Sheet metal bending machining precision detection and compensation method based on machine vision

Abstract

The application provides a sheet metal bending machining precision detection and compensation method based on machine vision, which comprises the following steps: step 1, a machine vision X, Y linear shaft positioning error detection system XYC is built; step 2, an ESSA-Elman positioning error prediction model for predicting X, Y linear shaft positioning error is established; step 3, a machine vision error detection system WJC is built; and step 4, neural network feedforward and machine vision real-time feedback are combined to compensate for size error, and machine vision real-time feedback is used to compensate for bending forming angle error. The application detects and compensates for the size error and the bending forming angle error, and improves the precision of the bending finished product.

Description

Technical Field

[0001] This invention belongs to the field of sheet metal bending processing accuracy detection and compensation technology, and particularly relates to a sheet metal bending processing accuracy detection and compensation method based on machine vision. Background Technology

[0002] Sheet metal bending, as an important part of the manufacturing industry, is characterized by its simple forming process and good mold versatility. This processing method uses the cooperation of the upper and lower molds of a bending machine to bend horizontally placed metal workpieces into the required shape. Among them, the bending machine, as the key equipment in this process, occupies a very important position in the sheet metal processing industry.

[0003] In the sheet metal bending process, robots and bending machines need to work together to complete each bending process. However, due to the positioning accuracy problem of the robot itself and the material characteristics of the metal bending parts, the accuracy of the bent product is affected. Summary of the Invention

[0004] The purpose of this invention is to provide a machine vision-based method for detecting and compensating for the accuracy of sheet metal bending processes. By detecting and compensating for dimensional errors and bending angle errors, the accuracy of the bent product is improved. To achieve the above objective, the following technical solution is adopted:

[0005] A machine vision-based method for detecting and compensating the accuracy of sheet metal bending processes includes the following steps:

[0006] Step 1: Build a machine vision XYC linear axis positioning error detection system to obtain the corner coordinates of the bent part and output the X and Y linear axis positioning error of the bent part at the current sampling position. This includes the following steps:

[0007] An industrial camera is positioned above the rectangular work area and connected to a PC; a robot is placed on one side of the rectangular work area to pick up the bent parts.

[0008] The long side of the rectangular working area is parallel to the Y-axis of the robot coordinate system, and the wide side is parallel to the X-axis of the robot coordinate system.

[0009] The rectangular working area is divided into several rectangular sampling positions P at fixed intervals. sample The length and width of each sampling position are parallel to the length and width of the rectangular working area, respectively;

[0010] Step 2: Establish the ESSA-Elman positioning error prediction model for predicting the positioning errors of the X and Y linear axes, which includes the following steps:

[0011] Step 21: Sampling is performed based on XYC. The X and Y linear axis positioning errors are output according to the identified corner coordinates of the bent parts.

[0012] To address the positioning errors of the robot's X and Y linear axes, the machine vision error detection system collected the X and Y axis positioning errors at each sampling position through multiple positioning error sampling experiments of the robot.

[0013] The X and Y axis positioning errors at each sampling location, along with the corresponding theoretical X and Y operating parameters of the robot, form a mapping dataset; where the theoretical X and Y operating parameters of the robot are the theoretical operating parameters P of the robot from the starting position to each sampling location. sample =(x sample ,y sample );

[0014] Step 22: Based on the sampled mapping dataset, use an optimization algorithm to optimize the neural network training and establish the ESSA-Elman localization error prediction model:

[0015] The ESSA-Elman positioning error prediction model is used to predict the positioning errors of the X and Y linear axes. The predicted X and Y linear axis positioning errors are used to correct the theoretical X and Y operating parameters of the robot. The corrected theoretical X and Y operating parameters of the robot are then input into the robot.

[0016] Step 3: Build the machine vision error detection system WJC, which specifically includes:

[0017] Industrial camera one is positioned on the bending machine, facing the front of the bent part to capture the rotation angle of the bent part; industrial camera two is positioned on the side of the bending machine, facing the side of the bent part to capture the forming angle in real time; the camera control device outputs the rotation angle error and forming angle error to the display module.

[0018] The PC feeds back the rotation angle error to the robot's end-rotor axis through the control module and robot control device, and feeds back the forming angle error to the upper die of the bending machine through the control module and bending machine control device for corresponding compensation.

[0019] The control module stores the ESSA-Elman positioning error prediction model;

[0020] Step 4: Compensate for dimensional errors using a combination of neural network feedforward and machine vision real-time feedback, and compensate for bending angle errors using machine vision real-time feedback. This includes the following steps:

[0021] The ESSA-Elman positioning error prediction model outputs positioning error values ​​based on the robot's theoretical X and Y operating parameters, compensates for the robot's theoretical X and Y operating parameters, and finally outputs the compensated robot's theoretical X and Y operating parameters to the robot.

[0022] WJC measures the rotation angle in real time and outputs the rotation angle error to the robot.

[0023] WJC measures the bending angle in real time and outputs the bending angle error to the robot.

[0024] Preferably, step 21 specifically includes:

[0025] With a fixed distance d sample To maintain a fixed sampling step size, several sampling positions are defined;

[0026] The robot performs zigzag sampling along the Y-axis on the rectangular work area plane, that is, it starts picking up the bent part from the starting position P0 and proceeds according to position P. 11 P 12 …P 1n After sampling sequentially, then from P 11 Distance d sample Position P 21 Start, along position P 21 P 22 …P 2n Sample the second row until the final position P is sampled. mn ;

[0027] Each sampling starts from the initial position P0 of the bent part, and the robot's movement is controlled by operating only the ideal X and Y operation parameters to reach each sampling position.

