Automatic Classification System and Method for Battery Electrode Wrinkle Defects

Abstract

This invention discloses an automatic classification system and method for battery electrode wrinkle defects, belonging to the field of battery inspection technology. The system includes: an edge detection module for extracting wrinkle contour features and delineating multiple wrinkle regions; a defect identification module for generating multiple wrinkle defect samples; a sorting and matching module for capturing sorting status parameters; and an automatic classification module for automatically classifying battery electrode wrinkle defects and constructing a sorting report. This application solves the technical problems of low classification efficiency, high false negative rate, and insufficient classification accuracy caused by reliance on manual experience in traditional battery electrode wrinkle defect detection. It achieves automatic and efficient classification of wrinkle defects through edge detection and multi-layer transfer learning, and dynamically optimizes the sorting strategy based on real-time sorting status parameters, thereby improving the efficiency and accuracy of automatic identification and classification of battery electrode wrinkle defects and reducing manual intervention.

Description

Technical Field

[0001] This invention relates to the field of battery testing technology, specifically to an automatic classification system and method for battery electrode wrinkling defects. Background Technology

[0002] In the manufacturing process of lithium-ion batteries, the electrode sheet, as a core component, directly determines the battery's performance, safety, and lifespan based on its surface quality. During production, electrode sheets are prone to various types of wrinkling defects due to uneven tension control or equipment vibration during processes such as coating, rolling, and slitting. These defects lead to uneven coating of the active material, which in turn causes serious problems such as internal short circuits, capacity decay, and thermal runaway. Therefore, accurate detection and automatic classification of wrinkling defects on the surface of battery electrodes are crucial for ensuring battery product quality.

[0003] Traditional electrode defect detection primarily relies on manual visual inspection or machine vision systems based on simple rules. Manual inspection is not only inefficient and labor-intensive, but also highly susceptible to missed and false detections due to visual fatigue and subjective judgment, making it difficult to guarantee consistent results. Early automated inspection systems were often designed for specific defect types, with poor algorithm generalization capabilities, making them ill-suited for handling complex, varied, and morphologically diverse real-world wrinkle defects. Furthermore, the scarcity of defect samples, especially rare ones, hinders the training of traditional machine learning models, limiting improvements in classification accuracy and robustness. Summary of the Invention

[0004] This application provides an automatic classification system and method for battery electrode wrinkle defects, aiming to solve the technical problems of low classification efficiency, high false negative rate and insufficient classification accuracy caused by the reliance on human experience in traditional battery electrode wrinkle defect detection. It achieves automatic and efficient classification of wrinkle defects through edge detection and multi-layer transfer learning, and dynamically optimizes the sorting strategy by combining real-time sorting status parameters, thereby improving the efficiency and accuracy of automatic identification and classification of battery electrode wrinkle defects and reducing the technical effect of manual intervention.

[0005] In view of the above problems, this application provides an automatic classification system and method for battery electrode wrinkling defects.

[0006] The first aspect disclosed in this application provides an automatic classification system for battery electrode wrinkling defects. This system includes: an edge detection module for acquiring images of the battery electrode surface, obtaining an electrode image dataset for edge detection, extracting wrinkle contour features, and delineating multiple wrinkle regions; a defect identification module for traversing the multiple wrinkle regions to perform defect identification and classification, obtaining defect classification results for multi-layer transfer learning, and generating multiple wrinkle defect samples; a sorting and matching module for sorting and matching based on the multiple wrinkle defect samples, formulating a sorting strategy for automatic sorting, and capturing sorting status parameters; and an automatic classification module for performing real-time analysis based on the sorting status parameters, obtaining real-time sorting anomaly parameters to update the sorting strategy, automatically classifying battery electrode wrinkling defects, and constructing a sorting report.

