Automatic grinding planning decision-making method for aircraft structural parts

Through the automatic grinding planning and decision-making method of aircraft structural parts, automatic grinding of aircraft structural parts is achieved, solving the problems of high dependence on labor and low efficiency in the existing technology, and improving grinding efficiency and quality.

CN119017201AActive Publication Date: 2024-11-26CHENGDU AIRCRAFT INDUSTRY GROUP

Patent Information

Application Number
CN202411311607.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-20
Publication Date
2024-11-26
Estimated Expiration
2044-09-20

AI Technical Summary

Technical Problem

The prior art is difficult to realize automatic polishing of aircraft structural parts, especially automatic polishing of complex structural parts, resulting in high dependence on labor and difficult to guarantee efficiency and quality.

Method used

A decision-making method for automatic grinding of aircraft structural parts, including positioning of structural parts to be polished, multi-imaging angle image combination judgment of single web multi-imaging angles, identification of transition area of ​​knife marks, scanning of area fine data, and acquisition of final web area grinding boundary data , first web grinding area planning, spatial interference-free distance calculation, and the remaining webs to be polished based on cycle planning methods, generate grinding procedures and perform grinding processing steps.

Benefits of technology

It realizes automatic grinding area identification and adaptive planning of aircraft structural parts, significantly reduces the generation of invalid grinding paths, improves the grinding efficiency and quality of aviation structural parts, reduces the deviation of single-view analysis, and ensures the reliability of automated grinding. .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119017201A_ABST
    Figure CN119017201A_ABST
Patent Text Reader

Abstract

The invention belongs to the field of aeronautical manufacturing, and particularly relates to an aircraft structural part automatic polishing planning decision method which comprises the following steps: positioning a structural part to be polished and a web region; single-web multi-imaging-angle image combination judgment is carried out; identifying and positioning a tool mark transition area; carrying out regional refined data scanning; and finally obtaining web region polishing boundary data. A first web grinding area is planned; space non-interference distance measurement and calculation; determining remaining to-be-polished webs based on a loop planning method; a polishing program is generated for polishing; carrying out polishing treatment; and rechecking the polishing area. Compared with a traditional manual grinding mode, according to the grinding area planning strategy provided by the method, the positions of the adjacent grinding areas and the corresponding attributes of the grinding characteristics can be fully considered, generation of invalid grinding paths can be obviously reduced, and the efficiency and quality of whole-process grinding of the aviation structural part are effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of aviation manufacturing, and in particular relates to an automatic grinding planning and decision-making method for aircraft structural parts. Background Art

[0002] Aircraft structural parts are the core structures that make up the skeleton and frame of the aircraft. The precise processing and manufacturing of structural parts based on the designed model is inseparable from a series of core manufacturing links, including the preparation of CNC machining programs for structural parts, the rough and fine machining of structural parts based on machine tools, and the grinding of the processed structural parts. However, the current grinding of structural parts still mainly relies on manual work. With the development and progress of society, the number of fitters willing to engage in structural part grinding has decreased significantly, which has brought challenges to the mass production and delivery of aircraft. Therefore, it is particularly urgent and necessary to develop and research technologies that can realize automatic grinding of structural parts.

[0003] At present, in the field of polishing, the technology for realizing automatic polishing is relatively mature and widely used in the automotive field. The main reason is that automobiles are mass-produced products, and the corresponding structural parts are of a single type, and the size, shape and other information are relatively fixed, which greatly reduces the difficulty of polishing. Batch polishing work can be completed based on pre-set polishing paths, parameters and trajectories. However, for the structural parts used on aircraft, there are many varieties and small batches. Conventional automotive polishing technology is difficult to apply to such scenarios. At the same time, due to the complexity of the configuration of aircraft structural parts, the difficulty of polishing is increased to a certain extent. Given that the current polishing technology in the automotive field is difficult to apply to the polishing of aircraft structural parts, it is urgent to design new and effective aircraft structural parts polishing methods and technologies.

[0004] Frames and beams account for a large proportion of aircraft structural parts, and their importance is higher than that of other types of auxiliary structural parts. There are a large number of groove features in frame and beam structural parts, and the workload of benchwork grinding these features is large. At the same time, the corresponding grinding path planning is the key technology to achieve automatic grinding priority solution.

[0005] Analysis of existing aircraft structural parts, such as the frame and beam structural parts, shows that manual grinding is still the mainstream grinding method. However, due to the small batch and multi-category characteristics of aircraft structural parts, common grinding strategies are difficult to apply to aircraft structural parts grinding scenarios. In order to break through the technical bottleneck of automatic grinding of aircraft structural parts, the dependence on manual work can be greatly reduced, and automatic and intelligent grinding of the entire process can be achieved. Accurate identification and planning of grinding areas are the key to realizing automated programming, and are also key technical issues that need to be solved in the entire aviation field. Summary of the invention

[0006] In order to solve the problems existing in the existing manual grinding, the present invention proposes an automatic grinding planning and decision-making method for aircraft structural parts, which can realize automatic identification and judgment of the grinding area of ​​aircraft structural parts. The identification result can provide a basis for the automatic planning of the grinding area, which is conducive to the realization of automatic grinding technology, especially the grinding of the web area of ​​the structural parts.

[0007] In order to achieve the above-mentioned invention object, the technical solution provided by this application is as follows:

[0008] An automatic grinding planning and decision-making method for aircraft structural parts comprises the following steps:

[0009] Step 1. Position the structural parts and web areas to be polished;

[0010] Step 2. Combination judgment of images of a single abdominal plate from multiple imaging angles;

[0011] Step 3. Identify and locate the transition area of ​​the cutter mark;

[0012] Step 4. Regional refined data scanning;

[0013] Step 5. Acquisition of final web area grinding boundary data;

[0014] Step 6. Planning of the first web grinding area;

[0015] Step 7. Spatial non-interference distance measurement;

[0016] Step 8. Determine the remaining webs to be polished based on the loop planning method;

[0017] Step 9. Generate a grinding program for grinding;

[0018] Step 10: Polishing;

[0019] Step 11. Recheck the polished area.

[0020] Furthermore, the positioning of the structural parts to be polished and the web area includes:

[0021] Based on the high-resolution three-channel fixed-position camera, the structure to be polished is imaged. The fixed-position camera is equipped with RGBD data acquisition capability, where D represents depth information and RGB represents image information. The original analysis image O is obtained. A global coarse positioning and recognition data model for the structure to be polished is constructed to complete the positioning of the structure to be polished in the original analysis image O. The global coarse positioning and recognition data model for the crane operator is:

[0022]

[0023] The goal of mathematical model learning is to minimize the objective function M({p i},{t i}), represents the logarithmic loss of target and non-target categories, represents the prediction box regression loss, N cls Represents the Minni-batch value, N reg represents the number of regression rectangles, λ represents the balance weight, and p i Indicates the predicted probability of Anchor as the target; With t i Respectively represent the coordinates of the ground truth and the coordinates of the prediction box. Represents the values ​​corresponding to the background and foreground. The loss function includes two parts: the classified tooling background and structural parts, as well as the loss of the category to be polished and the regression loss of the rough positioning border.

[0024] Furthermore, based on M({p i},{t i}) The corresponding image O is roughly identified and positioned, and the corresponding result coordinates are {t x ,t y ,t w ,t h} indicates that t x is the normalized central horizontal coordinate of the standard frame, t y is the normalized central ordinate of the standard frame, t w is the width of the box, t h is the height of the box; based on the resulting coordinates {t x ,t y ,t w ,t h}Further cropping and extraction of image O obtains a new image P; smoothing of the processed image P with different weights is performed, and after processing, the Gaussian difference D(x, y, δ) corresponding to the adjacent weight difference images is calculated, where x is the value of the image pixel in the X-axis direction, y is the value of the image pixel in the Y-axis direction, and δ is the corresponding standard deviation. The key point K in the image after Gaussian difference processing is found, and the Gaussian difference D(x, y, δ) is interpolated through Taylor series expansion to correct the key point and obtain the corrected point KP. Based on KP, the area to be polished in image P is precisely located, and the corresponding result of the same scale as image O is represented by Obj.

[0025] Further, in the step 1, the imaging cameras for the structural members to be polished are arranged to meet the following conditions: The cameras for imaging the structural members to be polished include a fixed-position camera, a horizontally adjustable moving-position camera 1, a horizontally adjustable moving-position camera 2, and a horizontally adjustable moving-position camera 3. The imaging axis center corresponding to the fixed-position camera is concentric with the center of the placement of the structural members to be polished and the geometric center of the structural members to be polished, with a distance deviation <dis-o>, where dis-o is the set distance deviation threshold between the center of the structural member placement and the geometric center of the structural members to be polished, and the fixed-position camera is located directly above the structural members to be polished; The horizontally adjustable moving-position camera 1 is arranged to move within the imaging plane P1, with the corresponding pitch angle α1 of the horizontally adjustable moving-position camera 1. There are num imaging docking points Point(i) on the circumference where the corresponding imaging plane P1 is located, that is, for the corresponding i ∈ [1, num], the circumferential radian interval between adjacent imaging docking points Point(i) and Point(i + 1) is equal to 360 / num, and each camera only performs imaging at the imaging docking point Point(i) with the corresponding pitch angle α1; The horizontally adjustable moving-position camera 2 is arranged to move within the imaging plane P2, with the corresponding pitch angle α2 of the horizontally adjustable moving-position camera 2. There are num imaging docking points Point(i) on the circumference where the corresponding imaging plane P2 is located, that is, for the corresponding i ∈ [1, num], the circumferential radian interval between adjacent imaging docking points Point(i) and Point(i + 1) is equal to 360 / num, and each camera only performs imaging at the imaging docking point Point(i) with the corresponding pitch angle α2; The horizontally adjustable moving-position camera 3 is arranged to move within the imaging plane P3, with the corresponding pitch angle α3 of the horizontally adjustable moving-position camera 3. There are num imaging docking points Point(i) on the circumference where the corresponding imaging plane P3 is located, that is, for the corresponding i ∈ [1, num], the circumferential radian interval between adjacent imaging docking points Point(i) and Point(i + 1) is equal to 360 / num, and each camera only performs imaging at the imaging docking point Point(i) with the corresponding pitch angle α3.

[0026] Furthermore, the placement strategy for the structural members to be polished is as follows: After the structural members are placed, the area to be polished corresponding to the structural members to be polished faces upward. If the structural members need to be polished on both sides, first select one side to face upward and polish it first, and then polish the other side of the model after completion; After placement, the geometric center of the structural members is fixed at the position of the entire workbench and is coaxial with the fixed-position camera to ensure the integrity of the structural member area obtained.