[0028] The bent parts arrive at each sampling point with the same posture as the starting position. Then, the X and Y axis positioning errors of the bent parts at the current sampling position are collected and output by an industrial camera and a PC.

[0029] The output positioning error is consistent with the unit of the robot's X and Y operation parameters, which is millimeters.

[0030] Preferably, the method for acquiring the X and Y linear axis positioning errors using machine vision in step 21 is as follows:

[0031] The image coordinate system is the opposite of the sampling coordinate system in step 1, so that the error value Δy of the robot's Y-axis corresponds to the Y-axis in step 1, and the error value Δx of the robot's X-axis corresponds to the X-axis in step 1.

[0032] Let the coordinates of corner points 2 and 4 of the bent part at a certain sampling location be P, respectively. t2 (x t2 ,y t2 ), P t4 (x t4 ,y t4 The coordinates of the midpoint of this side are...

[0033] The coordinates of the actual corner points 2 and 4 of the bent part, measured by the industrial camera, are P1 and P2 respectively.a2 (x a2 ,y a2 ), P a4 (x a4 ,y a4 The coordinates of the midpoint of this side are...

[0034] Error value Δy along the robot's Y-axis:

[0035]

[0036] P t and P a The Euclidean distance d between them:

[0037]

[0038] The error value Δx along the robot's X-axis is:

[0039]

[0040] The positioning error directions along the X and Y axes are determined by comparing P. t and P a The relationship between the magnitudes of the horizontal and vertical coordinates can be obtained.

[0041] Preferably, step 22 specifically includes:

[0042] Using the Elman neural network as the base network, the optimal initial weights and thresholds of the network are found iteratively by utilizing the sparrow search algorithm. Then, the network is trained with these weights and thresholds as initial values, which can achieve better prediction results.

[0043] Meanwhile, to address the issue of poor sparrow population diversity caused by the random generation of initial sparrow population locations in the sparrow search algorithm, the initial sparrow population locations are generated using the uniformity and uncertainty of the Tent chaotic mapping function. This preserves population diversity while enhancing the global search capability of the sparrow search algorithm, thus establishing the SSA-Elman positioning error prediction model ESSA-Elman optimized by the Tent chaotic mapping.

[0044] Preferably, step 4, the process of compensating for the robot's theoretical X and Y operating parameters, includes:

[0045] The actual positioning error E at each sampling location sample =(Δx) sample ,Δy sample );

[0046] The theoretical operating parameters P of the target location target The predicted positioning error E is obtained by inputting it into the prediction model. t ′arget ;

[0047] Finally, the theoretical operating parameters P for the target location. target The predicted positioning error E is superimposed in reverse. t ′ arget The compensated target position operation parameter P can then be obtained. modified The information is input into the robot to complete the compensation.

[0048] Preferably, in step 4, the rotation angle compensation process includes:

[0049] The theoretical angle of rotation required in the bending process is θ. First, the robot is controlled to rotate the bent part once by the angle θ.

[0050] After rotation, an industrial camera located directly above the rectangular working area is used to collect the compensation angle Δθ and compensation direction in real time. The compensation direction is either clockwise or counterclockwise.

[0051] The robot rotates the bent part twice according to the collected compensation angle Δθ and compensation direction to compensate for the error.

[0052] Preferably, the forming bending angle compensation process includes:

[0053] During the bending process, the slider pressure Y1 is first calculated according to the theoretical bending angle θ1 and the slider pressure calculation model, and Y1 is used to control the bending machine to bend once.

[0054] After the upper die of the bending machine is unloaded and the bent part has fully springed back, the bending forming angle θ2 after springback is collected in real time by an industrial camera located on the bending side of the bending machine. Then the springback angle Δθ = θ2 - θ1 at this time.

[0055] The compensation angle θ3 = θ1 - Δθ is obtained from the springback angle Δθ and the theoretical bending angle θ1. Based on θ3 and the slider pressure calculation model Y, the compensation slider pressure Y2 is calculated, and Y2 is used to control the bending machine for secondary bending compensation.

[0056]

[0057] V - Lower die opening width, mm;

[0058] R d -Lower die fillet radius, mm;

[0059] t - Thickness of the bent part when not under stress, in mm.

[0060] R - Bending radius, mm, derived from the empirical formula R = K v ·V is obtained;

[0061] K vThe value is taken as 0.156 according to the German industrial standard DIN6935-2010;

[0062] η - The thinning coefficient due to deformation caused by bending force, from Please obtain.

[0063] Preferably, the corner coordinates in step 21 are sub-pixel level corner coordinates.

[0064] Compared with the prior art, the advantages of the present invention are:

[0065] To address the dimensional and bending angle errors generated during the bending process, a machine vision error detection system was designed to detect these errors. Neural network feedforward compensation was used for the robot's X and Y linear axis positioning errors that caused dimensional errors. Machine vision was used to detect and compensate for the robot's rotational axis rotation angle errors and bending angle errors that caused dimensional errors in real time. The effectiveness of the compensation method was verified through bending comparison experiments before and after compensation. Attached Figure Description

[0066] Figure 1 This is a structural diagram of the sampling positions in the XYC region of a machine vision X and Y linear axis positioning error detection system.

