Conductive ink screen printing precision control method based on deep learning

Abstract

The invention discloses a conductive ink screen printing precision control method based on deep learning. The conductive ink screen printing precision control method comprises the following steps: acquiring microstructure image data of conductive ink on the surface of a screen printing substrate in real time; generating a thickness enhanced picture by adopting an improved competitive fuzzy C-means clustering algorithm; generating an edge enhancement picture by adopting an improved self-adaptive threshold function; carrying out non-subsampled Contourlet multi-scale decomposition to generate a microcrack enhanced picture; jointly inputting the enhanced pictures into a pre-trained defect recognition model, and extracting a multi-dimensional precision feature vector; dynamically generating a collaborative adjustment parameter set through the parameter prediction model, and carrying out iterative optimization to obtain an optimized parameter set; and according to the optimized parameter set, adjusting an execution mechanism of the screen printing equipment in real time through closed-loop feedback, and synchronously updating a weight coefficient of the defect identification model. According to the invention, real-time multi-dimensional perception of the microstructure of the ink is realized, the anti-interference capability of image processing is improved, a multi-parameter collaborative prediction model is established, and the microcrack detection sensitivity is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of printed electronics technology, and in particular to a conductive ink screen printing precision control method based on deep learning. Background Art

[0002] With the rapid development of printed electronics technology, conductive ink screen printing technology is increasingly used in flexible circuits, sensors and other fields. However, the printing accuracy control of conductive ink in existing technologies still faces multiple challenges: It is difficult to monitor microscopic morphological parameters in real time: Traditional methods rely on offline detection equipment (such as laser scanners), which cannot capture the dynamic changes of ink thickness distribution, edge diffusion and microcracks in the screen printing process in real time, resulting in delayed adjustment of process parameters; Insufficient anti-interference ability: Existing image processing algorithms (such as classic Canny edge detection) are prone to produce artifacts under complex noise (such as substrate texture interference, ink particle scattering noise), resulting in reduced ink area segmentation accuracy (error rate > 15%); Difficulty in co-optimizing parameters: The coupling mechanism of multiple parameters such as printing pressure, scraper angle, screen tension, etc. is complex, manual experience adjustment is inefficient (single debugging takes >30 minutes), and it is difficult to achieve global optimization; Low sensitivity in micro-crack detection: Conventional optical detection technology has an identification rate of less than 60% for sub-micron cracks, which can easily lead to the risk of circuit breakage in conductive lines.

[0003] In the existing technology, although some studies have tried to combine fuzzy clustering (such as traditional FCM algorithm) or wavelet transform (such as Haar wavelet) for image enhancement, there are problems such as fuzzy segmentation boundary (fixed fuzzy factor) and lack of multimodal feature fusion. In addition, the PID control method based on single-dimensional feedback has poor adaptability to dynamic process scenarios and is difficult to meet the needs of high-precision screen printing. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the present invention provides a conductive ink screen printing precision control method based on deep learning, which can realize real-time multi-dimensional perception of the ink micromorphology, improve the anti-interference ability of image processing, establish a multi-parameter collaborative prediction model, and enhance the sensitivity of microcrack detection.

[0005] To solve the above problems, the technical solution adopted by the present invention is as follows: A conductive ink screen printing precision control method based on deep learning, comprising the following steps: Real-time acquisition of microscopic image data of conductive ink on the surface of screen-printed substrate; The thickness enhanced images are generated by using the improved competitive fuzzy C-means clustering algorithm; An improved adaptive threshold function is used to generate edge-enhanced images; Generate microcrack enhancement images through non-subsampled Contourlet multi-scale decomposition; The enhanced images are jointly input into the pre-trained defect recognition model to extract multi-dimensional precision feature vectors; Dynamically generate collaborative adjustment parameter sets through parameter prediction models and iteratively optimize to obtain optimized parameter sets; The actuators of the screen printing equipment are adjusted in real time through closed-loop feedback according to the optimized parameter set, and the weight coefficients of the defect recognition model are updated synchronously.

[0006] As a preferred embodiment of the present invention, generating a thickness enhancement image includes: Mean shift filtering is used to preprocess the microscopic image data; Perform KL transform on the filtered image and select the first two principal components to construct a two-dimensional feature plane; An improved competitive fuzzy C-means clustering algorithm is used on the two-dimensional feature plane to dynamically adjust the competition radius of cluster centers to achieve unsupervised segmentation of ink areas. The segmented image is subjected to multiple erosion and dilation operations of 3×3 cross-shaped structural elements in sequence to optimize the boundary; The morphologically processed mask is fused with the original image to generate a pseudo-color thickness-enhanced image.