[0027] Furthermore, the corresponding mathematical model M({p i},{t i}), in order to prevent the influence of BX and GT position and size on the regression scale, where BX is the model prediction area and GT is the actual result area, the BX coordinate needs to be normalized, that is:

[0028]

[0029] Among them, p x is the x value of the center of the prediction box; p y is the y value of the center of the prediction box; is the x value of the center of the standard box GT; is the y value of the center of the standard box; b x is the x value of the center of the potential target area box; b y is the y value of the center of the prediction box; p w is the predicted box width; p h is the predicted box height value; b w is the center width of the prediction box; b h is the height value of the potential target area; is the standard box width value; is the standard frame height value; t x ,t y ,t w ,t h and Respectively represent the normalized center coordinates of BX and the standard frame and the width and height of the frame;

[0030] i is the index of the center of the prediction box; p i Indicates the predicted probability that the center of the prediction box is the target; Indicates the values ​​corresponding to the background and foreground, the background is 0 and the foreground is 1; With t i Respectively represent the coordinates of the GT enclosed and the coordinates of the prediction box, where t i is composed of {t x ,t y ,t w ,t h} composed of coordinate vectors; It represents the logarithmic loss corresponding to the identification and positioning of the workbench background and the area to be polished, that is: represents the prediction box coordinate regression deviation loss, and the corresponding mathematical model can be expressed as Where R represents the loss evaluation function, let In the formula Indicates that there is only a prospect There is a loss of coordinate position deviation;

[0031] The output results of the classification model and regression model are respectively expressed as {p i} and {t i}, while Ncls represents normalization processing, N reg Introduce the balance weight λ and implement normalization. cls performs normalization on N cls is the mini-batch value, and the reg term is normalized by N reg The value is the number of predicted box center positions, where the value corresponding to the balance weight λ is a positive integer;

[0032] The multi-classifier mathematical model is set as:

[0033]

[0034] The optimized processing method is to multiply the numerator by the constant C and the denominator by the constant C at the same time, that is:

[0035]

[0036] All values ​​add up to 1, which ensures that each output result can be transformed to the interval [0,1]; i represents the number of categories, j represents the number of categories output by the neural network, and f j Indicates the value corresponding to the jth category, y i represents the probability that the neural network predicts the output to be category i, f yi It indicates the value corresponding to the predicted i-th category, and x indicates the data input to the network;

[0037] Construct a mapping table Table(x,y) of the regional position and physical scale, and record the actual physical distance data Dis(point(x,y),C(x,y)) from each point point(x,y) in the images O and Obj to the geometric center / placement center C(x,y) of the structure to be polished.

[0038] Furthermore, the placement posture of the structural parts to be polished is jointly corrected with the design digital model. Based on the posture information of the structural parts, the placement posture of the structural parts in the design digital model is corrected to make the posture of the structural parts in the digital model consistent with the posture of the structural parts to be polished. The specific steps of the joint correction strategy are as follows: in the recognition result diagram Obj of the structural parts to be polished, 4 directions are set, namely vertically upward, downward, horizontally left and right. The reference point of the direction is the center of Obj, that is, (rows / 2, cols / 2) as a reference, where row represents the number of rows of Obj, and cols represents the number of columns of Obj; in the design digital model software, the digital model is adjusted based on the corresponding features of Obj vertically upward, downward, horizontally left and right in the recognition result diagram, so that the features presented on the screen are consistent with the features in Obj.

[0039] Furthermore, the web area that potentially needs to be polished is located. First, the web area is identified and judged in Obj. The corresponding specific steps are as follows: first, the RGB data obtained by the fixed-position camera is registered with the D data to ensure that the coordinate systems of the two types of data obtained coincide, and the camera depth D information acquisition function is started to obtain the D data information; based on the structural part boundary obtained in Obj, the structural part data and the background area (non-structural part) data are separated from the D information data; the background area part of the separated D information data is set to 0, while the data in the structural part area remains unchanged; since the obtained D data is in matrix form, the length and width of the matrix are the same as Obj, and the data Data(i,j) corresponding to each point of the structural part in the D data is read, where i represents the corresponding number of rows in the data D, and j represents the corresponding number of columns in the data D. All Data num The set of (i,j) is denoted as Set(Data num (i,j)), num represents the total number of data points corresponding to the structural parts in the D data; the calculation set Set(Data num Each Data in (i,j)) num (i,j) values ​​are sorted and the minimum value Min(Data num (i,j)) and the maximum value Max(Data num (i,j)); Set(Data num All Data in (i,j)) num (i,j) interval transformation, the corresponding transformation formula is: T(Data num (i,j))=255*|Min(Data num (i,j))-Data num (i,j)| / (Max(Data num (i,j))-Min(Data num (i,j))), the corresponding data after transformation is T-Set(Data num (i,j)); based on the first-order difference, complete T-Set(Data num The corresponding secondary transformation formula is: TData num (i,j)=(|TData num (i,j)-TData num (i,j+1)|+|TData num (i,j)-TData num (i+1,j)|) / 2, where TData num (i,j) represents the set T-Set (Data num(i,j)) corresponding to each point data; set the three-transformation data processing threshold T-threshold, if T-Set(Data num (i,j))≤T-threshold, then the corresponding T-Set(Data num (i, j))=0, otherwise it remains unchanged; the D data information after three transformations is transformed into an image to obtain an image Dimg, which is a single-channel grayscale image;

[0040] Based on the designed random assignment, different methods are used to extract the areas with the same grayscale in the image Dimg. The specific method is: read the grayscale value corresponding to each point in the image Dimg and change its value DV(i,j), DV(i,j)={Rv=(Num(random)%255),Gv=(Num(random)%255)and Gv!=Rv,Bv(Num(random)%255)and Bv!=Rv and Bv! =Gv}, if Gv=Rv, then regenerate Gv, if Bv=Rv, then regenerate Bv, if Bv=Gv, then regenerate Bv, until the obtained Rv, Gv, and Bv are all different, where Rv represents the value assigned to the newly added channel at the midpoint (i, j) of the image Dimg, where Gv represents the value assigned to the newly added channel at the midpoint (i, j) of the image Dimg, where Bv represents the value assigned to the newly added channel at the midpoint (i, j) of the image Dimg, and Num(random) represents the generation of a random positive integer; read the Rv, Gv, and Bv values ​​corresponding to each data point in the image Dimg, if the values ​​corresponding to the points separated by 1 unit distance are equal, then such data points can be classified into one category, extract the area composed of such points that satisfy a class relationship, and input the classification neural network for category division, the classification network belongs to the two categories, namely the web and the non-web; the area identified as the web by the network is the completed web area to be polished Area k (i, j) positioning step, k represents the number of successfully identified webs, and (i, j) represents the center coordinates of the corresponding web.

[0041] Furthermore, the combination of multiple imaging angles of a single web is specifically judged as follows: start the horizontally adjustable mobile camera 1, the horizontally adjustable mobile camera 2, and the horizontally adjustable mobile camera 3, and the corresponding imaging angles are α1, α2, and α3, respectively, and perform imaging on each Area at the num imaging stop points Point(i) corresponding to the imaging plane P1, the imaging plane P2, and the imaging plane P3. k(i, j) is imaged, and the number of image data obtained for each web corresponding to k is 3*num. The corresponding images are recorded as I1(h), I2(h), and I3(h), respectively, where h∈[1,num] represents the corresponding image number; the obtained I1(h), I2(h), and I3(h) totaling 3*num image data are evaluated based on the following designed formula:

[0042]

[0043] For I1(h), the larger the Value is, the more it indicates that there are knife marks waiting to be polished; analyze and compare the 3*num calculated Value values, and sort them from large to small. If there are β corresponding Value values ​​greater than the set lower limit, it indicates that the corresponding web area has knife marks and other features that need to be polished. The area to be polished is marked as G(m), where m represents the number of areas to be polished; cols and rows represent the number of columns and rows corresponding to I1(h), I2(h), and I3(h), respectively, i and j represent the subscripts of the pixel points along the row and column directions, respectively, the grayscale value corresponding to f at the (i, j) position, and x and y represent the horizontal and vertical directions, respectively.

[0044] Furthermore, the identification and positioning of the transition area to the tool mark is specifically as follows: based on the grinding area marked as G(m), the coordinates of the web area to be ground are determined. Since the acquired data D is obtained through a series of transformations to obtain G(m), D and G have the same image coordinate system, that is, G(m) determines that the web area to be ground has a one-to-one correspondence with D, and D and O are acquired based on the camera at the same position. Therefore, the coordinates of the web area to be ground in O can be obtained based on the G(m) mapping, and at the same time, the coordinates of the web to be ground in the digital model can be completed to obtain an accurate grinding tool path.

[0045] Furthermore, the regional refined data scanning is specifically as follows: G(m) determines the coordinates of the web area to be polished to complete the position information Loc(m) required for polishing in the digital model, and sets the refined scanning movement trajectory and path based on the position information Loc(m). According to the generated path, a circular scanning method is used to obtain the accurate data PData(m) of each scanning area. After processing m areas, the data PData(m) is analyzed, and the areas that do not need to be polished are removed from the m areas, and the areas that need to be polished are retained. The secondary judgment of the candidate areas of the web to be polished based on G(m) is completed, and the web area to be polished is finally determined by DM. a (i, j) represents, a represents the number of areas that need to be polished in the end, and (i, j) represents the coordinates of the corresponding area in G.

[0046] Furthermore, the final web area grinding boundary data acquisition is specifically as follows: based on the web area to be polished determined in the end, the DMa (i, j), complete the mapping of the polishing area a in the digital model, and calculate the boundary data Cdata corresponding to a in the digital model.

[0047] Furthermore, the first web grinding area planning is as follows: for each of the a areas to be ground, the distance dis from the geometric center to the initial grinding head is calculated. The grinding area with the smallest distance dis from the grinding head is the first grinding area. The diameter of the enclosed area corresponding to the grinding is calculated by the grinding head radius DR. The enclosed area calculation method is:

[0048]

[0049] Where r is the actual radius of the grinding head, r r represents the corresponding grinding head annular surface radius, w represents the grinding head side deflection angle, and λ represents the grinding head forward inclination angle; the number of optimized grinding path rows is designed, and the number of optimized grinding area path functions can be expressed as Among them, M i represents the grinding path length of Nv grinding areas, S j represents the length of the grinding path of the corresponding closed area boundary, F z Represents the length of the auxiliary path traversed by the grinding head; Nk is the number of grinding areas, i represents the measurement subscript of the grinding path length, j represents the measurement subscript of the grinding path length at the closed area boundary, and z represents the measurement subscript of the auxiliary path length traversed by the grinding head. The generation of the grinding head processing trajectory is completed based on the path optimization function, and the generation of the grinding head point data is based on the boundary data Cdata.