[0067] Figure 2 Here is a structural diagram of the WJC machine vision error detection system;

[0068] Figure 3 This is a flowchart of a machine vision-based method for detecting and compensating for the accuracy of sheet metal bending processes.

[0069] Figure 4 A schematic diagram illustrating the principle of a machine vision method for acquiring X and Y linear axis positioning errors.

[0070] Figure 5 A schematic diagram illustrating the principle of machine vision for acquiring rotation angle errors;

[0071] Figure 6 Schematic diagram of the training and learning process of the ESSA-Elman positioning error prediction model;

[0072] Figure 7 A schematic diagram illustrating the process of compensating for the theoretical X and Y operating parameters of a robot;

[0073] Figure 8 This is a schematic diagram of the rotation angle compensation process;

[0074] Figure 9 This is a schematic diagram of the forming and bending angle compensation process;

[0075] Figure 10 A calculation model for the slider's downward pressure;

[0076] Figure 11 This serves as the overall experimental platform for bending.

[0077] Figure 12 This is a flowchart illustrating the process of bending the first bend.

[0078] Figure 13 This is a diagram of the bending process for turning over a bent part.

[0079] Figure 14 This is a flowchart illustrating the process of making the second bend.

[0080] Figure 15 This is a measurement diagram for the precision of the bent finished product;

[0081] Figure 16 This is a schematic diagram for measuring dimensional errors. Detailed Implementation

[0082] The following will describe in more detail the machine vision-based sheet metal bending accuracy detection and compensation method of the present invention with reference to the schematic diagrams, which illustrate preferred embodiments of the invention. It should be understood that those skilled in the art can modify the invention described herein while still achieving the advantageous effects of the invention. Therefore, the following description should be understood as being of general knowledge to those skilled in the art and is not intended to limit the invention.

[0083] like Figures 1-16 A machine vision-based method for detecting and compensating the accuracy of sheet metal bending processes includes the following steps:

[0084] Step 1: Build a machine vision XYC linear axis positioning error detection system to obtain the corner coordinates of the bent part and output the X and Y linear axis positioning error of the bent part at the current sampling position. This includes the following steps:

[0085] An industrial camera is positioned above the rectangular work area and connected to a PC; a robot is placed on one side of the rectangular work area to pick up the bent parts.

[0086] The long side of the rectangular working area is parallel to the Y-axis of the robot coordinate system, and the wide side is parallel to the X-axis of the robot coordinate system.

[0087] The rectangular working area is divided into several rectangular sampling positions P at fixed intervals. sample The length and width of each sampling position are parallel to the length and width of the rectangular working area, respectively.

[0088] Step 2: Establish the ESSA-Elman positioning error prediction model for predicting the positioning errors of the X and Y linear axes, which includes the following steps:

[0089] Step 21: Sampling is performed based on XYC. The X and Y linear axis positioning errors are output according to the identified corner coordinates of the bent parts.

[0090] To address the positioning errors of the robot's X and Y linear axes, the machine vision error detection system collected the X and Y axis positioning errors at each sampling position through multiple positioning error sampling experiments of the robot.

[0091] The X and Y axis positioning errors at each sampling location, along with the corresponding theoretical X and Y operating parameters of the robot, form a mapping dataset; where the theoretical X and Y operating parameters of the robot are the theoretical operating parameters P of the robot from the starting position to each sampling location. sample =(x sample ,y sample ).

[0092] Step 21 specifically includes:

[0093] Using the robot's coordinate system as a reference, and a fixed distance d sample To use a fixed sampling step size;

[0094] The robot performs zigzag sampling along the Y-axis on the rectangular work area plane, that is, it starts picking up the bent part from the starting position P0 and proceeds according to position P. 11 P 12 …P 1n After sampling sequentially, then from P 11 Distance d sample Position P 21 Start, along position P 21 P 22 …P 2n Sample the second row until the final position P is sampled. mn ;

[0095] Each sampling starts from the initial position P0 of the bent part, and the robot's movement is controlled by operating only the ideal X and Y operation parameters to reach each sampling position.

[0096] The bent parts arrive at each sampling point with the same posture as the starting position. Then, the X and Y axis positioning errors of the bent parts at the current sampling position are collected and output by an industrial camera and a PC.

[0097] The output positioning error is consistent with the unit of the robot's X and Y operation parameters, which is millimeters.

[0098] The principle by which the camera control device acquires the X and Y linear axis positioning errors, i.e., the method by which machine vision acquires the X and Y linear axis positioning errors, is as follows:

[0099] The image coordinate system has the X-axis pointing horizontally to the right and the Y-axis pointing vertically downwards.

[0100] Let the coordinates of corner points 2 and 4 of the bent part at a certain sampling location be P, respectively. t2 (xt2 ,y t2 ), P t4 (x t4 ,y t4 The coordinates of the midpoint of this side are... The camera's coordinate system is not necessarily parallel to... Figure 1 The coordinate system in the middle.

[0101] The coordinates of the actual corner points 2 and 4 of the bent part, measured by the industrial camera, are P1 and P2 respectively. a2 (x a2 ,y a2 ), P a4 (x a4 ,y a4 The coordinates of the midpoint of this side are...

