Machine learning-based multi-equipartition correction method and system

Abstract

The invention relates to the technical field of PCB (Printed Circuit Board) production control, and provides a multi-equipartition deviation correction method and system based on machine learning, and particularly relates to the following steps: step 1, placing a copper substrate on an alignment platform, step 2, establishing a film target graph, and establishing plane coordinate models of an upper group of film targets and a lower group of film targets; 3, calculating offset distances between four alignment target points of the upper film and the lower film and alignment points of the substrate, adopting CCD imaging equipment with the pixel size of 3.45 microns to shoot and filter the film targets, extracting feature points in a gray level image, calculating contour distances between the film target points and the alignment points, obtaining an actual offset distance, and calculating the actual offset distance; the traditional alignment operation only realizes single-time single-face alignment, and the alignment operation is repeatedly adjusted for more than three times. Through algorithm innovation based on machine learning, the alignment precision reaches 10 [mu] m, and the precision requirement can be met by aligning plates in the same batch only twice.

Description

Technical Field

[0001] The present invention relates to the technical field of PCB production control, and particularly relates to a multiple equalization and deviation correction method and system based on machine learning. Background Art

[0002] In many industrial production and scientific research fields, precise positioning and measurement are crucial. Taking the electronics manufacturing industry as an example, during the production process of printed circuit boards (PCBs), it is necessary to accurately transfer the circuit patterns on the film to the substrate, which requires precise control of the positional relationship between the film target and the alignment points on the substrate. However, the traditional manual positioning and measurement methods are not only inefficient but also extremely vulnerable to human factors, resulting in large positioning errors and difficult to meet the requirements of modern high-precision production.

[0003] With the rise of machine vision technology, using CCD imaging to obtain target images has become an important means. However, in practical applications, many challenges are faced. On the one hand, the CCD imaging system is easily interfered by factors such as environmental light changes, its own noise, and optical system aberrations, resulting in poor image quality and blurred target features, seriously affecting the subsequent positioning and measurement accuracy. On the other hand, for the complex and diverse target pattern designs in different application scenarios, there is a lack of systematic methods, making it difficult to balance the easy recognition of the target, the high-precision measurement requirements, and the adaptability to the imaging system. Therefore, there is an urgent practical need to develop a complete technology for forming a film target map through CCD imaging. Summary of the Invention

[0004] Technical Problems to be Solved

[0005] Aiming at the above-mentioned shortcomings of the existing technology, the present invention provides a multiple equalization and deviation correction method and system based on machine learning, which can effectively solve the problems in the existing technology.

[0006] Technical Solutions

[0007] To achieve the above object, the present invention is realized through the following technical solutions:

[0008] The present invention provides a multiple equalization and deviation correction method based on machine learning, including the following steps:

[0009] Step 1: Place the copper substrate on the alignment platform.

[0010] Step 2: Establish a film target map and establish a plane coordinate model for the upper and lower groups of film targets.

[0011] Step 3: Calculate the offset distances between the four alignment target points on the upper and lower films and the substrate alignment points. Use a CCD imaging device with a pixel size of 3.45 μm to capture and filter the film target, extract the feature points in the grayscale image, calculate the contour spacing between the film target points and the alignment points, and obtain the actual offset distances;

[0012] Step 4: Based on the above offset data, establish a two-stage prediction algorithm model for machine learning. Select an optimization algorithm and training set data to train and iteratively optimize the model. In the second stage, based on the first stage, input the correction coefficient J(y) and output the predicted offset value P i ;

[0013] Step 5: After analyzing the image offset distances, the system controls the alignment platform to correct the alignment of the upper and lower films through a vacuum adsorption device;

[0014] Step 6: Collect the alignment correction data of the upper and lower films again, repeat the second stage in Step 4, and feed the compensation data back to the control system for secondary alignment.

[0015] Furthermore, in Step 2, for the upper and lower groups of film sheets, each group has four alignment target points, and the alignment platform surrounded by the eight film sheet targets in total is marked as the placement platform, and the remaining area is marked as the non-placement platform. Calculate the ratio T of the placement platform to the non-placement platform;

[0016] If the ratio T is less than the threshold, readjust the positions of the film sheet targets until the requirements are met.

