Reusability Evaluation Method and System for Nesting Tasks Based on Part Shape Similarity

Abstract

The present invention belongs to the field of optimized blanking of structural parts, and specifically discloses a method and system for evaluating the reusability of nesting tasks based on part shape similarity, which includes: obtaining a pre-trained twin neural network model; calculating the number of repeated parts in the target nesting task A and the historical nesting task B to obtain the nesting task repeatability; inputting the part graphics in task A and task B into the twin neural network model to obtain the shape similarity; correcting the shape similarity through the part size difference degree to obtain the part similarity matrix; pairing the parts in task A and task B according to the principle of maximum similarity, and calculating the nesting task similarity according to the pairing result; calculating the nesting task reusability according to the nesting task repeatability and the nesting task similarity. The method of the present invention can retrieve the historical task with the highest reusability for the target nesting task, lay a foundation for the transfer and reuse of nesting knowledge, and has good practical value.

Description

Technical Field

[0001] The present invention belongs to the field of optimized blanking of structural parts, and more specifically, relates to a method and system for evaluating the reusability of nesting tasks based on part shape similarity. Background Art

[0002] In the manufacturing industry, numerical control cutting technology occupies an important position and is widely used. During the numerical control cutting process, effectively optimizing nesting to improve the utilization rate of steel plates can reduce the manufacturing costs of enterprises and improve economic benefits.

[0003] Taking heavy industrial products represented by construction machinery as an example, the manufacturing production mode of these products is usually variable-variety batch production. Among them, products in the same series mostly adopt batch rotation production methods, and their parts also mostly adopt variant designs. Therefore, there is a certain degree of repeatability and similarity. However, the existing nesting algorithms cannot utilize this similarity, and each nesting requires recalculation. According to the above characteristics, in order to improve the efficiency of nesting, a reusable nesting mechanism can be introduced when designing the nesting for a new nesting task. That is, on the premise of repeated and similar nesting tasks, learn and reuse nesting knowledge such as the arrangement order of parts, rotation angle, and abutting position relationship. Before realizing the reuse of excellent past nesting knowledge and experience, the urgent problem to be solved is how to connect the current nesting task with the historical nesting task and evaluate the reusability of the front and back nesting tasks. Summary of the Invention

[0004] In view of the above defects or improvement requirements of the prior art, the present invention provides a method and system for evaluating the reusability of nesting tasks based on part shape similarity, aiming to reasonably evaluate the reusability of historical nesting tasks and meet the requirements for obtaining and reusing historical nesting knowledge in the reusable nesting mechanism.

[0005] To achieve the above object, according to one aspect of the present invention, a method for evaluating the reusability of nesting tasks based on part shape similarity is proposed, including the following steps:

[0006] S1. Obtain a pre-trained siamese neural network model;

[0007] S2. Calculate the number of repeated parts in the target nesting task A and the historical nesting task B to obtain the nesting task repeatability;

[0008] S3. Input the part graphics of the target nesting task A and the historical nesting task B into the pre-trained siamese neural network model in pairs to obtain the shape similarity; then correct the shape similarity through the part size difference degree to obtain the part similarity matrix;

[0009] S4. Based on the part similarity matrix, pair the parts in the target nesting task A and the historical nesting task B according to the principle of maximum similarity, and calculate the nesting task similarity based on the pairing results;

[0010] S5. Calculate the reusability of the nesting task based on the nesting task repeatability and the nesting task similarity.

[0011] As a further preference, the siamese neural network model is used for shape similarity measurement, which is constructed based on the convolutional neural network Vgg16 and pre-trained by the part shape data set.

[0012] As a further preference, when the siamese neural network model reads the part image, it first performs image preprocessing, and then performs image feature extraction and similarity measurement; the steps of image preprocessing are as follows:

[0013] Convert the part image into a binary image;

[0014] Calculate the zero-order origin moment and the first-order origin moment of the binary image, and calculate the image centroid coordinates (x c , y c );

[0015] Calculate the p+q order central moment W based on the image centroid coordinates pq :

[0016]

[0017] According to the p+q order central moment W pq , calculate the main axis direction angle θ of the image, and rotate the image clockwise by θ angle; the calculation formula of the main axis direction angle θ is as follows:

[0018]

[0019] where f(x,y) is the gray value of the image at the pixel coordinates (x,y), and p and q are the orders of the geometric moments.