[0102] The error value Δy along the robot's Y-axis: The following formulas are all existing technologies:

[0103]

[0104] P t and P a The Euclidean distance d between them:

[0105]

[0106] The error value Δx along the robot's X-axis is:

[0107]

[0108] The positioning error directions along the X and Y axes are determined by comparing P. t and P a The relationship between the magnitudes of the horizontal and vertical coordinates can be obtained.

[0109] Furthermore, the corner coordinates are sub-pixel level, which can achieve higher accuracy. After processing the source image acquired by the machine vision error detection system, the camera identifies the sub-pixel corner coordinates of the bent part and calculates the required size and angle error information. The source image processing includes: first, image preprocessing to obtain a binary image of the bent part; second, based on the coarse localization of the Canny edge detection algorithm, a more accurate sub-pixel edge is obtained through the Zernike moment sub-pixel detection algorithm, and then further subdivided based on the Harris pixel-level corner detection to obtain the sub-pixel corner coordinates.

[0110] Step 22: Based on the mapping dataset obtained from sampling ( Figure 6 Using sample data from [source], an optimization algorithm was employed to train a neural network and establish the ESSA-Elman positioning error prediction model.

[0111] The ESSA-Elman positioning error prediction model is used to predict the positioning errors of the X and Y linear axes. The predicted X and Y linear axis positioning errors are used to correct the theoretical X and Y operating parameters of the robot, and the corrected theoretical X and Y operating parameters of the robot are input to the robot.

[0112] Step 22 specifically includes:

[0113] Using the Elman neural network as the base network, the optimal initial weights and thresholds of the network are found iteratively by utilizing the sparrow search algorithm. Then, the network is trained with these weights and thresholds as initial values, which can achieve better prediction results.

[0114] Meanwhile, to address the issue of poor sparrow population diversity caused by the random generation of initial sparrow population locations in the sparrow search algorithm, the initial sparrow population locations are generated using the uniformity and uncertainty of the Tent chaotic mapping function. This preserves population diversity while enhancing the global search capability of the sparrow search algorithm, thus establishing the SSA-Elman positioning error prediction model ESSA-Elman optimized by the Tent chaotic mapping.

[0115] Step 3: Build the machine vision error detection system WJC, which specifically includes:

[0116] Industrial camera one is positioned on the bending machine, facing the front of the bent part to capture the rotation angle of the bent part; industrial camera two is positioned on the side of the bending machine, facing the side of the bent part to capture the forming angle in real time; the camera control device outputs the rotation angle error and forming angle error to the display module.

[0117] The principle by which the camera control device obtains the rotation angle error is as follows:

[0118] Another cause of dimensional errors is that the bending edge of the bent part is not parallel to the edge of the bending machine die, meaning there is a rotation angle error on the robot's end effector axis. Corner points 2 and 4 of the rectangular working area are parallel to the edge of the bending machine die. The machine vision system can calculate the required rotation angle compensation by detecting the coordinates of the current bent part relative to corner points 2 and 4 of the rectangular working area. The rectangular working area is always fixed and serves as an extension of the bending machine's edge; its four corner coordinates are fixed and can be measured in advance and stored in the program for later use. The calculation principle for the rotation angle error is as follows... Figure 5 As shown.

[0119] Let the coordinates of the current corner points 2 and 4 of the bent part be P and P respectively. a2 (x a2 ,y a2 ), P a4 (x a4 ,y a4 The coordinates of the two corner points corresponding to the edge of the rectangular working area are P1 and P2 respectively. t2(x t2 ,y t2 ), P t4 (x t4 ,y t4 According to P t2 and P t4 The coordinates can be used to obtain the slope k of the edge of the rectangular working area. t4 -y t2 ) / (x t4 -x t2 ).

[0120] x a2 <x a4 It is necessary to rotate clockwise to compensate for the error and obtain the P value. a4 After obtaining the equation of the line with slope k, the following model can be used to obtain the equation of line P. a2 The coordinate P of the foot of the perpendicular from the equation of the line. a0 (x a0 ,y a0 The following formulas are all existing technologies:

[0121]

[0122] With P a4 Let P be a vertex. a2 P a4 P a0 The error value Δθ of the angle between the three points can be obtained by the following formula, where A, B, and C are intermediate variables in the calculation. The following formulas are all existing techniques:

[0123]

[0124] x a2 >x a4 The calculation method for the rotation angle error is the same as that for clockwise rotation.

[0125] The PC feeds back the rotation angle error to the robot's end-effector rotation axis through the control module and robot control device, and feeds back the forming angle error to the upper die of the bending machine through the control module and bending machine control device for corresponding compensation.

[0126] The control module stores the ESSA-Elman positioning error prediction model.

[0127] Step 4: Compensate for dimensional errors using a combination of neural network feedforward and machine vision real-time feedback, and compensate for bending angle errors using machine vision real-time feedback. This includes the following steps:

[0128] The ESSA-Elman positioning error prediction model outputs positioning error values ​​based on the robot's theoretical X and Y operating parameters, compensates for the robot's theoretical X and Y operating parameters, and finally outputs the compensated robot's theoretical X and Y operating parameters to the robot.

[0129] WJC measures the rotation angle in real time and outputs the rotation angle error to the robot.

[0130] WJC measures the bending angle in real time and outputs the bending angle error to the robot.