[0007] Another aspect of this application discloses an automatic classification method for battery electrode wrinkling defects. This method includes: acquiring images of the battery electrode surface, obtaining an electrode image dataset for edge detection, extracting wrinkle contour features to delineate multiple wrinkle regions; traversing the multiple wrinkle regions for defect identification and classification, obtaining defect classification results for multi-layer transfer learning, and generating multiple wrinkle defect samples; sorting and matching based on the multiple wrinkle defect samples, formulating a sorting strategy for automatic sorting, and capturing sorting status parameters; performing real-time analysis based on the sorting status parameters, obtaining real-time sorting anomaly parameters to update the sorting strategy for automatic classification of battery electrode wrinkling defects, and constructing a sorting report.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] This method employs an edge detection module to acquire images of the battery electrode surface, obtaining an electrode image dataset for edge detection and extracting wrinkle contour features to delineate multiple wrinkle regions. A defect identification module traverses these wrinkle regions to classify defects, and the classification results are used for multi-layer transfer learning to generate multiple wrinkle defect samples. A sorting and matching module performs sorting matching based on these samples, formulates a sorting strategy for automatic sorting, and captures sorting status parameters. An automatic classification module analyzes these sorting status parameters in real time, obtains real-time sorting anomaly parameters, updates the sorting strategy, and automatically classifies battery electrode wrinkle defects, generating a sorting report. This approach solves the technical problems of low classification efficiency, high false negative rate, and insufficient classification accuracy caused by reliance on manual experience in traditional battery electrode wrinkle defect detection. It achieves automatic and efficient classification of wrinkle defects through edge detection and multi-layer transfer learning, dynamically optimizing the sorting strategy based on real-time sorting status parameters, improving the efficiency and accuracy of automatic identification and classification of battery electrode wrinkle defects, and reducing manual intervention.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0011] Figure 1 This application provides a schematic diagram of an automatic classification system for battery electrode wrinkle defects.

[0012] Figure 2 This application provides a flowchart illustrating an automatic classification method for battery electrode wrinkling defects.

[0013] Figure labeling: Edge detection module 11, Defect recognition module 12, Sorting and matching module 13, Automatic classification module 14. Detailed Implementation

[0014] This application provides an automatic classification system and method for battery electrode wrinkle defects, which solves the technical problems of low classification efficiency, high false negative rate and insufficient classification accuracy caused by the reliance on human experience in traditional battery electrode wrinkle defect detection. It achieves automatic and efficient classification of wrinkle defects through edge detection and multi-layer transfer learning, and dynamically optimizes the sorting strategy by combining real-time sorting status parameters, thereby improving the efficiency and accuracy of automatic identification and classification of battery electrode wrinkle defects and reducing the technical effect of manual intervention.

[0015] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be understood that this application is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0016] Example 1, as Figure 1 As shown in the embodiment of this application, an automatic classification system for battery electrode wrinkling defects is provided. The system includes:

[0017] The edge detection module 11 is used to acquire images of the battery electrode surface, obtain an electrode image dataset for edge detection, extract wrinkle contour features, and delineate multiple wrinkle regions.

[0018] Specifically, in the edge detection module 11, the surface of the battery electrode is continuously photographed by linear array cameras deployed at the start and end points of the transmission path to construct an electrode image dataset. Subsequently, the electrode image dataset is converted to grayscale to reduce the dimensionality of the image data, and noise suppression is applied to reduce the impact of background noise on subsequent processing. Then, an edge detection algorithm is used to perform edge detection and wrinkle contour extraction on the processed electrode image dataset. Based on the extracted contour features, multiple wrinkle regions are delineated according to their geometric shapes. These regions may contain different types of wrinkle defects, such as minor wrinkles, severe wrinkles, or edge defects, which will serve as the basis for subsequent defect identification and classification, providing effective input data for further defect analysis.

[0019] Furthermore, the edge detection module 11 includes:

[0020] The transmission path of the battery electrode is located, and the transmission start and end points are determined based on the transmission path. A first line scan camera is deployed based on the transmission start point, and a second line scan camera is deployed based on the transmission end point. The transmission speed parameters of the battery electrode are obtained by performing transmission calculations on the battery electrode based on the transmission path. The first and second line scan cameras are synchronously triggered to continuously acquire images of the battery electrode based on the transmission speed parameters, thereby obtaining multiple continuous surface images. The multiple continuous surface images are stitched and calibrated to obtain the electrode image dataset.