[0020] As a further preference, in step S2, the calculation method of the nesting task repeatability D R is as follows:

[0021]

[0022] where m is the number of parts in task A, n is the number of parts in task B, and c is the number of repeated parts in task A and task B.

[0023] As a further preference, in step S3, the calculation method of the part size difference degree r is as follows:

[0024]

[0025] Among them, s1 and s2 are the areas of two parts under actual dimensions, and s1 ≤ s2.

[0026] As a further optimization, in step S3, the shape similarity is corrected by the part size difference degree to obtain a part similarity matrix, specifically including:

[0027] Based on the part size difference degree, calculate the part size gap correction coefficient

[0028] Furthermore, obtain the part similarity E between two parts N = δ·E W , where E W is the shape similarity output by the siamese neural network model;

[0029] For each part a in task A i and each part b in task B j calculate the part similarity between each pair to obtain a part similarity matrix M of size m×n ij ; i = 1, 2,..., m, j = 1, 2,..., n, where m and n are the total numbers of parts in task A and task B respectively.

[0030] As a further optimization, in step S4, the specific method for pairing parts is:

[0031] Based on the part similarity matrix, pair part a in task A i with part b in task B j so that each part is paired with a part with a high similarity as much as possible;

[0032] When the number of parts in task A and task B is not equal, add pseudo-parts to the group with fewer parts to make up the number of parts. The similarity between the added pseudo-parts and any part is 0, and accordingly, convert the part similarity matrix M ij to M ij '.

[0033] As a further optimization, in step S4, calculate the nesting task similarity according to the pairing result specifically as:

[0034] Introduce a variable y ij , and let:

[0035]

[0036] Establish the following mathematical model:

[0037]

[0038]

[0039] The Hungarian algorithm is used to solve the above mathematical model, and the sum H of the part similarities based on the pairing results is obtained. max , and then the nesting task similarity D is calculated. s :

[0040]

[0041] As a further optimization, in step S5, the calculation method of the reusability D of the nesting task is as follows:

[0042] D = β·D R +(1 - β)·D S

[0043] where D R is the nesting task repeatability, D s is the nesting task similarity, and β is the weight of the nesting task repeatability, and β ∈ (0, 1).

[0044] According to another aspect of the present invention, a nesting task reusability evaluation system based on part shape similarity is provided, including a processor, and the processor is used to execute the above nesting task reusability evaluation method.

[0045] Generally speaking, compared with the prior art through the above technical solutions conceived by the present invention, the following technical advantages are mainly possessed:

[0046] 1. The present invention proposes an effective method for quantitatively evaluating the reusability of nesting tasks, including two quantitative indicators, namely the nesting task repeatability and the nesting task similarity, so as to evaluate the reusability of one or more historical tasks with similarity, laying a foundation for the transfer and reuse of knowledge under the reusable nesting mechanism.

[0047] 2. The twin neural network model established by the present invention realizes the measurement of the shape similarity of nesting parts, and corrects its rotation angle by calculating the main axis direction of the input image, enhancing the rotation invariance of the model. Specifically, through the calculation and adjustment of the main axis direction angle, the rotation angle of the input part graph is reset to a fixed orientation, so that the negative impact of the rotation of the part graph on the twin neural network model can be reduced to improve the rotation invariance of the model.

