A vision-based method for repairing vehicle sheet metal

Abstract

This invention provides a vision-based method for repairing vehicle sheet metal. The method includes: S1, constructing a 3D digital reference model of the sheet metal to be repaired, and obtaining the geometric dimensions, paint properties, and substrate mechanical parameters of intact sheet metal of the same model; S2, scanning the sheet metal to be repaired using a multimodal vision acquisition device, comparing the scanned data with the 3D digital reference model, and locating the coordinates of the area to be repaired and the type of defect; S3, starting a repair robot for targeted repair, collecting current repair data in real time during the repair process as visual feedback and core input for the reinforcement learning reward function, and dynamically adjusting the repair parameters to meet preset standards. This invention uses reinforcement learning to continuously align the repair operation with the 3D digital reference model, improving the consistency of repair quality; relying on precise scanning and positioning and real-time repair data feedback, it effectively avoids problems such as uneven paint thickness and can flexibly adapt to repair needs of different damage levels.

Description

Technical Field

[0001] This invention relates to the field of vehicle sheet metal repair technology, and specifically to a vehicle sheet metal repair method based on visual perception. Background Technology

[0002] During daily use, vehicle body panels are prone to damage from collisions, scratches, and other impacts. This damage not only affects the vehicle's appearance but can also lead to corrosion of the substrate due to paint damage, reducing the durability of the body structure. Currently, vehicle body panel damage repair mainly relies on manual operation or simple automated equipment, which has several technical limitations:

[0003] Traditional manual repair methods rely heavily on the technician's experience and judgment, resulting in low accuracy in defect location. They require visual estimation of dent depth, scratch range, and exposed substrate, which can easily lead to insufficient or excessive repair. For example, improper control of pulling force can cause the sheet metal to stretch and thin, uneven paint film thickness during touch-up can cause sagging and peeling, and paint mixing relies on experience, resulting in significant color differences and difficulty in matching the original factory paint finish.

[0004] While existing automated repair equipment has improved the accuracy of geometric repair to some extent—for example, Chinese patent document CN119076695A discloses a dent repair device for automobile repair—it still requires manual intervention for adjustment, cannot achieve fully automated closed-loop repair, and cannot guarantee that the body sheet metal can be repaired. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to propose a vehicle sheet metal repair method based on visual perception to solve the problems mentioned in the background section above.

[0006] This invention is achieved through the following technical solution:

[0007] S1. Construct a 3D digital baseline model of the sheet metal of the vehicle to be repaired: obtain the geometric dimensions, paint properties, and mechanical parameters of the substrate of the intact sheet metal of the same model vehicle;

[0008] S2. Use a multimodal vision acquisition device to perform a full-area scan of the vehicle sheet metal to be repaired, compare the scan data with the 3D digital reference model constructed in step S1 in multiple dimensions, and output the coordinates of the defect area to be repaired and the defect type by fusing defect detection algorithms.

[0009] S3. Start the repair robot, obtain the defect area and defect type according to step S2, and repair the defect area in a targeted manner through the repair robot; during the repair process, collect the current repair data of the defect area in real time through the multimodal vision acquisition device, use the current repair data as the real-time visual feedback result, and use it as the core input of the reward function of reinforcement learning;

[0010] The reward function includes a positive reward term and a constraint penalty term:

[0011] The core evaluation indicators for positive reward items are: the geometric conformity and paint property matching degree between the defective area and the standard area corresponding to the 3D digital benchmark model, and the smoothness of the repaired surface of the defective area;

[0012] The constraint penalty items target the operation angle deviation and operation force deviation of the repair robot. The gradient penalty coefficient is configured according to the degree of deviation. When the deviation exceeds the standard, the corresponding penalty is triggered.

[0013] Based on the output of the reward function, the repair parameters of the repair robot are dynamically adjusted until the repair morphology, paint properties, and substrate mechanical properties of the defective area all meet the preset standards of the 3D digital benchmark model.

