Circuit board production control method and system based on image recognition

Abstract

The invention relates to the technical field of circuit board production control, in particular to a circuit board production control method and system based on image recognition, and the method comprises the steps: obtaining sample image data and point cloud image data of an asymmetric target, and carrying out the preprocessing; constructing and training a data fusion model, and extracting a three-dimensional space information feature map sequence of the asymmetric target by using the preprocessed sample image data and point cloud image data; constructing and training a multi-layer circuit board production control model based on a generative adversarial network, extracting related features of an asymmetric target by using a three-dimensional space information feature map sequence, and outputting a preliminary control strategy; optimizing a multi-layer circuit board production control module by taking minimization of the influence of the multi-source interference factors as a target; and acquiring real-time image data of the asymmetric target, and outputting a multi-layer circuit board alignment control strategy by using the optimized multi-layer circuit board production control model. The objective of the invention is to realize adaptive dynamic adjustment of a multi-layer circuit board control strategy and improve the alignment precision of the multi-layer circuit board.

Description

Technical Field

[0001] The present invention relates to the technical field of circuit board production control, and in particular to a circuit board production control method and system based on image recognition. Background Art

[0002] In the field of modern electronic manufacturing, circuit boards are key carriers for carrying electronic components and achieving electrical connections. The accuracy and advancement of their production processes are directly related to the performance and quality of electronic products. Targets are a type of graphic mark with specific functions on circuit boards. They are used in many processes of the circuit board manufacturing process, such as exposure, drilling, mounting and other key nodes, and play an important role in providing accurate positioning benchmarks for automated processing equipment. Its precise position information guidance ensures that each process can be accurately executed according to the preset design plan at a microscopic scale, which is the cornerstone for ensuring the yield and consistency of circuit boards.

[0003] However, with the rapid iteration of electronic technology, the current circuit board design shows a trend of high complexity. On the one hand, special-shaped circuit boards and circuit boards with irregular functional partitions have gradually become the norm. The traditional target positioning technology based on image recognition is limited by its relatively fixed recognition mode and algorithm adaptability. It is difficult to accurately fit the geometric contour characteristics of such complex circuit boards, resulting in frequent positioning deviations during the processing process, and unable to meet the stringent requirements of refined production control. On the other hand, driven by the miniaturization and multifunctionality of electronic products, multi-layer circuit board architectures have emerged. Multi-layer circuit boards are formed by laminating and compounding single core boards with pre-prepared circuit patterns. This architecture greatly expands the freedom of circuit layout in a limited space and strongly supports the high-density output design requirements of electronic equipment. It is worth noting that the alignment accuracy of each single core board during the lamination process has become a bottleneck restricting the quality improvement of multi-layer circuit boards. Since the existing technology is difficult to achieve high-precision alignment calibration of each core board inside a multi-layer circuit board, the accumulation of small offsets between layers can easily cause a series of problems such as signal transmission interference and unstable electrical performance, which seriously hinders the development of multi-layer circuit boards towards higher integration and better performance. In order to optimize the alignment control of multi-layer circuit boards and improve their alignment accuracy, we proposed a circuit board production control method and system based on image recognition. Summary of the invention

[0004] The purpose of the present invention is to realize adaptive dynamic adjustment of multi-layer circuit board control strategy, improve the alignment accuracy of multi-layer circuit boards, and provide a circuit board production control method and system based on image recognition. The technical solutions adopted are as follows: The technical solution of the first aspect of the present invention provides a circuit board production control method based on image recognition, the method comprising: Acquire sample image data and point cloud image data of the asymmetric target and perform preprocessing; Construct and train a data fusion model, and use the preprocessed sample image data and point cloud image data to extract the three-dimensional spatial information feature map sequence of the asymmetric target; A multi-layer circuit board production control model is constructed and trained based on a generative adversarial network, and the asymmetric target-related features are extracted using a sequence of three-dimensional spatial information feature maps to output a preliminary control strategy. Optimize the multi-layer circuit board production control module with the goal of minimizing the impact of multi-source interference factors; Acquire real-time image data of asymmetric targets and use the optimized multi-layer circuit board production control model to output the multi-layer circuit board alignment control strategy.

[0005] Furthermore, obtaining sample image data and point cloud image data of the asymmetric target and performing preprocessing includes: Obtain sample image data and point cloud image data of a circuit board with an asymmetric target at a target process node, and mark the key areas of the asymmetric target, including key marking point areas for positioning and areas sensitive to shape changes; Based on the adaptive threshold bilateral filtering algorithm, the sample image data of the asymmetric target is preprocessed for denoising; The point cloud image data of the asymmetric target is sampled based on voxel grid downsampling, and outliers are removed according to the field density distribution of the point cloud image data of the asymmetric target after sampling.

[0006] Furthermore, obtaining sample image data and point cloud image data of the asymmetric target and performing preprocessing also includes: According to the preset compensation coefficient, the preprocessed sample image data is transformed, and its expression is: ; In the formula, Represents the transformed sample image data coordinates; It indicates the preset compensation coefficient, which is determined according to the uncertain factors in the production process of multi-layer circuit boards; Indicates lateral position; Indicates the vertical position.

[0007] Furthermore, obtaining sample image data and point cloud image data of the asymmetric target and performing preprocessing also includes: According to the production characteristics of multi-layer circuit boards, the pre-processed point cloud image data is transformed based on fractional Fourier transform, and its expression is:

[0008] In the formula, Represents the transformed point cloud image data coordinates; represents the fractional Fourier transform kernel function; Represents the data coordinates of the original point cloud image data in the polar coordinate system; represents radial distance; Indicates azimuth; Indicates the altitude angle; represents the integral variable; Among them, the expression of the fractional Fourier transform kernel function is:

[0009] In the formula, represents the angle parameter, , Indicates the fractional order, which is used to reflect the production interference at different angles; Represents an imaginary unit.