[0007] As a preferred embodiment of the present invention, an improved competitive fuzzy C-means clustering algorithm includes an objective function, a competitive radius and a membership update; Objective function: ; is the clustering objective function, is the total number of pixels, is the number of clusters, For the Pixel pair The class membership, is the fuzzy factor; is the competition weight coefficient; For the The feature vector of pixels, For the The cluster center of the class, is the competition radius.

[0008] As a preferred embodiment of the present invention, generating an edge enhanced image includes: Contourlet multi-scale decomposition is performed on the microscopic image data, and an improved adaptive threshold function is used to perform denoising on the directional sub-band coefficients; Construct a detail enhancement channel, generate a residual image by taking the difference between the original and denoised images, and perform piecewise linear enhancement; The self-guided filtering algorithm is used for edge-preserving optimization, and the Contourlet domain reconstructed image is fused with the filtered output by non-downsampling to generate an edge-enhanced image. Among them, the improved adaptive threshold function combines multimodal coupling mechanism, directional energy modulation function and cross-scale adaptation.

[0009] As a preferred implementation of the present invention, the improved adaptive threshold function is shown in Formula 4: (4); In the formula, For coordinates The directional subband threshold at Sub-band The local directional variance of For the , The absolute value of the subband coefficient of pixels, is the nonlinear modulation factor, For coordinates The directional energy ratio at is the mean and standard deviation of directional energy, is the scale adaptation coefficient, For Window The number of pixels inside is the decomposition series, Is the direction index.

[0010] As a preferred embodiment of the present invention, generating a microcrack enhancement image includes: Perform non-subsampled Contourlet multi-scale decomposition on the microscopic morphology image data to obtain the low-pass sub-band and high-frequency sub-band coefficients in multiple directions; performing adaptive gamma correction enhancement on the low-pass subband; Obtain the local directional variance of each high-frequency sub-band coefficient, set the threshold interval and classify the high-frequency sub-band coefficients; Perform inverse Contourlet transform on the enhanced low-pass subband and the processed high-frequency subband coefficients to reconstruct the image; The reconstructed image is refined three times along the main crack direction using 3×3 linear structural elements to generate a microcrack enhanced image. Among them, the non-subsampled Contourlet multi-scale decomposition includes: multi-scale pyramid decomposition, multi-directional filter bank decomposition and adaptive direction number optimization.

[0011] As a preferred implementation of the present invention, the number of adaptive directions is optimized, as shown in Formula 7: (7); In the formula, is the number of adaptive directions, For the High pass subband The maximum absolute value of all coefficients in , is the energy normalized benchmark.

[0012] As a preferred embodiment of the present invention, when extracting a multi-dimensional precision feature vector, it includes: The time-frequency domain joint features are extracted from the three-channel enhanced images, where the time-domain features include the contrast and energy indexes of the gray-level co-occurrence matrix; The improved CEEMDAN algorithm is used to decompose each channel image into 8-12 IMF components, and the weighted energy moment and multi-scale sample entropy of each component are calculated; Screen effective features based on the correlation coefficient method and retain the correlation coefficient with the target variable The features of the sample are used to construct a multi-domain feature set including time domain, frequency domain, energy moment and sample entropy; Using the UMAP algorithm to perform nonlinear dimensionality reduction on the multi-domain feature set; The reduced feature vector is input into the cascade attention mechanism module, and the final multi-dimensional precision feature vector is generated through spatial-channel dual attention weighting.

[0013] As a preferred embodiment of the present invention, an optimization parameter set is obtained, including: Construct a LS-WGAN-GP parameter generation model, whose generator G is composed of a temporal attention LSTM network, with a multi-dimensional feature vector sequence as input and an initial collaborative parameter set as output. The LSTM hidden layer introduces a gating mechanism to adaptively adjust the memory unit weights. The discriminator D is designed as a spatiotemporal joint convolutional network, which includes a 1D causal convolutional layer and a 2D asymmetric convolutional layer to evaluate the spatiotemporal consistency between the generated parameters and the real process data; Build a reinforcement learning optimizer based on the deep deterministic policy gradient algorithm that combines state space, action space, and reward function; Adopt elite retention strategy and dynamic exploration rate mechanism for parameter optimization; The model parameters are updated through the historical process database, and mixed precision training is used to accelerate convergence. When the continuous decrease rate of the mean square error is less than the threshold, the optimization is terminated and the optimized parameter set is output.

[0014] As a preferred embodiment of the present invention, real-time adjustment of the actuator and updating of the weight coefficient include: Construct an adaptive PID parameter expansion space and dynamically set the initial parameter range based on historical process data; A closed-loop control composite evaluation function combining time error integral and dynamic response index is used to calculate individual fitness in real time; Implement floating point coded genetic optimization algorithms, including: dynamic crossover mutation probability, arithmetic crossover operation, and directed Gaussian mutation; Deploy a closed-loop real-time verification mechanism, which executes periodically: inject the current optimal parameters into the PID controller, collect system output and calculate instantaneous errors, and update PID parameters; The defect recognition model weights are synchronously updated through the knowledge distillation loss function.

