Metal plate defect intelligent identification method based on big data
Through the intelligent big data identification method, the problem of insufficient preprocessing of image data and limited model generalization capabilities in metal sheet detection is solved, and efficient and accurate defect identification and real-time correction are achieved to adapt to the needs of different production environments.
Patent Information
- Application Number
- CN202411347008.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-26
- Publication Date
- 2025-01-07
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing metal sheet defect detection technology has insufficient image data preprocessing, lack of systematicity in abnormal target positioning and confidence correction, limited generalization capabilities of model, and lack of continuous optimization mechanism, resulting in difficulty in improving detection efficiency and accuracy.
Adopting intelligent recognition methods based on big data, through image data preprocessing, abnormal target positioning, confidence correction, abnormal identification and model construction and optimization steps, including grayscale processing, corner penalty factor correction, detection data mapping and random forest Algorithms, etc., build an intelligent identification model for metal sheet defects and optimize identification errors.
It improves the accuracy and accuracy of metal sheet defect identification, adapts to complex production environments, realizes intelligent identification of metal sheet defects and real-time confidence correction, saves resources and improves work efficiency.
Smart Images
Figure CN119273982A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of big data, and in particular to an intelligent metal plate defect recognition method based on big data. Background Art
[0002] In the field of metal processing and manufacturing, quality inspection of metal sheets is a key link to ensure product qualification and improve production efficiency. At present, the traditional inspection method commonly used in the market relies on manual visual inspection and simple mechanical measurement. This method is not only inefficient, but also easily affected by human factors, making it difficult to ensure the accuracy and consistency of the inspection results. In addition, with the rapid development of industrial automation and intelligent manufacturing, higher requirements are placed on the accuracy and speed of metal sheet defect detection.
[0003] In recent years, with the continuous advancement of image processing technology and big data analysis technology, automated defect recognition methods based on image processing and data analysis have gradually been applied to the field of metal sheet detection. However, most of the existing technical solutions have the following shortcomings: First, the preprocessing of image data is not sophisticated enough, and it is difficult to effectively remove noise and interference, which affects the accuracy of subsequent analysis; second, there is a lack of systematic methods for the positioning and confidence correction of abnormal targets, resulting in low credibility of recognition results; third, the constructed defect recognition model is often based on a limited data set, with limited generalization ability, and it is difficult to adapt to complex and changing production environments; fourth, there is a lack of continuous optimization mechanism for the recognition model, and it is impossible to automatically adjust the model parameters according to feedback in actual applications to further improve the recognition accuracy. Therefore, Kou is looking for new technologies to improve the detection efficiency and accuracy of metal sheet defect recognition. Summary of the invention
[0004] The purpose of the present invention is to provide an intelligent metal sheet defect identification method based on big data.
[0005] To achieve the above object, the present invention is implemented according to the following technical solutions:
[0006] The present invention comprises the following steps:
[0007] Collecting image data and detection data of the metal plate, and preprocessing the image data and the detection data;
[0008] Performing abnormal target positioning on the image data to obtain a problem image, and performing confidence correction and quantification on the problem image to obtain correction data;
[0009] Performing abnormal identification on the detection data according to the deviation to obtain defect data, and constructing an intelligent metal sheet defect identification model according to the correction data and the defect data;
[0010] The metal sheet defect intelligent recognition model is optimized according to the recognition error, the data to be recognized is input into the metal sheet defect intelligent recognition model, and the recognition result is output.
[0011] Furthermore, the method for locating abnormal targets in the image data to obtain a problem image includes:
[0012] Perform grayscale processing on the image data, obtain the grayscale threshold through the average grayscale value, roughly segment the pixels of the image data according to the grayscale threshold, and take the image data pixel area greater than the grayscale threshold as the key area;
[0013] Calculate critical area and defect length:
[0014]
[0015] The number of pixels in the key area is The length of the defect is g, the length and width of the minimum rectangle surrounding the defect are B and V respectively, the length and width of the key area are S and M respectively, and the number of pixels at coordinates w and h is k (wh);
[0016] Calculate the aspect ratio and moment of inertia of the minimum circumscribed matrix:
[0017]
[0018]
[0019] The aspect ratio of the minimum circumscribed matrix is φ, the moment of inertia is ζ, and the center of mass of the key area is (c w , c h );
[0020] Calculate the similarity between the sub-image and the key region:
[0021]
[0022] The aspect ratio of the sub-image is φ 2 , the moment of inertia of the sub-image is ζ 2 , the aspect ratio of the key area is φ 1 , the moment of inertia of the key area is ζ 1 , the pixel value of the sub-image at position (f, m) is C(f, m), the indexes of the sub-image in the horizontal direction and vertical direction are f and m respectively, the pixel value of the key area at (w+fh+m) is I(w+fh+m), and the length and width of the sub-image are G and R respectively;
[0023] The key areas with similarity greater than 0.975 are output as problem images.