[0010] Furthermore, a data fusion model is constructed and trained, and a sequence of three-dimensional spatial information feature maps of asymmetric targets is extracted using preprocessed sample image data and point cloud image data, including: Based on a multi-layer convolutional neural network, the edge contour and color texture features of the asymmetric target are extracted using the preprocessed sample image data to obtain a sample image feature map; Based on the three-dimensional convolution kernel, the spatial geometric features of the asymmetric target are extracted using the preprocessed point cloud image data to obtain the point cloud image feature map; Based on the attention mechanism, the sample image feature map and the point cloud image feature map are globally averaged pooled to obtain the sample image feature vector and the point cloud image feature vector; Use the fully connected layer to concatenate the sample image feature vector and the point cloud image feature vector to extract the attention weight matrix; The sample image feature map and the point cloud image feature map are fused using the attention weight matrix to obtain a sequence of three-dimensional spatial information feature maps of the asymmetric target.

[0011] Furthermore, a multi-layer circuit board production control model is constructed and trained based on a generative adversarial network, and the asymmetric target-related features are extracted using the three-dimensional spatial information feature map sequence. The output of the preliminary control strategy includes: The three-dimensional spatial information feature map sequence output by the data fusion model is used as the input of the generator; The generator is based on a long short-term memory network and outputs a sequence of feature vectors after time series fusion according to the time series dependencies in the production process of multi-layer circuit boards. Use fully connected layers to extract key information for multi-layer circuit board production control; The deconvolution layer is used to map the key information of multi-layer PCB production control back to the spatial dimension of the multi-layer PCB control strategy, and the preliminary control strategy matrix is ​​extracted.

[0012] Furthermore, a multi-layer circuit board production control model is constructed and trained based on a generative adversarial network, and the asymmetric target-related features are extracted using a sequence of three-dimensional spatial information feature maps. The output of the preliminary control strategy also includes: The preliminary control strategy matrix and the preset multi-layer circuit board production control strategy samples are used as the input of the discriminator, and the convolution layer is used to extract the circuit board positioning control strategy features; The circuit board positioning control strategy features are mapped to scalar values ​​using a fully connected layer, and the discriminator parameters are updated based on the adversarial loss.

[0013] Furthermore, the multi-layer circuit board production control module is optimized with the goal of minimizing the impact of multi-source interference factors, including: Based on kernel density estimation, the characteristic distribution law of the three-dimensional spatial information feature map sequence of the key area of ​​the asymmetric target under the target production state is extracted; Based on the preset structural points of the asymmetric target, the three-dimensional spatial information feature vector in the three-dimensional spatial information feature map sequence is divided into a number of sub-regions; The probability distribution of each sub-region feature vector is calculated, and the uncertainty entropy value is calculated according to the probability distribution of each sub-region feature vector.

[0014] Furthermore, the expression for minimizing the impact of multi-source interference factors is: ; In the formula, Represents the uncertainty entropy value; Indicates sub-area and The correlation coefficient matrix of represents the probability distribution of the sub-region feature vector; Represents the preset weight coefficient, ; Parameters that represent the degree of influence of the correlation of the regulator region on the entropy value; represents the total number of sub-regions; Indicates sub-area The preset weight coefficient of Indicates sub-area The preset weight coefficient of .

[0015] The technical solution of the second aspect of the present invention provides a circuit board production control system based on image recognition, using the circuit board production control method based on image recognition described in the technical solution of the first aspect of the present invention, the system includes: A data acquisition module, configured to acquire sample image data and point cloud image data of the asymmetric target and perform preprocessing; A data fusion module is configured to construct and train a data fusion model, and to extract a three-dimensional spatial information feature map sequence of an asymmetric target using the preprocessed sample image data and the point cloud image data; A production control module is configured to construct and train a multi-layer circuit board production control model based on a generative adversarial network, extract asymmetric target-related features using a three-dimensional spatial information feature map sequence, and output a preliminary control strategy; An optimization module configured to optimize a multi-layer circuit board production control module with the goal of minimizing the impact of multi-source interference factors; The circuit board alignment strategy output module is configured to obtain real-time image data of an asymmetric target and output a multi-layer circuit board alignment control strategy using an optimized multi-layer circuit board production control model.

[0016] The present invention has the following beneficial effects: The circuit board production control method based on image recognition provided by the present invention constructs a data fusion model, uses a large number of sample training to mine image data, and extracts a three-dimensional spatial information feature map sequence of an asymmetric target, wherein the three-dimensional spatial information feature map sequence includes the all-round spatial characteristics of the asymmetric target; then, a multi-layer circuit board production control model is constructed using a generative adversarial network, and a three-dimensional spatial information feature map sequence is input. The model can quickly lock the key features of the asymmetric target and output a preliminary control strategy; at the same time, the production control model is optimized considering the multi-source interference factors in the actual production of the multi-layer circuit board to ensure the accuracy of the alignment control strategy. Finally, the asymmetric target image is collected in real time, and the optimized production control model is input, and the accurate multi-layer circuit board alignment control strategy is output in real time, which dynamically adapts to the changes in the production process of the multi-layer circuit board, optimizes the alignment control strategy of the multi-layer circuit board according to the asymmetric target, overcomes the influence of multi-source interference factors, significantly improves the alignment accuracy of the multi-layer circuit board, ensures the production quality of the circuit board, and improves production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 A method flow chart of a circuit board production control method based on image recognition provided by one embodiment of the present invention; Figure 2A schematic diagram of the structure of a circuit board production control system based on image recognition provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the circuit board production control method and system based on image recognition proposed by the present invention, its specific implementation method, structure, features and effects, in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0020] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0021] The specific scheme of a circuit board production control method and system based on image recognition provided by the present invention is described in detail below in conjunction with the accompanying drawings.