[0014] Furthermore, in step S3, the mathematical formula for the reward function is as follows:

[0015] ;

[0016] In the formula, Represents a reward for geometric shape fit; The weighting coefficient representing the geometric fit bonus; Indicates the reward for matching paint attributes; This represents the weighting coefficient for the paint attribute matching degree reward; This indicates a reward for improving the smoothness of the repaired surface; This indicates the weighting coefficient for the surface smoothness repair bonus; Indicates a penalty for deviation in the working angle; The weighting coefficient representing the penalty for deviation in the working angle; This indicates a penalty for deviation in work effort; This represents the weighting coefficient for penalties imposed on deviations in work effort.

[0017] Furthermore, in step S3, the formula is further refined as follows:

[0018] ;

[0019] In the formula, The actual 3D point cloud data of the defect area; This represents the three-dimensional point cloud data corresponding to a standard region in a 3D digital model; This represents the root mean square deviation, used to calculate the average Euclidean distance between corresponding points in two point cloud sets; The reward coefficient, representing the degree of form fit, is used to adjust the sensitivity of the reward to changes in deviation. This represents an exponential function used to ensure reward values. Always within the range (0, 1);

[0020] ;

[0021] In the formula, Indicates the difference in paint thickness; The weighting of the paint thickness difference; Indicates the difference in gloss; The weights representing the differences in gloss; Indicates the paint sensitivity coefficient;

[0022] ;

[0023] In the formula, This represents the arithmetic mean roughness. This represents the flatness sensitivity coefficient;

[0024] ;

[0025] In the formula, This indicates the actual operating angle of the repair robot; Indicates the allowable angle deviation threshold; Indicates the critical value of angular deviation; , Represents the gradient penalty coefficient, corresponding to the penalty slope for slight and severe overshoot, respectively; This indicates an additional penalty value for exceeding the deviation limit;

[0026] ;

[0027] In the formula, This indicates the actual working force of the repair robot; Indicates the target workload; Indicates the allowable force deviation threshold; This indicates the critical value of the force deviation; , Represents the gradient penalty coefficient, corresponding to the penalty slope for slight and severe overshoot, respectively; This indicates an additional penalty value for exceeding the deviation limit.

[0028] Furthermore, the multimodal vision acquisition device includes a color 3D scanner, an eddy current thickness gauge, and a near-infrared camera.

[0029] Furthermore, step S1 also includes:

[0030] The 3D digital reference model is marked to obtain the regional coordinates of the sheet metal parts, including the doors, hood, fenders, roof, and trunk lid. The edge contour lines of each sheet metal part are extracted, and the area within the edge contour lines constitutes the scanning area.

[0031] Furthermore, step S2 also includes:

[0032] S21. Only the eddy current thickness gauge is activated, and the scanning area is scanned using a low-density scanning mode by the eddy current thickness gauge;

[0033] S22. In real time, compare the collected paint thickness data with the standard paint thickness of the corresponding area in the 3D digital benchmark model constructed in step S1. If any of the following conditions are met, it is determined to be a suspected impact area:

[0034] Firstly, the absolute value of the deviation between the paint thickness at a certain monitoring point and the standard thickness is greater than the first set threshold.

[0035] Secondly, the thickness fluctuation at two or more of the three consecutive detection points exceeds the second set threshold.

[0036] S23. When scanning to a suspected collision area, the multimodal vision acquisition device switches to high-density scanning mode and simultaneously starts the color 3D scanner, eddy current thickness gauge and near-infrared camera to collect multi-dimensional data.

[0037] S24. Perform fusion analysis on the collected multi-dimensional data and output the defect type;

[0038] S25. The repair robot automatically matches the corresponding repair tools and repair operation logic according to the determined defect type.

[0039] Furthermore, in step S24, the defect types include dent defects and scratch defects.

[0040] Furthermore, the repair operation logic includes a pulling stage, a painting stage, and a polishing stage, and each stage is assigned a different weighting coefficient, including: a weighting coefficient for geometric conformity bonus. Weighting coefficient of paint surface attribute matching degree reward Surface smoothness repair bonus weighting coefficient Weighting coefficient for working angle deviation penalty Weighting coefficient for work intensity deviation penalty .