[0048] 3. The present invention proposes a part size correction coefficient according to the difference in part area, and uses this coefficient to correct the part shape similarity, which can effectively solve the influence of part size on similarity discrimination in nesting tasks. Description of the Drawings

[0049] Figure 1It is the flowchart of the evaluation method for the reusability of nesting tasks based on part shape similarity measurement in the preferred embodiment of the present invention;

[0050] Figure 2 It is the schematic diagram of the structure of the Siamese neural network model in the preferred embodiment of the present invention;

[0051] Figure 3 In (a) - (c), it is the effect diagram of the graphic rotation direction correction of the Siamese neural network model in the preferred embodiment of the present invention;

[0052] Figure 4 In (a) - (f), it is the partial test results of the Siamese neural network model in the preferred embodiment of the present invention;

[0053] Figure 5 It is the schematic diagram of the calculation relationship between the part size difference degree r and the part size gap correction coefficient δ in the preferred embodiment of the present invention;

[0054] Figure 6 It is the part matching schematic diagram in two groups of nesting tasks in the preferred embodiment of the present invention;

[0055] Figure 7 It is the flowchart of the Hungarian algorithm adopted in the preferred embodiment of the present invention. Detailed implementation manners

[0056] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0057] A nesting task reusability evaluation method based on part shape similarity provided by an embodiment of the present invention. For the convenience of description, referring to Table 1, the meanings of the symbols introduced in the present invention are described as follows:

[0058] Table 1 Symbol definitions

[0059]

[0060] As Figure 1 shown, the nesting task reusability evaluation method includes the following steps:

[0061] S1. Establish a Siamese neural network model with the convolutional neural network Vgg16 as the core for shape similarity measurement, and construct a part shape data set to train the Siamese neural network model. As Figure 2As shown, the algorithm structure of the twin neural network model mainly includes three parts: graphic rotation direction correction, feature extraction network, and similarity measurement network.

[0062] S2. Input the part information in the target nesting task A and the historical nesting task B, count the number of duplicate parts in the target nesting task A and the historical nesting task B, and calculate the nesting task repeatability D R (A, B);

[0063] S3. Pairwise input the part drawings in task A and task B into the twin neural network model in step S1, and correct the shape similarity output by the model according to the part size difference to obtain the part similarity matrix M of the target nesting task A and the historical nesting task B;

[0064] S4. Based on the part similarity matrix M in step S3, pair the parts in the target task A and the historical task B according to the maximum similarity principle, and calculate the nesting task similarity D S (A, B);

[0065] S5. Combine the task repeatability D R (A, B) in step S2 and the nesting task similarity D S (A, B) in step S4 to calculate the nesting task reusability.

[0066] Furthermore, in step S1, for the twin neural network model used for part shape similarity measurement, its shape similarity measurement process includes the following sub-steps:

[0067] 1.1. Graphic rotation direction correction. The purpose of this step is to preprocess the input part drawing to enhance the rotation invariance of the model, as Figure 3 shown. The specific steps are as follows:

[0068] Step1: Read the part drawing and convert it into a binary image with a resolution of 128×128 pixels;

[0069] Step2: Calculate the centroid coordinates of the image; the specific method is:

[0070] For a single-channel grayscale image, its p+q-order geometric moment is:

[0071]

[0072] where f(x, y) is the grayscale value of the digital image at the pixel coordinates (x, y), and p and q are the orders of the geometric moments.

[0073] From equation (1), the zero-order origin moment and the first-order origin moment of the image can be obtained, and at the same time, the centroid coordinates of the image can be obtained:

[0074]

[0075] Among them, w 00 is the zero-order origin moment, w 10 and w 01 are the first-order origin moments, x c and y c are the coordinate values of the centroid of the image.

[0076] Step3: Calculate the (p + q)-order central moment W pq and the normalized central moment μ pq ;

[0077]

[0078]

[0079] Step4: Calculate the principal axis direction angle θ of the image, and rotate the image clockwise by θ degrees; the principal axis direction angle θ is calculated by equations (5) and (6):

[0080]

[0081]

[0082] After the calculation and adjustment of the principal axis direction angle, the rotation angle of the input part graph is reset to a fixed orientation. Therefore, the negative impact of the rotation of the part graph on the twin neural network model can be reduced to improve the rotation invariance of the model.

[0083] 1.2. Shape feature extraction. Shape feature extraction is completed by the feature extraction network G W The feature extraction network G W is a convolutional neural network based on the Vgg16 architecture.