[0022] See also Figure 1 , which shows a method flow chart of a circuit board production control method based on image recognition provided by an embodiment of the present invention, the method comprising: Step S100: acquiring sample image data and point cloud image data of an asymmetric target and performing preprocessing; Step S100 specifically includes: Step S110: Obtain sample image data and point cloud image data of a multi-layer circuit board with an asymmetric target under different interference factors, and mark the key areas of the asymmetric target, which include key marking point areas for positioning and shape change sensitive areas; specifically, according to different raw material batches, high-resolution industrial cameras and high-precision laser radars are deployed at key process nodes such as drilling, copper plating, and etching on the multi-layer circuit board production line; for asymmetric targets, positions such as the center of the positioning hole and the corner points of a specific pattern can be selected according to the multi-layer circuit board design specifications, and their coordinate ranges can be accurately marked by manual combination with professional image annotation software; the shape change sensitive area is determined based on experience and the parts prone to deformation in the production process; Step S120: Based on the adaptive threshold bilateral filtering algorithm, the sample image data is subjected to denoising preprocessing; specifically, the pixel coordinates in the sample image data are , whose gray value is , first calculate the spatial weight coefficient, its expression is: ; Calculate the grayscale weight coefficient, the expression is: ; In the formula, Represents pixel With the points in the neighborhood The Euclidean distance of; The spatial standard deviation representing the effect of controlling spatial proximity can be dynamically adjusted based on the target and overall board size and image resolution; Represents pixel Gray value of Represents pixel Gray value of It represents the standard deviation of the grayscale similarity control effect, which can be set according to the grayscale dynamic range of the image; Pixels after filtering Gray value The calculation formula is: ; In the formula, Represents the gray value of the pixel after filtering; Indicated in pixels The formula ensures that the new gray value of each pixel is the weighted average of the surrounding pixels, where the weight is determined by the spatial and gray similarity. This embodiment uses an adaptive threshold bilateral filtering algorithm to remove noise while retaining detail features, thereby improving the quality of sample image data.

[0023] Step S130: Sampling the point cloud image data based on voxel grid downsampling, and removing outliers according to the field density distribution of the sampled point cloud image data; specifically, firstly, based on voxel grid downsampling, the voxel grid side length is set according to the size and accuracy requirements of the multi-layer circuit board. , divide the three-dimensional space into For each point cloud in the grid, a representative point can be reserved using the centroid method. The coordinates of the point cloud points in the grid are , ( , represents the number of points in the grid), calculate the coordinates of the retained points ; Then, outliers are removed according to the field density distribution of the sampled point cloud image data, such as point ,That The set of neighbor points is The number of layers of a multi-layer circuit board is ,point The layer is Layer, define the influence coefficient of the target layer (It can be determined based on the impact of each layer on the overall alignment accuracy, and the value range is between 0 and 1). Considering the correlation of each layer of the multi-layer circuit board, the point cloud density can be expressed as:

[0024] In the formula, Indicate point Point cloud density; Indicates the number of neighboring points considered when calculating density; Represents the influence coefficient of the target layer; Indicates preset parameters, which can be selected according to the complexity of the multi-layer circuit board structure. For example, the structure is simple , complex structure or 3; Indicate point The k nearest neighbor point set of ; Indicate point and its neighbors The Euclidean distance between Then remove the outliers, if , It can be determined based on experience or the accuracy requirements of the multi-layer circuit board, and then it is judged as an outlier and removed, so that the point cloud data can more purely and accurately reflect the spatial characteristics of the multi-layer circuit board, which is beneficial to the subsequent optimization of the alignment control strategy based on the asymmetric target; this embodiment effectively reduces the amount of point cloud data by downsampling the voxel grid, while ensuring the retention of important features; and then accurately removes outliers by performing field density analysis on the sampled point cloud image data, ensuring that the point cloud data is more accurate, and can better reflect the spatial characteristics of the multi-layer circuit board, which helps to achieve more accurate and efficient multi-layer circuit board alignment control.

[0025] Step S140: transform the preprocessed sample image data according to a preset compensation coefficient, which is expressed as:

[0026] In the formula, Represents the transformed sample image data coordinates; Indicates the preset compensation coefficient, which can be determined according to the uncertain factors in the multi-layer circuit board production process; this step uses mathematical transformation to correct the image distortion caused by uncertain factors in the multi-layer circuit board production process, such as mechanical deformation, optical distortion, etc. For the deformation in the direction x, it can be set according to previous experience and the professional judgment of technicians The initial value of is set and continuously optimized; by performing precise geometric transformation in the x-direction on the preprocessed sample image data, the image distortion caused by various uncertain factors in the production process of multi-layer circuit boards can be effectively corrected, the image quality is further improved, and the accuracy of subsequent feature extraction and model training is ensured, thereby improving the recognition ability and positioning accuracy of the entire production control system for asymmetric targets.

[0027] Step S150: transform the preprocessed point cloud image data based on fractional Fourier transform according to the production characteristics of the multi-layer circuit board; specifically, the present embodiment uses fractional Fourier transform to transform the preprocessed point cloud image data, the purpose of which is to adapt to the characteristics of the multi-layer circuit board production process; fractional Fourier transform, as a generalized Fourier transform, can provide a continuous transition between time and frequency, and can effectively capture the angular deviation and rotational distortion that may be introduced in the production process of the multi-layer circuit board, and optimize feature extraction by selecting a suitable fractional order; first, the points in the point cloud image data are transformed into the image data; Convert to polar coordinates , Indicates radial distance, Indicates the azimuth, Represents the altitude angle; then the fractional Fourier transform kernel function is used to transform the data in polar coordinates, which can be expressed as:

[0028] In the formula, Represents the transformed point cloud image data coordinates; represents the fractional Fourier transform kernel function; Represents the data coordinates of the original point cloud image data in the polar coordinate system; represents radial distance; Indicates azimuth; Indicates the altitude angle; represents the integral variable; Among them, the expression of the fractional Fourier transform kernel function is:

[0029] In the formula, represents the angle parameter, , It represents the fractional order, and its value range is between 0 and 2. The specific value can be determined according to the experiment to reflect the production interference under different angles; Represents an imaginary unit.