[0015] Compared with the prior art, the present invention has the following beneficial effects: Realize real-time multi-dimensional perception of ink micro-morphology: Through high-resolution image acquisition and multi-scale fusion algorithm (non-subsampled Contourlet decomposition, improved competitive fuzzy C-means clustering), the ink thickness, edge diffusion and micro-crack density characteristics are extracted simultaneously to solve the limitation of single-dimensional detection of traditional methods; High precision of multi-dimensional feature fusion: Through the joint input of thickness enhanced images, edge enhanced images and micro-crack enhanced images, combined with UMAP nonlinear dimensionality reduction and cascade attention mechanism, the information retention rate of multi-domain features (time-frequency domain, energy moment, sample entropy) is increased to more than 90%, achieving a full range of ink morphology characterization; Excellent anti-interference performance: The improved competitive fuzzy C-means clustering algorithm effectively eliminates the interference of base texture through dynamic competition radius (error rate reduced to <5%) and morphological optimization (3×3 cross-shaped structural element iterative corrosion / expansion); the improved adaptive threshold function combined with multi-modal coupling mechanism (directional energy modulation function) can still maintain edge sharpening effect under noise PSNR ≥ 35dB; Closed-loop control dynamic optimization: Based on the LS-WGAN-GP parameter generation model and reinforcement learning algorithm (dynamic calibration of reward function), the coordinated optimization of printing pressure, scraper angle, and screen tension is achieved (adjustment time is shortened to <5 minutes), and parameter stability is improved by more than 30%, breaking through the bottleneck of manual experience adjustment efficiency; Breakthrough in micro-crack detection sensitivity: Non-subsampled Contourlet decomposition (adaptive optimization of the number of directions) combined with 3×3 linear structure morphological refinement enables detection sensitivity of sub-micron cracks to reach 95%, significantly reducing the risk of conductive line failure. The process efficiency is significantly improved: through adaptive PID parameter expansion space (dynamic cross-mutation probability) and mixed precision training (FP16 / FP32 accelerated convergence), the system response time is shortened to milliseconds, and the overall energy efficiency of the equipment is improved by 20%.

[0016] The present invention provides a full-process intelligent solution for high-precision conductive ink screen printing, can be widely used in the field of flexible electronic manufacturing, and has significant economic benefits and industrialization prospects.

[0017] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is the overall step diagram of the conductive ink screen printing precision control method; Figure 2 Generate a flowchart for thickness enhancement images; Figure 3 Generate a flowchart for edge-enhanced images. DETAILED DESCRIPTION

[0019] See also Figure 1 The conductive ink screen printing precision control method based on deep learning provided by the present invention comprises the following steps: Step S1: acquiring microscopic morphological image data of the conductive ink on the surface of the screen-printed substrate in real time through a high-resolution image acquisition device; Step S2: using an improved competitive fuzzy C-means clustering algorithm to process the microscopic image data to generate a thickness enhancement image; Step S3: using an improved adaptive threshold function to process the microscopic image data to generate an edge-enhanced image; Step S4: Processing the microscopic morphology image data by non-subsampled Contourlet multi-scale decomposition to generate a microcrack enhancement image; Step S5: jointly input the thickness enhanced image, the edge enhanced image and the microcrack enhanced image into a pre-trained defect recognition model to extract a multi-dimensional precision feature vector including ink thickness distribution, edge diffusion and microcrack density; Step S6: Based on the multi-dimensional precision feature vector, a collaborative adjustment parameter set of printing pressure, scraper angle and screen tension is dynamically generated through a parameter prediction model, and iterative optimization is performed through a reinforcement learning algorithm and a historical process database to generate an optimized parameter set; Step S7: adjusting the screen printing equipment actuator in real time through a closed-loop feedback control system according to the optimized parameter set, and synchronously updating the weight coefficient of the defect recognition model.

[0020] See also Figure 2 In the above step S2, generating a thickness enhanced image includes: The mean shift filter algorithm with a kernel function bandwidth of 3-8 pixels is used to preprocess the microscopic image data, which reduces the noise signal-to-noise ratio to below 30dB while retaining the sharpness of the boundaries between particles. Perform KL transformation on the filtered image, map the RGB color space to the feature space, and then select the first two principal components to construct a two-dimensional feature plane, with a cumulative contribution rate of ≥85%; An improved competitive fuzzy C-means clustering algorithm is used on the two-dimensional feature plane to achieve unsupervised segmentation of the ink area and the substrate by dynamically adjusting the competition radius of the cluster center. Morphological processing is performed on the segmented binary image, using a 3×3 cross-shaped structural element to perform 5-8 iterative corrosion operations to eliminate artifacts, and then performing 3-5 dilation operations to reconstruct the continuous boundary of the ink layer; The morphologically processed mask image is fused with the original image at pixel level to generate a pseudo-color mapping image with enhanced thickness features, in which the hue channel encodes the thickness gradient information and the saturation channel represents the regional confidence.