[0041] The beneficial effects of this invention are as follows: This invention uses reinforcement learning to make the repair operation continuously approach the benchmark model, which greatly improves the consistency of repair quality; relying on precise scanning and positioning and real-time repair data feedback, it effectively avoids problems such as uneven paint thickness and can flexibly adapt to the repair needs of different damage levels. Attached Figure Description

[0042] Figure 1 This is a flowchart of a vehicle sheet metal repair method based on visual perception according to the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the description of these embodiments is for the purpose of aiding understanding the present invention, but does not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0044] Reference Figure 1 As shown, a method for repairing vehicle sheet metal based on visual perception includes the following steps:

[0045] S1. Construct a 3D digital baseline model of the sheet metal of the vehicle to be repaired: obtain the geometric dimensions, paint properties, and mechanical parameters of the substrate of the intact sheet metal of the same model vehicle;

[0046] S2. Use a multimodal vision acquisition device to perform a full-area scan of the vehicle sheet metal to be repaired, compare the scan data with the 3D digital reference model constructed in step S1 in multiple dimensions, and output the coordinates of the defect area to be repaired and the defect type by fusing defect detection algorithms.

[0047] S3. Start the repair robot, obtain the defect area and defect type according to step S2, and repair the defect area in a targeted manner through the repair robot; during the repair process, collect the current repair data of the defect area in real time through the multimodal vision acquisition device, use the current repair data as the real-time visual feedback result, and use it as the core input of the reward function of reinforcement learning;

[0048] The reward function includes a positive reward term and a constraint penalty term:

[0049] The core evaluation indicators for positive reward items are: the geometric conformity and paint property matching degree between the defective area and the standard area corresponding to the 3D digital benchmark model, and the smoothness of the repaired surface of the defective area;

[0050] The constraint penalty items target the operation angle deviation and operation force deviation of the repair robot. The gradient penalty coefficient is configured according to the degree of deviation. When the deviation exceeds the standard, the corresponding penalty is triggered.

[0051] Based on the output of the reward function, the repair parameters of the repair robot are dynamically adjusted until the repair morphology, paint properties, and substrate mechanical properties of the defective area all meet the preset standards of the 3D digital benchmark model.

[0052] The core function of constructing the 3D digital benchmark model in step S1 is to provide a unified standard template for vehicle sheet metal repair. All subsequent scans, defect location, automated repair, and quality judgment are based on this model. In this process, the positions of headlights, windows, etc. need to be marked so that they can be avoided in subsequent scans.

[0053] In a preferred embodiment, step S1 can be performed by constructing the 3D digital reference model using the following formula:

[0054] S11. The geometric shape of the 3D digital reference model is characterized by a combination of point cloud data and parametric surfaces. The mathematical expression is as follows:

[0055] ;

[0056] In the formula, This represents the point cloud dataset of the 3D digital model. Indicates the first The three-dimensional coordinates of the measurement points Indicates the first The coordinates of each measurement point in a spatial rectangular coordinate system Indicates the total number of measurement points;

[0057] S12. Parametric fitting is performed using NURBS surfaces. The fitting formula is as follows:

[0058] ;

[0059] In the formula, Represent the equation of the NURBS surface. , For surface parameters; express direction B-spline basis functions; express direction B-spline basis functions; Indicates the first Weight factors for each control point; Indicates the first The three-dimensional coordinates of each control point; express The direction index can take values ​​from 0, 1, 2, ... ; express The direction index can take values ​​from 0, 1, 2, ... ;

[0060] S13. Mark the 3D digital reference model and obtain the regional coordinates of the sheet metal parts, including the doors, hood, fenders, roof, and trunk lid. Extract the edge contour lines of each sheet metal part, and the area within the edge contour lines constitutes the scanning area.

[0061] By planning the scanning area in step S13, all potential areas to be repaired are fully covered, while invalid scanning objects are excluded, reducing the idle travel of the equipment. After the scanning area is clear, the multimodal vision acquisition equipment can follow the planned optimal path, reducing the redundant storage and preprocessing workload of non-target data, and avoiding invalid actions of the repair robot in non-metallic areas.

[0062] The multimodal visual acquisition equipment includes a color 3D scanner, an eddy current thickness gauge, and a near-infrared camera.

[0063] Among them, the color 3D scanner can collect 3D point cloud data of the metal surface of the vehicle to be repaired, accurately record the surface geometry, including micro-geometric parameters such as the depth of the dent, the length and width of the scratch, and the height of the protrusion, and at the same time capture the color information of the metal paint surface.