[0030] This implementation can effectively capture and correct the angular deviation and rotational distortion that may occur during the production of multi-layer circuit boards by applying fractional Fourier transform to the pre-processed point cloud image data; this method is particularly suitable for dealing with the feature extraction problem of asymmetric targets at different angles, and by selecting the appropriate fractional order, it can better adapt to the interference situation in a specific production environment. This step not only improves the accuracy of subsequent feature extraction and model training, but also provides support for the realization of high-precision multi-layer circuit board alignment control, ensuring stability and reliability in complex and changing production environments.

[0031] Step S200: constructing and training a data fusion model, and extracting a three-dimensional spatial information feature map sequence of an asymmetric target using the preprocessed sample image data and point cloud image data; Step S200 specifically includes: Step S210: Based on a multi-layer convolutional neural network, the edge contour and color texture features of the asymmetric target are extracted using the preprocessed sample image data to obtain a sample image feature map; specifically, this embodiment constructs a convolutional neural network containing 6-8 convolutional layers; the size of the first convolution kernel is set to 7×7, the step size is 2, the number of input channels is set according to the number of channels of the sample image data, and the number of output channels is set to 64. For a certain pixel point in the input preprocessed multi-layer circuit board sample image data (indicates its coordinate position in the image), the eigenvalue calculation formula obtained after the first layer of convolution operation is: ; in, It represents the weight parameter of the first layer convolution kernel; represents the first bias term; represents the activation function; the size of the convolution kernel of the subsequent convolutional layers decreases successively, and the number of output channels changes reasonably as needed. Each layer performs convolution operations and activations according to the above formula, and finally obtains a sample image feature map reflecting the edge contour and color texture characteristics of the asymmetric target; at the same time, residual connections can be integrated into the network in view of the characteristics of multi-layer circuit boards. For example, the input of a certain layer is , the output after two layers of convolution transformation is , the output after residual connection is , so as to better retain the feature information of asymmetric targets on multi-layer circuit boards, avoid the problem of gradient vanishing, and make feature extraction more sufficient; use the preprocessed multi-layer circuit board sample image data as input, adjust the network weights through the back propagation algorithm, so that the network can learn the edge contour and color texture features of the asymmetric target. This embodiment uses a multi-layer convolutional neural network to extract the edge contour and color texture features of the asymmetric target from the preprocessed sample image data, and generates a high-quality sample image feature map. The network can not only capture fine-grained local features, but also prevent gradient vanishing through the residual connection mechanism, ensure the effective extraction of deep-level features, and accurately understand the asymmetric targets on multi-layer circuit boards, thereby helping to optimize the alignment control strategy.

[0032] Step S220: Based on the three-dimensional convolution kernel, the spatial geometric features of the asymmetric target are extracted using the preprocessed point cloud image data to obtain a point cloud image feature map; specifically, this embodiment uses a network architecture of 4-6 layers of three-dimensional convolution layers to process the point cloud image data after fractional Fourier transform; the size of the first layer of three-dimensional convolution kernel is set to 5×5×5, the step size is 2, the number of input channels is determined by the point cloud data dimension, and the number of output channels is set to 32. For a point in the point cloud image data , the eigenvalue calculation formula after the first layer of three-dimensional convolution is as follows: ; here, Represents the weight parameters of the first layer of three-dimensional convolution kernel; represents the second bias term; It represents the feature value of the corresponding position in the new feature map obtained after the first layer of three-dimensional convolution operation. The convolution kernel size of the subsequent convolution layer is gradually reduced to a smaller size, the step size is kept at 1, and the number of output channels is reasonably adjusted according to needs. The spatial geometric features of the asymmetric target are extracted through layers of three-dimensional convolution operations, and finally the point cloud image feature map is obtained. This embodiment takes into account that there may be spatial position correlations between different layers of a multi-layer circuit board that affect the point cloud features. A spatial attention mechanism is introduced after each layer of three-dimensional convolution to calculate the importance weights of each point in the point cloud at different spatial positions. Specifically, for point The spatial attention weight is calculated as: ; in, Represents the feature map output by the current 3D convolutional layer; Indicate point The corresponding eigenvector, represents the average pooling operation, Represents the maximum pooling operation; MLP is a multi-layer perceptron; represents the feature vector concatenation operation; Represents the spatial attention weight, which is used to perform weighted updates on the feature map, thereby focusing more on the key spatial geometric features of the asymmetric target embodied in the point cloud of the multi-layer circuit board; This implementation uses a multi-layer 3D convolutional neural network to effectively extract the spatial geometric features of asymmetric targets from the point cloud image data after fractional Fourier transformation, and generates a high-resolution point cloud image feature map. The network can not only capture the fine-grained local spatial structure, but also enhance the expressiveness of key features through the spatial attention mechanism, ensuring the effective extraction of deep-level features.