[0050] Furthermore, the spatial non-interference distance measurement method is as follows: based on the first obtained grinding area, the corresponding boundary data point Bdata(i,j) is calculated, and at the same time, the radius data GR corresponding to the rigid rotation area of ​​the grinding head in the radial direction during grinding is calculated, and the point Mpoint(i,j) where the center of the grinding head moves is directly calculated. The Euclidean distance DMB between the point Mpoint(i,j) and the point Bdata(i,j) is directly calculated. If the Euclidean distance DMB between the point Mpoint(i,j) and the point Bdata(i,j) in the plane satisfies DMB≤GR, the flag is marked as 0, and DMB>GR is marked as 1;

[0051] According to the optimal spatial non-interference distance measurement method, the workpiece interference is judged to correct the grinding path. If there is spatial grinding interference flag = 0, the path Mpoint (i, j) corresponding to the interference area is directly modified. If there is no interference flag = 1, there is no need to modify Mpoint (i, j).

[0052] Furthermore, the remaining webs to be polished are determined based on the cyclic planning method as follows: the first polishing area is taken as the starting point to plan a-1 areas to be polished globally, and the corresponding planning evaluation function is

[0053]

[0054] Among them, c1 represents the motion path weight coefficient, which corresponds to the importance evaluation of the grinding head motion path index, F(T k )The first item represents the total length of the grinding head motion path, c2 represents the regional grinding difficulty coefficient, Di represents the center clustering of two adjacent grinding areas, and Hi represents the grinding difficulty coefficient; global planning is performed and converted into the F(Tk) extreme value problem. The second grinding area is obtained based on the global optimal solution, and so on. h represents the subscript of the center clustering measurement of two adjacent grinding areas, and a represents the number of global grinding areas.

[0055] Furthermore, generating a grinding program for grinding specifically includes: generating a numerical control program that can control the movement of the grinding head.

[0056] Furthermore, the grinding process is specifically performed as follows: based on the generated grinding program, the grinding head is driven to complete the grinding process of a grinding area, the first grinding of each area is completed, and the evaluation is performed based on the re-inspection of the grinding area.

[0057] Furthermore, the re-inspection of the polished area is specifically as follows: the polished area is evaluated. If the parameters corresponding to the characteristics of the knife marks and step differences that need to be polished in the area do not meet the requirements, further polishing is required. The corresponding areas for further polishing are only selected from the a first polished areas that have not been polished in place, until the quality of all a areas meets the requirements.

[0058] The beneficial effects of the present invention are:

[0059] 1. Compared with the traditional manual grinding method, this method can realize the automatic identification of the grinding area and complete the adaptive planning decision of the grinding area based on the identification results, providing core technical support for the fully automatic grinding of aviation structural parts, especially complex structural parts. The proposed grinding area planning strategy can fully consider the position of adjacent grinding areas and the corresponding attributes of grinding features, which can significantly reduce the generation of invalid grinding paths and effectively improve the efficiency and quality of the whole process grinding of aviation structural parts. The judgment of the grinding area combined with the analysis results of multiple perspectives significantly reduces the deviation introduced by the single perspective analysis, improves the precision and accuracy of grinding, avoids quality problems caused by grinding, and ensures the reliability of automated grinding.

[0060] 2. The present invention proposes an effective method suitable for grinding aviation structural parts. The proposed grinding area planning strategy can fully consider the positions of adjacent grinding areas and the corresponding attributes of grinding features, which can significantly reduce the generation of invalid grinding paths and effectively improve the efficiency and quality of the whole process grinding of aviation structural parts. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 The figure is a flow chart of the method of the present invention.

[0062] Figure 2 Arrange imaging cameras for corresponding structural parts.

[0063] Figure 3 Corresponding grinding area and grinding head path planning in the grinding area. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is to explain the present invention rather than to limit the present invention. Based on the embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0065] The specific implementation method of the present invention is described below with reference to the accompanying drawings and examples, but the present invention is not limited to this embodiment.

[0066] Example 1

[0067] An automatic grinding planning and decision-making method for aircraft structural parts comprises the following steps:

[0068] Step 1. Position the structural parts and web areas to be polished;

[0069] Step 2. Combination judgment of images of a single abdominal plate from multiple imaging angles;

[0070] Step 3. Identify and locate the transition area of ​​the cutter mark;

[0071] Step 4. Regional refined data scanning;

[0072] Step 5. Acquisition of final web area grinding boundary data;

[0073] Step 6. Planning of the first web grinding area;

[0074] Step 7. Spatial non-interference distance measurement;

[0075] Step 8. Determine the remaining webs to be polished based on the loop planning method;

[0076] Step 9. Generate a grinding program for grinding;

[0077] Step 10: Polishing;

[0078] Step 11. Recheck the polished area.

[0079] The step 1 comprises:

[0080] Firstly, the structure to be polished is imaged based on a high-resolution three-channel fixed-position camera. The fixed-position camera is equipped with RGBD data acquisition capability, where D represents depth information and RGB represents image information, and the original analysis image O is obtained. A global coarse positioning and recognition data model for the structure to be polished is constructed to complete the positioning of the structure to be polished in the original analysis image O. The global coarse positioning and recognition data model for the crane operator is:

[0081]

[0082] The goal of mathematical model learning is to minimize the objective function M({p i},{t i}), represents the logarithmic loss of target and non-target categories, represents the prediction box regression loss, N cls Represents the Minni-batch value, N reg represents the number of regression rectangles, λ represents the balance weight, and p i Indicates the predicted probability of Anchor as the target; With t i Respectively represent the coordinates of the ground truth and the coordinates of the prediction box. Represents the values ​​corresponding to the background and foreground. The loss function includes two parts: the classified tooling background and structural parts, as well as the loss of the category to be polished and the regression loss of the rough positioning border.

[0083] Based on M({p i},{t i}) The corresponding image O is roughly identified and positioned, and the corresponding result coordinates are {t x ,t y ,t w ,t h} indicates that t x is the normalized central horizontal coordinate of the standard frame, t y is the normalized central ordinate of the standard frame, t w is the width of the box, t h is the height of the box; based on the resulting coordinates {t x ,t y ,t w ,t h}Further cropping and extraction of image O obtains a new image P; smoothing of the processed image P with different weights is performed, and after processing, the Gaussian difference D(x, y, δ) corresponding to the adjacent weight difference images is calculated, where x is the value of the image pixel in the X-axis direction, y is the value of the image pixel in the Y-axis direction, and δ is the corresponding standard deviation. The key point K in the image after Gaussian difference processing is found, and the Gaussian difference D(x, y, δ) is interpolated through Taylor series expansion to correct the key point and obtain the corrected point KP. Based on KP, the area to be polished in image P is precisely located, that is, the area to be polished is precisely located based on the image obtained by the fixed-position camera and the designed algorithm. The corresponding result of the same scale as image O is represented by Obj.

[0084] In Step 1, as described above, the imaging cameras corresponding to the structural members to be polished are arranged. As shown in Figure 2, the arrangement and imaging method meet the following conditions: The cameras for imaging the structural members to be polished include a fixed-position camera, a horizontally adjustable moving-position camera 1, a horizontally adjustable moving-position camera 2, and a horizontally adjustable moving-position camera 3. The imaging axis center corresponding to the fixed-position camera is concentric with the center of the placement of the structural members to be polished and the geometric center of the structural members to be polished, with a distance deviation <dis-o>. <dis-o> is the set distance deviation threshold between the center of the structural member placement and the geometric center of the structural members to be polished, and the fixed-position camera is located directly above the structural members to be polished; The horizontally adjustable moving-position camera 1 is configured to move within the imaging plane P1. The corresponding pitch angle of the horizontally adjustable moving-position camera 1 is α1. There are num imaging docking points Point(i) on the circumference where the corresponding imaging plane P1 is located, that is, for the corresponding i ∈ [1, num]. The circumferential radian interval between adjacent imaging docking points Point(i) and Point(i + 1) is equal to 360 / num, and each camera only images at the imaging docking point Point(i) corresponding to the pitch angle α1; The horizontally adjustable moving-position camera 2 is configured to move within the imaging plane P2. The corresponding pitch angle of the horizontally adjustable moving-position camera 2 is α2. There are num imaging docking points Point(i) on the circumference where the corresponding imaging plane P2 is located, that is, for the corresponding i ∈ [1, num]. The circumferential radian interval between adjacent imaging docking points Point(i) and Point(i + 1) is equal to 360 / num, and each camera only images at the imaging docking point Point(i) corresponding to the pitch angle α2; The horizontally adjustable moving-position camera 3 is configured to move within the imaging plane P3. The corresponding pitch angle of the horizontally adjustable moving-position camera 3 is α3. There are num imaging docking points Point(i) on the circumference where the corresponding imaging plane P3 is located, that is, for the corresponding i ∈ [1, num]. The circumferential radian interval between adjacent imaging docking points Point(i) and Point(i + 1) is equal to 360 / num, and each camera only images at the imaging docking point Point(i) corresponding to the pitch angle α3; The role of the fixed-position camera is to obtain the global image of the structural members to be polished to achieve the boundary region positioning of the structural members, while the roles of the horizontally adjustable moving-position camera 1, the horizontally adjustable moving-position camera 2, and the horizontally adjustable moving-position camera 3 are to obtain the images of the web regions to be polished at different angles, and based on the obtained image fusion, to judge whether the corresponding regions need to be polished. The distribution of the horizontally adjustable moving-position camera 1, the horizontally adjustable moving-position camera 2, and the horizontally adjustable moving-position camera 3 has the characteristic of evenly dividing the hemisphere latitude.

[0085] As mentioned above, the placement strategy of the corresponding structural parts to be polished is: after the structural parts are placed, the corresponding polishing area of ​​the structural parts to be polished faces upwards. If the structural parts need to be polished on both sides, one side is selected to be polished first, and then the other side is modeled after polishing is completed; after placement, the geometric center of the structural parts is fixed in the position of the entire workbench and is coaxial with the fixed position camera to ensure the integrity of the structural parts area.

