom of the handrail assembly and the frame base.

[0013] Preferably, the edge computing terminal is signal-connected to the image acquisition device, the hydraulic breaker, the lateral drive device, and the multi-stage telescopic arm.

[0014] A convenient method for sampling wood shavings cakes, based on a convenient wood shavings cake sampling device, includes the following steps:

[0015] Step 1: Use an electromagnetic chuck to lift the scrap steel briquettes to the top of the retractable box;

[0016] Step 2: The image acquisition device acquires images of the scrap steel surface and transmits them to the edge computing terminal;

[0017] Step 3: The edge computing terminal identifies crack and iron oxide defect areas through a composite attention mechanism and generates three-dimensional fracture path instructions;

[0018] Step 4: The edge computing terminal outputs instructions and controls the lateral drive device to move along the lateral slide rail, and impacts the defective area of ​​the scrap steel according to the path instructions through the cooperation of the multi-stage telescopic arm and the crushing pick.

[0019] Step 5: The crushed scrap steel falls into the retractable container to complete the sampling.

[0020] Preferably, regarding step two, the surface image of the scrap steel briquettes is acquired in real time using an RGB-D depth camera, input to the edge computing terminal, and the image is normalized using the following formula:

[0021]

[0022] Where μ k σ k For training set statistics, i: index in the height direction (1≤i≤H), j: index in the width direction (1≤j≤W), k: index in the channel direction (1≤k≤C).

[0023] Preferably, for step three, the defect features are extracted using a composite attention mechanism, using the following formula:

[0024] Spatial attention module calculation formula:

[0025] W s =Softmax(Γ(μ(F)||ε(F)))

[0026] Channel attention module calculation formula:

[0027] W c =σ(MLP(P_a(F))+MLP′(P_a(F)))

[0028] Where F is the input feature map, Γ(·) is a 3*3 depthwise separable convolution, || represents channel-dimensional concatenation, μ(·) and ξ(·) represent average pooling and max pooling along the channel dimensions, respectively, and σ() is the sigmoid function;

[0029] P_a() represents global hybrid pooling: α·AvgPool(·)+(1-α)·MaxPool(·), α∈[0.4,0.6]; AvgPool calculates the average color of each region of the photo, MaxPool is the brightest point in each region of the photo, and MLP and MLP' are dual fully connected layers with non-shared parameters;

[0030] When max(Fout⊙Ws)>θdefect, the maximum value after multiplying the feature map and the spatial weights exceeds the threshold, it is marked as a structurally weak area, where θdefect=0.7 is determined by the ROC curve;

[0031] Preferably, the process of generating the three-dimensional fracture path described in step three is as follows:

[0032] The scrap steel briquettes were divided into 10mm×10mm×10mm voxel units;

[0033] Calculate impact priority:

[0034] P(v)=α·D(v)+β·S(v)

[0035] Where: v is a voxel unit, generated based on depth camera point cloud data; D(v) is the distance from voxel v to the surface, obtained by the following formula:

[0036]

[0037] d max d(v) represents the maximum thickness of the compaction block, and d(v) represents the actual measured distance.

[0038] S(v) is the defect confidence score (0-1) from the output value of the attention module. A larger value indicates a higher defect probability. The spatial attention weight W s The output values, α = 0.6 and β = 0.4, are weighting coefficients, indicating that internal defects break down before surface defects, and are determined through experimental optimization.

[0039] Generate B-spline curve path:

[0040]

[0041] Where: B(t) is the parameterized curve coordinate control point, representing the motion trajectory of the crusher, P i Let N be the coordinates of the i-th control point, corresponding to the center of the highest priority voxel. i,p (t) is a p-th order B-spline basis function that determines the smoothness of the curve. The above function controls the impact of the crushing pick on the briquette along the planned path B(t).

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] 1. This invention uses an electromagnetic chuck to adsorb scrap steel cakes and automatically performs image acquisition, defect identification, and intelligent crushing. This process is fully automated, greatly reducing the need for manpower. Compared to manual operation, this invention can significantly improve the speed of scrap steel processing. From real-time acquisition of surface cracks and iron oxide distribution characteristics of scrap steel, to analysis based on a deep learning model and generation of three-dimensional crushing path planning instructions, and finally to automatic crushing and sampling, the entire process is efficient and fast. The inclusion of casters makes the device easier to push and move, making it more convenient.