[0024] Furthermore, the method of performing confidence correction and quantification on the problem image to obtain correction data includes:
[0025] The penalty factor is constructed based on the corner score difference and identifier difference of the problem image, and the expression is:
[0026]
[0027] The penalty factor based on the corner point score difference is The penalty factor based on the difference in corner identifiers is τ, and the impact of the difference is μ 1 , the difference influence is μ 2 , the score of the upper left corner is U tl , the score of the lower right corner point is U br , the identifier of the upper left corner is Z tl , the identifier of the lower right corner is Z br ;
[0028] The compensation coefficient of the penalty factor is constructed according to the target area, and the expression is:
[0029]
[0030] The penalty factor The compensation coefficient is The compensation coefficient of the penalty factor τ is β τ , the area threshold of the problem image box is D max , penalty factor The inhibitory factor is The suppression factor of the penalty factor τ is
[0031] The penalty factor is updated using the compensation coefficient, and the expression is:
[0032]
[0033] The updated penalty factor for The updated penalty factor τ is
[0034] The penalty factor of the corner point score difference and the penalty factor of the corner point identifier difference are introduced into the confidence of the problem image frame. The expression is:
[0035]
[0036] The minimum area of the problem image box is D min , the confidence of the problem image box is H bx ;
[0037] Invalid problem image frames are excluded by combining the center point information and corner point differences, and the predicted center point confidence is added to the problem image frame confidence. The expression is:
[0038]
[0039] The confidence of the adjusted problem image frame is The confidence level of the center point is H ct , optimize the penalty factor The expression is:
[0040]
[0041] The area of the problem image frame composed of the two corner points after the vector displacement is D sf , the area of the problem image frame composed of the two corner points before displacement is D thsd , the compensation coefficient of the penalty factor τ after vector displacement is β ssf , the optimized penalty factor is τ * ;
[0042] According to the optimized penalty factor τ * Correct the confidence of the problem image frame, the expression is
[0043]
[0044] The confidence of the corrected problem image frame is The control weight is η;
[0045] The correction is stopped until the confidence of the corrected problem image frame is greater than 0.752, otherwise the problem image is segmented twice;
[0046] The defect area is determined according to the confidence of the corrected problem image frame, and the defect ratio of the metal sheet is calculated according to the defect area:
[0047]
[0048] The defect ratio of metal sheet is The area of the metal sheet is X, and the number of rows of the defect area image is The number of columns of the defect area image is The pixel value of the i-th row and y-th column is I(i, y).
[0049] Furthermore, the method of obtaining defect data by performing abnormality identification on the detection data according to the deviation includes:
[0050] Map the detection data to a low-dimensional space to obtain the smallest hypersphere containing all the detection data. The expression is:
[0051]
[0052] The penalty coefficient is λ, and the sth relaxation factor is The center of the hypersphere is a, the radius of the hypersphere is K, and the number of detection data is The sth detection data is b s , the mapping function is u(·), the detection data b s The distance from the center a is ||u(b s )-a||, the objective function is
[0053] The Lagrange multiplier is used to adjust the distance from the detection data to the center of the sphere. The expression is:
[0054]
[0055] The sth Lagrange multiplier is χ s , the Lagrangian function is Detection data s The distance to the center of the sphere is d(b s );
[0056] The radius is determined based on the detection data on the hypersphere, and the expression is:
[0057]
[0058] The vth detection data is b v , the kernel function is Z(,), and the jth Lagrange multiplier is χ1;
[0059] Calculate the deviation of the detection data:
[0060]
[0061] The detection data b s The deviation is The detection data with a deviation greater than 0.217K is output as defect data.