[0021] In a preferred embodiment, on the battery electrode production line, the transport path of the battery electrodes is determined according to the length and direction of the conveyor belt. The starting and ending points of the transport are then determined based on this path. The starting point is typically where the battery electrodes enter the conveyor belt, where a first line-scan camera is positioned to capture images of the front end of the electrodes. The ending point is typically where the battery electrodes leave the conveyor belt or enter the next processing step, where a second line-scan camera is positioned to capture images of the rear end of the electrodes, ensuring complete monitoring of the entire electrode surface. Subsequently, the transport speed of the battery electrodes is calculated by recording the distance traveled per unit time. This transport speed parameter ensures that the camera is triggered promptly and accurately when the battery electrodes pass through the camera's field of view, avoiding misalignment in image acquisition due to speed changes. Subsequently, based on the calculated battery electrode transmission speed parameters, the time required for the battery electrode to move a preset distance is calculated and used as the acquisition time interval for each image. After each time interval, the first and second line-scan cameras are synchronously triggered to ensure that the two cameras can continuously capture image data of different areas as the battery electrode moves from the starting point to the ending point, thus ensuring continuous image acquisition during the battery electrode's movement and obtaining multiple continuous surface images. Then, feature point detection (using algorithms such as SIFT, SURF, or ORB) is performed on the acquired continuous images to find a stable set of feature points between adjacent images. Feature points are regions with distinct characteristics in the image, such as corners and edges, which facilitates the alignment of subsequent images. Then, using the matched feature point pairs, the homography matrix Homography algorithm is used to calculate the geometric transformation relationship between the images. During this process, the least squares method is used to optimize image registration, ensuring that geometric transformations such as translation, rotation, and scaling between adjacent images are accurately estimated. Furthermore, using the calculated transformation matrix, subsequent images are translated, rotated, or scaled to ensure feature alignment between different images and eliminate geometric errors. Then, a weighted average is used to process overlapping areas, avoiding noticeable transitions at seams. This allows the registered images to be stitched together, forming a seamless, complete image of the battery electrode surface. Finally, the stitched images undergo global calibration to further eliminate geometric distortions caused by camera angle differences or deformations. The calibrated images are then stored chronologically to form an electrode image dataset. This dataset contains all the detailed information of the battery electrode surface and will serve as the basis for subsequent defect detection, analysis, and classification, improving the accuracy and efficiency of detection.

[0022] Furthermore, the edge detection module 11 includes:

[0023] Grayscale processing is performed on the electrode image dataset to generate a grayscale image dataset of the electrode surface. Noise suppression is applied to the grayscale image dataset of the electrode surface to obtain a standard image dataset. Bidirectional calculation is performed on the dataset according to multiple sizes to obtain bidirectional gradient magnitudes. The bidirectional gradient magnitudes are nonlinearly fused to obtain an edge enhancement image set, which is then binarized to generate a binary edge image set for feature analysis to determine the wrinkle contour features. Contour tracking is performed by traversing the binary edge image set according to the wrinkle contour features to determine multiple connected edge contour information. It is then determined whether the multiple connected edge contour information is a closed contour. When the multiple connected edge contour information is a closed contour, the battery electrode is divided and identified to determine the multiple wrinkle regions. When the multiple connected edge contour information is an unclosed contour, the connection endpoints are extracted to perform a closing operation on the multiple connected edge contour information. Based on the closing result, the battery electrode is divided and identified to determine the multiple wrinkle regions.