[0064] Eddy current thickness gauges are based on the principle of electromagnetic induction and non-contactly collect the thickness of paint layers (including electrophoretic layer, intermediate layer, color paint layer, and clear coat layer) and substrate (mainly cold-rolled steel plate) of the sheet metal of the vehicle to be repaired. They determine whether the paint layer is exposed, whether each layer is intact, and whether the substrate has been stretched and thinned due to impact.

[0065] By utilizing the penetrating power of near-infrared cameras, the thickness of the wet film on the paint surface and the drying speed can be monitored in real time during the paint repair stage to determine whether there is a risk of sagging.

[0066] Step S2 also includes:

[0067] S21. Only the eddy current thickness gauge is activated, and the scanning area is scanned using a low-density scanning mode by the eddy current thickness gauge;

[0068] S22. In real time, compare the collected paint thickness data with the standard paint thickness of the corresponding area in the 3D digital benchmark model constructed in step S1. If any of the following conditions are met, it is determined to be a suspected impact area:

[0069] Firstly, if the absolute value of the deviation between the paint thickness at a certain monitoring point and the standard thickness is greater than the first set threshold, for example, if the absolute value of the deviation between the paint thickness at a certain monitoring point and the standard thickness is greater than 20%, it is determined to be a suspected collision area. That is, assuming that the standard thickness of the paint layer in a certain area of ​​the 3D digital reference model is 25μm, if the paint thickness at the monitoring point is less than 20μm or greater than 30μm, it is determined to be a suspected collision area.

[0070] Secondly, if the thickness fluctuation of two or more monitoring points out of three consecutive detection points exceeds the second set threshold, for example, if the thickness fluctuation of two or more monitoring points out of three consecutive detection points exceeds 15μm, it is determined to be a suspected impact area. If the thicknesses of the three consecutive detection points are 25μm, 12μm, and 30μm respectively, and the 12μm and 30μm fluctuate by 18μm, it is determined to be a suspected impact area.

[0071] S23. When scanning to a suspected collision area, the multimodal vision acquisition device switches to high-density scanning mode and simultaneously starts the color 3D scanner, eddy current thickness gauge and near-infrared camera to collect multi-dimensional data.

[0072] S24. Perform fusion analysis on the collected multi-dimensional data and output the defect type;

[0073] S25. The repair robot automatically matches the corresponding repair tools and repair operation logic according to the determined defect type.

[0074] Specifically, in step S24, the defect types include dent defects and scratch defects.

[0075] Specifically, scanning is performed based on the scanning area output in step S13. Initially, only an eddy current thickness gauge is used for scanning, which can quickly complete the preliminary screening of the paint thickness of the entire vehicle's sheet metal area and rapidly identify abnormal areas that require special attention. As a preferred embodiment, the eddy current thickness gauge performs detection at a density of 1 to 2 detection points per square centimeter, and the scanning speed is set to 10-15 mm / s for low-density scanning, thereby quickly identifying suspected collision areas.

[0076] For suspected impact areas, switch to high-density scanning mode, increase the density of measurement points to 5 to 8 per square centimeter, and reduce the scanning speed to less than 10 mm / s; the eddy current thickness gauge maintains a thickness measurement accuracy of ±5 μm, and the near-infrared camera uses a wavelength of 900nm-1700nm, with an image resolution of 1280×960 pixels and a frame rate of 15fps.

[0077] Geometric data collected by a color 3D scanner, eddy current thickness gauge, and near-infrared camera are fused and analyzed. Based on preset judgment rules, the defect type is clearly identified, as follows:

[0078] Depression-type defects: In the geometric data collected by the color 3D scanner, there are surface depressions with a depth > 0.1mm; the eddy current thickness gauge shows that there is no exposed substrate on the paint surface (the thickness deviation of each layer is ≤ 15%), and the thickness of the substrate deviates from the reference model by ≤ 10%; the near-infrared camera did not detect microcracks or substrate peeling under the paint surface, and it was determined to be a depression-type defect, and the three-dimensional parameters such as depression depth and diameter were output simultaneously.

