Method for rapid detection and repair of surface defects of a chemical conversion foil

Abstract

The application discloses a chemical conversion foil surface defect rapid detection and repair method and relates to the technical field of electronic material manufacturing. The application synchronously collects data through a multimodal sensing array composed of a high-speed linear array camera and a laser confocal sensor, judges the surface quality and positions defects after processing; the trajectory of the defects is predicted through Kalman filtering, a dynamic repair instruction is generated in combination with a time sequence adaptive control model, and the repair operation is calibrated according to high-precision position detection; the effect data set is formed by collecting data on the repair area for the second time, the model parameters are optimized through a feedback mechanism, the problems of detection and repair fragmentation, information transmission delay, inaccurate repair response and lack of continuous optimization in traditional quality control are effectively solved, efficient closed-loop processing from defect detection to repair is realized, the surface quality of the chemical conversion foil is significantly improved, the performance and service life of the final product are guaranteed, and the dynamic demand of a high-speed production line is adapted.

Description

Technical Field

[0001] This invention relates to the field of electronic material manufacturing technology, specifically a method for rapid detection and repair of surface defects in electroforming foil. Background Technology

[0002] As the cornerstone of high-end electronic components, the surface integrity of the electroplated foil directly determines the performance and lifespan of the final product. It is a crucial link in the industrial chain, and its production process has extremely stringent requirements for surface quality. Any tiny flaw may lead to device failure.

[0003] However, existing quality control processes often separate defect detection from repair, relying on offline sampling or multi-step processing. This discontinuous mode not only severely restricts production cycle time, but also causes delays and deviations in the transmission of defect information, making it difficult to respond accurately to repair measures. In particular, it is inadequate when dealing with dynamic changes on high-speed production lines. Summary of the Invention

[0004] The purpose of this invention is to provide a method for rapid detection and repair of surface defects on electroforming foils. By acquiring data through a multimodal sensor array and combining it with intelligent algorithms, the method achieves rapid detection, precise location, dynamic repair, and closed-loop optimization of surface defects on electroforming foils.

[0005] The objective of this invention can be achieved through the following technical solutions: This application provides a method for rapid detection and repair of surface defects on electrolytic foil, including the following steps: A multimodal sensing array consisting of a high-speed linear array camera and a laser confocal sensor is used to simultaneously acquire a continuous high-resolution image sequence and corresponding three-dimensional depth point cloud data of the high-speed moving foil surface to determine the quality of the foil surface. A three-dimensional convolutional neural network model is constructed to fuse and analyze the acquired image sequence and three-dimensional depth point cloud data, segment the micro-defect region from the complex background texture, and obtain the initial defect mask and its three-dimensional coordinates in the current sensing coordinate system. Based on the obtained defect mask, the geometric topology information of the region is extracted from the corresponding 3D depth point cloud data, multi-dimensional parameters are calculated, and a structured defect feature vector is generated. By combining the foil movement speed information obtained in real time from the encoder of the production line, the position information in the defect feature vector is used as the observation value and input into the Kalman filter to estimate the state of the defect's movement trajectory on the foil, and obtain the predicted position and velocity vector at the predetermined repair time point. The defect feature vector and the predicted location vector are input into the time-adaptive control model, and a dynamic repair operation instruction sequence is generated based on the pre-set repair knowledge base and the correlation between defect type, size and depth. A high-precision position detector is deployed at the repair station to continuously monitor the actual position of the foil. If the spatial error between the detected foil position and the predicted position vector of the defect is less than the preset synchronization threshold, a new dynamic repair operation instruction sequence is generated. By using another set of multimodal sensor arrays downstream of the repair station, secondary data acquisition is performed on the area where the repair operation has been performed to obtain the surface image and three-dimensional point cloud data after repair, forming a repair effect dataset.

[0006] Furthermore, judging the quality of the electrolytic foil surface specifically includes: A multimodal sensing array consisting of a high-speed linear array camera and a laser confocal sensor is used to simultaneously acquire high-resolution image sequences and three-dimensional depth point cloud data of the surface of the electroformed foil. The high-resolution image sequences are then denoised using a mean filtering algorithm to obtain a denoised image sequence. If the point cloud data contains outliers, outliers are removed by a statistical outlier removal algorithm to obtain clean point cloud data. Based on the denoised image sequence, the Canny edge detection algorithm is used to extract the edge features of the foil surface to generate an edge feature map. Three-dimensional depth features are extracted from clean point cloud data, surface height distribution is calculated, and a depth feature map is generated. If the spatial coordinates of the edge feature map and the depth feature map are aligned, the two are fused through coordinate transformation to generate a comprehensive feature map containing image texture and three-dimensional depth. If misalignment occurs, spatial registration is performed using an iterative nearest-point algorithm to obtain a comprehensive feature map after registration. Then, the geometric and texture parameters of the foil surface are calculated to generate a surface characteristic description. The surface characteristic description is classified using a preset threshold to determine the quality of the foil surface.