[0021] Furthermore, the mean shift filtering algorithm is shown in Formula 1: (1); In the formula, is the signal-to-noise ratio, which is used to measure the quality of the filtered image; The original image is at coordinates The pixel value at The coordinates of the filtered image are The pixel value at is the square of the L2 norm, which is used to calculate signal energy and noise energy; is the kernel function bandwidth (3-8 pixels), which controls the smoothness of the filter; is the Gaussian kernel function , used to weight neighborhood pixels; The original image is offset The pixel value at ; Termination condition: adjacent centroid offset pixel; , For the Second and The centroid coordinates of the iteration.

[0022] Furthermore, the KL transformation includes: covariance matrix calculation, eigenvalue decomposition, cumulative contribution rate and projection matrix.

[0023] Covariance matrix calculation: ; For the The RGB color vector of pixels (three-dimensional), is the mean of all pixel color vectors, is the covariance matrix (3×3), is the total number of pixels.

[0024] Eigenvalue decomposition: ; is the eigenvector matrix (3×3), and the column vectors are the orthogonal basis of the covariance matrix; is a diagonal matrix containing the eigenvalues .

[0025] Cumulative contribution rate: ; is the cumulative contribution rate of the first two principal components.

[0026] Projection matrix: ; is the projection matrix (3×2), which maps the RGB space to the two-dimensional feature plane; is the first principal component eigenvector, which dominates the global thickness distribution; is the second principal component eigenvector, capturing local gradient changes; is the set of real numbers.

[0027] Furthermore, the improved competitive fuzzy C-means clustering algorithm includes an objective function, a competitive radius and a membership update.

[0028] Objective function: ; is the clustering objective function, is the total number of pixels, is the number of clusters, For the Pixel pair The class membership, is the fuzzy factor, which controls the fuzziness of the membership distribution; is the competition weight coefficient, balancing the clustering density and competition penalty; For the feature vector of pixels (two-dimensional), For the Cluster centers of classes (2D), is the competition radius; Competition radius: ; For the The competition radius of the class dynamically adjusts the size of the clustering area; For the The standard deviation of pixels within the class to the prototype, is the maximum distance from all pixels in the feature space to the prototype, For the The number of valid pixels for the class.

[0029] Membership update: ; For the Pixel pair The fuzzy membership of the class, For the Cluster centers of classes (2D); For the The competition radius of the class.

[0030] Furthermore, the morphological processing includes: an erosion operation, a dilation operation, and a cross-shaped structural element.

[0031] Corrosion Operation: ; is a binary mask image (0 is background, 1 is ink area), is the structural element, is the corrosion operator; is the number of corrosion iterations; This is the image translation operation, which moves the image along the vector The result after reverse translation.

[0032] Expansion operation: ; is the number of expansion iterations, Image translation operation, moving the image along the vector The result after translation.

[0033] Cross-shaped structural element: , which is used to preserve the right-angle boundary characteristics of the ink; the cross-shaped structural elements are symmetrical in the horizontal / vertical directions to adapt to the uniform diffusion characteristics of screen-printed inks; 3×3 pixels correspond to an actual size of approximately 5μm (assuming an image resolution of 1.67μm / pixel), which matches the typical ink particle size (3-8μm).

[0034] Furthermore, pixel-level fusion includes: HSV channel mapping and RGB conversion matrix.

[0035] HSV channel mapping: ; For hue, is the thickness gradient, is saturation, is the confidence level, is brightness (brightness), is the original brightness, is the gradient amplitude.

[0036] RGB conversion matrix: ; , , For the output channel, is the matrix coefficient, , It is the sine and cosine values ​​of the hue angle, used to calculate the chromaticity component.

[0037] See also Figure 3 In the above step S3, the edge enhanced image is generated, including: Contourlet multi-scale decomposition is performed on the microscopic image data, and an improved adaptive threshold function is used to perform denoising on the directional sub-band coefficients, where the threshold is set to 0.1-0.3 times the amplitude of each sub-band coefficient, which retains the high-frequency edge information while increasing the noise PSNR to more than 35dB; A detail enhancement channel is constructed, and a difference operation is performed between the original image and the denoised image to obtain a residual image. A piecewise linear enhancement function is used to process the residual image, as shown in Formula 2: (2); In the formula, is the enhanced residual image, is the residual image, the first threshold ; Second threshold , is the maximum gray value of the image; The self-guided filtering algorithm is used to optimize the edge of the enhanced residual image and set the filtering radius Pixels and Normalization Factors , which guides the filter output, as shown in Formula 3: (3); In the formula, Pixel is the local window centered on For the input image at position The pixel value of is the covariance, is the variance, is the local mean, is the normalization factor, is the local linear coefficient, is the local bias term, is the guided filter output.