[0017] Furthermore, in Step 2, establish a data model of four groups of plane coordinates for the upper and lower groups of film sheets with the eight film sheet targets as the origin, and the unit is 10 μm.

[0018] Furthermore, in Step 3, the upper and lower film targets and the substrate alignment points are all completely located at the center of the camera's field of view, and the camera optical axis is perpendicular to the placement plane. Through the calculation formula: GR = 0.299R + 0.587G + 0.114B, where R, G, and B are the red, green, and blue component values of the image pixels respectively, convert the RGB color image to a grayscale image by weighted average to obtain the grayscale value GR, and set the standard deviation of the Gaussian kernel to 2 and the kernel size to 3×3 to filter the image, remove the noise generated by camera noise, environmental interference, etc., and stretch the image grayscale value range from [50, 150] to [0, 255].

[0019] Furthermore, the features of the target point and alignment point in the grayscale image are extracted, the center point and corner point of the contour are selected, and the Harris corner detection algorithm is used to set the threshold of the corner response function to 0.01 to detect the corner points in the image. The coordinate value of each pixel point is obtained in the plane coordinate system. For the circular target point, the average coordinates of all points on the contour are calculated to obtain the coordinates of the center of the circle; for the rectangular alignment point, the coordinates of the diagonal intersection are calculated to determine the center position. According to the coordinate values (X, Y), the distance formula is used to calculate the offset distance between the film target point and the substrate alignment point, and the actual offset distance is obtained through the pixel resolution of the camera.

[0020] Furthermore, the prediction algorithm model in step four includes a first stage and a second stage. There is no historical data accumulation in the first stage, and the second stage is an offset prediction, which is based on environmental factors, equipment factors and historical factors. The environmental factors in the second stage are based on the prediction calculation in the first round, and the results are corrected according to other factors.

[0021] Furthermore, in the first stage, the linear regression equation is selected as the pre-selected model, and the data is fitted by the least squares method to establish a linear equation of the displacement offset y and the influencing factors x1, x2, ..., xn: y = β0 + β1x1 + β2x2 + ... + βnxn + ε, where β is the coefficient and ε is the error term. The parameters of the model are initialized, the loss function and the optimization algorithm are established, the mean square error (MSE) is used as the loss function, the variant gradient descent optimization algorithm is selected, the model is trained using the training set data, the loss function is iteratively calculated and the model parameters are updated using the optimization algorithm. When the performance of the validation set no longer improves or begins to decline, the training is stopped, the first stage is completed, and the output data matrix Zt is expressed as Z t ={C1 t , C2 t , C3 t , C4 t , C5 t , C6 t , C7 t , C8 t , P t}, the integrated data matrix Z is

[0022] Furthermore, in the second stage, a prediction model is derived based on the algorithm of the first stage; environmental factors, equipment factors, and historical factors are queried to calculate the correction coefficient J(y), the environmental factors are input to obtain the prediction data, the historical data of similar environmental factors are queried, and the prediction results are corrected using the weighted method. The specific formula is: P i is the offset prediction value at time i, is the output value of the initialization model, θ is the residual influence coefficient, R jis the result value output for environmental factor C on the j-th tree, T y represents the annual degradation rate of the device, α is an important factor control measure for the degradation rate, J(y) represents the dust accumulation influence rate, β is an important factor control measure for the degradation rate, and the offset prediction value P of the second stage is obtained i .