[0007] Furthermore, obtaining the initial defect mask and its initial three-dimensional coordinates in the current sensing coordinate system specifically includes: Image sequences and 3D depth point cloud data are acquired. By preprocessing and aligning temporal and spatial information, standardized input data is obtained. A 3D convolutional neural network model is constructed, and the standardized data is input to perform feature extraction and data fusion to generate a fused feature map. Through segmentation processing, complex background textures and micro-defect regions are separated from the fused feature map to obtain an initial defect mask. Morphological operations are used to optimize the mask boundary to obtain an optimized defect mask. The three-dimensional coordinates of the corresponding regions are extracted from the three-dimensional depth point cloud data to obtain an initial three-dimensional coordinate set. The initial three-dimensional coordinate set is calibrated in the sensing coordinate system. The calibrated three-dimensional coordinate set is obtained through coordinate transformation. The calibrated three-dimensional coordinate set is processed by cluster analysis to merge adjacent defect regions and determine the final three-dimensional coordinates of the defect.

[0008] Furthermore, a structured defect feature vector is generated, specifically including: After obtaining the defect mask, a subset of point cloud data of a specified region is extracted from the 3D depth point cloud data to obtain the point cloud data of the defect region. The volume of the point cloud subset is calculated using a stereoscopic method to obtain the defect volume parameter. The maximum depth of the defect region relative to the reference plane is calculated using the depth coordinates of the point cloud subset to obtain the defect depth parameter. Based on the boundary points of the defect region, the least squares method is used to fit the boundary curve, calculate the curvature value, and obtain the contour curvature parameter. The defect volume parameter, depth parameter and contour curvature parameter are combined into a feature vector. Principal component analysis is used to reduce the dimensionality of the feature vectors, generating structured defect feature vectors. Then, k-means clustering is used to classify the structured defect feature vectors, resulting in a quantitative representation of the defect type.

[0009] Furthermore, its predicted position and velocity vector at the predetermined repair time point are obtained, specifically including: The movement speed data of foil is obtained from the encoder of the production line. Real-time acquisition data is generated through data preprocessing. The data is then smoothed to obtain a smooth speed sequence. Based on the smooth velocity sequence and the position information in the defect feature vector, the state transition matrix of the Kalman filter is constructed, the initial state value is determined, and the position information observation is iteratively updated through the Kalman filter to obtain the motion trajectory estimate of the defect; If the error of the motion trajectory estimation is less than the preset threshold, the predicted velocity vector at the specified repair time point is calculated based on the trajectory estimation. According to the predicted velocity vector and the motion trajectory estimation, the predicted position vector is calculated using a linear interpolation method. By combining the predicted position vector and the predicted velocity vector, a dynamic state description of the defect on the foil is generated.

[0010] Furthermore, a dynamic repair operation instruction sequence is generated, specifically including: Obtain the defect feature vector and the predicted location vector, and use a convolutional neural network to extract the defect type, size, and depth features to obtain the feature dataset; If the defect type in the feature dataset matches the type in the preset repair knowledge base, the corresponding repair tool parameter template is obtained from the knowledge base to obtain a preliminary tool parameter set. Based on the preliminary tool parameter set and the feature dataset, a long short-term memory network is used to generate time series of pulse energy, action duration, and spot focusing parameters to obtain a dynamic parameter sequence. A rule engine is used to constrain and validate the parameters. If a parameter exceeds a preset threshold, it is adjusted to the threshold range to obtain an optimized parameter sequence. This sequence is then input into the time-series adaptive control model to generate a dynamic repair instruction sequence, which drives the repair tool to perform the job and obtain the repair completion status.

[0011] Furthermore, continuous monitoring of the actual position of the foil material includes: The real-time position data of the foil is obtained by a high-precision position detector, the actual position vector of the foil is generated, and the predicted position vector of the defect on the foil is calculated by a preset defect prediction model to generate a trigger signal. The system acquires a pre-stored sequence of dynamic repair operation instructions. Through the instruction parsing module, the sequence of dynamic repair operation instructions is decomposed into a sequence of specific repair actions. The action execution module controls the repair station to perform the corresponding repair operations based on the decomposed repair action sequence, thereby obtaining the position data of the repaired foil. Position deviation information is extracted from the position data of the repaired foil, and the parameters of the defect prediction model are updated.

[0012] Furthermore, a dataset of repair results is generated, specifically including: Secondary data acquisition of the repair area is performed by a multimodal sensor array to obtain surface images and 3D point cloud data, generating an initial repair effect dataset. If the resolution of the surface image in the initial repair effect dataset is lower than a preset threshold, an interpolation algorithm is used to enhance the surface image to obtain an enhanced surface image. The surface texture features of the repair area are identified by a feature extraction algorithm to obtain a texture feature set. If the matching degree between the texture feature set and the preset standard texture features is lower than a preset threshold, the three-dimensional point cloud data is geometrically corrected by a point cloud registration algorithm to obtain corrected point cloud data. Based on the corrected point cloud data, a stereomicroscopic algorithm is used to reconstruct the three-dimensional surface model of the repair area to obtain surface model data. If the geometric deviation between the surface model data and the preset standard model exceeds a preset threshold, the repair parameters are adjusted through an iterative optimization algorithm to obtain optimized repair parameters, which are used to update the repair effect dataset and generate the final repair effect dataset.

[0013] Furthermore, after forming the repair effect dataset, the process also includes: obtaining the defect feature vector before repair and comparing it with the repair effect dataset to quantitatively evaluate the residual error and material modification degree of the repair operation, and using the evaluation result as a feedback signal to adjust the weight parameters inside the time-series adaptive control model online to adaptively optimize the repair strategy.