[0038] Edge-preserving denoising mechanism: When the input image And the enhanced residual image In the window When the area is highly correlated, the edge details are retained; when the area is flat noise, the output , achieving noise suppression.

[0039] The Contourlet domain reconstructed image is fused with the guided filter output in a non-downsampled manner to generate an edge-enhanced image, whose edge sharpness is 40-60% higher than that of the original image.

[0040] Furthermore, the improved adaptive threshold function combines the multimodal coupling mechanism, the directional energy modulation function and the cross-scale adaptation, as shown in Formula 4: (4); In the formula, For coordinates The directional subband threshold at Sub-band The local directional variance of For the , The absolute value of the subband coefficient of pixels, is the nonlinear modulation factor, For coordinates The directional energy ratio at is the mean and standard deviation of directional energy, is the scale adaptation coefficient, For Window The number of pixels inside is the decomposition series, Is the direction index.

[0041] Multimodal coupling mechanism: Numerator term: Based on the dynamic scaling threshold of local variance, the threshold is automatically raised in high texture areas to suppress noise; Denominator: Normalized directional energy aggregation term ,pass To achieve nonlinear regulation ( (while weakening the strong edge dominance effect).

[0042] Directional energy modulation function: Hyperbolic tangent term: Soft limit is applied to abnormal energy values ​​to avoid threshold mutation. When the nonlinear modulation factor Approaching 1.0 (protecting significant edges); when When the nonlinear modulation factor Approaching -1.0 (suppressing outlier noise).

[0043] Cross-scale adaptation: scale adaptation coefficient With the decomposition level Power growth ( ), higher levels (coarse scales) use higher thresholds to match the human eye's sensitivity to macroscopic structures.

[0044] In the above step S4, generating a microcrack enhancement image includes: Perform non-subsampled Contourlet multi-scale decomposition on the microscopic image data to obtain the low-pass sub-band and high-frequency sub-band coefficients in multiple directions , where the decomposition level , direction number ; For low pass subband Perform adaptive gamma correction enhancement, the correction function is: ,in And it is negatively correlated with local contrast; Get the coefficients of each high frequency subband The local directional variance , set the threshold interval: ; is the strong edge threshold, is the weak edge threshold, is the noise threshold; Classify the high-frequency subband coefficients based on the threshold interval: when When , it is determined as a strong edge and the original high-frequency sub-band coefficients are retained; when When , it is judged as a weak edge and directional gain is applied ; when When , it is judged as noise and the soft threshold function is used To process, is a sign function used to adjust the high frequency subband coefficients Polarity retention; For the enhanced low-pass subband And the processed high frequency subband coefficients Perform inverse Contourlet transform to reconstruct the image; The reconstructed image is morphologically optimized and three refinement operations are performed along the main crack direction using 3×3 linear structural elements to generate a microcrack enhanced image, with the crack signal-to-noise ratio (SNR) increased by 50-80%.

[0045] Furthermore, non-subsampled Contourlet multi-scale decomposition includes: multi-scale pyramid decomposition, multi-directional filter bank decomposition and adaptive direction number optimization; Multi-scale pyramid decomposition, as shown in Formula 5: (5); In the formula, , For the , Level low-pass sub-band, For the Level high pass subband, is the adaptive low-pass filter kernel, is the adaptive high-pass filter kernel, is the index parameter of the filter kernel, are image space coordinates.

[0046] The multi-directional filter bank decomposition is shown in Equation 6: (6); In the formula, For the Level Directional subband, is a directional filter bank, is the translation compensation parameter, is the direction index, is the direction angle, A collection of directions.

[0047] Adaptive direction number optimization, as shown in Formula 7: (7); In the formula, is the number of adaptive directions, For the High pass subband The maximum absolute value of all coefficients in , is the energy normalized benchmark.