[0062] Furthermore, a method for constructing an intelligent metal sheet defect recognition model based on the correction data and the defect data includes:
[0063] The objective function is constructed based on the correction data and defect data, and the expression is:
[0064]
[0065] The number of defect data is N, and the image defect weight is ρ 2 , the data defect weight is ρ 1 , the defect ratio of metal sheet is Detection data s The deviation is The actual defect value is The predicted defect value is Actual defect value and predicted defect value The loss function is
[0066] The intelligent identification model of metal sheet defects is constructed, including random forest algorithm, multimodal classification algorithm, BP neural network and SVM vector machine combined classifier;
[0067] The random forest algorithm randomly divides the input data into training and test sets in a ratio of 5:2;
[0068] The multimodal classification algorithm obtains the difference by integrating the information of different data sources of the training set, and classifies the training set according to the difference to obtain hierarchical data;
[0069] BP neural network and SVM vector machine combined classifier uses BP neural network to learn the features of hierarchical data, and then uses SVM to classify and decide the hierarchical data to calculate the defect value;
[0070] The test set is input into the intelligent recognition model of metal sheet defects, and the target value is obtained according to the objective function. The training is stopped when the loss function in the target value reaches the minimum.
[0071] Furthermore, the method for optimizing the intelligent recognition model of metal sheet defects according to the recognition error includes:
[0072] Use the Sobol sequence to generate the population, the expression is:
[0073]
[0074] The control parameter is ψ, and the position of the zth particle is The search space is capped at The search space is bounded by The random number generated by the Sobol sequence is x;
[0075] Calculate the fitness of a particle:
[0076]
[0077] The identification error of the zth particle is er z , the defect ratio of the zth particle is W z , the error weight is ω, and the proportional weight is υ;
[0078] The position of the particle with the largest fitness is taken as the target, and the position of the particle is updated according to the probability. The expression is:
[0079]
[0080] The position of the zth particle at the t+1th iteration is The position of the zth particle at the tth iteration is The number of particles is The goal of the tth iteration is The probability is p, and the random numbers from 0 to 1 are r 1 、r 2 、r 3 、r 4 ;
[0081] Adjust the defect ratio, the expression is:
[0082]
[0083] The escape ratio is W and the maximum number of iterations is t max ;
[0084] According to the escape ratio and Gaussian perturbation strategy, the particle position is updated to obtain the perturbation position. The expression is:
[0085]
[0086] The optimized position of the zth particle in the t+1th iteration is A random number between 0 and 1 is r 5 , the difference between the target and search subposition at the tth iteration is The variance is σ and the Gaussian function is The dominant position at the tth iteration is The perturbed position of the zth particle at the t+1th iteration is Iterate continuously until the maximum number of iterations is reached, then stop iterating, otherwise update the particle fitness and reselect the target.
[0087] The beneficial effects of the present invention are:
[0088] The present invention is a method for intelligently identifying defects in metal sheets based on big data. Compared with the prior art, the present invention has the following technical effects:
[0089] The present invention can improve the accuracy of intelligent identification of metal sheet defects in big data through preprocessing, abnormal target positioning, confidence correction and quantification, abnormal identification, model construction and identification error optimization steps, thereby improving the precision of metal sheet defect identification in big data. Optimizing the identification of metal sheet defects in big data can greatly save resources and improve work efficiency. It can realize the intelligent identification of metal sheet defects in big data, and perform confidence correction and error optimization on the intelligent identification of metal sheet defects in big data in real time. It is of great significance to the intelligent identification of metal sheet defects in big data, can adapt to the needs of intelligent identification of metal sheet defects in big data of different standards and intelligent identification of metal sheet defects in different big data, and has a certain universality. BRIEF DESCRIPTION OF THE DRAWINGS
[0090] Figure 1 This is a flowchart of the steps of a method for intelligently identifying metal sheet defects based on big data in the present invention. DETAILED DESCRIPTION
[0091] The present invention is further described below by means of specific embodiments. The illustrative embodiments and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.