[0014] Furthermore, the adaptive optimization of the repair strategy specifically includes: The defect feature vector and the post-repair effect dataset are obtained, and the comparison analysis method is used to compare them. The residual error value between the defect feature vector and the post-repair effect dataset is calculated, and the degree of material modification is determined by material property change analysis to obtain quantitative evaluation results. If the residual error value in the quantitative evaluation result exceeds the preset threshold, the error source is identified by the error analysis algorithm, a feedback signal value is generated, the weight parameter set of the time series control model is adjusted according to the feedback signal value, the weight parameters are updated by the gradient descent algorithm, and the updated weight parameter set is obtained. The time-series control model is optimized by updating the weight parameter set, a new repair strategy set is generated, the material data is reprocessed according to the new repair strategy set, and a new post-repair effect dataset is generated. If the residual error value of the new post-repair effect dataset is lower than the preset threshold, the repair strategy set is determined as the final optimization result.

[0015] The beneficial effects of this invention are as follows: By synchronously acquiring images and 3D point cloud data through a multimodal sensor array and combining it with a 3D convolutional neural network, rapid defect detection and localization are achieved. This solves the problems of delay and deviation in defect information transmission caused by offline sampling or step-by-step processing in the background technology. Data is acquired in real time using a high-speed linear array camera and a laser confocal sensor. After noise reduction, feature fusion and intelligent segmentation, micro-defects are accurately identified, making the detection response speed adaptable to the dynamic changes of high-speed production lines. This enables the instant capture and accurate transmission of defect information, providing timely and accurate basic data for subsequent repair, and breaking the discontinuous mode of detection and repair. By predicting defect trajectories using Kalman filtering, generating dynamic repair instructions using a time-adaptive control model, and achieving real-time synchronous calibration through high-precision position detection, this approach solves the problem of traditional repair measures failing to accurately respond to dynamic changes on high-speed production lines. It also utilizes production line speed information to predict defect repair locations, combines defect feature matching with a repair knowledge base to generate appropriate instructions, and dynamically adjusts instructions through position monitoring to ensure real-time synchronization between repair operations and defect locations. This significantly improves repair accuracy, avoids repair failures due to positional deviations, and allows dynamic defects on high-speed production lines to be repaired promptly and accurately. By acquiring repair effect datasets through secondary data collection and combining them with feedback mechanisms to optimize the parameters of the time-series adaptive control model, the problem of lack of closed-loop evaluation and continuous optimization of repair effect in existing quality control is solved. Then, the repaired area is subjected to secondary detection to quantify residual error and material modification degree. The evaluation results are fed back to the control model to adjust the weight parameters, realize the adaptive optimization of the repair strategy, continuously improve the adaptability and effectiveness of the repair process, continuously improve the surface quality of the formed foil, reduce the risk of device failure caused by defects, and ensure the performance and life of the final product. Attached Figure Description

[0016] To better understand and implement this application, the technical solution is described in detail below with reference to the accompanying drawings.

[0017] Figure 1A flowchart illustrating the rapid detection and repair method for surface defects in electrolytic foil provided in this application; Figure 2 A schematic diagram illustrating the process of obtaining the initial defect mask and its initial three-dimensional coordinates in the current sensing coordinate system for the rapid detection and repair method of surface defects on electroformed foil provided in this application. Figure 3 A flowchart illustrating the process of obtaining the predicted position and velocity vector of the surface defect of the electrolytic foil at a predetermined repair time point using the rapid detection and repair method for surface defects provided in this application. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description relating to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of methods and systems consistent with some aspects of this application as detailed in the appended claims.

[0019] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0020] The following detailed description of the specific implementation methods, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0021] Example 1 Please see Figures 1-3 This embodiment provides a method for rapid detection and repair of defects on the surface of electrolytic foil, including the following steps: S1. A multimodal sensing array consisting of a high-speed linear array camera and a laser confocal sensor is used to synchronously acquire a continuous high-resolution image sequence and corresponding three-dimensional depth point cloud data of the high-speed moving foil surface to determine the quality of the foil surface.

[0022] Furthermore, judging the quality of the electrolytic foil surface specifically includes: A multimodal sensing array consisting of a high-speed linear array camera and a laser confocal sensor is used to simultaneously acquire high-resolution image sequences and three-dimensional depth point cloud data of the surface of the electroformed foil. The high-resolution image sequences are then denoised using a mean filtering algorithm to obtain a denoised image sequence. If the point cloud data contains outliers, outliers are removed by a statistical outlier removal algorithm to obtain clean point cloud data. Based on the denoised image sequence, the Canny edge detection algorithm is used to extract the edge features of the foil surface to generate an edge feature map. Three-dimensional depth features are extracted from clean point cloud data, surface height distribution is calculated, and a depth feature map is generated. If the spatial coordinates of the edge feature map and the depth feature map are aligned, the two are fused through coordinate transformation to generate a comprehensive feature map containing image texture and three-dimensional depth. If misalignment occurs, spatial registration is performed using an iterative nearest-point algorithm to obtain a comprehensive feature map after registration. Then, the geometric and texture parameters of the foil surface are calculated to generate a surface characteristic description. The surface characteristic description is classified using a preset threshold to determine the quality of the foil surface.