[0044] 2. This invention, by combining a depth camera with a compound attention mechanism, can accurately identify cracks and iron oxide distribution on the surface of scrap steel, ensuring the accuracy of the crushing location. At the same time, through B-spline curve path planning, it achieves effective impact on the briquette, further improving the accuracy and reliability of sampling. Whether it is shavings briquette or particle steel briquette, this device can effectively complete the sample extraction and crushing work, and is applicable to various types of scrap steel materials, with strong versatility and practicality. Attached Figure Description

[0045] Figure 1 This is an isometric view of the overall structure of the present invention;

[0046] Figure 2 This is a schematic diagram of the retractable box of the present invention;

[0047] Figure 3 This is a flowchart of the process of the present invention.

[0048] In the diagram: 1. Chassis base; 2. Braked swivel casters; 3. Telescopic housing; 3a. Fixed housing; 3b. Movable housing; 3c. Folding linkage assembly; 3d. Hydraulic cylinder; 4. Chassis body; 5. Edge computing terminal; 6. Battery; 7. Handrail assembly; 8. Control handle; 9. Electric lifting mechanism; 10. Telescopic robotic arm; 11. Electromagnetic chuck; 12. Linear guide rail; 13. Hydraulic breaker pick; 14. Multi-stage telescopic arm; 15. Lateral drive device; 16. Image acquisition device; 17. Lateral slide rail fixing frame. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] To address the inconvenience and slowness of moving, crushing, sampling, and testing existing scrap steel briquettes and blocks, please refer to [link / reference needed]. Figure 1-3This embodiment provides the following technical solution:

[0051] A convenient wood shavings cake sampling device includes a frame base 1, a retractable box 3 on one side of the top of the frame base 1, a frame body 4 above the retractable box 3, a linear guide rail 12 on the side wall of the frame body 4 facing the frame base 1, the surface of the linear guide rail 12 being hardened and having a friction coefficient ≤0.15, an electric lifting mechanism 9 being installed on the linear guide rail 12, the electric lifting mechanism 9 being a scissor lift, a wall-mounted lift, or a boom lift, a retractable robotic arm 10 being installed on the electric lifting mechanism 9, the retractable robotic arm 10 having three degrees of freedom, namely telescopic motion, pitch motion, and rotational motion, and an electromagnetic chuck 11 being installed at the end of the retractable robotic arm 10.

[0052] The electromagnetic chuck 11 uses rare-earth permanent magnet material and is equipped with a power-off protection device. The power-off protection device includes a supercapacitor energy storage module, which can maintain the magnetic force for ≥5 seconds after a power outage. The electromagnetic chuck 11 is connected to the telescopic robotic arm 10 through a universal joint mechanism and has a multi-directional rotational freedom of ±30°. When the scrap steel is at a low position on the ground or on a non-horizontal surface, the electromagnetic chuck 11 can adaptively adjust the adsorption angle to ensure full contact with the scrap steel surface. During adsorption operations, the telescopic robotic arm 10 can drive the electromagnetic chuck 11 to a minimum height of 1.2m with a hook-like motion trajectory. After adsorption is completed, it is automatically reset to the appropriate working position by an electric push rod set at the connection between the electromagnetic chuck 11 and the telescopic robotic arm 10, with a reset accuracy of ±2°.

[0053] A transverse slide rail fixing frame 17 is provided on the upper part of the frame body 4. A transverse drive device 15 is installed on the bottom surface of the transverse slide rail fixing frame 17 via a slide rail. The transverse drive device 15, in conjunction with the slide rail, can achieve displacement at any point on the x-axis and y-axis. The XY worktable system used in laser cutting machines and 3D printers can be adopted.

[0054] The bottom of the lateral drive device 15 is equipped with a multi-stage telescopic arm 14, and a hydraulic crushing pick 13 is installed at the bottom end of the multi-stage telescopic arm 14. The hydraulic crushing pick 13 is made of high-carbon alloy steel to adapt to high-frequency impact. The bottom of the lateral drive device 15 is also equipped with an image acquisition device 16, which is located on one side of the multi-stage telescopic arm 14. It is used to collect the surface cracks and iron oxide distribution characteristics of scrap steel in real time and transmit the images to the edge computing terminal 5.