[0084] After preprocessing such as rotation of the principal axis direction and size scaling of the input shape picture X, it is converted into a binary image with a size of 96×96 pixels, and then processed by the feature extraction network G W to output a one-dimensional feature vector G W (X) with a length of 4608.

[0085] 1.3. Shape similarity measurement. Input the graphs X1 and X2 to be compared. After being processed by the feature extraction network G W , finally output the feature vectors G W (X1) and G W (X2). Compare G W (X1) and G WAfter subtracting each item and taking the absolute value, the similarity vector S between X1 and X2 can be obtained as shown in Equation (7):

[0086] S i = |G W (X1) i - G W (X2) i | (7)

[0087] The purpose of taking the absolute value is to avoid errors caused by the different input orders of X1 and X2. After obtaining the similarity vector S, a two-layer fully connected network is used to judge the similarity. There is only one neuron at the output end of the second-layer fully connected network. After processing the finally output data with the Sigmoid function, the value range is mapped between 0 and 1 as the similarity between X1 and X2.

[0088] Further, in step S2, the nesting task repeatability D R (A, B) is calculated as follows:

[0089]

[0090] where D R is the repeatability between nesting task A and nesting task B, m is the number of parts in nesting task A, n is the number of parts in nesting task B, and c is the number of repeated parts between task A and task B.

[0091] As the nesting task repeatability, D R has the following three properties:

[0092] a) When c = 0, that is, when the number of repeated parts is zero, D R = 0;

[0093] b) When m = n = c, that is, when all parts are repeated, D R = 100%;

[0094] c) D R (A, B) = D R (B, A).

[0095] Further, step S3 includes the following sub-steps:

[0096] 3.1. Calculate the part size difference degree r:

[0097]

[0098] where s1 and s2 are the areas of the two parts to be compared under the actual size, assuming s1 ≤ s2

[0099] 3.2. Based on the part size difference degree r, the calculation relationship of the part size gap correction coefficient δ is shown in Equation (10):

[0100]

[0101] Its function image is as Figure 5 shown. During the actual nesting process, the positioning algorithm is very sensitive to the part size changes. Some small-scale size changes may also cause large disturbances to the part positioning. Therefore, when the part size difference degree r starts to increase from 0, as a penalty mechanism, the size correction coefficient δ should also decrease at a relatively fast rate, and the functional relationship in Equation (10) can better meet this characteristic.

[0102] 3.3. Part similarity E N :

[0103] E N = δ·E W (11)

[0104] Among them, E W is the part shape similarity output by the Siamese neural network.

[0105] 3.4. Calculate the part similarity matrix M:

[0106] M ij = E N (a i ,b j ) (12)

[0107] Among them, a i is the part (i = 1, 2,..., m) in the nesting task A, and b j is the part (j = 1, 2,..., n) in the nesting task B.

[0108] Furthermore, in step S4, the calculation process of the nesting task similarity D S (A, B) is as follows:

[0109] As Figure 6 shown, based on the part similarity matrix M, each part a i in the nesting task A is compared with the part b jPair them up and find the best matching strategy to pair each part with a part with a high similarity as much as possible. In practical applications, the number of parts in the front and back nesting tasks is not necessarily equal. For the case where the number of parts on both sides is not equal, "pseudo-parts" need to be added to the group with fewer parts to make up the number of parts. The similarity between the added pseudo-parts and any part is 0. Here, it is assumed that m < n, that is, the number of parts in task A is less. At this time, pseudo-parts need to be added to task A, and the original part similarity matrix M of size m×n is transformed into a part similarity matrix M′ of size n×n; introduce a variable y with a value of 0 or 1 ij , and let:

[0110]

[0111] Establish the mathematical model of the problem as shown in Equation (14):

[0112]

[0113] The Hungarian algorithm can be used to solve the model in Equation (14) to obtain the sum H of part similarities under the best matching relationship max .