[0023] Specifically, this method addresses the problems of inaccurate and inefficient defect detection caused by single data acquisition methods, noise interference, and complex background textures in traditional detection methods. It removes image noise through mean filtering, cleans up outliers in point cloud data through statistical outlier removal algorithms, and extracts and registers features using Canny edge detection and iterative nearest point algorithms. Finally, it generates a comprehensive feature map and describes and classifies surface characteristics, which significantly improves the accuracy and reliability of defect detection on electroforming foil surfaces. This provides a more accurate basis for subsequent repair operations and improves production efficiency and product quality.

[0024] S2. Construct a three-dimensional convolutional neural network model, fuse and analyze the acquired image sequence with the three-dimensional depth point cloud data, segment the micro-defect region from the complex background texture, and obtain the initial defect mask and its three-dimensional coordinates in the current sensing coordinate system.

[0025] Furthermore, obtaining the initial defect mask and its initial three-dimensional coordinates in the current sensing coordinate system specifically includes: S21. Acquire image sequences and 3D depth point cloud data, align temporal and spatial information through preprocessing to obtain standardized input data, construct a 3D convolutional neural network model, input standardized data, perform feature extraction and data fusion, and generate a fused feature map; S22. Through segmentation processing, the complex background texture and micro-defect region are separated from the fused feature map to obtain the initial defect mask. If the boundary clarity of the initial defect mask is lower than the preset threshold, morphological operations are used to optimize the mask boundary to obtain the optimized defect mask. The three-dimensional coordinates of the corresponding region are extracted from the three-dimensional depth point cloud data to obtain the initial three-dimensional coordinate set. S23. In the sensing coordinate system, calibrate the initial three-dimensional coordinate set, obtain the calibrated three-dimensional coordinate set through coordinate transformation, process the calibrated three-dimensional coordinate set through cluster analysis, merge adjacent defect regions, and determine the final defect three-dimensional coordinates.

[0026] The construction of a 3D convolutional neural network model requires designing a network architecture adapted to the task requirements. The input layer receives standardized image sequences and 3D depth point cloud data aligned temporally and spatially. The feature extraction layer consists of multiple 3D convolutional blocks, extracting spatiotemporal features from the data from local to global through 3D convolution operations, batch normalization, and the ReLU activation function. The fusion layer uses an attention mechanism or feature stitching to effectively fuse the texture features of the image with the 3D structural features of the point cloud, highlighting the features of the defect region. The segmentation layer achieves upsampling through transposed convolution and supplements details with skip connections. After softmax activation, it outputs a defect probability mask, and then optimizes the mask boundary through morphological operations. Simultaneously, a labeled dataset containing various defects needs to be prepared, and preprocessed with denoising, normalization, and voxelization. After dividing the dataset into training, validation, and test sets, the model is trained using the Adam optimizer with a weighted loss of cross-entropy and mean squared error as the target. Data augmentation, early stopping strategies, and hyperparameter fine-tuning are combined to improve performance, ultimately resulting in a model that can accurately segment microscopic defect regions and output their 3D coordinates.

[0027] Specifically, by fusing and analyzing image sequences and 3D depth point cloud data using a 3D convolutional neural network model, precise segmentation and localization of micro-defects in complex backgrounds were achieved. This effectively solved the inaccuracy problem of traditional methods in defect identification and localization, significantly improved the accuracy and reliability of defect detection, provided accurate defect location and shape information for subsequent repair operations, and improved the efficiency and effectiveness of overall production quality control.

[0028] S3. Based on the obtained defect mask, extract the geometric topology information of the region from the corresponding 3D depth point cloud data, calculate multi-dimensional parameters including the volume, depth, and contour curvature of the defect, and generate a structured defect feature vector.

[0029] Furthermore, a structured defect feature vector is generated, specifically including: After obtaining the defect mask, a subset of point cloud data of a specified region is extracted from the 3D depth point cloud data to obtain the point cloud data of the defect region. The volume of the point cloud subset is calculated using a stereoscopic method to obtain the defect volume parameter. The maximum depth of the defect region relative to the reference plane is calculated using the depth coordinates of the point cloud subset to obtain the defect depth parameter. In this study, a stereochemical method was used to calculate the volume of a subset of point clouds. The stereochemical method is a volume estimation technique based on 3D point cloud data. By calculating the volume contribution of all points in the point cloud subset, the total volume of the defect region is obtained. The specific formula is as follows: ; , , These are three consecutive points in the point cloud subset. It is the total number of point cloud sub-centers.

[0030] Based on the boundary points of the defect region, the least squares method is used to fit the boundary curve, calculate the curvature value, obtain the contour curvature parameter, and combine the defect volume parameter, depth parameter and contour curvature parameter into a feature vector; The least squares method obtains the best-fit curve by minimizing the error between the fitted curve and the actual data points, and then calculates the curvature value of this curve to obtain the contour curvature parameter. The formula for calculating curvature is: , , It is the derivative of the curve. , It is the second derivative of the curve.