[0055] The edge computing terminal 5 is equipped with a deep learning model based on a convolutional neural network for target detection and has deployed a metal defect detection algorithm based on an attention mechanism. It is used to analyze the morphological features of scrap steel transmitted by the image acquisition device 16 in real time and generate three-dimensional crushing path planning instructions.

[0056] The telescopic housing 3 includes a fixed housing 3a, a movable housing 3b, a folding linkage group 3c, and a hydraulic cylinder 3d. The fixed housing 3a is fixed to the upper surface of the frame base 1 by bolts. The folding linkage group 3c consists of two sets of V-shaped hydraulic rods, one end of which is hinged to the inner side wall of the fixed housing 3a, and the other end is hinged to the top of the movable housing 3b. The hydraulic cylinder 3d is horizontally arranged between the fixed housing 3a and the frame body 4, and the end of the piston rod is hinged to the intersection point of the folding linkage group 3c.

[0057] The frame body 4 and the frame base 1 are integrated. The bottom of the frame body 4 and the connection between the frame base are hollowed out. The hollowed-out part is equipped with a fixed box 3a, a folding linkage group 3c, a hydraulic cylinder 3d, etc. The movable box 3b is detachable from the fixed box 3a.

[0058] A battery 6 is connected to the side of the frame body 4 away from the linear guide rail 12. The battery module 6 can work continuously for 4-6 hours when fully charged. A handrail assembly 7 is installed on the side of the battery 6 away from the frame body 4. The top of the handrail assembly 7 is equipped with a control handle 8 for operating the equipment and an edge computing terminal 5. The control panel 8 is equipped with an emergency stop button, a mode selection switch and a parameter adjustment knob. Braked casters 2 are added to the bottom of the handrail assembly 7 and the frame base 1.

[0059] The edge computing terminal 5 is connected to the image acquisition device 16, the hydraulic breaker 13, the lateral drive device 15, and the multi-stage telescopic arm 14 via signal connection.

[0060] A convenient method for sampling wood shavings cakes, based on a convenient wood shavings cake sampling device, includes the following steps:

[0061] Step 1: Use the electromagnetic chuck 11 to pick up the scrap steel patties and place them above the retractable box 3;

[0062] Step 2: Real-time images of the scrap steel briquettes are acquired using an RGB-D depth camera and input into the edge computing terminal 5. The images are then normalized using the following formula:

[0063]

[0064] Where μ k σ k For training set statistics, i: index in the height direction (1≤i≤H), j: index in the width direction (1≤j≤W), k: index in the channel direction (1≤k≤C);

[0065] Step 3: Edge computing terminal 5 identifies crack and iron oxide defect areas through a composite attention mechanism and generates three-dimensional fracture path instructions, specifically:

[0066] Defect features are extracted using a composite attention mechanism, employing the following formula:

[0067] Spatial attention module calculation formula:

[0068] W s =Softmax(Γ(μ(F)||ε(F)))

[0069] Channel attention module calculation formula:

[0070] W c =σ(MLP(P_a(F))+MLP′(P_a(F)))

[0071] Where F is the input feature map, Γ(·) is a 3*3 depthwise separable convolution, || represents channel-dimensional concatenation, μ(·) and ξ(·) represent average pooling and max pooling along the channel dimensions, respectively, and σ() is the sigmoid function;

[0072] P_a() represents global hybrid pooling: α·AvgPool(·)+(1-α)·MaxPool(·), α∈[0.4,0.6]; AvgPool calculates the average color of each region of the photo, MaxPool is the brightest point in each region of the photo, and MLP and MLP' are dual fully connected layers with non-shared parameters;

[0073] When max(Fout⊙Ws)>θdefect, the maximum value after multiplying the feature map and the spatial weights exceeds the threshold, it is marked as a structurally weak area. Here, θdefect = 0.7 is determined by the ROC curve.

[0074] The specific process of generating a 3D broken path is as follows:

[0075] The scrap steel briquettes were divided into 10mm×10mm×10mm voxel units;

[0076] Calculate impact priority:

[0077] P(v)=α·D(v)+β·S(v)

[0078] Where: v is a voxel unit, generated based on depth camera point cloud data; D(v) is the distance from voxel v to the surface, obtained by the following formula:

[0079]

[0080] d max d(v) represents the maximum thickness of the compaction block, and d(v) represents the actual measured distance.