[0086] As mentioned above, the corresponding mathematical model M({p i},{t i}), in order to prevent the influence of BX and GT position and size on the regression scale, where BX is the model prediction area and GT is the actual result area, the BX coordinate needs to be normalized, that is:

[0087]

[0088] Among them, p x is the x value of the center of the prediction box; p y is the y value of the center of the prediction box; is the x value of the center of the standard box GT; is the y value of the center of the standard box; b x is the x value of the center of the potential target area box; b y is the y value of the center of the prediction box; p w is the predicted box width; p h is the predicted box height value; b w is the center width of the prediction box; b h is the height value of the potential target area; is the standard box width value; is the standard frame height value; t x ,t y ,t w ,t h and Respectively represent the normalized center coordinates of BX and the standard frame and the width and height of the frame;

[0089] i is the index of the center of the prediction box; p i Indicates the predicted probability that the center of the prediction box is the target; Indicates the values ​​corresponding to the background and foreground, the background is 0 and the foreground is 1; With t i Respectively represent the coordinates of the GT enclosed and the coordinates of the prediction box, where t i is composed of {t x ,t y ,t w ,t h} composed of coordinate vectors; It represents the logarithmic loss corresponding to the identification and positioning of the workbench background and the area to be polished, that is: represents the prediction box coordinate regression deviation loss, and the corresponding mathematical model can be expressed as Where R represents the loss evaluation function, let In the formula Indicates that there is only a prospect There is a loss of coordinate position deviation;

[0090] The output results of the classification model and regression model are respectively expressed as {p i} and {t i}, while N cls represents normalization processing, N reg Introduce the balance weight λ and implement normalization. cls performs normalization on N cls is the mini-batch value, and the reg term is normalized by N reg The value is the number of predicted box center positions, where the value corresponding to the balance weight λ is a positive integer;

[0091] The multi-classifier mathematical model is set as:

[0092]

[0093] Since the above classifier formula is an exponential operation, if the value is very large, it may cause an exponential explosion problem. The optimized processing method is to multiply the numerator by the constant C and the denominator by the constant C at the same time, that is:

[0094]

[0095] All values ​​add up to 1, which ensures that each output result can be transformed to the interval [0,1]; i represents the number of categories, j represents the number of categories output by the neural network, and f j Indicates the value corresponding to the jth category, y i represents the probability that the neural network predicts the output to be category i, f yi It indicates the value corresponding to the predicted i-th category, and x indicates the data input to the network;

[0096] Construct a mapping table Table(x,y) between the area position and the physical scale. The table is used to record the actual physical distance data Dis(point(x,y),C(x,y)) from each point point(x,y) in the image O and Obj to the geometric center / center of the structure to be polished C(x,y). The value of the physical length can provide a basis for the positioning of the corresponding features in the subsequent polishing area and the design model, because the image O and Obj are in different coordinate systems with the design model, and the actual physical distance data Dis(point(x,y),C(x,y)) is the only feasible way to realize the association of the polishing area feature mapping. The mapping table is the key to realizing data mapping. The imaging area and size of the fixed position camera remain unchanged, the imaging angle of the camera remains unchanged, and the parameters such as resolution are also fixed. Therefore, it is ensured that the actual spatial distance corresponding to each image obtained by the fixed position camera is small and will not affect the actual feature mapping.

[0097] The placement posture of the structural parts to be polished is combined with the design digital model correction strategy. The purpose of this step is to correct the placement posture of the structural parts in the design digital model based on the posture information of the structural parts, so that the posture of the structural parts in the digital model is consistent with the posture of the structural parts to be polished, which is conducive to the subsequent positioning of the area to be polished in the digital model, and at the same time, it is convenient to generate an automatic polishing program that can control the movement of the polishing head. The specific steps of the joint correction strategy are: In the identification result diagram Obj of the structural parts to be polished, there are 4 directions set, namely vertically upward, downward, horizontally left and right, and the reference point of the direction is O The center of bj, i.e. (rows / 2, cols / 2), is used as a reference, where row represents the number of rows of Obj and cols represents the number of columns of Obj. In the design digital model software, the digital model is adjusted based on the corresponding features of Obj vertically upward, downward, horizontally left and right in the recognition result diagram, so that the features presented on the screen are consistent with the features in Obj. The consistent interpretation is that the contour features presented by the corresponding structural parts remain basically the same as the contour features in Obj, and the corresponding magnification, i.e. the size of the structural parts features presented in the screen direction, may be different.

[0098] After completing the identification and positioning of the boundary area of ​​the structural part to be polished, it is necessary to further locate the potential web area that needs to be polished. First, the identification and judgment of the web area is completed in Obj. The corresponding specific steps are: first, the RGB data and D data obtained by the fixed-position camera are registered to ensure that the coordinate systems of the two types of data obtained coincide, that is, after the registration, the corresponding coordinates of the RGB data and D data obtained by the same object in the two coordinate systems are consistent; because the fixed-position camera is equipped with the ability to collect RGBD data information, the camera depth D information collection function is started to obtain D data information; based on the structural part boundary obtained in Obj, the structural part data and the background area (non-structural part) data are separated from the D information data; the background area part of the separated D information data is set to 0, while the data in the structural part area remains unchanged; because the obtained D data is in matrix form, the length and width of the matrix are the same as Obj, and the data Data(i,j) corresponding to each point of the structural part in the D data is read, i represents the corresponding number of rows in the data D, j represents the corresponding number of columns in the data D, and all Data num The set of (i,j) is denoted as Set(Data num (i,j)), num represents the total number of data points corresponding to the structural parts in the D data; the calculation set Set(Data num Each Data in (i,j)) num (i,j) values ​​are sorted and the minimum value Min(Data num (i,j)) and the maximum value Max(Data num (i,j)); Set(Data num All Data in (i,j)) num (i,j) interval transformation, the corresponding transformation formula is: T(Data num (i,j))=255*|Min(Data num (i,j))-Data num (i,j)| / (Max(Data num (i,j))-Min(Data num (i,j))), the corresponding data after transformation is T-Set(Data num (i,j)); based on the first-order difference, complete T-Set(Data num The corresponding secondary transformation formula is: TData num (i,j)=(|TData num (i,j)-TData num (i,j+1)|+|TData num (i,j)-TData num(i+1,j)|) / 2, where TData num (i,j) represents the set T-Set (Data num (i,j)) corresponding to each point data; set the three-transformation data processing threshold T-threshold, if T-Set(Data num (i,j))≤T-threshold, then the corresponding T-Set(Data num (i, j))=0, otherwise it remains unchanged; the D data information after three transformations is transformed into an image to obtain an image Dimg, which is a single-channel grayscale image;

[0099] Based on the designed random assignment, different methods are used to extract the areas with the same grayscale in the image Dimg. The specific method is: read the grayscale value corresponding to each point in the image Dimg and change its value DV(i,j), DV(i,j)={Rv=(Num(random)%255),Gv=(Num(random)%255)and Gv!=Rv,Bv(Num(random)%255)and Bv!=Rv and Bv! =Gv}, if Gv=Rv, then regenerate Gv, if Bv=Rv, then regenerate Bv, if Bv=Gv, then regenerate Bv, until the obtained Rv, Gv, and Bv are all different, where Rv represents the value assigned to the newly added channel at the midpoint (i, j) of the image Dimg, where Gv represents the value assigned to the newly added channel at the midpoint (i, j) of the image Dimg, where Bv represents the value assigned to the newly added channel at the midpoint (i, j) of the image Dimg, and Num(random) represents the generation of a random positive integer; read the Rv, Gv, and Bv values ​​corresponding to each data point in the image Dimg, if the values ​​corresponding to the points separated by 1 unit distance are equal, then such data points can be classified into one category, extract the area composed of such points that satisfy a class relationship, and input the classification neural network for category division, the classification network belongs to the two categories, namely the web and the non-web; the area identified as the web by the network is the completed web area to be polished Area k (i, j) positioning step, k represents the number of successfully identified webs, and (i, j) represents the center coordinates of the corresponding web.

[0100] The specific judgment of the combination of multiple imaging angles of a single web is as follows: The purpose of this step is to further judge which webs need to be polished and which do not need to be polished based on the web areas that have been positioned, so as to improve the polishing efficiency of the structural parts. The specific implementation method steps are as follows: Start the horizontally adjustable mobile position camera 1, the horizontally adjustable mobile position camera 2, and the horizontally adjustable mobile position camera 3, and the corresponding imaging angle pitch angles are α1, α2, and α3 respectively, and the num imaging stop points Point(i) corresponding to the imaging plane P1, the imaging plane P2, and the imaging plane P3 are respectively imaged for each Area k (i, j) is imaged, and the number of image data obtained for each web corresponding to k is 3*num. The corresponding images are recorded as I1(h), I2(h), and I3(h), respectively, where h∈[1,num] represents the corresponding image number; the obtained I1(h), I2(h), and I3(h) totaling 3*num image data are evaluated based on the following designed formula:

[0101]

[0102] This formula can reflect and evaluate the high-frequency characteristics of the image due to large differences such as knife marks and step differences. Theoretically, for I1(h), the larger the Value, the more it indicates that there are knife marks waiting to be polished; analyze and compare the 3*num calculated Value values, and sort them from large to small. If there are β corresponding Value values ​​greater than the set lower limit, it indicates that the corresponding web area has knife marks and other features that need to be polished. The area to be polished is marked as G(m), where m represents the number of areas to be polished; cols and rows represent the number of columns and rows corresponding to I1(h), I2(h), and I3(h), respectively, i and j represent the subscripts of the pixel points along the row and column directions, respectively, the grayscale value corresponding to f at the (i, j) position, and x and y represent the horizontal and vertical directions, respectively.

[0103] Specifically, the identification and positioning of the transition area of ​​the tool mark is as follows: the role of this step is to determine the position information of the web area to be polished containing the tool mark in the polishing tool and the three-dimensional digital model. The position information on the polishing tool can ensure the accurate positioning of the web polishing area during polishing, and the position information in the three-dimensional digital model can ensure the accuracy of the polishing head movement program generated in the digital model. The implementation steps are: based on the polished area marked as G(m), the coordinates of the web area to be polished are determined. Since the acquired data D is obtained through a series of transformations to obtain G(m), D and G have the same image coordinate system, that is, G(m) determines that the web area to be polished has a one-to-one correspondence with D, and D and O are obtained based on the camera at the same position. Therefore, the coordinates of the web area to be polished in O can be obtained based on the G(m) mapping, and at the same time, the coordinates of the web to be polished in the digital model are completed to obtain an accurate polishing tool path.

[0104] The specific steps of regional refined data scanning are as follows: The purpose of this step is to re-judge the coordinate area of ​​the web area to be polished determined by G(m). While avoiding invalid scanning, it can increase the accuracy of the preliminary determination of the area to be polished. The implementation steps are as follows: G(m) determines the coordinates of the web area to be polished and completes the position information Loc(m) required for polishing in the digital model. Based on the position information Loc(m), the refined scanning movement trajectory and path are set. According to the generated path, a cyclic scanning method is used to obtain the accurate data PData(m) of each scanning area. After processing m areas, the data PData(m) is analyzed, and the areas that do not need to be polished are removed from the m areas, and the areas that need to be polished are retained. The secondary judgment of the candidate areas of the web to be polished based on G(m) is completed. The web area to be polished is finally determined by DM. a (i, j) represents, a represents the number of areas that need to be polished in the end, and (i, j) represents the coordinates of the corresponding area in G.

[0105] The final web area grinding boundary data acquisition is specifically as follows: based on the final determined web area to be ground, use DM a (i, j), complete the mapping of the grinding area a in the digital model, and calculate the boundary data Cdata corresponding to a in the digital model. The necessity of calculating the boundary data Cdata is that the movement boundary of the grinding head can be strictly limited based on the boundary data to avoid grinding the area that does not need to be polished, which will cause the size of the structural parts to exceed the tolerance and be scrapped.