[0033] Step S230: Based on the attention mechanism, perform global average pooling on the sample image feature map and the point cloud image feature map, respectively, to obtain a sample image feature vector and a point cloud image feature vector; In this embodiment, global average pooling (GAP) is applied to the sample image feature map and the point cloud image feature map, respectively, with the purpose of averaging all values ​​on each channel in each feature map to obtain a single value representing the average response of the channel on the entire feature map, which helps to reduce the spatial dimension of the feature map and retain important channel information; For the sample image feature map And point cloud image feature map , the feature map sizes obtained after three-dimensional convolution are , ,in , are the height and width respectively, is the number of channels; for the sample image feature map , calculate the average value of each channel to obtain a size of The sample image feature vector ; Similarly, for the point cloud image feature map Perform the same operation to obtain a The point cloud image feature vector ; Step S240: Use the fully connected layer to concatenate the sample image feature vector and the point cloud image feature vector to extract the attention weight matrix; specifically, concatenate the sample image feature vector and the point cloud image feature vector, and the concatenated vector dimension is , suppose the concatenated vector can be expressed as ; Then a fully connected layer is used to extract the attention weight matrix, and the weight matrix of the fully connected layer is set to , whose dimensions are , is the preset number of attention heads, and the bias term is set to , after the full connection layer and activation function, the attention weight matrix is ​​obtained , which can be expressed as: ; Step S250: Using the attention weight matrix to fuse the sample image feature map and the point cloud image feature map to obtain a sequence of three-dimensional spatial information feature maps of the asymmetric target; specifically, using the obtained attention weight matrix Sample image feature map And point cloud image feature map Fusion; First, the sample image feature map and the point cloud image feature map are flattened according to the channel dimension to obtain two two-dimensional matrices with sizes and ; Then concatenate them column by column into a large two-dimensional matrix , size is ; Then, through the attention weight matrix right Perform weighted summation to obtain the fused feature matrix, which is expressed as: ; In the formula, As a sequence of three-dimensional spatial information feature maps of asymmetric targets, it is used in the subsequent circuit board production control model; this embodiment combines sample images and point cloud data and introduces an attention mechanism to more accurately capture and process the key features of asymmetric targets on multi-layer circuit boards; global average pooling helps reduce redundant information, while the attention mechanism allows the model to focus on the most relevant parts, thereby improving the quality of feature representation. This method not only improves recognition accuracy, but also enhances the model's ability to adapt to complex environments, making quality control of circuit board production more reliable and efficient.

[0034] Step S300: constructing and training a multi-layer circuit board production control model based on a generative adversarial network, extracting asymmetric target-related features using a three-dimensional spatial information feature map sequence, and outputting a preliminary control strategy; Step S300 specifically includes: Step S310: The three-dimensional spatial information feature map sequence output by the data fusion model is used as the input of the generator; we use the three-dimensional spatial information feature map sequence extracted in step S200 as the initial input of the generator. This feature map sequence has integrated the key information of the multi-layer circuit board sample image and the point cloud image about the asymmetric target, and its dimension is , It is used to reflect the timing characteristics of the multi-layer circuit board production process. These data will serve as the input of the generator part in the Generative Adversarial Network (GAN); Step S320: Based on the long short-term memory network, the generator outputs a sequence of feature vectors after time series fusion according to the time series dependency in the multi-layer circuit board production process; specifically, a long short-term memory network module is constructed inside the generator. For the input feature map sequence, for each time step, the LSTM unit outputs a sequence of feature vectors after time series fusion according to the time series dependency in the multi-layer circuit board production process. The state at the previous moment outputs a new state, and the gating mechanism inside the LSTM (input gate , Forget Gate , output gate ) and memory cells Allows the model to learn and remember long-term dependencies. In this way, LSTM can capture the timing dependencies in the multi-layer circuit board production process and output a time-fused feature vector sequence, which helps to more accurately refine the dynamic characteristics of the production process; Step S330: extract key information of multi-layer circuit board production control by using a fully connected layer; immediately after the LSTM output, access a series of fully connected layers; specifically, in step S330, a series of fully connected layers are used to further process the feature vector sequence output by the LSTM, and the fully connected layer can extract key information of multi-layer circuit board production control from the feature vector sequence after time series fusion, thereby preparing for the next step of mapping; Step S340: using a deconvolution layer to map the key information of the multi-layer circuit board production control back to the spatial dimension of the multi-layer circuit board control strategy, and extracting a preliminary control strategy matrix; specifically, in step S340, the deconvolution layer is used to remap the key information extracted from the previous step back to the spatial dimension of the multi-layer circuit board control strategy, and the output result is a preliminary control strategy matrix, which contains strategy information for guiding the production of multi-layer circuit boards; Step S350: taking the preliminary control strategy matrix and the preset multi-layer circuit board production control strategy samples as inputs of the discriminator, and using the convolution layer to extract control strategy features; in step S350, the preliminary control strategy matrix generated by the generator and the preset multi-layer circuit board production control strategy samples are taken as inputs of the discriminator, and the discriminator is composed of a deep convolutional neural network, which gradually and finely extracts the differences in the true and false features of the control strategy through the convolution layer to help determine whether the generated control strategy is reasonable and effective; Step S360: Mapping the control strategy features to a scalar value using a fully connected layer, and updating the discriminator parameters based on the adversarial loss; In this embodiment, the extracted control strategy features are mapped to a scalar value through a fully connected layer, and the parameters of the discriminator can be updated based on the adversarial loss function, so that the discriminator can better distinguish between the real and generated control strategies, thereby prompting the generator to continuously improve its output, and ultimately achieve the purpose of optimizing the control strategy for multi-layer circuit board production; This embodiment integrates the advantages of image recognition technology and generative adversarial networks to achieve precise control of asymmetric targets in the production process of multi-layer circuit boards. By introducing LSTM to capture the timing dependencies in the production process and combining it with the adversarial training mechanism of the generative adversarial network, it not only improves the accuracy of the control strategy, but also enhances the learning ability of the model, enabling it to adapt to complex changes in the production environment.