[0023] Furthermore, in step three, there are eight CCD forming devices, with four upper and four lower films respectively

[0024] A multi-equalization rectification system based on machine learning, comprising:

[0025] A data storage module, based on the transfer interface of a data collection platform for all contour spacing data to be stored, obtains all contour data to be stored and actual offset data; including a color image library obtained by a CCD camera, as well as the image gray value converted therefrom and the image stretching gray range value obtained by denoising and filtering

[0026] An image data conversion module: converts the RGB image captured by the CCD camera into a gray image by weighted average, and uses the built-in Gaussian filtering algorithm to remove noise, noise points and stretch the gray range value to a certain range

[0027] An image feature extraction module: based on the gray image, according to the Harris corner detection algorithm, extracts the characteristics of the target points and alignment points on the gray image, detects the corners in the image, and obtains the coordinates of each pixel point in the coordinate system

[0028] An offset distance optimization processing module; the linear regression equation is a preselected model. In the first stage, a linear equation of the displacement offset and influencing factors is established, and the variant stochastic gradient descent optimization algorithm is selected for iterative optimization training, and the obtained correction coefficient is brought into P i formula to obtain the optimization result of the second stage

[0029] Beneficial effects

[0030] The technical solution provided by the present invention has the following beneficial effects compared with the known public technology:

[0031] The present invention innovates a pre-alignment algorithm. When a copper substrate is placed on the alignment platform, a target pattern of the film is established through CCD imaging. With a total of 8 targets as the origin points, a data model is established to calculate the offset distances between the four alignment target points of the upper and lower films and the alignment points of the substrate. A micron-level multiple equal division algorithm is designed to improve the alignment accuracy. At the same time, based on the prediction algorithm of machine learning, the alignment deviation will be included in the prediction algorithm to achieve high efficiency and high precision in advance alignment during the production process. Traditional alignment operations only achieve single-sided alignment once, and the alignment operation needs to be adjusted more than 3 times repeatedly. Through the algorithm innovation based on machine learning, the alignment accuracy reaches 10μm, and only two alignments are required for the same batch of plates to meet the accuracy requirements.

[0032] In the present invention, by reasonably selecting multiple CCD cameras, configuring high-quality optical lenses, and precisely setting imaging parameters, a high-resolution and low-distortion target image of the film can be obtained. Combined with advanced image processing and optimization algorithms, such as edge detection and contour extraction, the key features of the target pattern can be accurately extracted to achieve sub-pixel-level positioning accuracy. From image acquisition to target feature extraction and analysis, they can all be quickly processed by a computer, which greatly shortens the time required for positioning and measurement compared with traditional manual operations. In large-scale production scenarios, it can significantly improve production efficiency and reduce production costs. Technologies such as image filtering and distortion correction are used to process the original image, effectively removing interference factors such as noise and distortion, and enhancing the stability and reliability of the image. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0034] Figure 1 is the step flowchart of the present invention;

[0035] Figure 2 is the program block diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0037] The present invention will be further described below in conjunction with embodiments.

[0038] Embodiment: A multiple equalization and correction method based on machine learning. Referring to the attached Figure 1 as shown, it includes the following steps:

[0039] Step 1: Place the copper substrate on the alignment platform.

[0040] Step 2: Establish a film target map and establish a planar coordinate model for the upper and lower groups of film targets; for the upper and lower groups of films, each group has four alignment target points, and the alignment platform surrounded by a total of eight film targets is marked as the placement platform, and the remaining area is marked as the non-placement platform, and the ratio T of the placement platform to the non-placement platform is obtained by calculation.

[0041] If the ratio T is less than the threshold, readjust the position of the film target until the requirements are met; in Step 2, based on the eight film targets as the origin, a data model of the planar coordinates of the four groups of the upper and lower groups of films is established, with the unit of 10 μm.