[0024] In a preferred embodiment, to extract wrinkle features from battery electrode images, the acquired battery electrode surface image dataset is first processed to grayscale. Specifically, the RGB values ​​of each pixel are weighted and converted into grayscale values, resulting in a grayscale image dataset of the electrode surface. This simplifies subsequent image analysis and reduces computational complexity. Since the acquired battery electrode images may contain noise (such as equipment noise, surface dust, etc.), noise suppression is also required for the grayscale image dataset. Common noise suppression methods include Gaussian blur and median filtering. Gaussian blur applies Gaussian filtering to the image through convolution, smoothing the image and removing high-frequency noise. Median filtering uses the median of the pixel's neighborhood instead of the current pixel value, effectively removing salt-and-pepper noise. After noise suppression, a standard image dataset is generated for subsequent edge detection and feature extraction. Subsequently, based on the standard image dataset, the Sobel operator was used to perform convolution operations on the standard image dataset with convolution kernels of different sizes (such as 3×3 and 5×5). The changes in the horizontal and vertical directions of each pixel were calculated to obtain the horizontal and vertical gradients. By adding the horizontal and vertical gradients to a set, a bidirectional gradient magnitude was formed, which was used to represent the regions in the image with the most dramatic changes in brightness, i.e., the edges. Then, a non-linear fusion method was used to calculate the square root of the sum of the squares of the horizontal and vertical gradients in the bidirectional gradient magnitude, and histogram equalization (such as CLAHE) was used to enhance the edge contrast, resulting in an edge-enhanced image set, which made the wrinkled areas on the surface of the battery electrode more prominent. Next, the Otsu algorithm is used to set a threshold, and all pixels with gradient values ​​higher than this threshold are set to white and marked with 1, representing edges. Pixels with gradient values ​​lower than this threshold are set to black and marked with 0, representing background, thus forming a binary edge image set. Morphological analysis methods (such as region area, perimeter, shape factor, etc.) are then used to extract features from the binary edge image set to obtain wrinkle contour features. Then, an edge tracking algorithm (such as the contour lookup function in OpenCV) is used to track connected regions of the edges in the binary edge image set based on the wrinkle contour features, identifying connected edge contours in the image, thus forming multiple connected edge contour information. These connected edge contour information each contain multiple sets of coordinate points. Then, the multiple connected edge contour information are traversed to determine whether they form closed contours, that is, to check whether the Euclidean distance between the first and last points of the contour is less than the overlap distance. If the contour is determined to be closed, the closed region is directly marked as a wrinkle region, thus identifying multiple wrinkle regions. If the contour is determined to be unclosed, the endpoints of the unclosed contour will be extracted and connected using interpolation methods (such as line connection or curve fitting) to form a closed contour.After the closure operation, the battery electrode is re-divided into regions, and the closed contour region is marked as a wrinkled region, thus obtaining multiple wrinkled regions. These wrinkled regions will serve as the basis for subsequent defect analysis and classification, ensuring that wrinkled defects on the battery electrode can be effectively detected and identified.

[0025] The defect identification module 12 is used to traverse the multiple wrinkled regions to identify and classify defects, obtain the defect classification results, perform multi-level transfer learning, and generate multiple wrinkled defect samples.

[0026] Specifically, in the defect identification module 12, after obtaining the wrinkled areas on the battery electrode surface, geometric and texture features are extracted for each wrinkled area, and a high-dimensional region feature vector is constructed based on the extracted characteristics. Subsequently, based on these region feature vectors, a trained classification model (such as a Convolutional Neural Network (CNN), Support Vector Machine (SVM), or ensemble learning algorithm) is used to classify the wrinkled areas, outputting defect type labels, such as minor wrinkles, severe wrinkles, and multiple wrinkles. Furthermore, the classification model assigns a confidence label to the identified defect category to represent the probability that the defect belongs to a certain category. For example, if the probability of a wrinkled area being classified as a severe wrinkle is 0.92 and the probability of being classified as a minor wrinkle is 0.08, then the confidence level of the classification result is 92%. Afterwards, multi-layer transfer learning is performed on the defect classification results according to these confidence levels to obtain more stable and accurate multiple wrinkle defect samples. These samples are stored in a defect sample library for subsequent identification, sorting, and optimization.