[0035] Step S400: Optimizing the multi-layer circuit board production control module with the goal of minimizing the impact of multi-source interference factors; Step S400 specifically includes: Step S410: Based on kernel density estimation, extract the characteristic distribution law of the three-dimensional spatial information feature map sequence of the asymmetric target key area under the target production state; specifically, from the output of the data fusion model, accurately locate and extract the three-dimensional spatial information feature map sequence set containing only the asymmetric target key area, and record it as , the feature map sequence contains the comprehensive features of the key area of ​​the asymmetric target on the multi-layer circuit board under different interferences after the previous multi-step processing; the feature dimension of the key area is set to d dimension, which can be obtained by flattening the feature map sequence according to certain rules or extracting the key dimension information according to the structural characteristics of the multi-layer circuit board target; this embodiment uses the kernel density estimation method to select the Gaussian kernel function , find the optimal kernel bandwidth through cross-validation ; For each feature vector in the set , It means The first feature vectors, The value range is from 1 to M, where M is the number of eigenvectors contained in each feature graph sequence. The kernel density estimate in the d-dimensional space is calculated to obtain the kernel density estimation matrix D, where the matrix elements are This is the eigenvector Relative probability density estimate under the target production state. In this way, the characteristic distribution law of the key area under the target production state can be depicted; the characteristic distribution law of the key area of ​​the asymmetric target under the target production state can be accurately depicted. Through kernel density estimation, we can identify which characteristic vectors are more likely to appear under the ideal production state, thereby helping us to distinguish and reduce the impact of multi-source interference factors; and thus help to improve the robustness and accuracy of the multi-layer circuit board production control system, ensuring that high-quality production standards can be maintained even in the presence of multiple interferences.

[0036] Step S420: Based on the preset structure points of the asymmetric target, the three-dimensional spatial information feature vector in the three-dimensional spatial information feature map sequence is divided into several sub-regions; specifically, this embodiment divides the feature vector in the three-dimensional spatial information feature map sequence into several sub-regions based on the preset structure points of the asymmetric target; it is intended to pre-set specific structure points according to the design and production process requirements of the circuit board, combined with the parts that have a significant impact on the alignment accuracy of the multi-layer circuit board, and ensure that these structure points can be accurately identified during the image acquisition and processing process. The selection of structure points can be based on the design and process requirements of the circuit board, as well as those parts that have a greater impact on the alignment accuracy of the multi-layer circuit board, and select structure points such as the center of the positioning hole, the corner point of a specific shape, the center of the pattern on the asymmetric target; these preset structure point coordinates are converted from the original image coordinate system to the coordinate system of the three-dimensional spatial information feature map sequence to ensure that they can be correctly mapped to the spatial position of the three-dimensional spatial information feature map sequence; then, according to the actual production accuracy requirements and the resolution of the feature map, a suitable radius or side length is defined for each structure point to determine the range of the divided sub-region. Taking each structural point as the center, the three-dimensional spatial information feature vector is divided into several sub-areas according to the radius or side length range defined above. , indicating the sub-regions, and the value range is from 1 to K; it should be noted that, according to actual needs, different sub-region shapes can be selected, such as circle, rectangle, etc., to adapt to different application scenarios; this embodiment decomposes the complex three-dimensional spatial information feature vector into small units-sub-regions that are easier to manage and analyze; this method not only simplifies the subsequent data processing process, but also ensures that each sub-region contains sufficient feature information, thereby better reflecting the actual structural characteristics of the circuit board; and in this way, the production process of the multi-layer circuit board can be more finely controlled, especially in terms of positioning accuracy, which significantly improves production efficiency and product quality.

[0037] Step S430: Calculate the probability distribution of each sub-region feature vector, and calculate the uncertainty entropy value according to the probability distribution of each sub-region feature vector; specifically, for each sub-region , count the eigenvectors contained therein The frequency of the sub-region feature vector is calculated by combining the overall feature distribution law obtained by kernel density estimation. , It represents any eigenvector in the sub-region, and its calculation formula is as follows:

[0038] In the formula, It represents the feature vector In sub-area The frequency of occurrence; It is a coefficient used to balance the contribution of frequency and kernel density estimation, and its value range is between 0.1 and 0.5; It represents the probability density value of the feature vector in the corresponding sub-region based on the kernel density estimation; Represents the feature vector In sub-area The frequency of occurrence; Indicates sub-area Eigenvector The probability density value obtained based on the kernel density estimation; Then consider the entropy value of the sub-region correlation. First, calculate the correlation coefficient matrix R between the sub-regions. The elements in this matrix are Indicates the sub-area and The correlation between them can be obtained by calculating the covariance of the sub-region feature vector and normalizing it; the expression of the uncertainty entropy value is: ; In the formula, Represents the uncertainty entropy value; Indicates sub-area and The correlation coefficient matrix is ​​used to measure the mutual relationship or dependence between different sub-regions; represents the probability distribution of the sub-region feature vector; Represents the preset weight coefficient, which can be pre-set according to the importance of the sub-area to the alignment accuracy of the multi-layer circuit board. ; The parameter that indicates the influence of the correlation of the regulatory sub-region on the entropy value, and its value range is between 0.05 and 0.2; represents the total number of sub-regions; Indicates sub-area The preset weight coefficient of Indicates sub-area The preset weight coefficient of Represents the feature vector Belongs to sub-area ; In the actual production process, the data of asymmetric targets is collected in real time, and the uncertainty entropy value under the current production state is calculated according to the above process. , define the optimization objective function: the parameters of the multi-layer circuit board production control model (including the generator and the discriminator) are recorded as , the optimization objective function is defined as follows: In the actual production process, the data of the asymmetric target is collected in real time, and then the uncertainty entropy value under the current production state is calculated according to the above process The parameters of the multi-layer circuit board production control model (which consists of two parts, the generator and the discriminator) are recorded as , the optimization objective function is defined as follows:

[0039] In the formula, represents the objective function, Represents the regularization parameter, which ranges from 0.001 to 0.01 and is used to prevent the model from overfitting; Represents the model parameter vector Among them The L2 norm square of the parameters is used to constrain the parameters of the model; Then use stochastic gradient descent (SGD) or an adaptive optimization algorithm (such as the Adam algorithm) to minimize As a goal, the parameters of the multi-layer circuit board production control model are updated, and the update formula is as follows: ; In the formula, Represents the updated model parameters; represents the current model parameters, Represents the learning rate, and its initial value is set to 0.001, and it gradually decreases as the training progresses; by continuously iterative optimization, the model can output a more accurate multi-layer circuit board alignment control strategy when facing multi-source interference factors, thereby effectively minimizing the impact of multi-source interference factors; this embodiment provides a systematic means to evaluate and optimize the uncertainty in the production process of multi-layer circuit boards; by calculating the probability distribution of the feature vector of each sub-region and the corresponding uncertainty entropy value, we can quantify the uncertainty level of different sub-regions and adjust the production control strategy based on this; this method not only helps to improve the alignment accuracy, but also enhances the robustness and adaptability of the system, so that it can still maintain high-precision production standards in the presence of multiple interferences. In addition, by introducing the correlation between sub-regions, the entropy value calculation is further optimized, so that the model can consider the global information more comprehensively, thereby outputting a more accurate control strategy, and ultimately achieving the goal of effectively minimizing the impact of multi-source interference factors.