[0092] The present invention provides a method for intelligently identifying metal sheet defects based on big data, comprising the following steps:
[0093] like Figure 1 As shown, in this embodiment, the following steps are included:
[0094] Collecting image data and detection data of the metal plate, and preprocessing the image data and the detection data;
[0095] In the actual evaluation, the metal sheet data of XXX production plant within 5 years was used to randomly select 2 metal sheets as the research objects;
[0096] The test data include temperature, atmosphere, time, oxidation weight gain, oxidation rate, oxide film morphology and composition, iron content, carbon content, silicon content, manganese content, phosphorus content, sulfur content, alloy element content, strength, hardness, toughness, surface flatness, roughness, scratches, cracks, rust, length, width, thickness, corrosion resistance, grain size;
[0097] Performing abnormal target positioning on the image data to obtain a problem image, and performing confidence correction and quantification on the problem image to obtain correction data;
[0098] In the actual evaluation, the number of problematic images of the two metal sheets was 11 and 7 respectively;
[0099] Performing abnormal identification on the detection data according to the deviation to obtain defect data, and constructing an intelligent metal sheet defect identification model according to the correction data and the defect data;
[0100] In the actual evaluation, the defect data of the two metal sheets were 0.392 and 0.251 respectively;
[0101] The metal sheet defect intelligent recognition model is optimized according to the recognition error, the data to be recognized is input into the metal sheet defect intelligent recognition model, and the recognition result is output.
[0102] In this embodiment, the method for locating abnormal targets in the image data to obtain a problem image includes:
[0103] Perform grayscale processing on the image data, obtain the grayscale threshold through the average grayscale value, roughly segment the pixels of the image data according to the grayscale threshold, and take the image data pixel area greater than the grayscale threshold as the key area;
[0104] Calculate critical area and defect length:
[0105]
[0106] The number of pixels in the key area is The length of the defect is g, the length and width of the minimum rectangle surrounding the defect are B and V respectively, the length and width of the key area are S and M respectively, and the number of pixels at coordinates w and h is k (wh);
[0107] Calculate the aspect ratio and moment of inertia of the minimum circumscribed matrix:
[0108]
[0109] The aspect ratio of the minimum circumscribed matrix is φ, the moment of inertia is ζ, and the center of mass of the key area is (c w , c h );
[0110] Calculate the similarity between the sub-image and the key region:
[0111]
[0112] The aspect ratio of the sub-image is φ 2 , the moment of inertia of the sub-image is ζ 2 , the aspect ratio of the key area is φ 1 , the moment of inertia of the key area is ζ 1, the pixel value of the sub-image at position (f, m) is C(f, m), the indexes of the sub-image in the horizontal direction and vertical direction are f and m respectively, the pixel value of the key area at (w+fh+m) is I(w+fh+m), and the length and width of the sub-image are G and R respectively;
[0113] The key areas with similarity greater than 0.975 are output as problem images.
[0114] In this embodiment, the method of performing confidence correction and quantification on the problem image to obtain correction data includes:
[0115] The penalty factor is constructed based on the corner score difference and identifier difference of the problem image, and the expression is:
[0116]
[0117] The penalty factor based on the corner point score difference is The penalty factor based on the difference in corner identifiers is τ, and the impact of the difference is μ 1 , the difference influence is μ 2 , the score of the upper left corner is U tl , the score of the lower right corner point is U br , the identifier of the upper left corner is Z tl , the identifier of the lower right corner is Z br ;
[0118] The compensation coefficient of the penalty factor is constructed according to the target area, and the expression is:
[0119]
[0120] The penalty factor The compensation coefficient is The compensation coefficient of the penalty factor τ is β τ , the area threshold of the problem image box is D max , penalty factor The inhibitory factor is The suppression factor of the penalty factor τ is
[0121] The penalty factor is updated using the compensation coefficient, and the expression is:
[0122]
[0123] The updated penalty factor for The updated penalty factor τ is
[0124] The penalty factor of the corner point score difference and the penalty factor of the corner point identifier difference are introduced into the confidence of the problem image frame. The expression is:
[0125]
[0126] The minimum area of the problem image box is D min , the confidence of the problem image box is H bx ;
[0127] Invalid problem image frames are excluded by combining the center point information and corner point differences, and the predicted center point confidence is added to the problem image frame confidence. The expression is:
[0128]
[0129] The confidence of the adjusted problem image frame is The confidence level of the center point is H ct , optimize the penalty factor The expression is:
[0130]
[0131] The area of the problem image frame composed of the two corner points after the vector displacement is D sf , the area of the problem image frame composed of the two corner points before displacement is D thsd , the compensation coefficient of the penalty factor τ after vector displacement is β ssf , the optimized penalty factor is τ * ;
[0132] According to the optimized penalty factor τ * Correct the confidence of the problem image frame, the expression is
[0133]
[0134] The confidence of the corrected problem image frame is The control weight is η;
[0135] The correction is stopped until the confidence of the corrected problem image frame is greater than 0.752, otherwise the problem image is segmented twice;
[0136] The defect area is determined according to the confidence of the corrected problem image frame, and the defect ratio of the metal sheet is calculated according to the defect area:
[0137]
[0138] The defect ratio of metal sheet is The area of the metal sheet is X, and the number of rows of the defect area image is The number of columns of the defect area image is The pixel value of the i-th row and y-th column is I(i, y).