[0106] The first web grinding area planning is as follows: The purpose of this step is to determine the first web to be ground and calculate the DM of each grinding area a (i,j) The distance data between the motion coordinates of the grinding head and the structural parts can avoid the spatial interference problem caused by the unreasonable setting of the grinding path. The first grinding area planning method is: the distance dis from the geometric center of a area to be polished to the initial grinding head, where the grinding area with the smallest distance dis from the grinding head is the first grinding area. This planning method can ensure that the invalid stroke of the grinding head is as short as possible, thereby improving the grinding efficiency. The design of the grinding head movement method corresponding to the grinding area aims to be efficient and short the total movement path of the grinding head, thereby reducing the grinding time consumption while ensuring the grinding quality, and improving the grinding quality. The specific implementation method is: the diameter of the enclosed area corresponding to the grinding is calculated by the grinding head radius DR. The enclosed area calculation method is

[0107]

[0108] Where r is the actual radius of the grinding head, r rrepresents the corresponding grinding head annular surface radius, w represents the grinding head side deflection angle, and λ represents the grinding head forward inclination angle; the number of optimized grinding path rows is designed, and the number of optimized grinding area path functions can be expressed as Among them, M i represents the grinding path length of Nv grinding areas, S j represents the length of the grinding path of the corresponding closed area boundary, F z Represents the length of the auxiliary path traversed by the grinding head; Nk is the number of grinding areas, i represents the measurement subscript of the grinding path length, j represents the measurement subscript of the grinding path length at the closed area boundary, and z represents the measurement subscript of the auxiliary path length traversed by the grinding head. The generation of the grinding head processing trajectory is completed based on the path optimization function, and the generation of the grinding head point data is based on the boundary data Cdata.

[0109] The method for measuring the spatial non-interference distance is as follows: based on the first obtained grinding area, the corresponding boundary data point Bdata(i,j) is calculated, and at the same time, the radius data GR corresponding to the rigid rotation area of ​​the grinding head in the radial direction during grinding is calculated, and the point Mpoint(i,j) where the center of the grinding head moves is directly calculated. The Euclidean distance DMB between the point Mpoint(i,j) and the point Bdata(i,j) is directly calculated. If the Euclidean distance DMB between the point Mpoint(i,j) and the point Bdata(i,j) in the plane is less than or equal to GR, the flag is marked as 0, and DMB>GR is marked as 1.

[0110] According to the optimal spatial non-interference distance measurement method, the workpiece interference is judged to correct the grinding path. If there is spatial grinding interference flag = 0, the path Mpoint (i, j) corresponding to the interference area is directly modified. If there is no interference flag = 1, there is no need to modify Mpoint (i, j).

[0111] The remaining webs to be polished are determined based on the cyclic planning method. The function of this step is to select the remaining web areas to be polished and ensure that the sum of the paths of the polishing head holes composed of the selected web areas is as small as possible. The specific method steps are: take the first polishing area as the starting point to plan the global a-1 areas to be polished. The corresponding planning evaluation function is

[0112]

[0113] Among them, c1 represents the motion path weight coefficient, which corresponds to the importance evaluation of the grinding head motion path index, F(T k)The first item represents the total length of the grinding head motion path, c2 represents the regional grinding difficulty coefficient, Di represents the center clustering of two adjacent grinding areas, and Hi represents the grinding difficulty coefficient; global planning is performed and converted into the F(Tk) extreme value problem. The second grinding area is obtained based on the global optimal solution, and so on. h represents the subscript of the center clustering measurement of two adjacent grinding areas, and a represents the number of global grinding areas.

[0114] Generate a grinding program for grinding: This step is to generate a numerical control program that can control the movement of the grinding head in the software according to the order of the evaluated grinding areas and the corresponding grinding head movement mode of each area, so as to accurately control the grinding head to move in a predetermined manner. The corresponding grinding area and grinding head path planning of the grinding area, such as Figure 3 shown.

[0115] The specific grinding process is as follows: based on the generated grinding program, the grinding head is driven to complete the grinding process of a grinding area, and the first grinding of each area is completed. After the first grinding, there may be problems such as insufficient grinding and incomplete grinding, so further grinding is required, and judgment is made based on the re-inspection of the grinding area.

[0116] The re-inspection of the polished area is as follows: the re-inspection steps are the same as those mentioned above, and the polished area is evaluated. If the parameters corresponding to the characteristics of the knife marks and step differences in the area that need to be polished do not meet the requirements, further polishing is required. The corresponding areas for further polishing are only selected from the a first polishing areas that have not been polished in place, until the quality of all a areas meets the requirements.

[0117] Example 2

[0118] 1. A method for identifying and locating the boundary area of ​​a structure to be polished. The main function of the steps of this method is to locate the boundary of the structure to be polished based on the structure image obtained by the camera. The specific steps are: first, the structure to be polished is imaged based on a high-resolution three-channel fixed-position camera. The fixed-position camera is equipped with RGBD data acquisition capability, D represents depth information, RGB represents image information, and the original analysis image O is obtained; a global coarse positioning recognition data model for the structure to be polished is constructed to complete the positioning of the structure to be polished in the original analysis image O. The global coarse positioning recognition data model is

[0119]

[0120] The goal of mathematical model learning is to minimize the objective function M({p i},{t i}), the loss function includes two parts, namely, the classification of tooling background and structural parts, and the regression loss of the category to be polished and the rough positioning border; based on M({p i},{t i}) The corresponding image O is roughly identified and positioned, and the corresponding result coordinates are {t x ,t y ,t w ,t h}; based on the result coordinates, {t x ,t y ,t w ,t h}Further cropping and extraction of image O obtains a new image P; the processed image P is smoothed with different weights, and after processing, the Gaussian difference D(x, y, δ) corresponding to the adjacent weight difference images is calculated, and the key point K in the image after Gaussian difference processing is found, and the Gaussian difference D(x, y, δ) is interpolated by Taylor series expansion to correct the key point and obtain the corrected point KP. Based on KP, the area to be polished in image P is precisely located, that is, the area where the structural parts to be polished are located is accurately located through the image obtained by the fixed-position camera and based on the designed algorithm. The corresponding result of the same scale as image O is represented by Obj.

[0121] 1.1 As described above, the imaging cameras for the corresponding structural members to be polished are arranged as shown in Figure 2. The arrangement and imaging method meet the following conditions: The cameras for imaging the structural members to be polished include a fixed-position camera, a horizontally adjustable moving-position camera 1, a horizontally adjustable moving-position camera 2, and a horizontally adjustable moving-position camera 3. The imaging axis center corresponding to the fixed-position camera is concentric with the center where the structural members to be polished are placed and the geometric center of the structural members to be polished, and there may be a distance deviation <dis-o>. The fixed-position camera is located directly above the structural members to be polished; The horizontally adjustable moving-position camera 1 is configured to move within the imaging plane P1. The corresponding pitch angle of the horizontally adjustable moving-position camera 1 is α1. There are num imaging docking points Point(i) on the circumference where the corresponding imaging plane P1 is located, that is, the corresponding i ∈ [1, num]. The circumferential radian interval between adjacent imaging docking points Point(i) and Point(i + 1) is equal and is 360 / num. And each camera only images at the imaging docking point Point(i) with the corresponding pitch angle α1; The horizontally adjustable moving-position camera 2 is configured to move within the imaging plane P2. The corresponding pitch angle of the horizontally adjustable moving-position camera 2 is α2. There are num imaging docking points Point(i) on the circumference where the corresponding imaging plane P2 is located, that is, the corresponding i ∈ [1, num]. The circumferential radian interval between adjacent imaging docking points Point(i) and Point(i + 1) is equal and is 360 / num. And each camera only images at the imaging docking point Point(i) with the corresponding pitch angle α2; The horizontally adjustable moving-position camera 3 is configured to move within the imaging plane P3. The corresponding pitch angle of the horizontally adjustable moving-position camera 3 is α3. There are num imaging docking points Point(i) on the circumference where the corresponding imaging plane P3 is located, that is, the corresponding i ∈ [1, num]. The circumferential radian interval between adjacent imaging docking points Point(i) and Point(i + 1) is equal and is 360 / num. And each camera only images at the imaging docking point Point(i) with the corresponding pitch angle α3; The function of the fixed-position camera is to obtain the global image of the structural members to be polished to achieve the boundary region positioning of the structural members. The functions of the horizontally adjustable moving-position camera 1, the horizontally adjustable moving-position camera 2, and the horizontally adjustable moving-position camera 3 are to obtain the images of the web regions to be polished at different angles, and based on the obtained image fusion, it is judged whether the corresponding region needs to be polished. The distribution of the horizontally adjustable moving-position camera 1, the horizontally adjustable moving-position camera 2, and the horizontally adjustable moving-position camera 3 has the characteristic of evenly dividing the hemisphere latitude.

[0122] 1.2 As mentioned above, the placement strategy of the corresponding structural parts to be polished is: after the structural parts are placed, the corresponding polishing area of ​​the structural parts to be polished faces upwards. If the structural parts need to be polished on both sides, one side is selected to be polished first, and then the other side is modeled after polishing is completed; after placement, the geometric center of the structural parts is fixed in the position of the entire workbench and is coaxial with the fixed position camera to ensure the integrity of the structural parts area.

[0123] 1.3 As mentioned above, the corresponding mathematical model M({p i},{t i}), in order to prevent the influence of BX and GT position and size on the regression scale, the BX coordinate needs to be normalized, that is:

[0124]

[0125] Among them, p x is the x value of the center of the prediction box; p y is the y value of the center of the prediction box; is the x value of the center of the standard box GT; is the y value of the center of the standard box; b x is the x value of the center of the potential target area box; b y is the y value of the center of the prediction box; p w is the predicted box width; p h is the predicted box height value; b w is the center width of the prediction box; b h is the height value of the potential target area; is the standard box width value; is the standard frame height value; t x ,t y ,t w ,t h and Represents the normalized center coordinates of BX and the standard box and the width and height of the box respectively. i is the index of the center of the predicted box; p i Indicates the predicted probability that the center of the prediction box is the target; Indicates the values ​​corresponding to the background and foreground, the background is 0 and the foreground is 1; With t i Respectively represent the coordinates of the GT enclosed and the coordinates of the prediction box, where t i is composed of {t x ,t y ,t w ,t h} composed of coordinate vectors; It represents the logarithmic loss corresponding to the identification and positioning of the workbench background and the area to be polished, that is: represents the prediction box coordinate regression deviation loss, and the corresponding mathematical model can be expressed as Where R represents the loss evaluation function, let In the formula Indicates that there is only a prospect There is a loss in coordinate position deviation. The output results of the classification model and the regression model are respectively expressed as {p i} and {t i}, while N cls represents normalization processing, N reg Introduce the balance weight λ and implement normalization. cls performs normalization on N cls is the mini-batch value, and the reg term is normalized by N reg The value is the number of predicted box center positions, where the value corresponding to the balance weight λ is a positive integer.

[0126] The multi-classifier mathematical model is set as:

[0127]

[0128] Since the above classifier formula is an exponential operation, if the value is very large, it may cause an exponential explosion problem. The optimized processing method is to multiply the numerator by the constant C and the denominator by the constant C at the same time, that is:

[0129]

[0130] All values ​​add up to 1, which ensures that each output result can be transformed to the [0,1] interval.