[0040] Step S500: Acquire real-time image data of the asymmetric target, and use the optimized multi-layer circuit board production control model to output the multi-layer circuit board alignment control strategy; specifically, perform denoising preprocessing on the collected real-time image data according to the adaptive threshold bilateral filtering algorithm in step S120; input the preprocessed real-time image data and point cloud image data into the data fusion model constructed and trained in step S200, first extract the edge contour and color texture features of the asymmetric target from the image data through a multi-layer convolutional neural network to obtain a sample image feature map, extract spatial geometric features from the point cloud data through a three-dimensional convolution kernel to obtain a point cloud image feature map, then perform global average pooling based on the attention mechanism, splice to extract the attention weight matrix, and finally use the weight matrix fusion to obtain the asymmetric target. A real-time three-dimensional spatial information feature map sequence of symmetric targets, which integrates the key features of the multi-modal asymmetric targets on the multi-layer circuit board in real time; then the real-time three-dimensional spatial information feature map sequence obtained above is input into the multi-layer circuit board production control model (including generator and discriminator) optimized by step S400; in the generator part, according to the processing method based on long short-term memory network (LSTM) in step S320, the real-time timing dependency of the multi-layer circuit board production process is considered, and the feature vector sequence after timing fusion is output, and then the key information of the multi-layer circuit board production control is extracted through the fully connected layer of step S330, and then the deconvolution layer of step S340 is used to map the key information back to the spatial dimension of the multi-layer circuit board control strategy, and extract the real-time preliminary control strategy matrix.

[0041] In addition, in this embodiment, a correction coefficient based on the difference between the real-time feature and the historical average feature is introduced, and the difference correction coefficient is defined as: ; In the formula, represents the difference correction factor; Represents a real-time feature vector sequence; Represents the average of the historical feature vector sequence; by multiplying each element of the preliminary control strategy matrix by Make corrections to better adapt to the differences between real-time production status and historical status; Afterwards, the corrected preliminary control strategy matrix and the preset multi-layer circuit board production control strategy samples are used as the input of the discriminator. The control strategy features are extracted using the convolutional layer and then mapped to scalar values ​​through the fully connected layer. After the discriminator makes judgments based on adversarial losses and optimizes the model internally, the multi-layer circuit board alignment control strategy is finally output. The strategy contains the alignment adjustment parameters such as the accurate displacement and rotation angle of each layer of the circuit board in the current real-time state, which directly guides the production equipment to perform corresponding operations. In summary, this embodiment is based on the data processing, model building and optimization of multi-layer circuit boards and asymmetric targets in each step. By collecting and preprocessing real-time data, it effectively integrates real-time features and uses innovative correction mechanisms to adapt to changes in production status. With the help of the optimized production control model to output a positioning control strategy that fits the actual situation, real-time and precise control of the multi-layer circuit board production process is achieved. This method significantly improves the positioning accuracy of multi-layer circuit boards, reduces production errors caused by interference factors, and ensures product quality and production efficiency. In addition, the introduced difference correction coefficient further enhances the robustness and adaptability of the system, so that high-precision production standards can be maintained even in the presence of multiple interferences.

[0042] See also Figure 2 , which shows a schematic diagram of the structure of a circuit board production control system based on image recognition provided by an embodiment of the present invention, the system comprising: A data acquisition module, configured to acquire sample image data and point cloud image data of the asymmetric target and perform preprocessing; A data fusion module is configured to construct and train a data fusion model, and to extract a three-dimensional spatial information feature map sequence of an asymmetric target using the preprocessed sample image data and the point cloud image data; A production control module is configured to construct and train a multi-layer circuit board production control model based on a generative adversarial network, extract asymmetric target-related features using a three-dimensional spatial information feature map sequence, and output a preliminary control strategy; An optimization module configured to optimize a multi-layer circuit board production control module with the goal of minimizing the impact of multi-source interference factors; The circuit board alignment strategy output module is configured to obtain real-time image data of an asymmetric target and output a multi-layer circuit board alignment control strategy using an optimized multi-layer circuit board production control model.

[0043] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0044] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

Claims

1. A circuit board production control method based on image recognition, characterized in that: The method comprises: Acquire sample image data and point cloud image data of the asymmetric target and perform preprocessing; Construct and train a data fusion model, and use the preprocessed sample image data and point cloud image data to extract the three-dimensional spatial information feature map sequence of the asymmetric target; A multi-layer circuit board production control model is constructed and trained based on a generative adversarial network, and the asymmetric target-related features are extracted using a sequence of three-dimensional spatial information feature maps to output a preliminary control strategy. Optimize the multi-layer circuit board production control module with the goal of minimizing the impact of multi-source interference factors; Acquire real-time image data of asymmetric targets and use the optimized multi-layer circuit board production control model to output the multi-layer circuit board alignment control strategy.

2. The circuit board production control method based on image recognition according to claim 1, characterized in that: Acquiring sample image data and point cloud image data of an asymmetric target and preprocessing them include: Obtain sample image data and point cloud image data of a circuit board with an asymmetric target at a target process node, and mark the key areas of the asymmetric target, including key marking point areas for positioning and areas sensitive to shape changes; Based on the adaptive threshold bilateral filtering algorithm, the sample image data of the asymmetric target is preprocessed for denoising; The point cloud image data of the asymmetric target is sampled based on voxel grid downsampling, and outliers are removed according to the field density distribution of the point cloud image data of the asymmetric target after sampling.