[0048] In the above step S5, when extracting the multi-dimensional precision feature vector, it includes: The time-frequency domain joint features are extracted from the three-channel enhanced images respectively. The time-domain features include the contrast and energy index of the gray-level co-occurrence matrix (GLCM), and the frequency-domain features are extracted by fast curvelet transform to extract the energy proportion of the directional sub-band. The improved CEEMDAN algorithm is used to decompose each channel image into 8-12 IMF components, and the weighted energy moment of each component is calculated. and multi-scale sample entropy , where the exponential decay weight exponential decay; For the The energy intensity of the order IMF, is the total number of IMF components ( ), which is adaptively determined by CEEMDAN; is the IMF level index; ; is the embedding dimension, is the similarity tolerance, for The number of matching templates, for Dimension matching template number; Screen effective features based on the correlation coefficient method and retain the correlation coefficient with the target variable (thickness / diffusion / crack density) The features of the sample are used to construct a multi-domain feature set including time domain, frequency domain, energy moment and sample entropy; The UMAP algorithm is used to perform nonlinear dimensionality reduction on the multi-domain feature set, and its objective function is shown in Formula 8: (8); In the formula, is the objective function value, Index the data point pairs and traverse all unique data point pairs; For point With point The strength of the association in high-dimensional space, For point With point The correlation strength in low-dimensional space has a mapping dimension of 3-5 dimensions; is the logarithm of the similarity ratio, is the logarithm of the dissimilarity ratio; The reduced feature vector is input into the cascade attention mechanism module, and the final multi-dimensional precision feature vector is generated through spatial-channel dual attention weighting. , its feature dimension compression ratio reaches 8:1 and the information retention rate is ≥90%; It is the characteristic value of ink thickness distribution, which is used to characterize the average thickness and uniformity of the ink layer; It is the edge diffusion characteristic value, which is used to measure the diffusion width and clarity of the ink edge; is the characteristic value of microcrack density, which is used to characterize the number and length of cracks per unit area.

[0049] Furthermore, the contrast and energy indexes of the gray level co-occurrence matrix are shown in Formula 9 and Formula 10 respectively: (9); In the formula, is the contrast, is the gray value of two adjacent pixels in the image, For the distance , direction angle The gray value is and The probability of a pixel pair appearing is is the grayscale difference square, is the pixel pitch, is the direction angle, is the gray level; (10); In the formula, is the energy value; Furthermore, the improved CEEMDAN algorithm includes: The amplitude of Gaussian white noise is optimized through an adaptive noise injection strategy, and the noise intensity is dynamically adjusted to adapt to the local characteristics of images in different channels, as shown in Formula 11: (11); In the formula, For the moment After adding noise The secondary disturbance signal, For the moment Next, we take each channel image. The noise amplitude is adaptive and dynamically adjusted according to the local signal-to-noise ratio (SNR); For the moment Next Gaussian white noise added; A multi-stage noise residual iterative decomposition is adopted to ensure that the residual component of each decomposition contains only valid information that has not been extracted, as shown in Formula 12 and Formula 13: (12); (13); In the formula, For the moment Next The integrated average IMF component of order, For the moment Next Sub-noise disturbance signal After EMD decomposition, the The order IMF component, , For the moment Next , The residual component after the decomposition of order, Add times for noise; The number of IMF components (8-12) is controlled by dynamic termination conditions to avoid over-decomposition or under-decomposition, as shown in Formula 14: (14); In the formula, For the Residual Component The Frobenius norm squared, is the total energy of each channel image, is the dynamic termination threshold.

[0050] In the above step S6, an optimization parameter set is generated, including: Construct the LS-WGAN-GP parameter generation model, whose generator G is composed of a temporal attention LSTM network and the input is a multi-dimensional feature vector sequence , is the time window length, and the output is the initial collaborative parameter set , is the initial value of the pressure-related parameters, is the initial value of the angle-related parameters, is the initial value of the tension-related parameters, where the LSTM hidden layer introduces a gating mechanism to adaptively adjust the memory unit weights; The discriminator D is designed as a spatiotemporal joint convolutional network, which includes a 1D causal convolutional layer (core length = 7, dilation factor = 3) and a 2D asymmetric convolutional layer (core size = 5×1), which is used to evaluate the spatiotemporal consistency between the generated parameters and the real process data and output the discriminant confidence. ; Based on the deep deterministic policy gradient (DDPG) algorithm, a reinforcement learning optimizer combining state space, action space and reward function is constructed, where: State space: a sequence of multidimensional feature vectors and device status code ; Action Space: Parameter Adjustment ; Reward function: , For the moment The reward value, is the defect suppression weight, is the confidence weight for determination, is the parameter stability weight, where , dynamic calibration through historical process data; Adopting the adaptive optimization mechanism of elite retention strategy, each round of iteration retains the Top-K ( In each round of iteration, the five parameter sets with the highest fitness are retained) and the optimal parameter set is used as the priority experience playback sample, and the optimal parameter set is used according to the number of iterations. Dynamically adjust the exploration rate: , For the The exploration rate of the round iteration, is the maximum exploration rate, is the minimum exploration rate, is the attenuation coefficient, where ; The model parameters are updated online through the historical process database, and mixed precision training (FP16 / FP32) is used to accelerate convergence. When the mean square error (MSE) of the parameter set on the validation set decreases by 3 consecutive rounds When the optimization is terminated, the optimization parameter set is output .