[0031] Principal component analysis (PCA) is used to reduce the dimensionality of the feature vectors, generating structured defect feature vectors. K-means clustering is then used to classify these structured feature vectors, resulting in a quantitative representation of the defect type. K-means clustering divides data points into k clusters, ensuring that data points within clusters are as close as possible to each other and data points between clusters are as far apart as possible. Ultimately, each defect feature vector is assigned to a cluster, thus obtaining a quantitative representation of the defect type.

[0032] The process involves using Principal Component Analysis (PCA) to reduce the dimensionality of the eigenvectors. PCA reduces the dimensionality of the eigenvectors by projecting the data onto the principal component directions, while retaining the most important information. Specific steps include: Calculate the covariance matrix of the eigenvectors; Find the eigenvalues ​​and eigenvectors of the covariance matrix; Select the eigenvectors corresponding to the largest eigenvalues ​​as principal components; The original feature vectors are projected onto the principal component directions to obtain the dimensionality-reduced feature vectors.

[0033] Specifically, the method extracts the geometric topological information of the defect region from 3D depth point cloud data and calculates multi-dimensional parameters (such as volume, depth, and contour curvature) to generate a structured defect feature vector. This process solves the problem of incomplete and inaccurate description of defect features in traditional methods, not only improving the accuracy and efficiency of defect feature extraction, but also providing more accurate and reliable data support for subsequent defect analysis, classification, and repair, significantly enhancing the overall performance and intelligence level of defect detection on electrolytic capacitor foil surfaces.

[0034] S4. Combining the foil movement speed information obtained in real time from the encoder of the production line, the position information in the defect feature vector is input into the Kalman filter as the observation value to estimate the state of the defect's movement trajectory on the foil and obtain its predicted position and velocity vector at the predetermined repair time point.

[0035] Furthermore, its predicted position and velocity vector at the predetermined repair time point are obtained, specifically including: S41. Obtain foil movement speed data from the production line encoder, generate real-time acquisition data through data preprocessing, and if the timestamp of the real-time acquisition data is complete and the speed value is within the preset threshold range, then smooth the data to obtain a smooth speed sequence. S42. Based on the smooth velocity sequence and the position information in the defect feature vector, construct the state transition matrix of the Kalman filter, determine the initial state value, and iteratively update the position information observation through the Kalman filter to obtain the motion trajectory estimate of the defect. S43. If the error of the motion trajectory estimation is less than the preset threshold, the predicted velocity vector at the specified repair time point is calculated based on the trajectory estimation. The predicted position vector is calculated using a linear interpolation method based on the predicted velocity vector and the motion trajectory estimation. The dynamic state description of the defect on the foil is generated by combining the predicted position vector and the predicted velocity vector.

[0036] Specifically, by combining the real-time speed information of the foil obtained by the encoder on the production line, the Kalman filter is used to estimate the state of the defect's trajectory on the foil, thereby obtaining the predicted position and velocity vector of the defect at the predetermined repair time point. This process solves the problem of inaccurate repair caused by the inability to accurately predict the dynamic position of the defect in traditional methods. It not only improves the accuracy and reliability of defect position prediction, but also provides accurate time and space information for subsequent dynamic repair operations, significantly improving the accuracy and efficiency of repair, and ensuring high quality and high efficiency of chemically formed foil surface repair.

[0037] S5. Input the defect feature vector and the predicted position vector into the time-adaptive control model. Based on the preset repair knowledge base and the correlation between defect type, size and depth, the model generates a set of dynamic repair operation instruction sequences containing repair tool pulse energy, action duration and spot focusing parameters.

[0038] Furthermore, a dynamic repair operation instruction sequence is generated, containing the repair tool pulse energy, action duration, and spot focusing parameters, specifically including: Obtain the defect feature vector and the predicted location vector, and use a convolutional neural network to extract the defect type, size, and depth features to obtain the feature dataset; If the defect type in the feature dataset matches the type in the preset repair knowledge base, the corresponding repair tool parameter template is obtained from the knowledge base to obtain a preliminary tool parameter set. Based on the preliminary tool parameter set and the feature dataset, a long short-term memory network is used to generate time series of pulse energy, action duration, and spot focusing parameters to obtain a dynamic parameter sequence. A rule engine is used to constrain and validate the parameters. If a parameter exceeds a preset threshold, it is adjusted to the threshold range to obtain an optimized parameter sequence. This sequence is then input into the time-series adaptive control model to generate a dynamic repair instruction sequence, which drives the repair tool to perform the job and obtain the repair completion status.

[0039] The time-adaptive control model constructs a basic framework, taking defect feature vectors (including parameters such as type, size, and depth) and predicted position vectors (including the position and speed at the repair time point) as inputs, and the dynamic sequence of pulse energy, action duration, and spot focusing parameters of the repair tool as outputs. It incorporates a time-dependent relationship processing module (such as a long short-term memory network layer) to capture the correlation between parameters over time. Furthermore, it builds a preset repair knowledge base, collects the mapping relationship between defect types and optimal tool parameters in historical repair cases, forms a parameter template library, and sets parameter constraint thresholds (such as upper limit of pulse energy and lower limit of action duration) based on the performance of production line equipment (such as the energy range and focusing accuracy of the repair tool) as the verification basis for the rule engine. Next, the training dataset is prepared: historical defect features, predicted locations, and corresponding optimal repair parameter sequences are labeled, standardized, and then divided into training and validation sets. The model is trained using gradient descent algorithm with parameter sequence prediction error (e.g., mean squared error) as the loss function. The mapping pattern between defect features and dynamic parameter sequences is learned through a long short-term memory network. At the same time, a rule engine module is embedded to verify the legality of parameters in real time and adjust them within the threshold to ensure that the output parameters meet the equipment safety and repair requirements. Finally, the repair effect (e.g., defect residual rate) of the model-generated parameters is evaluated through the validation set. The network weights and knowledge base templates are iteratively optimized to ultimately form a time-adaptive control model that can dynamically generate adaptive repair instruction sequences based on real-time defect features and locations.