[0131] 2. Construct a mapping table Table(x,y) between the regional position and the physical scale. The table is used to record the actual physical distance data Dis(point(x,y),C(x,y)) from each point point(x,y) in the images O and Obj to the geometric center / center of the structure to be polished C(x,y). The value of the physical length can provide a basis for the positioning of the corresponding features in the subsequent polishing area and the design model, because the image O and Obj are in different coordinate systems with the design model, and the actual physical distance data Dis(point(x,y),C(x,y)) is the only feasible way to realize the association of the polishing area feature mapping. The mapping table is the key to realizing data mapping. The imaging area and size of the fixed position camera remain unchanged, the imaging angle of the camera remains unchanged, and the resolution and other parameters are also fixed. Therefore, it is ensured that the actual spatial distance corresponding to each image obtained by the fixed position camera is small and will not affect the actual feature mapping.

[0132] 3. The joint correction strategy of the placement posture of the structural parts to be polished and the design digital model. The purpose of this step is to correct the placement posture of the structural parts in the design digital model based on the posture information of the structural parts, so that the posture of the structural parts in the digital model is consistent with the posture of the structural parts to be polished, which is conducive to the subsequent positioning of the area to be polished in the digital model, and at the same time, it is convenient to generate an automatic polishing program that can control the movement of the polishing head. The specific steps of the joint correction strategy are: In the identification result diagram Obj of the structural parts to be polished, 4 directions are set, namely vertically upward, downward, horizontally left and right, and the reference point of the direction is The center of Obj, i.e. (rows / 2, cols / 2), is used as a reference, where row represents the number of rows of Obj, and cols represents the number of columns of Obj. In the design digital model software, the digital model is adjusted based on the corresponding features of Obj vertically upward, downward, horizontally left and right in the recognition result diagram, so that the features presented on the screen are consistent with the features in Obj. The consistent interpretation is that the contour features presented by the corresponding structural parts remain basically the same as the contour features in Obj, and the corresponding magnification, i.e. the size of the structural parts features presented in the screen direction, may be different.

[0133] 4. After completing the identification and positioning of the boundary area of ​​the structural part to be polished, it is necessary to further locate the potential web area that needs to be polished. First, complete the identification and judgment of the web area in Obj. The corresponding specific steps are: first, align the RGB data and D data obtained by the fixed-position camera to ensure that the coordinate systems of the two types of data obtained coincide, that is, after the alignment, the corresponding coordinates of the RGB data and D data obtained by the same object in the two coordinate systems are consistent; because the fixed-position camera is equipped with the ability to collect RGBD data information, start the camera's depth D information collection function to obtain D data information; based on the structural part boundary obtained in Obj, separate the structural part data and the background area (non-structural part) data from the D information data; set the background area part of the separated D information data to 0, while the data in the structural part area remains unchanged; because the obtained D data is in matrix form, the length and width of the matrix are the same as Obj, read the data Data(i,j) corresponding to each point of the structural part in the D data, i represents the corresponding number of rows in the data D, j represents the corresponding number of columns in the data D, and all Data num The set of (i,j) is denoted as Set(Data num (i,j)), num represents the total number of data points corresponding to the structural parts in the D data; the calculation set Set(Data num Each Data in (i,j)) num (i,j) values ​​are sorted and the minimum value Min(Data num (i,j)) and the maximum value Max(Data num (i,j)); Set(Data numAll Data in (i,j)) num (i,j) interval transformation, the corresponding transformation formula is: T(Data num (i,j))=255*|Min(Data num (i,j))-Data num (i,j)| / (Max(Data num (i,j))-Min(Data num (i,j))), the corresponding data after transformation is T-Set(Data num (i,j)); based on the first-order difference, complete T-Set(Data num The corresponding secondary transformation formula is: TData num (i,j)=(|TData num (i,j)-TData num (i,j+1)|+|TData num (i,j)-TData num (i+1,j)|) / 2, where TData num (i,j) represents the set T-Set (Data num (i,j)) corresponding to each point data; set the three-transformation data processing threshold T-threshold, if T-Set(Data num (i,j))≤T-threshold, then the corresponding T-Set(Data num (i,j))=0, otherwise it remains unchanged; the D data information after three transformations is transformed into an image to obtain the image Dimg, which is a single-channel grayscale image; based on the designed random assignment, different methods are used to extract the areas with the same grayscale in the image Dimg. The specific method is: read the grayscale value corresponding to each point in the image Dimg and change its value DV(i,j), DV(i,j)={Rv=(Num(random)%255),Gv=(Num(random)%255)and Gv!=Rv,Bv(Num(random)

[0134] %255)and Bv!=Rv and Bv!=Gv}, if Gv=Rv, then regenerate Gv, if Bv=Rv, then regenerate Bv, if Bv=Gv, then regenerate Bv, until the obtained Rv, Gv, and Bv are all different, where Rv represents the value assigned to the newly added channel at the midpoint (i, j) of the image Dimg, where Gv represents the value assigned to the newly added channel at the midpoint (i, j) of the image Dimg, where Bv represents the value assigned to the newly added channel at the midpoint (i, j) of the image Dimg, and Num(random) represents the generation of a random positive integer; read the Rv, Gv, and Bv values ​​corresponding to each data point in the image Dimg, if the values ​​corresponding to the points separated by 1 unit distance are equal, then such data points can be classified into one category, extract the area composed of such points that satisfy a class relationship, and input the classification neural network for category division, the classification network belongs to the two categories, namely the web and the non-web; the area identified as the web by the network is the completed web area to be polished Area k (i, j) positioning step, k represents the number of successfully identified webs, and (i, j) represents the center coordinates of the corresponding web.

[0135] 5. Judgment of the combination of images of multiple imaging angles of a single web. The purpose of this step is to further judge which webs need to be polished and which do not need to be polished based on the web areas that have been positioned, so as to improve the polishing efficiency of the structural parts. The specific implementation method steps are as follows: start the horizontally adjustable mobile position camera 1, the horizontally adjustable mobile position camera 2, and the horizontally adjustable mobile position camera 3, and the corresponding imaging angle pitch angles are α1, α2, and α3 respectively, and the num imaging stop points Point(i) corresponding to the imaging plane P1, the imaging plane P2, and the imaging plane P3 are respectively imaged for each Area k (i, j) is imaged, that is, the number of image data obtained for each web corresponding to k is 3*num, and the corresponding images are recorded as I1(h), I2(h), and I3(h), respectively, where h∈[1,num] represents the corresponding image number; the obtained I1(h), I2(h), and I3(h) totaling 3*num image data are evaluated based on the following designed formula:

[0136]

[0137] This formula can reflect and evaluate the high-frequency characteristics of the image due to large differences such as tool marks and step differences. Theoretically, for I1(h), the larger the Value, the more it indicates that there are tool marks waiting to be polished. Analyze and compare the 3*num calculated Value values, and sort them from large to small. If there are β corresponding Value values ​​greater than the set lower limit, it indicates that the corresponding web area has tool marks and other features that need to be polished. The area to be polished is marked as G(m), where m represents the number of areas that need to be polished.

[0138] 6. Identification and positioning of the transition area of ​​the cutting mark. The purpose of this step is to determine the position information of the web area to be polished containing the cutting mark in the polishing tool and the three-dimensional digital model. The position information on the polishing tool can ensure the accurate positioning of the web polishing area during polishing, and the position information in the three-dimensional digital model can ensure the accuracy of the polishing head movement program generated in the digital model. The implementation steps are: based on the polishing area marked as G(m), the coordinates of the web area to be polished are determined. Since the acquired data D is obtained through a series of transformations to obtain G(m), D and G have the same image coordinate system, that is, G(m) determines that the web area to be polished has a one-to-one correspondence with D, and D and O are acquired based on the camera at the same position. Therefore, the coordinates of the web area to be polished in O can be obtained based on the G(m) mapping, and at the same time, the coordinates of the web to be polished in the digital model are completed to obtain an accurate polishing tool path.

[0139] 7. Refined regional data scanning. The purpose of this step is to re-judge the coordinate area of ​​the web area to be polished determined by G(m). While avoiding invalid scanning, it can increase the accuracy of the preliminary determination of the area to be polished. The implementation steps are: G(m) determines the coordinates of the web area to be polished and completes the position information Loc(m) required for polishing in the digital model. Based on the position information Loc(m), the refined scanning movement trajectory and path are set. According to the generated path, a cyclic scanning method is used to obtain the accurate data PData(m) of each scanning area. After processing m areas, the data PData(m) is analyzed, and the areas that do not need to be polished are removed from the m areas, and the areas that need to be polished are retained. The secondary judgment of the candidate areas of the web to be polished based on G(m) is completed. The web area to be polished is finally determined by DM. a (i, j) represents, a represents the number of areas that need to be polished in the end, and (i, j) represents the coordinates of the corresponding area in G.

[0140] 8. The final web area grinding boundary data is obtained, and the web area to be polished is determined by DM. a(i, j), complete the mapping of the grinding area a in the digital model, and calculate the boundary data Cdata corresponding to a in the digital model. The necessity of calculating the boundary data Cdata is that the movement boundary of the grinding head can be strictly limited based on the boundary data to avoid grinding the area that does not need to be polished, which will cause the size of the structural parts to exceed the tolerance and be scrapped.

[0141] 9. The first web grinding area planning, the purpose of this step is to determine the first grinding web and calculate the DM of each grinding area a (i,j) The distance data between the motion coordinates of the grinding head and the structural parts can avoid the spatial interference problem caused by the unreasonable setting of the grinding path. The first grinding area planning method is: the distance dis from the geometric center of a area to be polished to the initial grinding head, where the grinding area with the smallest distance dis from the grinding head is the first grinding area. This planning method can ensure that the invalid stroke of the grinding head is as short as possible, thereby improving the grinding efficiency. The design of the grinding head movement method corresponding to the grinding area aims to be efficient and short the total movement path of the grinding head, thereby reducing the grinding time consumption while ensuring the grinding quality, and improving the grinding quality. The specific implementation method is: the radius DR of the grinding head is used to calculate the diameter of the enclosed area corresponding to the grinding. The enclosed area calculation method is as follows:

[0142]

[0143] Where r is the actual radius of the grinding head, r r represents the corresponding grinding head annular surface radius, w represents the grinding head side deflection angle, and λ represents the grinding head forward inclination angle; the number of optimized grinding path rows is designed, and the number of optimized grinding area path functions can be expressed as Where Mi represents the grinding path length of Nv grinding areas, Sj represents the grinding path length of the corresponding closed area boundary, and Fz represents the auxiliary path length traversed by the grinding head; the generation of the grinding head processing trajectory is completed based on the path optimization function, and the generation of the grinding head point data is achieved based on the boundary data Cdata.