[0027] Furthermore, the defect identification module 12 includes:

[0028] Geometric feature analysis is performed on the multiple folded regions to obtain folded geometric features; texture feature analysis is performed on the multiple folded regions to obtain folded texture features; vector calculation is performed on the multiple folded regions based on the folded geometric features and the folded texture features to generate a region feature vector set; defect identification analysis is performed based on the region feature vector set to obtain defect classification results; confidence analysis is performed based on the defect classification results to determine confidence values ​​for multiple defect classes; multi-level transfer learning is performed on the defect classification results according to the confidence values ​​for multiple defect classes to generate the multiple folded defect samples.

[0029] In a preferred embodiment, for the obtained multiple wrinkled regions, the multiple wrinkled regions are traversed. For each traversed wrinkled region, geometric feature analysis is performed to extract wrinkle geometric features including area, perimeter, aspect ratio, shape factor, and boundary curvature. The area and perimeter can be calculated based on the region's contour point set; the aspect ratio can be calculated using the minimum bounding rectangle method; the shape factor quantifies the regularity of the wrinkle shape using indicators such as roundness and rectangularity; and the boundary curvature is calculated through the gradient changes of edge pixels. Furthermore, texture feature analysis is performed to extract wrinkle texture features including gray-level statistics, gray-level co-occurrence matrix, local binary pattern, and frequency domain features. Gray-level statistics include indicators such as the region's gray-level histogram, mean, variance, skewness, and kurtosis; the gray-level co-occurrence matrix includes texture parameters such as energy, contrast, correlation, and homogeneity; the local binary pattern is determined by extracting the LBP histogram of the 8-neighborhood or circular neighborhood; and the frequency domain features are extracted using Fast Fourier Transform (FFT) or Wavelet Transform. Subsequently, the geometric and texture features are normalized (e.g., Min-Max normalization) and combined in a fixed order to form a high-dimensional region feature vector, creating a region feature vector set for subsequent defect identification and analysis. This set is then input into a classification model to identify wrinkled regions. This model can be constructed using a convolutional neural network (CNN). During construction, the region feature vectors of sample images labeled with defect types and confidence levels are input into the CNN. Iterative training is performed through forward propagation, loss calculation, backpropagation, and parameter optimization. The model is validated and adjusted using data not used for training to ensure the accuracy of the constructed classification model meets the preset accuracy. After receiving the region feature vector set, the classification model analyzes each region feature vector to determine the defect type of each region feature vector, forming a defect classification result. A confidence level is assigned to each defect type, resulting in multi-class defect confidence values. Then, the defect classification results are sorted in descending order according to the generated multi-class defect confidence values. Wrinkled regions with confidence values ​​higher than the threshold are selected as reliable samples for multi-level transfer learning to generate multiple final wrinkled defect samples. These samples are stored in the defect sample library for further optimization and generation of subsequent sorting strategies.

[0030] Furthermore, the defect identification module 12 includes:

[0031] The defect classification results are sorted in descending order according to the confidence values ​​of the multiple defect categories to generate a defect category sequence. Multiple confidence thresholds are set based on the defect classification results, and the confidence values ​​of the multiple defect categories are matched and compared according to the multiple confidence thresholds. The confidence values ​​of the multiple defect categories that are higher than the multiple confidence thresholds are extracted and marked as high confidence values. The high confidence values ​​are used as indexes to retrieve the defect classification results and extract high confidence samples. Multi-layer transfer learning is performed based on the high confidence samples to generate the multiple wrinkle defect samples.

[0032] In one feasible implementation, after obtaining the defect classification results and corresponding multi-class defect confidence values ​​for the wrinkled region, the defect classification results are arranged from high to low confidence values ​​to generate a defect category sequence. Subsequently, multiple confidence thresholds are pre-set based on the defect classification, such as 0.95, 0.90, and 0.85, and the multi-class defect confidence values ​​are compared with the corresponding confidence thresholds. Confidence values ​​higher than the thresholds are extracted and marked as high-confidence values. Then, using the high-confidence values ​​as indexes, the sorted defect category sequence is retrieved to quickly locate and extract the corresponding classification results, forming a high-confidence sample set. This high-confidence sample set, due to its high judgment reliability, can serve as high-quality input data for transfer learning. Then, multi-layer transfer learning is performed on the high-confidence samples in the high-confidence sample set through noise perturbation, linear interpolation and other methods to expand the data scale, thereby generating multiple new wrinkle defect samples, and storing these samples in the defect sample library to support subsequent defect identification and sorting optimization.