[0040] Specifically, by inputting the defect feature vector and the predicted location vector into the time-series adaptive control model, a dynamic repair operation instruction sequence is generated based on a preset repair knowledge base and defect features. This solves the problems of unsatisfactory repair results and poor adaptability caused by the lack of dynamic adjustment and intelligent decision-making in traditional repair methods. It not only improves the accuracy and adaptability of the repair process but also achieves intelligent and automated repair, significantly enhancing the quality and efficiency of electrolytic capacitor foil surface repair.

[0041] S6. Deploy a high-precision position detector at the repair station to continuously monitor the actual position of the foil. If the spatial error between the detected foil position and the predicted position vector of the defect is less than the preset synchronization threshold, the dynamic repair operation instruction sequence generated in the previous step will be triggered immediately.

[0042] Furthermore, continuous monitoring of the actual position of the foil material includes: The real-time position data of the foil is obtained by a high-precision position detector, and the actual position vector of the foil is generated. A preset defect prediction model is used to calculate the predicted position vector of the defect on the foil. If the spatial error between the actual position vector of the foil and the predicted position vector of the defect is less than the preset synchronization threshold, the triggering condition is determined to be met and a triggering signal is generated. The system acquires a pre-stored sequence of dynamic repair operation instructions. Through the instruction parsing module, the sequence of dynamic repair operation instructions is decomposed into a sequence of specific repair actions. The action execution module controls the repair station to perform the corresponding repair operations based on the decomposed repair action sequence, thereby obtaining the position data of the repaired foil. Position deviation information is extracted from the position data of the repaired foil, and the parameters of the defect prediction model are updated.

[0043] Specifically, it solves the problems of repair delay and inaccuracy caused by inaccurate location information in traditional repair methods. The real-time monitoring and feedback mechanism not only improves the accuracy of repair, but also realizes the automation and intelligence of the repair process, significantly improving production efficiency and product quality.

[0044] S7. Through another set of multimodal sensor arrays downstream of the repair station, secondary data acquisition is performed on the area where the repair operation has been performed to obtain the surface image and three-dimensional point cloud data after repair, forming a repair effect dataset.

[0045] Furthermore, a dataset of repair results is generated, specifically including: Secondary data acquisition of the repair area is performed by a multimodal sensor array to obtain surface images and 3D point cloud data, generating an initial repair effect dataset. If the resolution of the surface image in the initial repair effect dataset is lower than a preset threshold, an interpolation algorithm is used to enhance the surface image to obtain an enhanced surface image. The surface texture features of the repair area are identified by a feature extraction algorithm to obtain a texture feature set. If the matching degree between the texture feature set and the preset standard texture features is lower than a preset threshold, the three-dimensional point cloud data is geometrically corrected by a point cloud registration algorithm to obtain corrected point cloud data. Based on the corrected point cloud data, a stereomicroscopic algorithm is used to reconstruct the three-dimensional surface model of the repair area to obtain surface model data. If the geometric deviation between the surface model data and the preset standard model exceeds a preset threshold, the repair parameters are adjusted through an iterative optimization algorithm to obtain optimized repair parameters, which are used to update the repair effect dataset and generate the final repair effect dataset.

[0046] Specifically, it solves the problem of the lack of accurate evaluation and optimization of the repair effect in traditional methods. It not only achieves accurate evaluation of the repair effect, but also optimizes the repair parameters through a feedback mechanism to ensure that the repair quality meets the standards, and significantly improves the accuracy and stability of the repair.

[0047] Furthermore, after forming the repair effect dataset, the process also includes: obtaining the defect feature vector before repair and comparing it with the repair effect dataset to quantitatively evaluate the residual error and material modification degree of the repair operation, and using the evaluation result as a feedback signal to adjust the weight parameters inside the time-series adaptive control model online to adaptively optimize the repair strategy.

[0048] Furthermore, the adaptive optimization of the repair strategy specifically includes: The defect feature vector and the post-repair effect dataset are obtained, and the comparison analysis method is used to compare them. The residual error value between the defect feature vector and the post-repair effect dataset is calculated, and the degree of material modification is determined by material property change analysis to obtain quantitative evaluation results. If the residual error value in the quantitative evaluation result exceeds the preset threshold, the error source is identified by the error analysis algorithm, a feedback signal value is generated, the weight parameter set of the time series control model is adjusted according to the feedback signal value, the weight parameters are updated by the gradient descent algorithm, and the updated weight parameter set is obtained. The time-series control model is optimized by updating the weight parameter set, a new repair strategy set is generated, the material data is reprocessed according to the new repair strategy set, and a new post-repair effect dataset is generated. If the residual error value of the new post-repair effect dataset is lower than the preset threshold, the repair strategy set is determined as the final optimization result.