[0131] The process of compensating for the robot's theoretical X and Y operating parameters includes:

[0132] The actual positioning error E at each sampling location sample =(Δx) sample ,Δy sample );

[0133] Target location (similar) Figure 2 The theoretical operating parameters P of the sampling position in the sample (in the sample) target The predicted positioning error E is obtained by inputting it into the prediction model. t ′ arget ;

[0134] Finally, the theoretical operating parameters P for the target location. target The predicted positioning error E′ is superimposed in reverse. target The compensated target position operation parameter P can then be obtained. modified The information is input into the robot to complete the compensation.

[0135] The rotation angle compensation process includes:

[0136] The theoretical angle of rotation required in the bending process is θ. First, the robot is controlled to rotate the bent part once by the angle θ.

[0137] After rotation, an industrial camera located directly above the rectangular working area is used to collect the compensation angle Δθ and compensation direction in real time. The compensation direction is either clockwise or counterclockwise.

[0138] The robot rotates the bent part twice according to the collected compensation angle Δθ and compensation direction to compensate for the error.

[0139] The forming and bending angle compensation process includes:

[0140] During the bending process, the slider pressure Y1 is first calculated according to the theoretical bending angle θ1 and the slider pressure calculation model, and Y1 is used to control the bending machine to bend once.

[0141] After the upper die of the bending machine is unloaded and the bent part has fully springed back, the bending forming angle θ2 after springback is collected in real time by an industrial camera located on the bending side of the bending machine. Then the springback angle Δθ = θ2 - θ1 at this time.

[0142] The compensation angle θ3 = θ1 - Δθ is obtained from the springback angle Δθ and the theoretical bending angle θ1. Based on θ3 and the slider pressure calculation model Y (existing technology), the compensation slider pressure Y2 is calculated, and Y2 is used to control the bending machine for secondary bending compensation.

[0143]

[0144] V - Lower die opening width, mm;

[0145] R d -Lower die fillet radius, mm;

[0146] t - Thickness of the bent part when not under stress, in mm.

[0147] R - Bending radius, mm, derived from the empirical formula R = K v ·V is obtained;

[0148] K v The value is taken as 0.156 according to the German industrial standard DIN6935-2010;

[0149] η - The thinning coefficient due to deformation caused by bending force, from Please obtain.

[0150] Bending process error compensation experiment:

[0151] (1) Setting up a bending test platform: The bending machine is controlled by the bending machine control cabinet to press down the upper die of a small bending machine to achieve bending and forming. The robot control cabinet controls a five-degree-of-freedom bending auxiliary robot to grasp and feed the bending parts and move accordingly. The air pump inflates and deflates to pick up and put down the bending parts. The rectangular working area extends the edge of the bending machine and is parallel to the side of the bending machine. The coordinates of its four corner points (the origin is adopted) are set as follows. Figure 1 The origin point (in the experiment) is measured in advance and stored in the program (control module) for later use.

[0152] During processing, the robot's X and Y linear axis positioning errors are compensated by the ESSA-Elman error prediction and compensation model to control the bending dimensions. The PC displays the real-time detection status of two industrial cameras. Basler industrial camera 1 is positioned above the bent part. Basler industrial camera 1 and its control unit work together to output the rotation angle error: using the corner coordinates of the rectangular working area as a reference, the rotation angle error of the bent part is compared with the corner coordinates of the bent part (control module) to obtain the error, and then fed back to the robot's end effector for B-axis compensation. Basler industrial camera 2 is located on the sides of the upper and lower dies of the bending machine to collect the current bending forming angle and feeds it back to the upper die of the bending machine to compensate for the forming angle error.

[0153] When bending, the lower die of the small bending machine is fixed, while the upper die is driven by two coupled servo motors to press down and form the bending angle.

[0154] (2) Design the bending process flow:

[0155] To conduct actual bending experiments before and after compensation, bending group A, compensated using the compensation method of this invention, and uncompensated bending group B were set up, and the experiments were carried out according to the following bending process flow:

[0156] 1. Robot loading and single-sided bending:

[0157] Obtain the length L0 and bending angle θ0 of the first step.

[0158] The robot's end effector picks up and delivers the bent part from the robot's zero position (sampling start position) to a specific position in the working area of ​​the CNC bending machine to determine the length of the first step.

[0159] When bending group A is feeding material, the error is predicted and compensated by the ESSA-Elman positioning error prediction model, while bending group B is not compensated.

[0160] The first bending step is performed according to θ0. Bending group A undergoes secondary compensation through the bending forming angle feedback compensation method, while bending group B is not compensated.

[0161] 2. The bent part is turned around and bent again on one side:

[0162] The robot rotates the workpiece 180°. Bending group A compensates for the rotation angle error by collecting the error through machine vision, while bending group B does not compensate.

[0163] Obtain the dimensions L1 and bending angle θ1 for the second step.

[0164] Bending group A compensates for the dimensional length and bending angle of the second step in the same way as in step (1).

[0165] Bending group B was not compensated.

[0166] 3. Robot unloading: The robot picks up the bent finished product and leaves the working area of ​​the bending machine.

[0167] 4. Precision Measurement: The dimensional error of the bent product is measured by a precision vernier caliper, and the forming angle error of the bent product is measured by a precision universal angle gauge. The dimensional error and forming angle error of bending groups A and B are compared and analyzed.

[0168] (3) Bending process compensation experiment:

[0169] An actual bending process experiment was conducted using a 304 stainless steel rectangular bending part with dimensions of 400mm×100mm×1mm.

