Automatic sorting method of ferrovanadium

Abstract

The application discloses an automatic sorting method for vanadium-iron alloy, and relates to the technical field of automatic control, which comprises the following steps: training a deep learning model for identifying whether the vanadium-iron alloy is qualified and outputting an initial bounding box containing the position and confidence of the vanadium-iron alloy; adaptively adjusting the bounding box according to the confidence score output by the deep learning model, dynamically determining a scaling coefficient through pre-generated lookup table or function calculation, and correcting the width and height of the initial bounding box; implementing coordinate conversion to convert the center pixel coordinates of the target bounding box into grabbing coordinates in the base coordinate system of the mechanical arm; and the mechanical arm performs grabbing action according to a double-stop confirmation mechanism to grab the unqualified vanadium-iron alloy. The application aims to realize high-precision and fully-automatic sorting of vanadium-iron alloy, improve production efficiency, and ensure personnel health.

Description

Technical Field

[0001] This invention relates to the field of automation control technology, and in particular to an automatic sorting method for ferrovanadium alloys. Background Technology

[0002] Ferrovanadium alloys are key raw materials for steel smelting and special alloy manufacturing, and their grading directly affects the quality of downstream products. Currently, most domestic enterprises still use manual visual sorting. This method has the following technical drawbacks: (1) High health risks: The dust concentration at the sorting site is high, and long-term operation can easily induce occupational diseases such as pneumoconiosis.

[0003] (2) The manual judgment standards are inconsistent, and long-term operation can easily lead to visual fatigue, resulting in a high rate of missed picking and mispicking. In addition, the processing capacity per unit time is limited and cannot meet the needs of large-scale production.

[0004] (3) Some existing general material sorting systems (such as machine vision-based parts sorting lines) are usually designed for industrial parts with regular shape and single color. The fixed-size detection frame and simple coordinate correspondence method used are difficult to adapt to the characteristics of ferrovanadium alloy with extremely irregular shape, complex surface texture and variable reflective properties, resulting in a very high failure rate of grasping in actual application and failing to meet the reliability requirements of industrial production.

[0005] Therefore, developing a high-precision, high-efficiency, and high-reliability automated sorting method is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the problems existing in the prior art, the purpose of this invention is to provide an automatic sorting method for ferrovanadium alloys, aiming to achieve high-precision, fully automatic sorting of ferrovanadium alloys, improve production efficiency, and protect personnel health.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is: an automatic sorting method for ferrovanadium alloys, comprising the following steps: Step 1: Train a deep learning model to identify whether ferrovanadium alloy is qualified and output an initial bounding box containing the location and confidence level of the ferrovanadium alloy. Step 2: Adaptive bounding box adjustment. Based on the confidence score output by the deep learning model, the scaling factor is dynamically determined through a pre-generated lookup table or function calculation to correct the width and height of the initial bounding box. Step 3: Perform coordinate transformation to convert the center pixel coordinates of the target bounding box into the grasping coordinates in the robot arm base coordinate system; Step 4: The robotic arm performs a gripping action according to the dual stop confirmation mechanism to grip the unqualified ferrovanadium alloy.

[0008] As a further improvement of the present invention, in step 1, images of ferrovanadium alloy under different working conditions, angles, and lighting conditions are collected to construct a dedicated dataset and train the YOLO v8 deep learning model.

[0009] As a further improvement of the present invention, in step 2, the width and height of the initial bounding box are modified as follows: If the initial bounding box has a width of w and a height of h, then the corrected bounding box has a height w' = w × α and a width h' = h × β; where α and β are scaling factors.

[0010] As a further improvement of the present invention, in step 3, the conversion method adopts a hand-eye calibration algorithm to pre-calculate the rotation matrix R and translation vector T; or adopts a preset mapping table method to obtain the correspondence between visual coordinates and robotic arm motion parameters through experimental calibration, and uses linear interpolation for precise matching.

[0011] As a further improvement to the present invention, the hand-eye calibration algorithm is as follows: A checkerboard calibration board is placed within the working range of the robotic arm, and the robotic arm is controlled to move to multiple poses. Simultaneously, the corner coordinates of the calibration board in the camera image and the pose of the robotic arm's end effector in the base coordinate system are recorded. The rotation matrix R and translation vector T from the camera coordinate system to the robotic arm's base coordinate system are calculated using the `calibrateHandEye()` function in the OpenCV library. During real-time transformation, the center pixel coordinates (u, v) of the target bounding box are extracted. Combined with the camera intrinsic parameter matrix K and the depth Z_c, the 3D coordinates in the robotic arm's base coordinate system are calculated: P_robot = R × (K... -1 ×[u, v,1] T ×Z_c)+T.