[0049] Specifically, it solves the problem of the lack of real-time feedback and dynamic optimization mechanisms in traditional repair methods. This closed-loop feedback and dynamic optimization mechanism significantly improves the accuracy and adaptability of repair, ensures that the repair effect meets quality standards, and improves the overall repair efficiency and product quality.

[0050] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for rapid detection and repair of surface defects on electrolytic foil, characterized in that: The steps include the following: A multimodal sensing array consisting of a high-speed linear array camera and a laser confocal sensor is used to simultaneously acquire a continuous high-resolution image sequence and corresponding three-dimensional depth point cloud data of the high-speed moving foil surface to determine the quality of the foil surface. A three-dimensional convolutional neural network model is constructed to fuse and analyze the acquired image sequence and three-dimensional depth point cloud data, segment the micro-defect region from the complex background texture, and obtain the initial defect mask and its three-dimensional coordinates in the current sensing coordinate system. Based on the obtained defect mask, the geometric topology information of the region is extracted from the corresponding 3D depth point cloud data, multi-dimensional parameters are calculated, and a structured defect feature vector is generated. By combining the foil movement speed information obtained in real time from the encoder of the production line, the position information in the defect feature vector is used as the observation value and input into the Kalman filter to estimate the state of the defect's movement trajectory on the foil, and obtain the predicted position and velocity vector at the predetermined repair time point. The defect feature vector and the predicted location vector are input into the time-adaptive control model, and a dynamic repair operation instruction sequence is generated based on the pre-set repair knowledge base and the correlation between defect type, size and depth. A high-precision position detector is deployed at the repair station to continuously monitor the actual position of the foil. If the spatial error between the detected foil position and the predicted position vector of the defect is less than the preset synchronization threshold, a new dynamic repair operation instruction sequence is generated. By using another set of multimodal sensor arrays downstream of the repair station, secondary data acquisition is performed on the area where the repair operation has been performed to obtain the surface image and three-dimensional point cloud data after repair, forming a repair effect dataset.

2. The method for rapid detection and repair of surface defects on electroformed foil according to claim 1, characterized in that: Judging the quality of the electrolytic foil surface specifically includes: A multimodal sensing array consisting of a high-speed linear array camera and a laser confocal sensor is used to simultaneously acquire high-resolution image sequences and three-dimensional depth point cloud data of the surface of the electroformed foil. The high-resolution image sequences are then denoised using a mean filtering algorithm to obtain a denoised image sequence. If the point cloud data contains outliers, outliers are removed by a statistical outlier removal algorithm to obtain clean point cloud data. Based on the denoised image sequence, the Canny edge detection algorithm is used to extract the edge features of the foil surface to generate an edge feature map. Three-dimensional depth features are extracted from clean point cloud data, surface height distribution is calculated, and a depth feature map is generated. If the spatial coordinates of the edge feature map and the depth feature map are aligned, the two are fused through coordinate transformation to generate a comprehensive feature map containing image texture and three-dimensional depth. If misalignment occurs, spatial registration is performed using an iterative nearest-point algorithm to obtain a registered comprehensive feature map. Then, the geometric and texture parameters of the foil surface are calculated to generate a surface characteristic description. The surface characteristic description is classified using a preset threshold to determine the quality of the foil surface.

3. The method for rapid detection and repair of surface defects on electroformed foil according to claim 1, characterized in that: Obtaining the initial defect mask and its initial three-dimensional coordinates in the current sensing coordinate system specifically includes: Image sequences and 3D depth point cloud data are acquired. By preprocessing and aligning temporal and spatial information, standardized input data is obtained. A 3D convolutional neural network model is constructed, and the standardized data is input to perform feature extraction and data fusion to generate a fused feature map. Through segmentation processing, complex background textures and micro-defect regions are separated from the fused feature map to obtain an initial defect mask. Morphological operations are used to optimize the mask boundary to obtain an optimized defect mask. The three-dimensional coordinates of the corresponding regions are extracted from the three-dimensional depth point cloud data to obtain an initial three-dimensional coordinate set. The initial three-dimensional coordinate set is calibrated in the sensing coordinate system. The calibrated three-dimensional coordinate set is obtained through coordinate transformation. The calibrated three-dimensional coordinate set is processed by cluster analysis to merge adjacent defect regions and determine the final three-dimensional coordinates of the defect.

4. The method for rapid detection and repair of surface defects on electroformed foil according to claim 1, characterized in that: Generate structured defect feature vectors, specifically including: After obtaining the defect mask, a subset of point cloud data of a specified region is extracted from the 3D depth point cloud data to obtain the point cloud data of the defect region. The volume of the point cloud subset is calculated using a stereoscopic method to obtain the defect volume parameter. The maximum depth of the defect region relative to the reference plane is calculated using the depth coordinates of the point cloud subset to obtain the defect depth parameter. Based on the boundary points of the defect region, the least squares method is used to fit the boundary curve, calculate the curvature value, and obtain the contour curvature parameter. The defect volume parameter, depth parameter and contour curvature parameter are combined into a feature vector. Principal component analysis is used to reduce the dimensionality of the feature vectors, generating structured defect feature vectors. Then, k-means clustering is used to classify the structured defect feature vectors, resulting in a quantitative representation of the defect type.