[0170] Five bending pieces were each placed in bending group A and bending group B, and the bending was carried out in sequence according to the process flow.

[0171] Bending group A and bending group B are set with the same bending parameters: the first step dimension length L0 = 10mm and the bending angle θ0 = 120°, and the second step dimension length L1 = 10mm and the bending angle θ1 = 90°.

[0172] 1. Make the first bend

[0173] The sampling start position is the robot's zero position. The robot picks up the bent part from the zero position, and the robot's state at this time is as follows: Figure 12 (a).

[0174] To determine the first step dimension length (i.e., the length to be bent) L0 = 10mm, the theoretical operating parameters (-378, 315) were input into the ESSA-Elman model to obtain the corresponding positioning error, and the compensated coordinates (-377.542, 314.282) were input into the robot. The theoretical operating parameters, obtained from previous experiments, refer to the robot's position after moving from zero position to the bending machine and feeding the bent part into the first step dimension of 10mm.

[0175] The dimensions for the first step are determined, and the robot begins operation. At this point, the robot's state is as follows: Figure 12 As shown in (b).

[0176] Substituting θ0 = 120° into the slider compression calculation model Y, the slider compression amount for one bend is calculated to be Y1 = 1.812 mm. The bending machine bends once with Y1 = 1.812 mm, and the robot performs corresponding follow-up movements. At this time, the robot state is as follows. Figure 12 As shown in (c).

[0177] through Figure 11 The Basler industrial camera 2 measured the springback forming angle θ2 = 121.362°. The original image and edge image are shown below. Figure 12(d) and Figure 12 As shown in (e), Figure 4 and Figure 5 The sub-pixel corner coordinates and forming angle errors identified by the measurement principle are shown in Table 1. At this time, the springback angle Δθ=θ2-θ0=1.362°, the compensation angle θ3=θ0-Δθ=118.638°, and the secondary bending slider pressure Y2=1.849mm is obtained from the θ3 slider pressure calculation model Y. The secondary bending is compensated by Y2=1.849mm.

[0178] Table 1 Measurement results of the first curve angle

[0179]

[0180]

[0181] Among them, the bottom point refers to Figure 12 (e) The points on the straight line where the crease is located, the two side points refer to the two points connecting the bottom point, and the straight line where the side points are located forms the end of the bent piece.

[0182] 2. Turn the bent part around and compensate for the rotation angle.

[0183] The robot picks up the bent part and moves back to the rectangular working area. Since the rectangular working area is fixed, its four sides are parallel to the four sides of the bending machine and extend from the machine's edges. The coordinates of the four corner points of the rectangular working area are pre-measured and stored in the program for later use. The robot picks up the bent part and moves its Y-axis back to the camera's view. Its A-axis rotates to make the bent part face upwards for easier camera measurement. The robot's end effector rotates along its B-axis. At this point, the robot's state is as follows: Figure 13 As shown in (a). Figure 11 The original image and edge image obtained by measuring with a Basler industrial camera 1 are as follows: Figure 13 (b) and Figure 13 As shown in (c).

[0184] The sub-pixel corner coordinates and compensation angles obtained from the detection are shown in Table 2. At this time, the robot's B-axis is controlled to rotate counterclockwise to compensate for the rotation angle error.

[0185] Table 2 Measurement results of rotation angle error

[0186]

[0187] The coordinates of the upper right point refer to Figure 1 The coordinates of corner point 2; the coordinates of the bottom right point refer to Figure 1 The coordinates of corner point 4 in the diagram.

[0188] 3. Make the second bend

[0189] The second step dimension L1 is set to 10mm. The robot's theoretical operating parameters (-385, 359) are input into the ESSA-Elman model to obtain the corresponding positioning error. After compensation, the coordinates (-383.933, 357.560) are input into the robot. The second step dimension is then determined, and the robot's state is as follows. Figure 14 As shown in (a), substituting θ1 = 90° into the slider compression calculation model Y, the slider compression amount for one bending operation is calculated. The bending machine performs one bending operation, and the robot follows accordingly. The robot's state at this time is as follows: Figure 14 As shown in (b). Figure 11 The Basler industrial camera 2 measured the springback forming angle as θ2 = 91.595°. The original image and edge image are shown below. Figure 14 (c) and Figure 14 As shown in (d). Similarly, the theoretical operating parameters were obtained from previous experiments.

[0190] by Figure 4 and Figure 5 The sub-pixel corner coordinates and angle errors identified by the measurement principle are shown in Table 3. At this time, the rebound angle Δθ=θ2-θ1=1.595° and the compensation angle θ3=θ1-Δθ=88.405°. Substituting into the slider pressure calculation model Y, the secondary bending slider pressure Y2=2.717mm is obtained, and the secondary bending with Y2=2.717mm is used for compensation.

[0191] Table 3. Measurement results of the second curve angle.

[0192]

[0193] Among them, the bottom point refers to Figure 14 In Figure (d), the points on the straight line where the crease is located, the two side points refer to the two points connecting the bottom point, and the straight line where the side points are located forms the end of the bent part.

[0194] (4) Robot unloading

[0195] The robot picks up the bent finished product, removes it from the bending work area, and places it in the finished product area.