[0139] In this embodiment, the method for obtaining defect data by performing abnormality identification on the detection data according to the deviation includes:
[0140] Map the detection data to a low-dimensional space to obtain the smallest hypersphere containing all the detection data. The expression is:
[0141]
[0142] The penalty coefficient is λ, and the sth relaxation factor is The center of the hypersphere is a, the radius of the hypersphere is K, and the number of detection data is The sth detection data is b s , the mapping function is u(·), the detection data b s The distance from the center a is ||u(b s )-a||, the objective function is
[0143] The Lagrange multiplier is used to adjust the distance from the detection data to the center of the sphere. The expression is:
[0144]
[0145] The sth Lagrange multiplier is χ s , the Lagrangian function is Detection data s The distance to the center of the sphere is d(b s );
[0146] The radius is determined based on the detection data on the hypersphere, and the expression is:
[0147]
[0148] The Vth detection data is b v , the kernel function is Z(,), and the j-th Lagrange multiplier is χj;
[0149] Calculate the deviation of the detection data:
[0150]
[0151] The detection data b s The deviation is The detection data with a deviation greater than 0.217K is output as defect data.
[0152] In this embodiment, the method for constructing an intelligent metal sheet defect recognition model based on the correction data and the defect data includes:
[0153] The objective function is constructed based on the correction data and defect data, and the expression is:
[0154]
[0155] The number of defect data is N, and the image defect weight is ρ 2 , the data defect weight is ρ 1 , the defect ratio of metal sheet is Detection data s The deviation is The actual defect value is The predicted defect value is Actual defect value and predicted defect value The loss function is
[0156] The intelligent identification model of metal sheet defects is constructed, including random forest algorithm, multimodal classification algorithm, BP neural network and SVM vector machine combined classifier;
[0157] The random forest algorithm randomly divides the input data into training and test sets in a ratio of 5:2;
[0158] The multimodal classification algorithm obtains the difference by integrating the information of different data sources of the training set, and classifies the training set according to the difference to obtain hierarchical data;
[0159] BP neural network and SVM vector machine combined classifier uses BP neural network to learn the features of hierarchical data, and then uses SVM to classify and decide the hierarchical data to calculate the defect value;
[0160] The test set is input into the intelligent recognition model of metal sheet defects, and the target value is obtained according to the objective function. The training is stopped when the loss function in the target value reaches the minimum.
[0161] In this embodiment, the method for optimizing the intelligent recognition model of metal sheet defects according to the recognition error includes:
[0162] Use the Sobol sequence to generate the population, the expression is:
[0163]
[0164] The control parameter is ψ, and the position of the zth particle is The search space is capped at The search space is lower bounded by The random number generated by the Sobol sequence is x;
[0165] Calculate the fitness of a particle:
[0166]
[0167] The identification error of the zth particle is er z , the defect ratio of the zth particle is W z , the error weight is ω, and the proportional weight is υ;
[0168] The position of the particle with the largest fitness is taken as the target, and the position of the particle is updated according to the probability. The expression is:
[0169]
[0170] The position of the zth particle at the t+1th iteration is The position of the zth particle at the tth iteration is The number of particles is The goal of the tth iteration is The probability is p, and the random numbers from 0 to 1 are r 1 、r 2 、r 3 、r 4 ;
[0171] Adjust the defect ratio, the expression is:
[0172]
[0173] The escape ratio is W and the maximum number of iterations is t max ;
[0174] According to the escape ratio and Gaussian perturbation strategy, the particle position is updated to obtain the perturbation position. The expression is:
[0175]
[0176] The optimized position of the zth particle in the t+1th iteration is The random number from 0 to 1 is r 5 , the difference between the target and search subposition at the tth iteration is The variance is σ and the Gaussian function is The dominant position at the tth iteration is The perturbed position of the zth particle at the t+1th iteration is
[0177] Iterate continuously until the maximum number of iterations is reached, then stop iterating, otherwise update the particle fitness and reselect the target.