[0079] Scratch-related defects: Linear grooves (length > 5mm, width < 4mm) are present in the geometric data collected by the color 3D scanner; combined with eddy current thickness gauge data, the defects are further subdivided according to the paint damage level: ① Clear coat scratches: only the clear coat thickness deviation is > 20%, the intermediate coat and below are intact; ② Color coat scratches: the color coat thickness deviation is > 20%, the intermediate coat is intact; ③ Exposed substrate scratches: the electrophoretic layer or substrate is exposed, and the corresponding layer thickness detection value is close to 0; the near-infrared camera did not detect any hidden cracks in the direction of the groove extension, and it was determined to be a scratch-related defect of the corresponding layer.

[0080] Furthermore, in step S3, the mathematical formula for the reward function is as follows:

[0081] ;

[0082] In the formula, Represents a reward for geometric shape fit; The weighting coefficient representing the geometric fit bonus; Indicates the reward for matching paint attributes; This represents the weighting coefficient for the paint attribute matching degree reward; This indicates a reward for improving the smoothness of the repaired surface; This indicates the weighting coefficient for the surface smoothness repair bonus; Indicates a penalty for deviation in the working angle; The weighting coefficient representing the penalty for deviation in the working angle; This indicates a penalty for deviation in work effort; This represents the weighting coefficient for penalties imposed on deviations in work effort.

[0083] Furthermore, the reward function is further refined as follows:

[0084] ;

[0085] In the formula, The actual 3D point cloud data of the defect area; This represents the three-dimensional point cloud data corresponding to a standard region in a 3D digital model; This represents the root mean square deviation, used to calculate the average Euclidean distance between corresponding points in two point cloud sets; The reward coefficient, representing the degree of form fit, is used to adjust the sensitivity of the reward to changes in deviation. This represents an exponential function used to ensure reward values. Always within the range (0, 1);

[0086]

[0087] In the formula, This represents the difference in paint thickness, calculated by the deviation between the actual paint thickness during repair and the standard paint thickness of the corresponding area in the 3D digital reference model. The weighting of the paint thickness difference; The gloss difference is calculated by the deviation between the gloss of the paint surface during repair and the gloss of the corresponding area in the 3D digital reference model. The weights representing the differences in gloss; Indicates the paint sensitivity coefficient;

[0088] ;

[0089] In the formula, This represents the arithmetic mean roughness. This represents the flatness sensitivity coefficient;

[0090] ;

[0091] In the formula, This indicates the actual operating angle of the repair robot; Indicates the allowable angle deviation threshold; Indicates the critical value of angular deviation; , Represents the gradient penalty coefficient, corresponding to the penalty slope for slight and severe overshoot, respectively; This indicates an additional penalty value for exceeding the deviation limit;

[0092] ;

[0093] In the formula, This indicates the actual working force of the repair robot; Indicates the target workload; Indicates the allowable force deviation threshold; This indicates the critical value of the force deviation; , Represents the gradient penalty coefficient, corresponding to the penalty slope for slight and severe overshoot, respectively; This indicates an additional penalty value for exceeding the deviation limit.

[0094] Once the multimodal vision acquisition device outputs a clear defect type, the repair robot automatically matches the corresponding repair tools and operational logic based on the determined defect type. Specifically:

[0095] 1. Geometric morphology quantification determination:

[0096] Initial point cloud data of the area to be repaired was acquired using a color 3D scanner. 3D point cloud data corresponding to the standard area in the 3D digital model Calculate the root mean square deviation A geometric anomaly is preliminarily identified when the following conditions are met:

[0097] (1) The maximum local depression depth (point cloud Z-axis deviation) is > 0.1 mm, and

[0098] If it is greater than 0.05mm, then it corresponds to Insufficient fit in shape;

[0099] (2) Integrity judgment of paint surface: The thickness of each layer of paint (electrophoretic layer, intermediate layer, color paint layer, clear coat layer) is measured by eddy current thickness gauge. Based on the standard thickness of the corresponding area, different preset parameters are used to determine whether it is a dent, a scratch in the clear coat, a scratch in the paint coat, or a scratch that exposes the substrate.

[0100] 2. Automatic tool matching:

[0101] (1) Defects such as dents: through and Based on the determined recess area, select the appropriate micro sheet metal drawing head specification; for example: if d≤20mm, a 5mm diameter drawing head is suitable; if 20mm<d≤50mm, a 5mm diameter drawing head is suitable, etc. After drawing, automatically switch to a 5000-grit fine polishing wheel for surface flatness optimization.