[0012] As a further improvement of the present invention, the preset mapping table method is specifically as follows: A series of calibration points with known coordinates are pre-placed within the visual field. The visual coordinates of each calibration point are recorded, including the Euclidean distance *d* from the origin and the angle *θ* between the point and the origin, as well as the motion parameters required for the robotic arm to actually grasp that point. Simultaneously, a mapping table is established, through... Calculate the Euclidean distance from point (x1, y1) to the origin (0, 0) using the formula... , Calculate the angle between point (x1, y1) and origin (0, 0); during actual operation, calculate the distance d and angle θ between the center point of the vanadium-iron alloy and the origin, find the corresponding interval in the table, use the linear interpolation formula to calculate the precise distance X and angle Y, and then look up the robotic arm motion data in reverse.

[0013] As a further improvement of the present invention, in step 4, the dual stop confirmation mechanism is specifically as follows: First step: The photoelectric sensor at the ferrovanadium alloy positioning station is triggered, and the PLC controls the conveyor belt to stop; Second step: The vision system confirms that the ferrovanadium alloy has stopped at the gripping position through continuous frame image analysis.

[0014] The beneficial effects of this invention are: 1. High sorting accuracy: The visual model trained on a dedicated dataset achieves a recognition accuracy of over 96%, and combined with adaptive bounding boxes, the success rate of grasping is significantly improved.

[0015] 2. High degree of automation: PLC controls the entire process, enabling unattended operation, greatly improving production efficiency and reducing labor costs.

[0016] 3. High reliability: Hand-eye calibration technology and dual stop confirmation mechanism ensure the accuracy of coordinate transformation and the reliability of grasping action.

[0017] 4. High adaptability: Specifically optimized for the shape, texture, and luster characteristics of ferrovanadium alloys, solving the problem of poor adaptability of general sorting equipment.

[0018] 5. High safety: It completely replaces manual sorting and eliminates the health hazards of dust. Attached Figure Description

[0019] Figure 1 This is a flowchart of an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the acquisition of image samples and their annotation data in an embodiment of the present invention; Figure 3 This is a graph showing the recognition fitting curve of the deep learning model in an embodiment of the present invention. Detailed Implementation

[0020] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] Example

[0022] like Figure 1 As shown, an automatic sorting method for ferrovanadium alloys includes: (1) Visual recognition based on deep learning: By collecting images of ferrovanadium alloy under different working conditions, angles and lighting conditions, a dedicated dataset is constructed and a YOLO v8 deep learning model is trained. This model can identify qualified and unqualified products and output an initial bounding box containing the location and confidence level of the ferrovanadium alloy.

[0023] (2) Adaptive bounding box adjustment: This invention proposes an adaptive adjustment algorithm to address the issue that the initial bounding box output by YOLO v8 may not perfectly match the irregular vanadium-iron alloy contour. Based on the confidence score output by the model, the algorithm dynamically determines the scaling factors α and β through a pre-generated lookup table or function calculation, and corrects the width (w) and height (h) of the bounding box (e.g., w'=w×0.9, h'=h×0.85), so that it tightly wraps around the vanadium-iron alloy body, providing a precise positioning basis for subsequent grasping.

[0024] (3) Implement coordinate transformation to convert the center pixel coordinates of the target bounding box into the grasping coordinates in the coordinate system of the robotic arm base. The transformation method can be the classic hand-eye calibration algorithm ("eye-out" method), which pre-calculates the rotation matrix R and translation vector T; or it can be the preset mapping table method, which obtains the correspondence between visual coordinates and robotic arm motion parameters through experimental calibration, and uses linear interpolation for accurate matching.

[0025] (4) To ensure accurate grasping timing, this invention designs a dual stop confirmation mechanism. First, the ferrovanadium alloy triggers the photoelectric sensor at the positioning station, and the PLC controls the conveyor belt to stop; second, the vision system confirms that the ferrovanadium alloy has stopped at the grasping position through continuous frame image analysis (e.g., the center coordinate change is <2 pixels for 5 consecutive frames). The robotic arm only performs the grasping operation when both signals are satisfied, effectively preventing misgrabbing or empty grasping.