5. The method for rapid detection and repair of surface defects on electroformed foil according to claim 1, characterized in that: The predicted position and velocity vector at the predetermined repair time point are obtained, specifically including: The movement speed data of foil is obtained from the encoder of the production line. Real-time acquisition data is generated through data preprocessing. The data is then smoothed to obtain a smooth speed sequence. Based on the smooth velocity sequence and the position information in the defect feature vector, the state transition matrix of the Kalman filter is constructed, the initial state value is determined, and the position information observation is iteratively updated through the Kalman filter to obtain the motion trajectory estimate of the defect; If the error of the motion trajectory estimation is less than the preset threshold, the predicted velocity vector at the specified repair time point is calculated based on the trajectory estimation. According to the predicted velocity vector and the motion trajectory estimation, the predicted position vector is calculated using a linear interpolation method. By combining the predicted position vector and the predicted velocity vector, a dynamic state description of the defect on the foil is generated.

6. The method for rapid detection and repair of surface defects on electroformed foil according to claim 1, characterized in that: Generate a dynamic repair job instruction sequence, specifically including: Obtain the defect feature vector and the predicted location vector, and use a convolutional neural network to extract the defect type, size, and depth features to obtain the feature dataset; If the defect type in the feature dataset matches the type in the preset repair knowledge base, the corresponding repair tool parameter template is obtained from the knowledge base to obtain a preliminary tool parameter set. Based on the preliminary tool parameter set and the feature dataset, a long short-term memory network is used to generate time series of pulse energy, action duration, and spot focusing parameters to obtain a dynamic parameter sequence. A rule engine is used to constrain and validate the parameters. If a parameter exceeds a preset threshold, it is adjusted to the threshold range to obtain an optimized parameter sequence. This sequence is then input into the time-series adaptive control model to generate a dynamic repair instruction sequence, which drives the repair tool to perform the job and obtain the repair completion status.

7. The method for rapid detection and repair of surface defects on electroformed foil according to claim 1, characterized in that: Continuous monitoring of the actual position of the foil, specifically including: The real-time position data of the foil is obtained by a high-precision position detector, the actual position vector of the foil is generated, and the predicted position vector of the defect on the foil is calculated by a preset defect prediction model to generate a trigger signal. The system acquires a pre-stored sequence of dynamic repair operation instructions. Through the instruction parsing module, the sequence of dynamic repair operation instructions is decomposed into a sequence of specific repair actions. The action execution module controls the repair station to perform the corresponding repair operations based on the decomposed repair action sequence, thereby obtaining the position data of the repaired foil. Position deviation information is extracted from the position data of the repaired foil, and the parameters of the defect prediction model are updated.

8. The method for rapid detection and repair of surface defects on electroformed foil according to claim 1, characterized in that: The dataset of repair results is formed, specifically including: Secondary data acquisition of the repair area is performed by a multimodal sensor array to obtain surface images and 3D point cloud data, generating an initial repair effect dataset. If the resolution of the surface image in the initial repair effect dataset is lower than a preset threshold, an interpolation algorithm is used to enhance the surface image to obtain an enhanced surface image. The surface texture features of the repair area are identified by a feature extraction algorithm to obtain a texture feature set. If the matching degree between the texture feature set and the preset standard texture features is lower than a preset threshold, the three-dimensional point cloud data is geometrically corrected by a point cloud registration algorithm to obtain corrected point cloud data. Based on the corrected point cloud data, a stereomicroscopic algorithm is used to reconstruct the three-dimensional surface model of the repair area to obtain surface model data. If the geometric deviation between the surface model data and the preset standard model exceeds a preset threshold, the repair parameters are adjusted through an iterative optimization algorithm to obtain optimized repair parameters, which are used to update the repair effect dataset and generate the final repair effect dataset.

9. The method for rapid detection and repair of surface defects on electroformed foil according to claim 8, characterized in that: After forming the repair effect dataset, the process also includes: obtaining the defect feature vector before repair and comparing it with the repair effect dataset to quantitatively evaluate the residual error and material modification degree of the repair operation, and using the evaluation result as a feedback signal to adjust the weight parameters inside the time-series adaptive control model online to adaptively optimize the repair strategy.

10. The method for rapid detection and repair of surface defects on electroformed foil according to claim 9, characterized in that: The adaptive optimization of the repair strategy specifically includes: The defect feature vector and the post-repair effect dataset are obtained, and the comparison analysis method is used to compare them. The residual error value between the defect feature vector and the post-repair effect dataset is calculated, and the degree of material modification is determined by material property change analysis to obtain quantitative evaluation results. If the residual error value in the quantitative evaluation result exceeds the preset threshold, the error source is identified by the error analysis algorithm, a feedback signal value is generated, the weight parameter set of the time series control model is adjusted according to the feedback signal value, the weight parameters are updated by the gradient descent algorithm, and the updated weight parameter set is obtained. The time-series control model is optimized by updating the weight parameter set, a new repair strategy set is generated, the material data is reprocessed according to the new repair strategy set, and a new post-repair effect dataset is generated. If the residual error value of the new post-repair effect dataset is lower than the preset threshold, the repair strategy set is determined as the final optimization result.

Images