[0178] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for intelligent identification of metal sheet defects based on big data, characterized in that: The following steps are involved: Collecting image data and detection data of the metal plate, and preprocessing the image data and the detection data; Performing abnormal target positioning on the image data to obtain a problem image, and performing confidence correction and quantification on the problem image to obtain correction data; Performing abnormal identification on the detection data according to the deviation to obtain defect data, and constructing an intelligent metal sheet defect identification model according to the correction data and the defect data; The metal sheet defect intelligent recognition model is optimized according to the recognition error, the data to be recognized is input into the metal sheet defect intelligent recognition model, and the recognition result is output.
2. According to claim 1, a method for intelligently identifying metal sheet defects based on big data is characterized in that: The method for performing abnormal target positioning on the image data to obtain a problem image comprises: Perform grayscale processing on the image data, obtain the grayscale threshold through the average grayscale value, roughly segment the pixels of the image data according to the grayscale threshold, and take the image data pixel area greater than the grayscale threshold as the key area; Calculate critical area and defect length: The number of pixels in the critical area is l, the length of the defect is g, the number of pixels of the length and width of the minimum rectangle surrounding the defect are B and V respectively, the length and width of the critical area are S and M respectively, and the number of pixels at coordinates w and h is k(w, h); Calculate the aspect ratio and moment of inertia of the minimum circumscribed matrix: The aspect ratio of the minimum circumscribed matrix is φ, the moment of inertia is ζ, and the center of mass of the key area is (c w , c h );Calculate the similarity between the sub-image and the key area: The aspect ratio of the sub-image is φ2, the moment of inertia of the sub-image is ζ2, the aspect ratio of the key area is φ1, the moment of inertia of the key area is ζ1, the pixel value of the sub-image at position (f, m) is C(f, m), the indexes of the sub-image in the horizontal and vertical directions are f and m respectively, the pixel value of the key area at (w+fh+m) is I(w+fh+m), and the length and width of the sub-image are G and R respectively; The key areas with similarity greater than 0.975 are output as problem images.
3. According to the big data-based intelligent metal sheet defect identification method of claim 1, it is characterized in that: The method of performing confidence correction and quantification on the problem image to obtain correction data comprises: The penalty factor is constructed based on the corner score difference and identifier difference of the problem image, and the expression is: The penalty factor based on the difference in corner score is θ, the penalty factor based on the difference in corner identifier is τ, the difference influence is μ1, the difference influence is μ2, and the score of the upper left corner is U tl , the score of the lower right corner point is U br , the identifier of the upper left corner is Z tl , the identifier of the lower right corner is Z br ; The compensation coefficient of the penalty factor is constructed according to the target area, and the expression is: The compensation coefficient of the penalty factor θ is β θ , the compensation coefficient of the penalty factor τ is β τ , the area threshold of the problem image box is D max , the suppression factor of the penalty factor θ is The suppression factor of the penalty factor τ is The penalty factor is updated using the compensation coefficient, and the expression is: The updated penalty factor θ is The updated penalty factor τ is The penalty factor of the corner point score difference and the penalty factor of the corner point identifier difference are introduced into the confidence of the problem image frame. The expression is: The minimum area of the problem image box is D m in, the confidence of the problem image box is Hb x ; Invalid problem image frames are excluded by combining the center point information and corner point differences, and the predicted center point confidence is added to the problem image frame confidence. The expression is: The confidence of the adjusted problem image frame is The confidence level of the center point is H ct , optimize the penalty factor The expression is: The area of the problem image frame composed of the two corner points after the vector displacement is D sf , the area of the problem image frame composed of the two corner points before displacement is D thsd , the compensation coefficient of the penalty factor τ after vector displacement is β ssf , the optimized penalty factor is τ*; The confidence of the problem image frame is corrected according to the optimized penalty factor T*, and the expression is: The confidence of the corrected problem image frame is The control weight is η; The correction is stopped until the confidence of the corrected problem image frame is greater than 0.752, otherwise the problem image is segmented twice; The defect area is determined according to the confidence of the corrected problem image frame, and the defect ratio of the metal sheet is calculated according to the defect area: The defect ratio of metal sheet is The area of the metal sheet is X, and the number of rows of the defect area image is The number of columns of the defect area image is The pixel value of the i-th row and y-th column is I(i, y).