[0081] S(v) is the defect confidence score (0-1) from the output value of the attention module. A larger value indicates a higher defect probability. The spatial attention weight W sThe output values, α = 0.6 and β = 0.4, are weighting coefficients, indicating that internal defects break down before surface defects, and are determined through experimental optimization.

[0082] Generate B-spline curve path:

[0083]

[0084] Where: B(t) is the parameterized curve coordinate control point, representing the motion trajectory of the crusher 13, P i Let N be the coordinates of the i-th control point, corresponding to the center of the highest priority voxel. i,p (t) is a p-th order B-spline basis function, which determines the smoothness of the curve. The above function controls the impact of the crushing pick 13 on the briquette along the planned path B(t).

[0085] Step 4: The edge computing terminal 5 outputs instructions and controls the lateral drive device 15 to move along the lateral slide rail, and impacts the defective area of ​​the scrap steel according to the path instructions through the cooperation of the multi-stage telescopic arm 14 and the crushing pick 13.

[0086] Step 5: The crushed scrap steel falls into the retractable box 3 to complete the sampling.

[0087] In steps two and three, a crushing model needs to be established to improve the crushing generalization ability. The specific methods are as follows:

[0088] S1. By collecting image data of different scrap steel briquettes and blocks on-site and image data of scrap steel briquettes and blocks piled up on-site, a scrap steel briquettes and blocks dataset is established; by collecting images of different scrap steel briquettes and blocks after they are crushed on-site, a scrap steel crushing dataset is established.

[0089] S2. Use labelimg to identify the scrap steel patties in each image of the scrap steel patty and block dataset from step S1.

[0090] S3. Use labelimg to identify the scrap steel crushing dataset in step S1, and label the information as the scrap steel crushing status in the image.

[0091] S4. Use Jetson-OPTIMIZED CutMix technology to perform image enhancement on the scrap steel briquette and block dataset identified in step S2 and the scrap steel crushed dataset identified in step S3 to improve the diversity of samples and enhance the generalization ability of the recognition model.

[0092] S5. Build a deep learning framework for object detection based on convolutional neural networks on a PC. Train the scrap steel briquette and block dataset and the scrap steel crushed dataset after image enhancement in step S4 in multiple rounds to obtain the scrap steel briquette and block recognition model and the scrap steel crushed recognition model, respectively.

[0093] S6. Save the scrap steel briquette and briquette recognition model and the scrap steel crushing recognition model from step S5 in the SD card storage module of the edge computing terminal (17).

[0094] In step S4, the CutMix image enhancement generation formula is as follows:

[0095]

[0096] In the formula: For the input vector, Let (xi, yi) and (xj, yj) be the data labels, where (xi, yi) and (xj, yj) are two randomly selected samples and their corresponding labels from the same batch. M is a binary mask matrix (a 0 / 1 matrix) generated using a CUDA kernel function, where ° represents pixel-wise multiplication accelerated by Tensor Cores, and N(0, 0.1) is the set of values ​​for each sample. 2 ) represents Gaussian noise with a standard deviation σ = 0.1, and λ is a number randomly sampled from the Beta distribution, λ ∈ Beta [0, 1].

[0097] Step S5 uses the Focal Loss loss function during training, and its expression is:

[0098]

[0099] In the formula: P t The model's predicted probability for the true class:

[0100] When the sample is a positive class (region requiring fragmentation): P t =P

[0101] When the sample is a negative class (background / non-target): P t =1-P

[0102] α t is the category weight coefficient, γ is the hard sample focusing factor, ∈=0.1κ=1.5 is the higher-order balance term.

[0103] The methods for identifying scrap steel briquettes and blocks in step S2 and scrap steel crushing in step S3 are as follows: A target detection model based on a convolutional neural network is used as the backbone network, and BCELoss is used as the classification loss function. For each category, the predicted probabilities are p and 1-p. The BCELoss function calculates both the probability of belonging to this category and the probability of not belonging to this category, ensuring that the loss calculated in both cases is not zero.

[0104]

[0105] N is the number of samples, yi is the true label of the i-th sample, and pi is the probability that the model predicts the i-th sample to be a positive sample (the base of the log is e).