[0026] The following is a further explanation of this embodiment: (1) Model Training: Over 100,000 images of vanadium-iron alloy were collected, covering various working conditions such as dust, dust-free, humid, and dry, as well as various angles such as frontal, side, top, and tilt angles, and various lighting conditions such as weak light, strong light, side light, and top light. Labeling was performed using Labelimg, labeling the images as qualified and unqualified, and dividing them into training and test sets at an 8:2 ratio. Examples of collected image samples and their labeled data are shown below. Figure 2 As shown. The model was trained using the YOLO v8 deep learning framework, and the final model achieved an average recognition accuracy of 96.5% on the test set. The fitting accuracy is as follows. Figure 3 As shown.

[0027] (2) Adaptive bounding box adjustment: YOLO v8 outputs the initial bounding box (including center coordinates x, y, width w, height h, and confidence score), and determines the scaling factors α and β based on the confidence score by looking up a table. For example: When Score > 0.95, α=0.9, β=0.85; When 0.85 < Score ≤ 0.95, α=0.88, β=0.82; When Score ≤ 0.85, α=0.85, β=0.78.

[0028] The adjusted bounding box: w' = w × α, h' = h × β. These coefficients were determined through extensive gripping tests to ensure that the bounding box tightly encloses the vanadium-iron alloy body.

[0029] (3) Coordinate transformation (either method can be used): Hand-eye calibration (Method 1): The "eye outside the hand" calibration method is adopted. A checkerboard calibration board is placed within the working range of the robotic arm. The robotic arm is controlled to move to multiple poses, while simultaneously recording the corner coordinates of the calibration board in the camera image and the pose of the robotic arm's end effector in the base coordinate system. The rotation matrix R and translation vector T from the camera coordinate system to the robotic arm's base coordinate system are calculated using the `calibrateHandEye()` function in the OpenCV library. During real-time conversion, the center pixel coordinates (u,v) of the target bounding box are extracted. Combined with the camera intrinsic matrix K and the depth Z_c (obtained through laser ranging or a preset height), the 3D coordinates in the robotic arm's base coordinate system are calculated: P_robot = R × (K... -1 ×[u, v,1] T ×Z_c)+T.

[0030] Pre-set mapping table (Method 2): A series of calibration points with known coordinates are pre-placed within the visual field. The visual coordinates of each calibration point (Euclidean distance d from the origin and the angle θ between the point and the origin) and the motion parameters required for the robotic arm to actually grasp that point are recorded. Simultaneously, a mapping table is established, through... Calculate the Euclidean distance from point (x1, y1) to the origin (0, 0) using the formula... , The angle between the calculated point (x1, y1) and the origin (0, 0) is shown in Table 1 (X is the distance interval setting value, and Y is the angle interval setting value). During actual operation, the distance d and angle θ between the center point of the ferrovanadium alloy and the origin are calculated. The corresponding interval is found in the table, and the accurate Y value is calculated using the linear interpolation formula. Then, the robot arm motion data is retrieved in reverse.

[0031] Table 1

[0032] (4) Double stop confirmation: When the leading edge of the ferrovanadium alloy triggers the positioning photoelectric sensor, the PLC immediately sends a first stop signal, cutting off the power to the conveyor belt drive motor. Subsequently, the vision system begins to perform a stationary confirmation judgment: continuously monitoring the image sequence, when the change in the center coordinates of the ferrovanadium alloy in five consecutive frames is less than 2 pixels, it determines that the ferrovanadium alloy has come to a complete stop and sends a second stop signal to the PLC.

[0033] The PLC's internal control logic is configured such that it only sends a permission to grasp to the robotic arm controller after receiving the first and second stop signals in sequence. If the second stop signal is not received within the specified time, the PLC issues an alarm and pauses the process.

[0034] Grasping: The PLC sends the gripping coordinates to the robotic arm controller. The robotic arm moves to the target point according to the planned path, closes the gripper, grabs the defective ferrovanadium alloy, and transfers it to the defective material bin. Reset: After the robotic arm returns to its original position, the PLC controls the conveyor belt to restart, starting the next cycle.

[0035] Exception handling is divided into three parts: (1) Recognition failure: If the confidence level is lower than the preset threshold (e.g., 0.75), it is marked as "to be inspected" and picked up by the robotic arm and moved to the re-inspection area.