[0033] Furthermore, the defect identification module 12 includes:

[0034] Based on the high-confidence samples, high-dimensional features are extracted through random selection; the high-dimensional features are perturbed by Gaussian noise through a feature layer to generate a first type of sample; the high-dimensional features are linearly interpolated through a feature transformation layer to generate a second type of sample; incremental transfer learning is performed using the first type of sample and the second type of sample to generate the multiple wrinkle defect samples.

[0035] In one optional implementation, after obtaining high-confidence samples, a certain number of samples are randomly selected from the high-confidence sample set. For each selected sample, multi-layer feature extraction is performed to obtain a high-dimensional feature representation. This high-dimensional feature not only contains the geometric information of the folds but also incorporates their texture details, comprehensively characterizing the feature distribution of the fold defect region. Subsequently, the high-dimensional feature vector is used as the input to the feature layer, and Gaussian noise with zero mean and adjustable variance is introduced into each feature dimension. Through multiple perturbation operations, samples similar to the original features but with a certain degree of randomness can be generated. These samples are defined as the first type of samples, which can simulate the random changes of fold defects caused by factors such as illumination, shooting angle, and surface reflection in the actual detection environment. Simultaneously, a linear interpolation operation is performed on the high-dimensional features through a feature transformation layer. Specifically, feature vectors from two randomly selected high-confidence samples are weighted and combined according to a certain ratio. The interpolation ratio can be a fixed value (e.g., 0.5) or randomly selected within a certain range. The new feature vector generated by the interpolation is mapped back to the sample space, resulting in a second type of sample. This second type serves as a transition between different wrinkle features, thus filling the feature gaps between categories. Then, the first and second types of samples are used together as an expanded training set. Incremental transfer learning is performed in conjunction with the original high-confidence samples. This involves loading the current network parameters onto an existing pre-trained model (e.g., an adversarial network) and inputting the expanded training set into the network in batches, gradually updating the parameters so that the model can simultaneously learn the feature distributions of both the original and expanded samples. Then, the learning rate and regularization parameters are dynamically adjusted over multiple training cycles to prevent the model from overfitting to new samples. After training, the model after incremental transfer learning generates multiple wrinkle defect samples, which are stored in a defect sample library to support subsequent defect identification and sorting optimization.

[0036] The sorting and matching module 13 is used to sort and match the multiple wrinkle defect samples, formulate sorting strategies for automatic sorting, and capture sorting status parameters.

[0037] Specifically, in the sorting and matching module 13, after obtaining multiple wrinkle defect samples, these samples are analyzed to determine the defect category of each sample. A sorting strategy is then formulated based on these defect categories. For example, electrodes with minor wrinkle defects are assigned to the rework area, electrodes with severe wrinkle defects to the waste area, and defect-free electrodes to the qualified product area. Subsequently, the sorting strategy is sent to the execution mechanism. When the electrodes are transported to the sorting position, the robotic arm or conveyor belt automatically completes the sorting action based on the classification results, transporting electrodes of different categories to their corresponding positions, achieving fully automated processing. During the sorting process, sensors monitor the sorting status in real time, capturing sorting status parameters such as whether the sorting action is completed, sorting time and cycle time, whether the sorting result matches the classification result, and the operating status of the execution mechanism. These parameters are used for dynamic optimization and anomaly correction of subsequent sorting strategies, thereby achieving efficient, accurate, and intelligent sorting of battery electrode wrinkle defects.

[0038] Furthermore, the sorting and matching module 13 includes:

[0039] Based on the analysis of the multiple wrinkle defect samples, multiple wrinkle defect category information is obtained; battery electrodes are dynamically matched according to the multiple wrinkle defect sample information to formulate a sorting strategy; the sorting strategy is decomposed and transformed to generate a sorting control instruction sequence; the sorting control instruction sequence is sent to the automatic sorting device to perform automatic sorting operation to track the position of the battery electrodes and generate electrode position stream data; the sorting action of the battery electrodes is recorded according to the electrode position stream data, and the sorting status parameters are captured.