[0102] (2) Scratch-type defects: The defect area is polished by automatically switching to a 1000-grit fine polishing wheel to remove surface dust, etc. Then, the paint gun is used to repair the defects layer by layer in the order of electrophoretic layer → intermediate layer → color paint layer → clear varnish. The UV curing lamp is switched to cure the defects. Finally, the 5000-grit fine polishing wheel is automatically switched to polish the defect area.

[0103] 3. Improve robot operation parameter settings, taking dent defects as an example:

[0104] When it is a concave defect, the target operating intensity This is the pulling force:

[0105] like If it is less than 0.1mm, then The value range is 10-25N;

[0106] like If it is greater than 0.1mm, then The value range is 25-40N;

[0107] Allowable angle deviation threshold At this point, it needs to be perpendicular to the reference plane of the recessed area. Set to 5°.

[0108] During the repair process, the multimodal vision acquisition device acquires the actual 3D point cloud data of the repair area every 300ms. and calculate :

[0109] like This increases the pulling force of the repair robot, increases the pulling frequency, and accelerates the rebound of the dent.

[0110] like Then, the working angle of the repair robot is finely adjusted to reduce the pulling force and pulling frequency;

[0111] like If the pulling stops, the machine automatically switches to a 5000-grit fine polishing wheel and enters the surface flatness optimization stage.

[0112] Among them, the penalty for deviation in the working angle Presets can be made for different tools, for example:

[0113] ;

[0114] That is, if the angle deviation of the drawing head is greater than 5° during operation, a gradient penalty is triggered and automatic calibration is performed until... .

[0115] Similarly, penalties for deviations in work intensity Presets can be made for different tools, for example:

[0116] ;

[0117] That is, if the force deviation of the drawing head is greater than 5N when it is working, the force should be reduced immediately and paused for 0.5s. After calibration, it should be resumed to avoid damage to the substrate.

[0118] Preferably, the repair operation logic includes a pulling stage, a painting stage, and a polishing stage, with different weighting coefficients set for each stage, including a weighting coefficient for geometric conformity bonus. Weighting coefficient of paint surface attribute matching degree reward Surface smoothness repair bonus weighting coefficient Weighting coefficient for working angle deviation penalty Weighting coefficient for work intensity deviation penalty Specific examples are as follows:

[0119] I. The drawing stage is shown in the table below:

[0120]

[0121] II. The painting stage is shown in the table below:

[0122]

[0123] III. Polishing stages are shown in the table below:

[0124]

[0125] By setting the above weighting coefficients, we can achieve full-dimensional consistency between the repaired area and the original factory condition, fully meeting the user's core demand for the repaired area to be as close as possible to the original shape and paint color.

[0126] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0127] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" and "second" may explicitly or implicitly include one or more of that feature.