[0114] Specifically, as Figure 7 shown, the specific steps of the Hungarian algorithm are as follows:

[0115] Step1: Row reduction; that is, subtract the minimum element of each row in matrix M′ from each row;

[0116] Step2: Column reduction; that is, subtract the minimum element of each column from each column after the row reduction transformation;

[0117] Step3: Try to match; after row reduction and column reduction, there are 0 elements in each row and each column of matrix M′. If there is only one independent 0 element in each row and each column, the matching relationship at the position of the 0 element is output as the optimal solution, otherwise go to Step4;

[0118] Step4: Draw lines covering 0s; draw the fewest horizontal and vertical lines to cover all 0 elements, subtract the minimum value from all the elements not covered by the lines, and add this minimum value to the elements at the intersections of the lines to update the matrix, and then go to step3.

[0119] Furthermore, according to the sum H of part similarities max , calculate the nesting task similarity:

[0120]

[0121] As the nesting task similarity, D s has the following three properties:

[0122] a) When the parts in task A and task B are completely dissimilar, the part similarity is 0, and H max = 0. At this time, D s = 0;

[0123] b) When all the similar parts in task A and task B are duplicate parts, since the similarity of duplicate parts is 1, at this time, H max is equal to the number of duplicate parts in the task,

[0124] The similarity of the nesting task is equivalent to the duplication degree of the nesting task. When all parts are completely duplicate, D S = D R = 100%;

[0125] c) D S (A, B) = D S (B, A).

[0126] Furthermore, in step S5, the reusability of the nesting task is calculated as follows:

[0127] Because duplicate parts and similar parts have different values for the reuse of nesting knowledge, when evaluating the reusability of the historical nesting task for the current nesting task, it is necessary to perform a weighted sum of the duplication degree and similarity of the historical nesting task and the current nesting task. Let the weight of the nesting task duplication degree be β, β ∈ (0, 1), and denote the reusability of the nesting task as D. Then there is:

[0128] D = β · D R + (1 - β) · D S (16)

[0129] When the nesting task needs to inherit and reuse the nesting knowledge in the historical task, the reusability can be calculated through formula (16), and the reusability of one or more historical tasks with similarity can be quantitatively evaluated to determine the reusability degree of the historical task for the current task, which is used as a reference for the transfer and reuse of nesting knowledge.

[0130] The following are specific embodiments:

[0131] In step S1, the feature extraction network parameters in the Siamese neural network model are shown in Table 2:

[0132] Table 2 Feature extraction network structure

[0133]

[0134] In step S2, the part information in the target nesting task A and the historical nesting task B is shown in Table 3 and Table 4 respectively:

[0135] Table 3 Part information in nesting task A

[0136]

[0137] Part information in nesting task B of Table 4

[0138]

[0139] From the above part information, it can be known that the number of parts m in the target nesting task A is 111, the number of parts n in the historical nesting task B is 107, and the number of repeated parts c is 107. The nesting task repeatability D of task A and task B can be obtained R (A, B) = 34.86%.

[0140] In step S3, some test results of the siamese neural network model are as Figure 4 shown.

[0141] In step S4, the Hungarian algorithm is used to solve and obtain the best matching relationship as follows

[0142] Table 5 Part similarity matching results in tasks A and B

[0143]

[0144] According to the data in Table 5, the nesting task similarity of nesting task A and nesting task B is D S (A, B) = 81.54%.

[0145] In step S5, take the weight β = 0.3, and calculate the reusability D of the historical nesting task B for the current nesting task A = 67.54%.

[0146] It can be seen from the results that because it contains a large number of repeated parts and similar parts, the nesting task B has a high reusability for the nesting task A. The reusability of other historical nesting tasks relative to the target task can be calculated by the same method, and the historical task with the highest reusability can be found as the object for nesting knowledge reuse and learning, and the nesting knowledge such as the arrangement order, adjacency relationship, and rotation angle of the repeated parts and similar parts is transferred to the target task, thereby improving the efficiency of nesting calculation.