4. According to the big data-based intelligent metal sheet defect identification method of claim 1, it is characterized in that: The method for obtaining defect data by performing abnormality identification on the detection data according to the deviation includes: Map the detection data to a low-dimensional space to obtain the smallest hypersphere containing all the detection data. The expression is: The penalty coefficient is λ, and the sth relaxation factor is The center of the hypersphere is a, the radius of the hypersphere is K, and the number of detection data is The sth detection data is b s , the mapping function is u(·), the detection data b s The distance from the center a is ||u(b s )-a||, the objective function is The Lagrange multiplier is used to adjust the distance from the detection data to the center of the sphere. The expression is: The sth Lagrange multiplier is χ s , the Lagrangian function is Detection data s The distance to the center of the sphere is d(b s ); The radius is determined based on the detection data on the hypersphere, and the expression is: The Vth detection data is b v , the kernel function is Z(,), and the j-th Lagrange multiplier is χj; Calculate the deviation of the detection data: The detection data b s The deviation is The detection data with a deviation greater than 0.217K is output as defect data.
5. According to the big data-based intelligent metal sheet defect identification method of claim 1, it is characterized in that: The method for constructing an intelligent metal sheet defect recognition model according to the correction data and the defect data comprises: The objective function is constructed based on the correction data and defect data, and the expression is: The number of defect data is N, the image defect weight is ρ2, the data defect weight is ρ1, and the metal sheet defect ratio is Detection data s The deviation is The actual defect value is The predicted defect value is Actual defect value and predicted defect value The loss function is The intelligent identification model of metal sheet defects is constructed, including random forest algorithm, multimodal classification algorithm, BP neural network and SVM vector machine combined classifier; The random forest algorithm randomly divides the input data into training and test sets in a ratio of 5:2; The multimodal classification algorithm obtains the difference by integrating the information of different data sources of the training set, and classifies the training set according to the difference to obtain hierarchical data; BP neural network and SVM vector machine combined classifier uses BP neural network to learn the features of hierarchical data, and then uses SVM to classify and decide the hierarchical data to calculate the defect value; The test set is input into the intelligent recognition model of metal sheet defects, and the target value is obtained according to the objective function. The training is stopped when the loss function in the target value reaches the minimum.
6. The method for intelligently identifying metal sheet defects based on big data according to claim 1 is characterized in that: The method for optimizing the intelligent recognition model of metal sheet defects according to the recognition error includes: Use the Sobol sequence to generate the population, the expression is: The control parameter is ψ, and the position of the zth particle is The search space is capped at The search space is lower bounded by The random number generated by the Sobol sequence is x; Calculate the fitness of a particle: The identification error of the zth particle is er z , the defect ratio of the zth particle is W z , the error weight is ω, and the proportional weight is υ; The position of the particle with the largest fitness is taken as the target, and the position of the particle is updated according to the probability. The expression is: The position of the zth particle at the t+1th iteration is The position of the zth particle at the tth iteration is The number of particles is The goal of the tth iteration is The probability is p, and the random numbers from 0 to 1 are r1, r2, r3, and r4 respectively; Adjust the defect ratio, the expression is: The escape ratio is W and the maximum number of iterations is t max ; According to the escape ratio and Gaussian perturbation strategy, the particle position is updated to obtain the perturbation position. The expression is: The optimized position of the zth particle in the t+1th iteration is The random number between 0 and 1 is r5, and the difference between the target and search subpositions at the tth iteration is The variance is σ and the Gaussian function is The dominant position at the tth iteration is The perturbed position of the zth particle at the t+1th iteration is Iterate continuously until the maximum number of iterations is reached, then stop iterating, otherwise update the particle fitness and reselect the target.