[0042] Example 2: Calculate the offset distances between the four alignment target points of the upper and lower films and the substrate alignment points. Use a CCD imaging device with a pixel size of 3.45 μm to photograph and filter the film target, extract the feature points in the grayscale image, calculate the contour spacing between the film target points and the alignment points, and obtain the actual offset distances. In step 3, the upper and lower film targets and the substrate alignment points are all completely centered in the camera's field of view, the camera optical axis is perpendicular to the placement plane, and there are eight CCD forming devices, with four set for each of the upper and lower films. Through the calculation formula: GR = 0.299R + 0.587G + 0.114B, where R, G, and B are the red, green, and blue component values of the image pixels respectively, convert the RGB color image to a grayscale image by weighted average to obtain the grayscale value GR, and set the standard deviation of the Gaussian kernel to 2 and the kernel size to 3×3 to filter the image and remove the noise points generated by camera noise, environmental interference, etc. Stretch the image grayscale value range from [50, 150] to [0, 255]. Extract the features of the target points and alignment points in the grayscale image, select the center points and corner points of the contour, and use the Harris corner detection algorithm to set the threshold of the corner response function to 0.01 to detect the corner points in the image. In the plane coordinate system, obtain the coordinate values of each pixel point. For the circular target points, calculate the average coordinates of all points on the contour to obtain the center coordinates of the circle. For the rectangular alignment points, calculate the intersection coordinates of the diagonals to determine the center position. According to the coordinate values (X, Y), use the distance formula to calculate the offset distance between the film target points and the substrate alignment points, and know the actual offset distance through the pixel resolution of the camera. When the copper substrate is placed on the alignment platform, establish a film target map through CCD imaging. With a total of 8 targets as the origin, establish a data model to calculate the offset distances between the four alignment target points of the upper and lower films and the substrate alignment points. Design a micron-level multiple equal division algorithm to improve the alignment accuracy. At the same time, based on the prediction algorithm of machine learning, the alignment deviation will be included in the prediction algorithm to achieve high efficiency and high precision pre-alignment in the production process. Traditional alignment operations only achieve single-sided alignment once, and the alignment operation needs to be adjusted more than 3 times repeatedly. Through the algorithm innovation based on machine learning, the alignment accuracy reaches 10 μm, and only two alignments are required for the same batch of plates to meet the accuracy requirements.

[0043] Example 3: Based on the above offset data, establish a prediction algorithm model with two stages of machine learning, select an optimization algorithm and training set data to train and iteratively optimize the model. In the second stage, based on the first stage, input the correction coefficient J(y) and output the predicted offset value P i ; The prediction algorithm model in step 4 includes the first stage and the second stage. There is no historical data accumulation in the first stage, and the second stage is offset prediction, which is predicted based on environmental factors, equipment factors, and historical factors. The environmental factors in the second stage are predicted and calculated based on the first round, and the results are corrected according to other factors.

[0044] In the first stage, a linear regression equation is selected as the preselected model, and the data is fitted by the least squares method to establish a linear equation for the displacement offset y and the influencing factors x1, x2, ..., xn: y = β0 + β1x1 + β2x2 + … + βnxn + ε, where β is the coefficient and ε is the error term. Then, the parameters of the model are initialized, and a loss function and an optimization algorithm are established. The mean squared error (MSE) is used as the loss function, and the formula is: where yi is the true value, is the predicted value, n is the number of samples. Through the formula group: ① m_t = β_1·m_(t - 1) + (1 - β_1)·g; ② v_t = B_2·v_(t - 1) + (1 - β_2)·g^2; ③ m_tHat = m_t / (1 - B_1^t); ④: v_tHat = v_t / (1 - β_2^t); ⑤: wt = w(t - 1) - n·mtHat / (√v tHat + ε); where, m_t and v_t respectively represent the exponentially weighted average, g is the current gradient, β1 and β_2 are hyperparameters, n is the learning rate, ε is a very small positive number, m_tHat and v_tHat are the adjusted exponentially weighted averages, and w_t is the current weight parameter. The variant stochastic gradient descent optimization algorithm is selected, and the training set data is used to train the model. The loss function is iteratively calculated and the model parameters are updated using the optimization algorithm. When the performance of the validation set no longer improves or begins to decline, the training is stopped, and the first stage is completed, and the data matrix Zt is output: denoted as Z t ={C1 t , C2 t , C3 t , C4 t , C5 t , C6 t , C7 t , C8 t , P t}, and the data matrix Z is integrated into

[0045]