[0106] Furthermore, DFLLoss and EIoULoss were combined as the bounding box regression loss function. DFLLoss dynamically adjusts the weights of targets of different sizes, making the model more accurate in detecting targets of different sizes; while EIoULoss decouples width and height errors based on CIoULoss, further improving the localization accuracy of broken areas and improving the accuracy of bounding box regression by independently optimizing the regression terms in the width and height directions.

[0107]

[0108] yi and yi+1 are the left and right integer values ​​of the floating-point value y, S is the output distribution, DFL enables the network to focus more quickly on values ​​near the target y and increase their probability.

[0109]

[0110] ρ 2 (b, bgt) represents the squared Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box, ρ 2 (w,w gt ρ represents the squared difference between the predicted width and the actual width. 2 (h,h gt ) represents the squared difference between the predicted height and the actual height, and c represents the diagonal distance of the smallest closure region that can simultaneously contain both the predicted and actual bounding boxes. The BCELoss function is used to accurately distinguish the crushed block area from the background, avoiding the crusher from accidentally hitting non-target areas. The DFL function is used to dynamically focus on the edge features of the crushed block, reducing the error value of boundary delineation. The EIoU Loss function is used to independently optimize the width / height regression, accurately select the defect area, and reduce the error in locating the voids in the crushed block.

[0111] Working Principle: The portable scrap steel patty sampling device focuses on automated sampling and intelligent crushing, achieving efficient processing through a three-tiered technology of visual analysis, defect localization, and intelligent crushing. First, the electromagnetic chuck 11 adsorbs the scrap steel patty and places it above the retractable housing 3. Next, the image acquisition device 16 collects real-time data on the surface cracks and iron oxide distribution characteristics of the scrap steel and transmits this information to the edge computing terminal 5. The edge computing terminal is equipped with a deep learning model based on a convolutional neural network-based target detection framework and deploys an attention-based metal defect detection algorithm to analyze image data and identify defect areas such as cracks and iron oxide, generating a three-dimensional crushing path planning instruction. Subsequently, according to the instruction, the lateral drive device 15 moves along the lateral slide rail fixing frame 17, and the multi-stage retractable arm 14, in conjunction with the hydraulic crushing pick 13, precisely impacts and crushes the scrap steel according to the planned path, prioritizing the processing of internal defect areas. Throughout the process, the crushed scrap steel automatically falls into the retractable housing 3 to complete the sampling. Furthermore, the braked casters 2 and the handrail assembly 7 on the chassis base 1 facilitate the overall movement and operation control of the equipment.

[0112] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0113] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

Claims

1. A portable flake press cake sampling device, characterized by: The utility model relates to a vehicle frame, including the frame base (1), one side of the top of frame base (1) is equipped with telescopic box (3), is provided with the frame main part (4) above telescopic box (3), is equipped with linear guide rail (12) on the side wall towards frame base (1) of frame main part (4), is provided with electric lifting mechanism (9) on linear guide rail (12), is installed with telescopic mechanical arm (10) on electric lifting mechanism (9), is equipped with electromagnetic chuck (11) at the end of telescopic mechanical arm (10), is provided with horizontal slide rail fixed frame (17) above frame main part (4), is installed with horizontal drive arrangement (15) through slide rail on the bottom surface of horizontal slide rail fixed frame (17), is provided with multistage telescopic arm (14) at the bottom of horizontal drive arrangement (15), is installed hydraulic breaking pick head (13) at the bottom end of multistage telescopic arm (14), the bottom of horizontal drive arrangement (15) is also provided with image acquisition device (16), and image acquisition device (16) is located at one side of multistage telescopic arm (14); The image acquisition device (16) is used for real-time acquisition of surface cracks and iron oxide distribution characteristics of scrap steel, and image transmission to an edge computing terminal (5); The edge computing terminal (5) carries a deep learning model based on a target detection framework of a convolutional neural network, and is deployed with a metal defect detection algorithm based on an attention mechanism, which is used for real-time analysis of scrap steel morphology characteristics transmitted by the image acquisition device (16) to generate three-dimensional crushing path planning instructions; The defect features are extracted through the attention mechanism, and the following formula is used: The spatial attention module calculation formula is as follows: W s = Softmax(Γ(μ(F) || ε(F))) The channel attention module calculation formula is as follows: W c = σ(MLP(P_a(F)) + MLP ′ (P_a(F))) Wherein F is an input feature map, Γ (·) is a 3*3 depth separable convolution, || represents channel dimension splicing, μ (·) and ξ (·) represent average pooling and maximum pooling along the channel dimension respectively, and sigma () is a sigmoid function; P_a () represents global mixed pooling: alpha. AvgPool (·) + (1-alpha). MaxPool (·), alpha belongs to [0.4, 0.6]; AvgPool is the average color of each region of the photo, MaxPool is the brightest point of each region of the photo, MLP and MLP' are double-path fully connected layers with shared parameters. When the maximum value of the feature map multiplied by the spatial weight exceeds the threshold, the feature map is marked as a structural weak area, wherein theta defect=0.7 is determined by the ROC curve.