[0144] 10. Spatial non-interference distance measurement method: Based on the first grinding area obtained, calculate its corresponding boundary data point Bdata(i,j), and at the same time calculate the radius data GR corresponding to the rigid rotation area of ​​the grinding head in the radial direction during grinding, and the point Mpoint(i,j) where the center of the grinding head moves. Directly calculate the Euclidean distance DMB between point Mpoint(i,j) and point Bdata(i,j). If the Euclidean distance DMB between point Mpoint(i,j) and point Bdata(i,j) in the plane is DMB≤GR, the flag is marked as 0, and DMB>GR is marked as 1.

[0145] 11. According to the optimal spatial non-interference distance measurement method, the workpiece interference is judged to correct the grinding path. If there is spatial grinding interference flag = 0, the path Mpoint (i, j) corresponding to the interference area is directly modified. If there is no interference flag = 1, there is no need to modify Mpoint (i, j).

[0146] 12. Determine the remaining webs to be polished based on the cyclic planning method. The function of this step is to select the remaining web areas to be polished and ensure that the sum of the paths of the polishing head holes composed of the selected web areas is as small as possible. The specific method steps are: take the first polishing area as the starting point to plan a-1 areas to be polished globally. The corresponding planning evaluation function is

[0147]

[0148] Among them, c1 represents the motion path weight coefficient, and its corresponding meaning is the importance evaluation of the grinding head motion path index. The first item of F(Tk) represents the total length of the grinding head motion path, c2 represents the regional grinding difficulty coefficient, Di represents the center cluster of two adjacent grinding areas, and Hi represents the grinding difficulty coefficient. Global planning is performed and converted into the F(Tk) extreme value problem. The second grinding area is obtained based on the global optimal solution, and so on.

[0149] 13. Generate a grinding program for grinding. The purpose of this step is to generate a CNC program that can control the movement of the grinding head in the software based on the order of the evaluated grinding areas and the corresponding grinding head movement mode of each area, so as to accurately control the grinding head to move in a predetermined manner. The corresponding grinding area and grinding head path planning of the grinding area, such as Figure 3 shown.

[0150] 14. Perform grinding processing. Drive the grinding head based on the generated grinding program to complete the grinding processing of a grinding area and complete the first grinding of each area. After the first grinding, there may be problems such as insufficient grinding and incomplete grinding, so further grinding is required and judgment is made based on the re-inspection of the grinding area.

[0151] 15. Re-inspection of the polished area. The re-inspection steps are the same as those mentioned above. The polished area is evaluated. If the parameters corresponding to the features to be polished, such as the knife marks and step differences in the area, do not meet the requirements, further polishing is required. The corresponding areas for further polishing are only selected from the a first polishing areas that are not polished in place, until the quality of all a areas meets the requirements.

[0152] The above description is only a preferred embodiment of the present invention and does not constitute any form of hindrance to the present invention. Any simple modification or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A decision-making method for automatic grinding planning of aircraft structural parts, characterized in that: It includes the following steps: Step 1. Position the structural member to be polished and the web area; Step 2. Combine and judge images of a single web at multiple imaging angles; Step 3. Identify and position the transition area of the tool joint mark; Step 4. Scan the refined data of the area; Step 5. Obtain the grinding boundary data of the final web area; Step 6. Plan the first web grinding area; Step 7. Calculate the non-interfering distance in space; Step 8. Determine the remaining webs to be polished based on the cyclic planning method; Step 9. Generate a grinding program for grinding; Step 10. Carry out the grinding process; Step 11. Re-inspect the grinding area.

2. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 1 is characterized in that: The positioning of the structural member to be polished and the web area includes: Image the structural member to be polished based on a high-resolution three-channel fixed-position camera. The fixed-position camera is configured with the ability to collect RGBD data, where D represents depth information and RGB represents image information, to obtain the original analysis image O; construct a global rough positioning and recognition data model for the structural member to be polished to complete the positioning of the structural member to be polished in the original analysis image O. The global rough positioning and recognition data model for the rigger is: The goal of mathematical model learning is to minimize the objective function M({p i },{t i }), represents the logarithmic loss of target and non-target categories, represents the prediction box regression loss, N cls Represents the Minni-batch value, N reg represents the number of regression rectangles, λ represents the balance weight, and p i Indicates the predicted probability of Anchor as the target; With t i Respectively represent the coordinates of the ground truth and the coordinates of the prediction box. Represents the values ​​corresponding to the background and foreground. The loss function includes two parts: the classified tooling background and structural parts, as well as the loss of the category to be polished and the regression loss of the rough positioning border.

3. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 2 is characterized in that: Based on M({p i },{t i }) The corresponding image O is roughly identified and positioned, and the corresponding result coordinates are {t x ,t y ,t w ,t h } indicates that t x is the normalized central horizontal coordinate of the standard frame, t y is the normalized central ordinate of the standard frame, t w is the width of the box, t h is the height of the box; based on the resulting coordinates {t x ,t y ,t w ,t h }Further cropping and extraction of image O obtains a new image P; smoothing of the processed image P with different weights is performed, and after processing, the Gaussian difference D(x, y, δ) corresponding to the adjacent weight difference images is calculated, where x is the value of the image pixel in the X-axis direction, y is the value of the image pixel in the Y-axis direction, and δ is the corresponding standard deviation. The key point K in the image after Gaussian difference processing is found, and the Gaussian difference D(x, y, δ) is interpolated through Taylor series expansion to correct the key point and obtain the corrected point KP. Based on KP, the area to be polished in image P is precisely located, and the corresponding result of the same scale as image O is represented by Obj.

4. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 1 is characterized in that: In Step 1, the imaging cameras for the corresponding structural member to be polished are arranged to meet the following conditions: the imaging cameras for the structural member to be polished include a fixed-position camera, a horizontally adjustable moving-position camera 1, a horizontally adjustable moving-position camera 2, and a horizontally adjustable moving-position camera 3. The imaging axis center corresponding to the fixed-position camera is concentric with the center where the structural member to be polished is placed and the geometric center of the structural member to be polished, with a distance deviation <dis-o>, where dis-o is the set distance deviation threshold between the center where the structural member is placed and the geometric center of the structural member to be polished, and the fixed-position camera is located directly above the structural member to be polished; The horizontally adjustable mobile camera 1 satisfies the requirement of moving within the imaging plane P1, and the corresponding horizontally adjustable mobile camera 1 pitch angle α1, there are num imaging stop points Point(i) on the circumference of the corresponding imaging plane P1, that is, the corresponding i∈[1,num], the circumferential arc interval between adjacent imaging stop points Point(i) and Point(i+1) is equal to 360 / num, and each camera only performs imaging at the imaging stop point Point(i) corresponding to the pitch angle α1; the horizontally adjustable mobile camera 2 satisfies the requirement of moving within the imaging plane P2, and the corresponding horizontally adjustable mobile camera 2 pitch angle α2, there are num imaging stop points Point(i) on the circumference of the corresponding imaging plane P2, that is, the corresponding i∈[1, num], the circumferential arc interval between adjacent imaging stop points Point(i) and Point(i+1) is equal to 360 / num, and each camera only performs imaging at the imaging stop point Point(i) corresponding to the pitch angle α2; the horizontally adjustable movable camera 3 satisfies the movement within the imaging plane P3, and the corresponding horizontally adjustable movable camera 3 pitch angle α3, there are num imaging stop points Point(i) on the circumference of the corresponding imaging plane P3, that is, the corresponding i∈[1,num], the circumferential arc interval between adjacent imaging stop points Point(i) and Point(i+1) is equal to 360 / num, and each camera only performs imaging at the imaging stop point Point(i) corresponding to the pitch angle α3.

5. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 4 is characterized in that: The corresponding placement strategy for the structural parts to be polished is as follows: after the structural parts are placed, the corresponding polishing area of ​​the structural parts to be polished faces upwards. If the structural parts need to be polished on both sides, one side is selected to be polished first, and then the other side is modeled after polishing is completed. After placement, the geometric center of the structural parts is fixed in the position of the entire workbench and is coaxial with the fixed position camera to ensure the integrity of the structural parts area.

6. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 4 is characterized in that: The corresponding mathematical model M({p i },{t i }), in order to prevent the influence of BX and GT position and size on the regression scale, where BX is the model prediction area and GT is the actual result area, the BX coordinate needs to be normalized, that is: Among them, p x is the x value of the center of the prediction box; p y is the y value of the center of the prediction box; is the x value of the center of the standard box GT; is the y value of the center of the standard box; b x is the x value of the center of the potential target area box; b y is the y value of the center of the prediction box; p w is the predicted box width; p h is the predicted box height value; b w is the center width of the prediction box; b h is the height value of the potential target area; is the standard box width value; is the standard frame height value; t x ,t y ,t w ,t h and Respectively represent the normalized center coordinates of BX and the standard frame and the width and height of the frame; i is the index of the center of the prediction box; p i Indicates the predicted probability that the center of the prediction box is the target; Indicates the values ​​corresponding to the background and foreground, the background is 0 and the foreground is 1; With t i Respectively represent the coordinates of the GT enclosed and the coordinates of the prediction box, where t i is composed of {t x ,t y ,t w ,t h } composed of coordinate vectors; It represents the logarithmic loss corresponding to the identification and positioning of the workbench background and the area to be polished, that is: It represents the prediction box coordinate regression deviation loss, and the corresponding mathematical model can be expressed as Where R represents the loss evaluation function, let In the formula Indicates that there is only a prospect There is a loss of coordinate position deviation; The output results of the classification model and regression model are respectively expressed as {p i } and {t i }, while N cls represents normalization processing, N reg Introduce the balance weight λ and implement normalization. cls performs normalization on N cls is the mini-batch value, and the reg term is normalized by N reg The value is the number of predicted box center positions, where the value corresponding to the balance weight λ is a positive integer; The multi-classifier mathematical model is set as: The optimized processing method is to multiply the numerator by the constant C and the denominator by the constant C at the same time, that is: All values ​​add up to 1, which ensures that each output result can be transformed to the interval [0,1]; i represents the number of categories, j represents the number of categories output by the neural network, and f j Indicates the value corresponding to the jth category, y i represents the probability that the neural network predicts the output to be category i, f yi It indicates the value corresponding to the predicted i-th category, and x indicates the data input to the network; Construct a mapping table Table(x,y) of the regional position and physical scale, and record the actual physical distance data Dis(point(x,y),C(x,y)) from each point point(x,y) in the images O and Obj to the geometric center / placement center C(x,y) of the structure to be polished.

7. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 6 is characterized in that: The placement posture of the structural parts to be polished is jointly corrected with the design digital model. Based on the posture information of the structural parts, the placement posture of the structural parts in the design digital model is corrected to make the posture of the structural parts in the digital model consistent with the posture of the structural parts to be polished. The specific steps of the joint correction strategy are as follows: in the recognition result diagram Obj of the structural parts to be polished, 4 directions are set, namely vertically upward, downward, horizontally left and right. The reference point of the direction is the center of Obj, that is, (rows / 2, cols / 2) as the reference, where row represents the number of rows of Obj, and cols represents the number of columns of Obj; in the design digital model software, the digital model is adjusted based on the corresponding features of Obj vertically upward, downward, horizontally left and right in the recognition result diagram, so that the features presented on the screen are consistent with the features in Obj.

8. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 7 is characterized in that: To locate the potential web area that needs to be polished, first complete the identification and judgment of the web area in Obj. The corresponding specific steps are: first, align the RGB data and D data obtained by the fixed-position camera to ensure that the coordinate systems of the two types of data obtained coincide, start the camera depth D information acquisition function, and obtain D data information; based on the structural part boundary obtained in Obj, separate the structural part data and the background area (non-structural part) data from the D information data; set the background area part of the separated D information data to 0, while the data in the structural part area remains unchanged; since the obtained D data is in matrix form, the length and width of the matrix are the same as Obj, read the data Data(i,j) corresponding to each point of the structural part in the D data, i represents the corresponding number of rows in the data D, j represents the corresponding number of columns in the data D, and all Data num The set of (i,j) is denoted as Set(Data num (i,j)), num represents the total number of data points corresponding to the structural parts in the D data; the calculation set Set(Data num Each Data in (i,j)) num (i,j) values ​​are sorted and the minimum value Min(Data num (i,j)) and the maximum value Max(Data num (i,j)); Set(Data num All Data in (i,j)) num (i,j) interval transformation, the corresponding transformation formula is: T(Data num (i,j))=255*|Min(Data num (i,j))-Data num (i,j)| / (Max(Data num (i,j))-Min(Data num (i,j))), the corresponding data after transformation is T-Set(Data num (i,j)); based on the first-order difference, complete T-Set(Data num The corresponding secondary transformation formula is: TData num (i,j)=(|TData num (i,j)-TData num (i,j+1)|+|TData num (i,j)-TData num (i+1,j)|) / 2, where TData num (i,j) represents the set T-Set (Data num (i,j)) corresponding to each point data; set the three-transformation data processing threshold T-threshold, if T-Set(Data num (i,j))≤T-threshold, then the corresponding T-Set(Data num (i, j))=0, otherwise it remains unchanged; the D data information after three transformations is transformed into an image to obtain an image Dimg, which is a single-channel grayscale image; Based on the designed random assignment, different methods are used to extract the areas with the same grayscale in the image Dimg. The specific method is: read the grayscale value corresponding to each point in the image Dimg and change its value DV(i,j), DV(i,j)={Rv=(Num(random)%255),Gv=(Num(random)%255)and Gv!=Rv,Bv(Num(random)%255)and Bv!=Rv and Bv! =Gv}, if Gv=Rv, then regenerate Gv, if Bv=Rv, then regenerate Bv, if Bv=Gv, then regenerate Bv, until the obtained Rv, Gv, and Bv are all different, where Rv represents the value assigned to the newly added channel at the midpoint (i, j) of the image Dimg, where Gv represents the value assigned to the newly added channel at the midpoint (i, j) of the image Dimg, where Bv represents the value assigned to the newly added channel at the midpoint (i, j) of the image Dimg, and Num(random) represents the generation of a random positive integer; read the Rv, Gv, and Bv values ​​corresponding to each data point in the image Dimg, if the values ​​corresponding to the points separated by 1 unit distance are equal, then such data points can be classified into one category, extract the area composed of such points that satisfy a class relationship, and input the classification neural network for category division, the classification network belongs to the two categories, namely the web and the non-web; the area identified as the web by the network is the completed web area to be polished Area k (i, j) positioning step, k represents the number of successfully identified webs, and (i, j) represents the center coordinates of the corresponding web.

9. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 1 is characterized in that: The specific judgment of the combination of multiple imaging angles of a single web is as follows: start the horizontally adjustable mobile camera 1, the horizontally adjustable mobile camera 2, and the horizontally adjustable mobile camera 3, and the corresponding imaging angles are α1, α2, and α3, and the num imaging stop points Point(i) corresponding to the imaging plane P1, the imaging plane P2, and the imaging plane P3 are respectively imaged for each Area k (i, j) is imaged, and the number of image data obtained for each web corresponding to k is 3*num. The corresponding images are recorded as I1(h), I2(h), and I3(h), respectively, where h∈[1,num] represents the corresponding image number; the obtained I1(h), I2(h), and I3(h) totaling 3*num image data are evaluated based on the following designed formula: For I1(h), the larger its value is, the more it indicates that there is a feature of cutting marks waiting to be polished; Analyze and compare the 3*num calculated Value values, and sort them from large to small. If there are β corresponding Value values ​​greater than the set lower limit, it indicates that the corresponding web area has features such as knife marks and needs to be polished. The area to be polished is marked as G(m), where m represents the number of areas to be polished; cols and rows represent the number of columns and rows corresponding to I1(h), I2(h), and I3(h), respectively, i and j represent the subscripts of the pixel points along the row and column directions, respectively, f corresponds to the grayscale value at the (i, j) position, and x and y represent the horizontal and vertical directions, respectively.

10. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 1, characterized in that: The identification and positioning of the transition area to the tool mark is specifically as follows: based on the grinding area marked as G(m), the coordinates of the web area to be ground are determined. Since the acquired data D is obtained through a series of transformations to obtain G(m), D and G have the same image coordinate system, that is, G(m) determines that the web area to be ground has a one-to-one correspondence with D, and D and O are acquired based on the camera at the same position. Therefore, the coordinates of the web area to be ground in O can be obtained based on G(m) mapping, and at the same time, the coordinates of the web to be ground in the digital model can be completed to obtain an accurate grinding tool path.

11. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 1, characterized in that: The specific process of regional refined data scanning is as follows: G(m) determines the coordinates of the web area to be polished and completes the position information Loc(m) required for polishing in the digital model. Based on the position information Loc(m), the refined scanning movement trajectory and path are set. According to the generated path, a circular scanning method is used to obtain the accurate data PData(m) of each scanning area. After processing m areas, the data PData(m) is analyzed, and the areas that do not need to be polished are removed from the m areas, and the areas that need to be polished are retained. The secondary judgment of the candidate areas of the web to be polished based on G(m) is completed, and the web area to be polished is finally determined by DM. a (i, j) represents, a represents the number of areas that need to be polished in the end, and (i, j) represents the coordinates of the corresponding area in G.

12. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 1, characterized in that: The final web area grinding boundary data acquisition is specifically as follows: based on the final determined web area to be ground, use DM a (i, j), complete the mapping of the polishing area a in the digital model, and calculate the boundary data Cdata corresponding to a in the digital model.

13. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 1, characterized in that: The first web grinding area planning is as follows: for each of the a areas to be ground, the distance dis from the geometric center to the initial grinding head is calculated. The grinding area with the smallest distance dis from the grinding head is the first grinding area. The diameter of the enclosed area corresponding to the grinding is calculated by the grinding head radius DR. The enclosed area calculation method is: Where r is the actual radius of the grinding head, r r represents the corresponding grinding head annular surface radius, w represents the grinding head side deflection angle, and λ represents the grinding head forward inclination angle; the number of optimized grinding path rows is designed, and the number of optimized grinding area path functions can be expressed as Among them, M i represents the grinding path length of Nv grinding areas, S j represents the length of the grinding path of the corresponding closed area boundary, F z Represents the length of the auxiliary path traversed by the grinding head; Nk is the number of grinding areas, i represents the measurement subscript of the grinding path length, j represents the measurement subscript of the grinding path length at the closed area boundary, and z represents the measurement subscript of the auxiliary path length traversed by the grinding head. The generation of the grinding head processing trajectory is completed based on the path optimization function, and the generation of the grinding head point data is based on the boundary data Cdata.

14. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 1, characterized in that: The method for measuring the spatial non-interference distance is as follows: based on the first obtained grinding area, the corresponding boundary data point Bdata(i,j) is calculated, and at the same time, the radius data GR corresponding to the rigid rotation area of ​​the grinding head in the radial direction during grinding is calculated, and the point Mpoint(i,j) where the center of the grinding head moves is directly calculated. The Euclidean distance DMB between the point Mpoint(i,j) and the point Bdata(i,j) is directly calculated. If the Euclidean distance DMB between the point Mpoint(i,j) and the point Bdata(i,j) in the plane is less than or equal to GR, the flag is marked as 0, and DMB>GR is marked as 1. According to the optimal spatial non-interference distance measurement method, the workpiece interference is judged to correct the grinding path. If there is spatial grinding interference flag = 0, the path Mpoint (i, j) corresponding to the interference area is directly modified. If there is no interference flag = 1, there is no need to modify Mpoint (i, j).

15. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 1, characterized in that: The remaining webs to be polished are determined based on the cyclic planning method as follows: the first polishing area is taken as the starting point to plan a-1 areas to be polished globally. The corresponding planning evaluation function is: Among them, c1 represents the motion path weight coefficient, which corresponds to the importance evaluation of the grinding head motion path index, F(T k )The first item represents the total length of the grinding head motion path, c2 represents the regional grinding difficulty coefficient, Di represents the center clustering of two adjacent grinding areas, and Hi represents the grinding difficulty coefficient; global planning is performed and converted into the F(Tk) extreme value problem. The second grinding area is obtained based on the global optimal solution, and so on. h represents the subscript of the center clustering measurement of two adjacent grinding areas, and a represents the number of global grinding areas.

16. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 1, characterized in that: Generating a grinding program for grinding specifically includes: generating a numerical control program that can control the movement of a grinding head.

17. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 1, characterized in that: The grinding process is specifically performed as follows: based on the generated grinding program, the grinding head is driven to complete the grinding process of a grinding area, the first grinding of each area is completed, and the evaluation is performed based on the re-inspection of the grinding area.

18. The automatic grinding planning and decision-making method for aircraft structural parts according to claim 1, characterized in that: The re-inspection of the polished area is as follows: the polished area is evaluated. If the parameters corresponding to the characteristics of the knife marks and step differences that need to be polished in the area do not meet the requirements, further polishing is required. The corresponding areas for further polishing are only selected from the a first polishing areas that have not been polished in place, until the quality of all a areas meets the requirements.

Citation Information

Patent Citations

  • Automatic free-form surface polishing and grinding system with industrial robot and method

    CN106799738A

  • Robot flexible intelligent grinding system based on six-dimensional force sensor

    CN110281108A

  • Grinding path planning method based on 3D vision

    CN117408049A

  • Adaptive recognition and analysis method for vibration signals of numerical control machining tool with any tooth number

    CN117862953A

  • Systems and methods for sanding a surface of a structure

    US20200171620A1

Cited By

  • Robot dynamic polishing path planning method, device and equipment, storage medium and program product

    CN120206526A

  • Polishing control method and equipment for casting part and medium

    CN120551858A

  • Multi-process autonomous control method and system for factory based on Internet of Things

    CN120921278A