[0046] Then, in the second stage, a prediction model is obtained based on the algorithm of the first stage; the correction coefficient J(y) is calculated by querying the environmental factors, equipment factors, and historical factors. The environmental factors are input to obtain the prediction data, and the historical data of similar environmental factors is queried, and the prediction result is corrected by the weighted method. The specific formula is: P i is the offset prediction value at time i, is the output value of the initialized model, θ is the residual influence coefficient, R j is the result value output for the environmental factor C on the jth tree, T yrepresents the annual degradation rate of the device, α is an important factor control measure for the degradation rate, J(y) represents the influence rate of ash deposition, β is an important factor control measure for the degradation rate, and the offset prediction value P of the second stage is obtained. i By reasonably selecting multiple CCD cameras, configuring high-quality optical lenses, and precisely setting imaging parameters, a high-resolution and low-distortion film target image can be obtained. With the cooperation of advanced image processing and optimization algorithms, such as edge detection and contour extraction, the key features of the target pattern can be accurately extracted, achieving sub-pixel level positioning accuracy. From image acquisition to target feature extraction and analysis, all can be quickly processed by a computer, greatly shortening the time required for positioning and measurement compared with traditional manual operations.

[0047] Embodiment 4: After analyzing the image offset distance, the system controls the alignment platform to correct the upper and lower film alignment through a vacuum adsorption device.

[0048] Step 6: Collect the upper and lower film alignment correction data again, repeat the second stage in Step 4, and feedback the compensation data to the control system for secondary alignment. In large-scale production scenarios, it can significantly improve production efficiency and reduce production costs. Technologies such as image filtering and distortion correction are used to process the original image, effectively removing interference factors such as noise and distortion, and enhancing the stability and reliability of the image.

[0049] Embodiment 5: Refer to the appendix Figure 2 As shown, a multiple equalization alignment correction system based on machine learning includes:

[0050] A data storage module, based on the transfer interface of the data collection platform for all contour spacing data to be stored, obtains all contour data to be stored and actual offset data; including a color image library obtained by CCD camera shooting, as well as the converted image gray value and the stretched gray range value of the image obtained by denoising and filtering; an image data conversion module: converts the RGB image captured by the CCD camera into a gray image by weighted average, and the internal Gaussian filtering algorithm removes noise and noise points and stretches the gray range value to a certain range; an image feature extraction module: based on the gray image, according to the Harris corner detection algorithm, extracts the characteristics of the target points and alignment points on the gray image, detects the corner points in the image, and obtains the coordinates of each pixel point in the coordinate system; an offset distance optimization processing module; a linear regression equation is a preselected model. In the first stage, a linear equation of displacement offset and influencing factors is established, and a variant stochastic gradient descent optimization algorithm is selected for iterative optimization training, and the obtained correction coefficient is brought into the P i formula to obtain the optimization result of the second stage.

[0051] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multiple equalization and correction method based on machine learning, characterized in that, It includes the following steps: Step 1: Place the copper substrate on the alignment platform. Step 2: Establish a film target map and establish a planar coordinate model for the upper and lower sets of film targets. Step 3: Calculate the offset distances between the four alignment target points of the upper and lower films and the substrate alignment points. Use a CCD imaging device with a pixel size of 3.45 μm to photograph and filter the film targets, extract the feature points in the grayscale image, calculate the contour spacing between the film target points and the alignment points, and obtain the actual offset distances. Step 4: Based on the above offset data, establish a prediction algorithm model for two stages of machine learning. Select an optimization algorithm and training set data to train and iteratively optimize the model. In the second stage, based on the first stage, input the correction coefficient J(y) and output the predicted offset value P i ; Step 5: After analyzing the image offset distances, the system controls the alignment platform to correct the alignment of the upper and lower films through the vacuum adsorption device. Step 6: Collect the alignment correction data of the upper and lower films again, repeat the second stage in Step 4, and feedback the compensation data to the control system for secondary alignment.

2. The multi-equalization correction method based on machine learning according to claim 1, wherein In Step 2, for the upper and lower sets of film sheets, with four alignment target points in each set, the alignment platform surrounded by the eight film targets is marked as the placement platform, and the remaining area is marked as the non-placement platform, and the ratio T of the placement platform to the non-placement platform is calculated. If the ratio T is less than the threshold, readjust the positions of the film targets until the requirements are met.