[0036] (2) Grabbing Failure: If the force value fed back by the pressure sensor at the end of the robotic arm is lower than the preset threshold, it is determined to be a no-grab. The system records the position of the ferrovanadium alloy and attempts to grab the ferrovanadium alloy again in the next cycle. If the same ferrovanadium alloy fails to be grabbed twice in a row, an alarm will be issued to prompt manual intervention.

[0037] (3) Communication failure: If the communication between the PLC and the vision system or the robotic arm controller is interrupted for more than 5 seconds, the system will automatically stop and alarm.

[0038] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. An automatic sorting method for ferrovanadium alloys, characterized in that, Includes the following steps: Step 1: Train a deep learning model to identify whether ferrovanadium alloy is qualified and output an initial bounding box containing the location and confidence level of the ferrovanadium alloy. Step 2: Adaptive bounding box adjustment. Based on the confidence score output by the deep learning model, the scaling factor is dynamically determined through a pre-generated lookup table or function calculation to correct the width and height of the initial bounding box. Step 3: Perform coordinate transformation to convert the center pixel coordinates of the target bounding box into the grasping coordinates in the robot arm base coordinate system; Step 4: The robotic arm performs a gripping action according to the dual stop confirmation mechanism to grip the unqualified ferrovanadium alloy.

2. The automatic sorting method for ferrovanadium alloys according to claim 1, characterized in that, In step 1, images of vanadium-iron alloy under different working conditions, angles, and lighting conditions are collected to construct a dedicated dataset and train the YOLO v8 deep learning model.

3. The automatic sorting method for ferrovanadium alloys according to claim 1, characterized in that, In step 2, the width and height of the initial bounding box are corrected as follows: If the initial bounding box has a width of w and a height of h, then the corrected bounding box has a height w' = w × α and a width h' = h × β; where α and β are scaling factors.

4. The automatic sorting method for ferrovanadium alloys according to claim 1 or 3, characterized in that, In step 3, the conversion method adopts a hand-eye calibration algorithm to pre-calculate the rotation matrix R and translation vector T; or adopts a preset mapping table method to obtain the correspondence between visual coordinates and robotic arm motion parameters through experimental calibration, and uses linear interpolation for accurate matching.

5. The automatic sorting method for ferrovanadium alloys according to claim 4, characterized in that, The hand-eye calibration algorithm is as follows: A checkerboard calibration board is placed within the working range of the robotic arm, and the robotic arm is controlled to move to multiple poses. Simultaneously, the corner coordinates of the calibration board in the camera image and the pose of the robotic arm's end effector in the base coordinate system are recorded. The rotation matrix R and translation vector T from the camera coordinate system to the robotic arm's base coordinate system are calculated using the `calibrateHandEye()` function in the OpenCV library. During real-time transformation, the center pixel coordinates (u, v) of the target bounding box are extracted. Combined with the camera intrinsic parameter matrix K and the depth Z_c, the 3D coordinates in the robotic arm's base coordinate system are calculated: P_robot = R × (K... -1 ×[u, v,1] T ×Z_c)+T.

6. The automatic sorting method for ferrovanadium alloys according to claim 4, characterized in that, The specific method for setting the mapping table is as follows: A series of calibration points with known coordinates are pre-placed within the visual field. The visual coordinates of each calibration point are recorded, including the Euclidean distance *d* from the origin and the angle *θ* between the point and the origin, as well as the motion parameters required for the robotic arm to actually grasp that point. Simultaneously, a mapping table is established, through... ; Calculate the Euclidean distance from point (x1, y1) to the origin (0, 0) using the formula... , Calculate the angle between point (x1, y1) and origin (0, 0); during actual operation, calculate the distance d and angle θ between the center point of the vanadium-iron alloy and the origin, find the corresponding interval in the table, use the linear interpolation formula to calculate the precise distance X and angle Y, and then look up the robotic arm motion data in reverse.

7. The automatic sorting method for ferrovanadium alloys according to claim 1, characterized in that, In step 4, the dual stop confirmation mechanism is specifically as follows: First step: The photoelectric sensor at the ferrovanadium alloy positioning station is triggered, and the PLC controls the conveyor belt to stop; Second step: The vision system confirms that the ferrovanadium alloy has stopped at the gripping position through continuous frame image analysis.

Images