[0128] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for repairing vehicle sheet metal based on visual perception, characterized in that, The method comprises the following steps: S1, constructing a 3D digital reference model of the iron sheet of the vehicle to be repaired: obtaining the geometric size data, paint attribute parameters and substrate mechanical parameters of the intact iron sheet of the same model vehicle; S2, using a multi-modal visual acquisition device to perform global scanning on the iron sheet of the vehicle to be repaired, comparing the scanning data with the 3D digital reference model constructed in step S1 in multiple dimensions, and outputting the coordinates of the defect area to be repaired and the defect type through a defect detection algorithm; S3, starting the repair robot, obtaining the defect area and defect type according to step S2, and performing targeted repair on the defect area through the repair robot; in the repair process, the multi-modal visual acquisition device acquires current repair data of the defect area in real time, takes the current repair data as visual real-time feedback results, and takes the current repair data as the core input of the reward function of reinforcement learning; In step S3, the mathematical function formula of the reward function is as follows: ； In the formula, represents a geometry fit reward; represents a weight coefficient of the geometry fit reward; represents a paint attribute matching reward; represents a weight coefficient of the paint attribute matching reward; represents a repair surface flatness reward; represents a weight coefficient of the repair surface flatness reward; represents a work angle deviation penalty; represents a weight coefficient of the work angle deviation penalty; represents a work force deviation penalty; represents a weight coefficient of the work force deviation penalty; In step S3, the above parameters are calculated by the following formula: ； wherein actual three-dimensional point cloud data of the defect area; three-dimensional point cloud data representing a corresponding standard area in the 3D digital model; denotes the root mean square deviation, which is used to calculate the average value of the Euclidean distance of corresponding points between two point cloud sets; denotes the reward coefficient of the shape fit degree, which is used to adjust the sensitivity of the reward to the deviation; denotes the exponential function, which is used to ensure that the reward value is always in the range of (0, 1]. ； wherein represents a difference in paint thickness; represents a weight of a difference in paint thickness; represents a difference in gloss; represents a weight of a difference in gloss; represents a paint sensitivity coefficient; ； In the formula, represents the arithmetic average roughness; represents the flatness sensitivity coefficient; ； In the formula, represents the actual working angle of the repair robot; represents the allowable angle deviation threshold value; represents the angle deviation critical value; , represents the gradient penalty coefficient, and the penalty slope corresponding to slight over-standard and serious over-standard, respectively; represents the deviation over-standard additional penalty value; ； In the formula, represents the actual working strength of the repair robot; represents the target working strength; represents the allowable strength deviation threshold; represents the strength deviation critical value; , represents the gradient penalty coefficient, respectively corresponding to the penalty slope of slight over-standard and serious over-standard; represents the deviation over-standard additional penalty value.

2. The method of claim 1, wherein: The reward function includes a positive reward term and a constraint penalty term: The core evaluation index of the positive reward term is: the geometric shape fit degree of the defect area and the corresponding standard area of the 3D digital reference model, the paint attribute matching degree, and the repair surface flatness of the defect area; The constraint penalty term is for the operation angle deviation and operation force deviation of the repair robot, and a gradient penalty coefficient is configured according to the deviation degree, and the corresponding penalty is triggered when the deviation exceeds the standard; Based on the output results of the reward function, the repair parameters of the repair robot are dynamically adjusted until the repair shape, paint attribute and substrate mechanical properties of the defect area meet the preset standards of the 3D digital reference model.

3. The method of claim 1, wherein: The multi-modal visual acquisition device comprises a color three-dimensional scanner, an eddy current thickness gauge and a near-infrared camera.

4. The method of claim 3, wherein: In step S1, the following steps are further included: The 3D digital reference model is labeled to obtain the coordinates of the iron sheet parts, which include the doors, the engine cover, the fender, the roof and the trunk lid, and the edge contour line of each iron sheet part is extracted, and the area within the edge contour line constitutes a scanning area.

5. The method of claim 4, wherein: In step S2, the following steps are further included: S21, only starting the eddy current thickness gauge, and scanning the scanning area by the eddy current thickness gauge in a low-density scanning mode; S22, comparing the collected paint thickness data with the standard paint thickness of the corresponding area in the 3D digital reference model constructed in step S1 in real time, and determining a suspected collision area when any of the following conditions is met: First, the absolute value of the deviation between the paint thickness of a certain monitoring point and the standard thickness is greater than a first set threshold; Second, the thickness fluctuation of two or more monitoring points in the continuous three detection points is greater than a second set threshold; S23, when the suspected collision area is scanned, the multi-modal visual acquisition device switches to a high-density scanning mode, and simultaneously starts the color three-dimensional scanner, the eddy current thickness gauge and the near-infrared camera to perform multi-dimensional data acquisition; S24, performing fusion analysis on the collected multi-dimensional data to output the defect type; S25, the repair robot automatically matches corresponding repair tools and repair operation logic according to the determined defect type.

6. The method of claim 5, wherein: In the step S24, the defect type includes a recess defect and a scratch defect.

7. The method of claim 6, wherein: The repair operation logic includes a pulling stage, a paint spraying stage and a polishing stage, and different weight coefficients are set for each stage, including: a weight coefficient of a geometric shape fitting degree reward , a weight coefficient of a paint surface attribute matching degree reward , a repair surface flatness reward weight coefficient , a weight coefficient of an operation angle deviation penalty , a weight coefficient of an operation force deviation penalty .

Images