3. A multiple equalization and correction method based on machine learning according to claim 2, characterized in that, In Step 2, based on the eight film targets as the origin, establish a data model of the four sets of planar coordinates of the upper and lower sets of film sheets, with the unit of 10 μm.

4. A multiple equalization correction method based on machine learning according to claim 3, characterized in that In Step 3, the upper and lower film targets and the substrate alignment points are all completely centered in the camera's field of view. The camera optical axis is perpendicular to the placement plane. The RGB color image is converted to a grayscale image by weighted averaging to obtain the grayscale value GR. Set the standard deviation of the Gaussian kernel to 2 and the kernel size to 3×3, filter the image to remove the noise points generated by camera noise, environmental interference, etc., and stretch the image grayscale value range from [50, 150] to [0, 255].

5. A multiple equalization and correction method based on machine learning according to claim 4, characterized in that, Extract the features of the target points and alignment points in the grayscale image. Select the center points and corner points of the contour. Through the Harris corner detection algorithm, set the threshold of the corner response function to 0.01 to detect the corner points in the image. In the planar coordinate system, obtain the coordinate values of each pixel point. For the circular target points, calculate the average coordinate of all points on the contour to obtain the center coordinate of the circle; for the rectangular alignment points, calculate the intersection coordinate of the diagonals to determine the center position. According to the coordinate values (X, Y), use the distance formula to calculate the offset distance between the film target points and the substrate alignment points, and know the actual offset distance through the pixel resolution of the camera.

6. A multiple equalization correction method based on machine learning according to claim 1, characterized in that The prediction algorithm model in Step 4 includes a first stage and a second stage. In the first stage, there is no historical data accumulation. The second stage is offset prediction, which is predicted based on environmental factors, equipment factors, and historical factors. The environmental factors in the second stage are predicted and calculated based on the first round, and the results are corrected according to other factors.

7. A multiple equalization and correction method based on machine learning according to claim 6, characterized in that, In the first stage, the linear regression equation is selected as the pre-selected model, and the data is fitted by the least squares method to establish a linear equation between the displacement offset y and the influencing factors x1, x2, ..., xn: and initialize the model parameters, establish the loss function and optimization algorithm, use the mean square error (MSE) as the loss function, select the variant gradient descent optimization algorithm, use the training set data to train the model, iteratively calculate the loss function and use the optimization algorithm to update the model parameters. When the performance of the validation set no longer improves or begins to decline, stop training, complete the first stage, and output the data matrix Zt: expressed as , the integrated data matrix Z is .

8. A multiple equalization and correction method based on machine learning according to claim 6, characterized in that, In the second stage, based on the algorithm in the first stage, obtain the prediction model; query the environmental factors, equipment factors, and historical factors to calculate the correction coefficient J(y), input the environmental factors to obtain the prediction data, query the historical data of similar environmental factors, and correct the prediction results by the weighted method.

9. A multiple equalization correction method based on machine learning according to claim 1, characterized in that There are eight CCD forming devices set in Step 3, with four set for each of the upper and lower films.

10. A multiple equalization and correction system based on machine learning, the correction system is used to execute the method of any one of claims 1-9, characterized in that, It includes: The data storage module obtains all the contour data to be stored and the actual offset data based on the transfer interface of the data collection platform for all the contour spacings to be stored; it includes the color image library obtained by the CCD camera shooting, as well as the image gray value converted therefrom and the image stretching gray range value obtained by denoising filtering. The image data conversion module: converts the RGB image captured by the CCD camera into a gray image by weighted average, and the built-in Gaussian filtering algorithm removes noise and noise points and stretches the gray range value to a certain range. The image feature extraction module: based on the gray image, according to the Harris corner detection algorithm, extracts the characteristics of the target points and alignment points on the gray image, detects the corners in the image, and obtains the coordinates of each pixel point in the coordinate system. The offset distance optimization processing module; The linear regression equation is a preselected model. In the first stage, a linear equation of the displacement offset and influencing factors is established, and a variant stochastic gradient descent optimization algorithm is selected for iterative optimization training. The obtained correction coefficients are substituted into the P i formula to obtain the optimization result of the second stage.

Images