[0196] (5) Precision measurement

[0197] The dimensions of each step in the bending process of the finished product are measured using precision vernier calipers, with a measurement accuracy of 0.02 mm. Figure 15 As shown in (a), a precision universal angle gauge is used to measure the forming angle error of the bent product. The measurement accuracy is as follows: Figure 15 As shown in (b). The bent product is as follows. Figure 15 As shown in (c), the dimensions, angles, and errors of each step are shown in Table 4.

[0198] Table 4. Error data for a certain bent product in bending group A.

[0199]

[0200] The above describes the processing procedure for a single compensated bent part. All bending groups A were compensated according to the above process, while bending groups B were not compensated.

[0201] Among them, such as Figures 15-16 As shown, taking size 1 as an example, the two sides connected to the crease (and) Figure 16 (The bending edge connection in the middle), where the actual length of both sides is uniformly represented by dimension 1. The difference between the length of one side and the standard step dimension L0 is ΔL1 / mm, and the difference between the length of the other side and the standard step dimension L0 is ΔL2 / mm.

[0202] The above is caused by the existence of rotation angle error, which makes the crease tilted relative to the two sides. Figure 16 In this context, a vernier caliper measures dimension 1, which is the length of the line segment from the bend to the crease.

[0203] Table 5. Experimental results of bending group A after compensation.

[0204]

[0205]

[0206] Table 6. Experimental Results of Experimental Group B Before Compensation

[0207]

[0208] The experimental results show that after compensation, the average values ​​of dimensional errors ΔL1 and ΔL2 of the five bent products are 0.17 mm and 0.22 mm, respectively, and the average values ​​of dimensional errors ΔL1 and ΔL2 are 0.27 mm and 0.30 mm, respectively. Compared with the uncompensated values ​​of 1.23 mm and 1.24 mm for dimensional error ΔL1, and 1.33 mm and 1.63 mm for dimensional error ΔL2, the compensated values ​​reduce dimensional errors ΔL1 and ΔL2 by 86.2% and 82.3%, respectively, and 79.7% and 81.6%, respectively. Furthermore, the compensated dimensional errors meet the national standard requirement of ±0.3 mm for bending processing as described in Section 2.2.

[0209] After compensation, the average angle 1 error of the five bent products was 0.20°, and the average angle 2 error was 0.23°. Compared with the average angle 1 error of 0.73° and the average angle 2 error of 0.72° before compensation, the angle 1 error was reduced by 72.6% and the angle 2 error was reduced by 68.1% after compensation. Moreover, the angle error after compensation meets the national standard requirement of ±0.75° for bending processing described in Section 2.2.

[0210] In summary, the combined compensation method of ESSA-Elman model feedforward compensation and machine vision feedback compensation proposed in this invention improves the accuracy of the finished bent parts and enhances the overall accuracy of the bending process.

[0211] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. A method for detecting and compensating the accuracy of sheet metal bending processes based on machine vision, characterized in that, Includes the following steps: Step 1: Build a machine vision XYC linear axis positioning error detection system to obtain the corner coordinates of the bent part and output the X and Y linear axis positioning error of the bent part at the current sampling position. This includes the following steps: An industrial camera is positioned above the rectangular work area and connected to a PC; a robot is placed on one side of the rectangular work area to pick up the bent parts. The long side of the rectangular working area is parallel to the Y-axis of the robot coordinate system, and the wide side is parallel to the X-axis of the robot coordinate system. The rectangular working area is divided into several rectangular sampling positions at fixed intervals. The length and width of each sampling position are parallel to the length and width of the rectangular working area, respectively; Step 2: Establish the ESSA-Elman positioning error prediction model for predicting the positioning errors of the X and Y linear axes, which includes the following steps: Step 21: Sampling is performed based on XYC. The X and Y linear axis positioning errors are output according to the identified corner coordinates of the bent parts. To address the positioning errors of the robot's X and Y linear axes, the machine vision error detection system collected the X and Y axis positioning errors at each sampling position through multiple positioning error sampling experiments of the robot. The X and Y axis positioning errors at each sampling location, along with the corresponding theoretical X and Y operating parameters of the robot, form a mapping dataset; where the theoretical X and Y operating parameters of the robot are the theoretical operating parameters of the robot from the starting position to each sampling location. ; Step 22: Based on the mapping dataset obtained from sampling, an optimization algorithm is used to optimize the neural network training and establish the ESSA-Elman positioning error prediction model: The ESSA-Elman positioning error prediction model is the SSA-Elman model optimized by the Tent chaotic mapping. The ESSA-Elman positioning error prediction model is used to predict the positioning errors of the X and Y linear axes. The predicted X and Y linear axis positioning errors are used to correct the theoretical X and Y operating parameters of the robot. The corrected theoretical X and Y operating parameters of the robot are then input into the robot. Step 3: Build the machine vision error detection system WJC, which specifically includes: Industrial camera one is positioned on the bending machine, facing the front of the bent part to capture the rotation angle of the bent part; industrial camera two is positioned on the side of the bending machine, facing the side of the bent part to capture the forming angle in real time; the camera control device outputs the rotation angle error and forming angle error to the display module. The PC feeds back the rotation angle error to the robot's end-rotor axis through the control module and robot control device, and feeds back the forming angle error to the upper die of the bending machine through the control module and bending machine control device for corresponding compensation. The control module stores the ESSA-Elman positioning error prediction model; Step 4: Compensate for dimensional errors using a combination of neural network feedforward and machine vision real-time feedback, and compensate for bending angle errors using machine vision real-time feedback. This includes the following steps: The ESSA-Elman positioning error prediction model outputs positioning error values ​​based on the robot's theoretical X and Y operating parameters, compensates for the robot's theoretical X and Y operating parameters, and finally outputs the compensated robot's theoretical X and Y operating parameters to the robot. WJC measures the rotation angle in real time and outputs the rotation angle error to the robot. WJC measures the bending angle in real time and outputs the bending angle error to the upper die of the bending machine to compensate for the forming angle error.