[0051] In the above step S7, the actuator is adjusted in real time and the weight coefficient is updated, including: Build an adaptive PID parameter expansion space and dynamically set the initial parameter range based on historical process data. As shown in Formula 15: (15); In the formula, is the PID reference parameter, is the device bandwidth factor, is the proportional term expansion increment, is the time decay constant of the integral term, is the differential oscillation frequency, The cumulative time of system operation; The closed-loop control composite evaluation function combining the time error integral and the dynamic response index is used to calculate the individual fitness in real time, as shown in Formula 16: (16); In the formula, is the individual fitness, is the overshoot percentage, To allow the maximum overshoot, To adjust the time, For reference adjustment time, To adjust the time The accumulation of time-weighted absolute errors, is the error function; Implemented floating point encoding genetic optimization algorithm; Deploy a closed-loop real-time verification mechanism that executes periodically at a preset time: inject the current optimal parameters into the PID controller, collect system outputs, and calculate instantaneous errors , For the moment The instantaneous error, For the moment The expected output value of The actual output of the system, update the PID parameters, as shown in formula 17: (17); In the formula, is the updated PID parameter, is the PID parameter before updating, is the learning rate, is the performance indicator function Parameters The partial derivative of Among them, the learning rate , the update threshold is set to the gradient modulus ; The defect recognition model weights are updated synchronously through the knowledge distillation loss function, as shown in Formula 18: (18); In the formula, is the total loss function, For real data By current model parameters No. Layer output, For simulated data Through the historical optimal model weight No. Layer output, for The square of the norm, is the cross entropy loss, is the number of layers; in, The teacher model is updated every 1000 iterations.

[0052] Furthermore, the floating point encoding genetic optimization algorithm includes: dynamic crossover mutation probability, arithmetic crossover operation and directed Gaussian mutation. Dynamic crossover mutation probability: Crossover probability , mutation probability , For the current individual The fitness value of is the maximum fitness value in the current population; Arithmetic Crossover Operation: Offspring Parameters , where the cross-weight parameter Uniform distribution ; and are the original parameter values ​​of the two parent individuals; Directed Gaussian mutation: Limit the variation of the selected parameters: , is the variation range of the parameter, which obeys Gaussian distribution , The mean is 0 and the standard deviation is Gaussian distribution, is the value of the current parameter, which is used to dynamically adjust the variation range.

[0053] The above-mentioned embodiments are only preferred embodiments of the present invention and cannot be used to limit the scope of protection of the present invention. Any non-substantial changes and substitutions made by technicians in this field on the basis of the present invention shall fall within the scope of protection required by the present invention.

Claims

1. A conductive ink screen printing precision control method based on deep learning, characterized in that: The following steps are involved: Real-time acquisition of microscopic image data of conductive ink on the surface of screen-printed substrate; The thickness enhanced images are generated by using the improved competitive fuzzy C-means clustering algorithm; An improved adaptive threshold function is used to generate edge-enhanced images; Generate microcrack enhancement images through non-subsampled Contourlet multi-scale decomposition; The enhanced images are jointly input into the pre-trained defect recognition model to extract multi-dimensional precision feature vectors; Dynamically generate collaborative adjustment parameter sets through parameter prediction models and iteratively optimize to obtain optimized parameter sets; The actuators of the screen printing equipment are adjusted in real time through closed-loop feedback according to the optimized parameter set, and the weight coefficients of the defect recognition model are updated synchronously.

2. The conductive ink screen printing precision control method based on deep learning according to claim 1 is characterized in that: Generate thickness-enhanced images, including: Mean shift filtering is used to preprocess the microscopic image data; Perform KL transform on the filtered image and select the first two principal components to construct a two-dimensional feature plane; An improved competitive fuzzy C-means clustering algorithm is used on the two-dimensional feature plane to dynamically adjust the competition radius of cluster centers to achieve unsupervised segmentation of ink areas. The segmented image is subjected to multiple erosion and dilation operations of 3×3 cross-shaped structural elements in sequence to optimize the boundary; The morphologically processed mask is fused with the original image to generate a pseudo-color thickness-enhanced image.

3. The conductive ink screen printing precision control method based on deep learning according to claim 2 is characterized in that: Improved competitive fuzzy C-means clustering algorithm, including objective function, competitive radius and membership update; Objective function: ; is the clustering objective function, is the total number of pixels, is the number of clusters, For the Pixel pair The class membership, is the fuzzy factor; is the competition weight coefficient; For the The feature vector of pixels, For the The cluster center of the class, is the competition radius.