[0040] In a preferred embodiment, for the obtained multiple wrinkle defect samples, the same method described above is used to analyze these wrinkle defect samples using a classification model to determine multiple wrinkle defect category information. Subsequently, based on the analysis results, dynamic matching is performed on the actually detected battery electrodes. In this process, the Euclidean distance between the feature vector of each battery electrode and the feature vectors of multiple wrinkle defect samples is calculated. If the Euclidean distance between the feature vector of a battery electrode and a wrinkle defect sample is less than or equal to a preset similarity distance, it indicates that the battery electrode and the wrinkle defect category information corresponding to the wrinkle defect sample are highly matched. At this time, the wrinkle defect category information is used to mark the battery electrode, and a sorting strategy is formulated for the battery electrode based on the marked defect type. Then, the sorting strategy rules are mapped to specific action parameters, such as sorting position, execution time, and target area, thereby generating a set of ordered sorting control instruction sequences. Each sorting control instruction sequence contains the operation information of the actuator, such as the gripping coordinates of the robotic arm, the injection timing of the pneumatic device, and the diversion point of the conveyor belt. Then, the sorting control command sequence is sent to the automated sorting device in real time. During the battery electrode transfer process, the automated sorting device uses sensor positioning, visual positioning, and other methods to track the position of the battery electrodes, generating electrode position stream data. This electrode position stream data records the displacement, speed, and arrival time of the electrode on the transfer path. Finally, during the sorting process, the sorting actions of the battery electrodes are recorded in conjunction with the electrode position stream data, forming sorting status parameters. These sorting status parameters include the completion time of the sorting action, position deviation, and the number of sorting actions per operation. These parameters are used for subsequent optimization of the sorting strategy and anomaly correction, ensuring high precision and high reliability in sorting battery electrode wrinkles and defects.

[0041] The automatic classification module 14 is used to perform real-time analysis based on the sorting status parameters, obtain real-time sorting anomaly parameters to update the sorting strategy, automatically classify battery electrode wrinkle defects, and construct a sorting report.

[0042] Specifically, in the automatic sorting module 14, the sorting status parameters are analyzed in real time to determine whether they meet the corresponding status parameter threshold range. This allows for the analysis of abnormal situations during the sorting process, such as action delays, offsets, and incomplete actions, thereby generating real-time sorting anomaly parameters. Subsequently, the sorting strategy is updated based on the generated real-time sorting anomaly parameters. The updated sorting strategy is immediately applied to the automatic sorting device, realizing an automatic sorting and sorting closed loop. This dynamically optimizes the sorting and sorting accuracy and generates a sorting report. This sorting report includes data such as the sorting quantity and classification statistics of each batch of battery electrode sheets, the quantity, proportion, and confidence analysis of various crease defects, sorting success rate, anomaly rate, and average sorting time, ensuring the efficiency, accuracy, and traceability of sorting battery electrode sheet crease defects.

[0043] Furthermore, the automatic classification module 14 includes:

[0044] Multiple state parameter threshold ranges are set for battery electrode sheets. It is determined whether the sorting state parameter is within the multiple state parameter threshold ranges. When the sorting state parameter is within the multiple state parameter threshold ranges, the battery electrode sheet is determined to be successfully sorted, and a first classification result is generated. Sorting anomaly detection is performed based on the first classification result, and a first real-time sorting anomaly parameter is generated. When the sorting state parameter is not within the multiple state parameter threshold ranges, the battery electrode sheet is determined to be sorted to be unsuccessful, and a second real-time sorting anomaly parameter is generated.