2. The method for detecting and compensating sheet metal bending accuracy based on machine vision according to claim 1, characterized in that, Step 21 specifically includes: At a fixed distance To maintain a fixed sampling step size, several sampling positions are defined; The robot performs zigzag sampling along the Y-axis on the rectangular work area plane, starting from the initial position. Begin picking up the bent parts, according to their position. After sampling in sequence, then from with Distance is Location Begin, along the position Sample the second row until the final position is reached. ; Each sampling is performed at the initial position of the bent part. Starting from the desired position, the robot's movement is controlled by manipulating the ideal X and Y parameters to reach each sampling position. The bent parts arrive at each sampling point with the same posture as the starting position. Then, the X and Y axis positioning errors of the bent parts at the current sampling position are collected and output by an industrial camera and a PC. The output positioning error is consistent with the unit of the robot's X and Y operation parameters, which is millimeters.

3. The method for detecting and compensating sheet metal bending accuracy based on machine vision according to claim 2, characterized in that, The method for machine vision to collect X and Y linear axis positioning errors in step 21 is as follows: The image coordinate system is opposite to the coordinate system used for sampling in step 1, which causes the error value of the robot's Y-axis to be... The error value along the robot's X-axis corresponding to the Y-axis in step 1. Corresponding to the X-axis in step 1; Let the coordinates of corner points 2 and 4 of the bent part at a certain sampling location be respectively... , The coordinates of the midpoint of this side are ; The coordinates of the actual corner points 2 and 4 of the bent part, measured by the industrial camera, are as follows: , The coordinates of the midpoint of this side are ; Error value of robot Y-axis : ； and Euclidean distance between ; ； Then the error value of the robot's X-axis : ； The positioning error directions along the X and Y axes are determined by comparison. and The relationship between the magnitudes of the horizontal and vertical coordinates can be obtained.

4. The method for detecting and compensating sheet metal bending accuracy based on machine vision according to claim 1, characterized in that, Step 22 specifically includes: Using the Elman neural network as the base network, the optimal initial weights and thresholds of the network are found iteratively by utilizing the sparrow search algorithm. Then, the network is trained with these weights and thresholds as initial values, which can achieve better prediction results. Meanwhile, to address the issue of poor sparrow population diversity caused by the random generation of initial sparrow population locations in the sparrow search algorithm, the initial sparrow population locations are generated using the uniformity and uncertainty of the Tent chaotic mapping function. This preserves population diversity while enhancing the global search capability of the sparrow search algorithm, thus establishing the SSA-Elman positioning error prediction model ESSA-Elman optimized by the Tent chaotic mapping.

5. The method for detecting and compensating sheet metal bending accuracy based on machine vision according to claim 1, characterized in that, Step 4, the process of compensating for the robot's theoretical X and Y operating parameters, includes: Actual positioning error at each sampling location ; The theoretical operating parameters of the target location Inputting the data into the prediction model yields the predicted positioning error value. ; Finally, the theoretical operating parameters for the target location. The predicted positioning error values ​​are superimposed in reverse. The compensated target position operation parameters can then be obtained. The information is input into the robot to complete the compensation.

6. The method for detecting and compensating sheet metal bending accuracy based on machine vision according to claim 1, characterized in that, Step 4, the rotation angle compensation process includes: The theoretical angle of rotation required during bending is: First, by angle Control the robot to rotate the bent part once; After rotation, an industrial camera positioned directly above the rectangular working area is used to collect compensation angle data in real time. And the compensation direction, which is clockwise or counterclockwise; The robot compensates for the collected angles. The bending part is rotated twice in the direction of compensation to compensate for the error.

7. The method for detecting and compensating sheet metal bending accuracy based on machine vision according to claim 1, characterized in that, The forming and bending angle compensation process includes: During bending, first bend according to the theoretical bending angle. Calculate the slider downward pressure using the slider downward pressure calculation model. ,by Control the bending machine to perform one bending operation; After the upper die of the bending machine is unloaded and the bent part has fully springed back, an industrial camera located on the bending side of the bending machine is used to collect the bending forming angle in real time after the springback. Then the rebound angle at this time ; From the rebound angle and theoretical bending angle Determine the compensation angle ,based on The compensation slider downward pressure is calculated using the slider downward pressure calculation model Y. and with Controlling compensation for secondary bending of the bending machine; ； - Lower mold opening width, mm; -Lower die fillet radius, mm; -Thickness of the bent part when not under stress, in mm - Bending radius, mm, determined by empirical formula Request; The value is taken as 0.156 according to the German industrial standard DIN6935-2010; -The thinning coefficient due to deformation caused by the force applied to the bent part, by Please obtain.

8. The method for detecting and compensating sheet metal bending accuracy based on machine vision according to claim 1, characterized in that, In step 21, the corner coordinates are sub-pixel level corner coordinates.

Images