4. The conductive ink screen printing precision control method based on deep learning according to claim 1 is characterized in that: Generate edge-enhanced images, including: Contourlet multi-scale decomposition is performed on the microscopic image data, and an improved adaptive threshold function is used to perform denoising on the directional sub-band coefficients; Construct a detail enhancement channel, generate a residual image by taking the difference between the original and denoised images, and perform piecewise linear enhancement; The self-guided filtering algorithm is used for edge-preserving optimization, and the Contourlet domain reconstructed image is fused with the filtered output by non-downsampling to generate an edge-enhanced image. Among them, the improved adaptive threshold function combines multimodal coupling mechanism, directional energy modulation function and cross-scale adaptation.

5. The conductive ink screen printing precision control method based on deep learning according to claim 4 is characterized in that: The improved adaptive threshold function is shown in Formula 4: （4）； In the formula, For coordinates The directional subband threshold at Sub-band The local directional variance of For the , The absolute value of the subband coefficient of pixels, is the nonlinear modulation factor, For coordinates The directional energy ratio at is the mean and standard deviation of directional energy, is the scale adaptation coefficient, For Window The number of pixels inside is the decomposition series, Is the direction index.

6. The conductive ink screen printing precision control method based on deep learning according to claim 1, characterized in that: Generate microcrack enhancement images, including: Perform non-subsampled Contourlet multi-scale decomposition on the microscopic morphology image data to obtain the low-pass sub-band and high-frequency sub-band coefficients in multiple directions; performing adaptive gamma correction enhancement on the low-pass subband; Obtain the local directional variance of each high-frequency sub-band coefficient, set the threshold interval and classify the high-frequency sub-band coefficients; Perform inverse Contourlet transform on the enhanced low-pass subband and the processed high-frequency subband coefficients to reconstruct the image; The reconstructed image is refined three times along the main crack direction using 3×3 linear structural elements to generate a microcrack enhanced image. Among them, the non-subsampled Contourlet multi-scale decomposition includes: multi-scale pyramid decomposition, multi-directional filter bank decomposition and adaptive direction number optimization.

7. The conductive ink screen printing precision control method based on deep learning according to claim 6 is characterized in that: Adaptive direction number optimization, as shown in Formula 7: （7）； In the formula, is the number of adaptive directions, For the High pass subband The maximum absolute value of all coefficients in , is the energy normalized benchmark.

8. The conductive ink screen printing precision control method based on deep learning according to any one of claims 2, 4 and 6, characterized in that: When extracting multi-dimensional precision feature vectors, it includes: The time-frequency domain joint features are extracted from the three-channel enhanced images, where the time-domain features include the contrast and energy indexes of the gray-level co-occurrence matrix; The improved CEEMDAN algorithm is used to decompose each channel image into 8-12 IMF components, and the weighted energy moment and multi-scale sample entropy of each component are calculated; Screen effective features based on the correlation coefficient method and retain the correlation coefficient with the target variable The features of the sample are used to construct a multi-domain feature set including time domain, frequency domain, energy moment and sample entropy; Using the UMAP algorithm to perform nonlinear dimensionality reduction on the multi-domain feature set; The reduced feature vector is input into the cascade attention mechanism module, and the final multi-dimensional precision feature vector is generated through spatial-channel dual attention weighting.

9. The conductive ink screen printing precision control method based on deep learning according to claim 8, characterized in that: Get the optimized parameter set, including: Construct a LS-WGAN-GP parameter generation model, whose generator G is composed of a temporal attention LSTM network, with a multi-dimensional feature vector sequence as input and an initial collaborative parameter set as output. The LSTM hidden layer introduces a gating mechanism to adaptively adjust the memory unit weights. The discriminator D is designed as a spatiotemporal joint convolutional network, which includes a 1D causal convolutional layer and a 2D asymmetric convolutional layer to evaluate the spatiotemporal consistency between the generated parameters and the real process data; Build a reinforcement learning optimizer based on the deep deterministic policy gradient algorithm that combines state space, action space, and reward function; Adopt elite retention strategy and dynamic exploration rate mechanism for parameter optimization; The model parameters are updated through the historical process database, and mixed precision training is used to accelerate convergence. When the continuous decrease rate of the mean square error is less than the threshold, the optimization is terminated and the optimized parameter set is output.

10. The conductive ink screen printing precision control method based on deep learning according to claim 9, characterized in that: Real-time adjustment of actuators and update of weight coefficients, including: Construct an adaptive PID parameter expansion space and dynamically set the initial parameter range based on historical process data; A closed-loop control composite evaluation function combining time error integral and dynamic response index is used to calculate individual fitness in real time; Implement floating point coded genetic optimization algorithms, including: dynamic crossover mutation probability, arithmetic crossover operation, and directed Gaussian mutation; Deploy a closed-loop real-time verification mechanism, which executes periodically: inject the current optimal parameters into the PID controller, collect system output and calculate instantaneous errors, and update PID parameters; The defect recognition model weights are synchronously updated through the knowledge distillation loss function.

Images