2. A portable pellet press sampling device according to claim 1, wherein: The telescopic box (3) includes a fixed box (3a), a movable box (3b), a folding linkage group (3c), and a hydraulic cylinder (3d). The fixed box (3a) is bolted to the upper surface of the frame base (1). The folding linkage group (3c) is composed of two V-shaped hydraulic rods, one end of which is hinged to the inner side wall of the fixed box (3a), and the other end is hinged to the top of the movable box (3b). The hydraulic cylinder (3d) is horizontally arranged between the fixed box (3a) and the frame main body (4), and the piston rod end is hinged to the intersection point of the folding linkage group (3c).

3. A portable pellet press sampling device according to claim 2, wherein: The side away from the linear guide rail (12) of the frame body (4) is connected with a storage battery (6), the side away from the frame body (4) of the storage battery (6) is installed with a handrail assembly (7), the top of the handrail assembly (7) is provided with a control handle (8) and an edge computing terminal (5), the bottom of the handrail assembly (7) and the frame base (1) is installed with a brake universal wheel (2).

4. A portable pellet press sampling device according to claim 3, wherein: The edge computing terminal (5) is signal connected with the image acquisition device (16), the hydraulic breaking pick head (13), the transverse driving device (15) and the multi-stage telescopic arm (14).

5. A method for sampling flake press cakes on the fly, implemented on the basis of a device for sampling flake press cakes on the fly according to any one of claims 1 to 4, characterized in that: Comprise the following steps: Step one: through the electromagnetic suction cup (11) adsorbs the scrap steel cake to the telescopic box (3) above; Step two: the image acquisition device (16) acquires the scrap steel surface image and transmits to the edge computing terminal (5); Step three: the edge computing terminal (5) identifies the crack and iron oxide defect area through the composite attention mechanism, generates three-dimensional breaking path instruction; Step four: the edge computing terminal (5) outputs the instruction and controls the transverse driving device (15) to move along the transverse slide rail, and impacts the scrap steel defect area through the cooperation of the multi-stage telescopic arm (14) and the breaking pick head (13) according to the path instruction; Step five: the broken scrap steel material falls into the telescopic box (3), and the sampling is completed.

6. A method of portable pellet press sampling according to claim 5, wherein: For step two, the RGB-D depth camera is used to collect the surface image of the scrap steel cake in real time, which is input into the edge computing terminal (5), and the image is normalized by the following formula: where μ k , σ k are the training set statistics, i: index in the height direction (1≤i≤H), j: index in the width direction (1≤j≤W), k: index in the channel direction (1≤k≤C).

7. A method of portable pellet press sampling according to claim 6, wherein: For step three, the generation process of three-dimensional breaking path is as follows: The scrap steel cake is divided into 10mm×10mm×10mm voxel units; Calculate the impact priority: P(v)=α·D(v)+β·S(v) Where: v is the voxel unit, which is generated based on the depth camera point cloud data; D(v) is the distance from the voxel v to the surface, which is obtained by the following formula: d max d is the maximum thickness of the briquette, d(v) is the actual measured distance; S(v) is the defect confidence (0-1) output value from the attention module, the larger the value, the higher the defect probability, spatial attention weight W s output value, α = 0.6, β = 0.4 are weight coefficients, indicating that internal defects are preferred to surface defects, which are determined by experimental optimization; Generate B-spline curve path: wherein: B(t) is a parametric curve coordinate control point, representing the movement trajectory of the breaking pick (13), P i is the i-th control point coordinate, corresponding to the highest priority voxel center, N i,p (t) is a p-th B-spline basis function, determining the smoothness of the curve, controlling the breaking pick (13) to impact the briquetting cake along the planned path B(t) through the above-mentioned B-spline curve path function.

Images