3. The circuit board production control method based on image recognition according to claim 2, characterized in that: Acquiring sample image data and point cloud image data of an asymmetric target and preprocessing the data also includes: According to the preset compensation coefficient, the preprocessed sample image data is transformed, and its expression is: ; In the formula, Represents the transformed sample image data coordinates; It indicates the preset compensation coefficient, which is determined according to the uncertain factors in the production process of multi-layer circuit boards; Indicates lateral position; Indicates the vertical position.

4. The circuit board production control method based on image recognition according to claim 2, characterized in that: Acquiring sample image data and point cloud image data of an asymmetric target and preprocessing the data also includes: According to the production characteristics of multi-layer circuit boards, the pre-processed point cloud image data is transformed based on fractional Fourier transform, and its expression is: In the formula, Represents the transformed point cloud image data coordinates; represents the fractional Fourier transform kernel function; Represents the data coordinates of the original point cloud image data in the polar coordinate system; represents radial distance; Indicates azimuth; Indicates the altitude angle; represents the integral variable; Among them, the expression of the fractional Fourier transform kernel function is: In the formula, Represents the angle parameter, , Indicates the fractional order, which is used to reflect the production interference at different angles; Represents an imaginary unit.

5. The circuit board production control method based on image recognition according to claim 1, characterized in that: Construct and train a data fusion model, and use the preprocessed sample image data and point cloud image data to extract the three-dimensional spatial information feature map sequence of the asymmetric target, including: Based on a multi-layer convolutional neural network, the edge contour and color texture features of the asymmetric target are extracted using the preprocessed sample image data to obtain a sample image feature map; Based on the three-dimensional convolution kernel, the spatial geometric features of the asymmetric target are extracted using the preprocessed point cloud image data to obtain the point cloud image feature map; Based on the attention mechanism, the sample image feature map and the point cloud image feature map are globally averaged pooled to obtain the sample image feature vector and the point cloud image feature vector; Use the fully connected layer to concatenate the sample image feature vector and the point cloud image feature vector to extract the attention weight matrix; The sample image feature map and the point cloud image feature map are fused using the attention weight matrix to obtain a sequence of three-dimensional spatial information feature maps of the asymmetric target.

6. The circuit board production control method based on image recognition according to claim 5, characterized in that: A multi-layer circuit board production control model is constructed and trained based on a generative adversarial network. The asymmetric target-related features are extracted using a sequence of three-dimensional spatial information feature maps. The output of the preliminary control strategy includes: The three-dimensional spatial information feature map sequence output by the data fusion model is used as the input of the generator; The generator is based on a long short-term memory network and outputs a sequence of feature vectors after time series fusion according to the time series dependencies in the production process of multi-layer circuit boards. Use fully connected layers to extract key information for multi-layer circuit board production control; The deconvolution layer is used to map the key information of multi-layer PCB production control back to the spatial dimension of the multi-layer PCB control strategy, and the preliminary control strategy matrix is ​​extracted.

7. The circuit board production control method based on image recognition according to claim 6, characterized in that: A multi-layer circuit board production control model is constructed and trained based on a generative adversarial network. The asymmetric target-related features are extracted using a sequence of three-dimensional spatial information feature maps. The output of the preliminary control strategy also includes: The preliminary control strategy matrix and the preset multi-layer circuit board production control strategy samples are used as the input of the discriminator, and the convolution layer is used to extract the circuit board positioning control strategy features; The circuit board positioning control strategy features are mapped to scalar values ​​using a fully connected layer, and the discriminator parameters are updated based on the adversarial loss.

8. The circuit board production control method based on image recognition according to any one of claims 1 to 7, characterized in that: The control module for optimizing the production of multi-layer circuit boards with the goal of minimizing the impact of multi-source interference factors includes: Based on kernel density estimation, the characteristic distribution law of the three-dimensional spatial information feature map sequence of the key area of ​​the asymmetric target under the target production state is extracted; Based on the preset structural points of the asymmetric target, the three-dimensional spatial information feature vector in the three-dimensional spatial information feature map sequence is divided into a number of sub-regions; The probability distribution of each sub-region feature vector is calculated, and the uncertainty entropy value is calculated according to the probability distribution of each sub-region feature vector.

9. The circuit board production control method based on image recognition according to claim 8, characterized in that: The expression for minimizing the impact of multi-source interference factors is: ; In the formula, Represents the uncertainty entropy value; Indicates sub-area and The correlation coefficient matrix of represents the probability distribution of the sub-region feature vector; Represents the preset weight coefficient, ; Parameters that represent the degree of influence of the correlation of the regulator region on the entropy value; represents the total number of sub-regions; Indicates sub-area The preset weight coefficient of Indicates sub-area The preset weight coefficient of .

10. A circuit board production control system based on image recognition, characterized in that: The circuit board production control method based on image recognition according to any one of claims 1 to 9 is adopted, and the system comprises: A data acquisition module, configured to acquire sample image data and point cloud image data of the asymmetric target and perform preprocessing; A data fusion module is configured to construct and train a data fusion model, and to extract a three-dimensional spatial information feature map sequence of an asymmetric target using the preprocessed sample image data and the point cloud image data; A production control module is configured to construct and train a multi-layer circuit board production control model based on a generative adversarial network, extract asymmetric target-related features using a three-dimensional spatial information feature map sequence, and output a preliminary control strategy; An optimization module configured to optimize a multi-layer circuit board production control module with the goal of minimizing the impact of multi-source interference factors; The circuit board alignment strategy output module is configured to obtain real-time image data of an asymmetric target and output a multi-layer circuit board alignment control strategy using an optimized multi-layer circuit board production control model.

Images