[0147] Those skilled in the art can easily understand that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for evaluating the reusability of nesting tasks based on part shape similarity, characterized in that It includes the following steps: S1. Obtain a pre-trained siamese neural network model; S2. Calculate the number of repeated parts in the target nesting task A and the historical nesting task B to obtain the nesting task repeatability; S3. Pairwise input the part drawings in the target nesting task A and the historical nesting task B into the pre-trained siamese neural network model to obtain the shape similarity; furthermore, correct the shape similarity through the part size difference degree to obtain the part similarity matrix; The calculation method of the part size difference degree r is as follows: Where s1 and s2 are the areas of two parts under the actual size, and s1 ≤ s2; Correcting the shape similarity through the part size difference degree to obtain the part similarity matrix specifically includes: Calculate the correction coefficient of part size gap based on the difference degree of part sizes Furthermore, the part similarity E between the two parts is obtained. N = δ·E W , where E W is the shape similarity output by the Siamese neural network model; For each part a in task A i and each part b in task B j calculate the similarity between each pair of parts to obtain a part similarity matrix M of size m×n ij ; where i = 1, 2,..., m, j = 1, 2,..., n, and m and n are the total numbers of parts in task A and task B respectively; S4. Based on the part similarity matrix, pair the parts in the target nesting task A and the historical nesting task B according to the principle of maximum similarity, and calculate the nesting task similarity according to the pairing result; S5. Calculate the nesting task reusability according to the nesting task repeatability and the nesting task similarity.

2. The method for evaluating the reusability of nesting tasks based on part shape similarity according to claim 1, characterized in that, The siamese neural network model is used for shape similarity measurement, which is constructed based on the convolutional neural network Vgg16, and the siamese neural network model is pre-trained through the part shape data set.

3. The evaluation method for the reusability of nesting tasks based on part shape similarity according to claim 2, wherein When the siamese neural network model reads the part image, it first performs image preprocessing, and then performs image feature extraction and similarity measurement; the steps of image preprocessing are as follows: Convert the part image into a binary image; Calculate the zero-order moment about the origin and the first-order moment about the origin of the binary image, and obtain the centroid coordinates (x c , y c ) of the image; Calculating the p+q-th central moment W based on the centroid coordinates of the image pq : According to the central moment \(W\) of order \(p + q\) pq , calculate the principal axis direction angle \(\theta\) of the image, and rotate the image clockwise by \(\theta\) angle; the calculation formula for the principal axis direction angle \(\theta\) is as follows: Where f(x, y) is the gray value of the image at the pixel coordinates (x, y), and p and q are the orders of the geometric moments.

4. The method for evaluating the reusability of nesting tasks based on part shape similarity according to claim 1, characterized in that, In step S2, the duplication degree D of the nesting task R is calculated as follows: Where m is the number of parts in task A, n is the number of parts in task B, and c is the number of repeated parts in task A and task B.

5. The method for evaluating the reusability of nesting tasks based on part shape similarity according to claim 1, wherein In step S4, the specific method for pairing the parts is: Based on the part similarity matrix, part a in task A i is paired with part b in task B j so that each part is paired with a part with high similarity as much as possible; When the number of parts in Task A and Task B is not equal, add dummy parts to the group with fewer parts to make up the number of parts. The similarity between the added dummy parts and any part is 0, and based on this, convert the part similarity matrix M ij to M ij '.

6. The method for evaluating the reusability of nesting tasks based on part shape similarity according to claim 5, wherein In step S4, calculating the nesting task similarity according to the pairing result is specifically: Introduce variable y ij , and let: Establish the following mathematical model: The above mathematical model is solved using the Hungarian algorithm to obtain the sum H of part similarities based on the pairing results max , and then the nesting task similarity D is calculated s :

7. The evaluation method for the reusability of nesting tasks based on part shape similarity according to any one of claims 1-6, characterized in that, In step S5, the calculation method of the nesting task reusability D is as follows: D = β·D R +(1 - β)·D S Among them, D R is the nesting task repeatability, D s is the nesting task similarity, and β is the weight of the nesting task repeatability, where β ∈ (0, 1).

8. A nesting task reusability evaluation system based on part shape similarity, characterized in that, It includes a processor, and the processor is used to execute the nesting task reusability evaluation method according to any one of claims 1-7.

Images