[0045] In a preferred embodiment, firstly, threshold ranges for multiple battery electrode sorting status parameters are set according to business needs, including but not limited to sorting action completion time thresholds, position deviation thresholds, and single sorting action count thresholds for the actuator. Then, the sorting status parameters are compared with their corresponding thresholds. If all sorting status parameters are within the threshold range, the battery electrode is determined to have been successfully sorted, and a first classification result is generated. For battery electrodes determined to have been successfully sorted, sorting anomaly detection is also performed. That is, it mainly checks whether the sorted category is consistent with the original classification result. If the sorting result perfectly matches the classification category, it is recorded as normal. If an error exists, such as a slightly wrinkled electrode being sorted into a heavily wrinkled area, a first real-time sorting anomaly parameter is generated to mark the abnormal sorting result. If a sorting status parameter is not within the threshold range, the battery electrode is determined to have failed to be sorted. In this case, the abnormal sorting status parameter is recorded to form a second real-time sorting anomaly parameter, used to identify anomalies in the sorting stage, providing refined anomaly monitoring and optimization basis for automatic sorting.

[0046] Furthermore, the automatic classification module 14 includes:

[0047] Based on the first real-time sorting anomaly parameter, sorting statistics are performed to generate a sorting success rate; sorting time is calculated according to the sorting success rate to obtain an average sorting time parameter; sorting priority analysis is performed based on the sorting success rate and the average sorting time parameter, and the sorting strategy is updated based on the analysis results to determine the sorting optimization result; battery electrode wrinkling defects are automatically classified a second time according to the sorting optimization result to construct the sorting report.

[0048] In one feasible implementation, based on the obtained first real-time sorting anomaly parameters, the sorting results of each batch of battery electrode sheets are statistically analyzed. The ratio of the number of successfully sorted electrode sheets to the total number of sorted sheets is calculated to obtain the sorting success rate of that batch. This success rate is used to evaluate the current sorting strategy and its effectiveness. A higher success rate indicates better classification accuracy and execution precision. Then, based on the time record from detection to sorting completion for each battery electrode sheet, the duration of each sorting action is statistically analyzed to obtain the total sorting time. Combining this with the sorting success rate, the total sorting time is divided by the number of successfully sorted electrode sheets to obtain the average sorting time, reflecting the processing efficiency in a real production environment. Subsequently, for defect categories with low sorting success rates or long processing times, the sorting strategy is adjusted preferentially, including optimizing the robotic arm's motion path and adjusting the sorting sequence, thereby generating sorting optimization results to guide the next batch of sorting operations. After strategy optimization, battery electrode wrinkling defects are automatically reclassified. During this process, the updated sorting strategy is used to reclassify and re-sort the remaining unsorted or abnormally sorted electrodes, automatically recording the results, success rate, time taken, and any anomalies to generate a complete sorting report. This process achieves a closed-loop operation from real-time anomaly parameter statistics, sorting efficiency assessment, strategy optimization to secondary classification and report generation, making battery electrode wrinkling defect sorting both efficient and traceable.

[0049] In summary, the automatic classification system for battery electrode wrinkling defects provided in this application has the following technical effects:

[0050] The edge detection module 11 is used to acquire images of the battery electrode surface, obtain an electrode image dataset for edge detection, extract wrinkle contour features, and delineate multiple wrinkle regions. The defect identification module 12 is used to traverse the multiple wrinkle regions for defect identification and classification, obtain defect classification results, perform multi-layer transfer learning, and generate multiple wrinkle defect samples. The sorting and matching module 13 is used to sort and match based on the multiple wrinkle defect samples, formulate sorting strategies for automatic sorting, and capture sorting status parameters. The automatic classification module 14 is used to perform real-time analysis based on the sorting status parameters, obtain real-time sorting anomaly parameters, update the sorting strategy, automatically classify battery electrode wrinkle defects, and construct a sorting report. Through the above steps, the technical problems of low classification efficiency, high false negative rate, and insufficient classification accuracy caused by reliance on human experience in traditional battery electrode wrinkle defect detection are solved. It achieves automatic and efficient classification of wrinkle defects through edge detection and multi-layer transfer learning, and dynamically optimizes the sorting strategy by combining real-time sorting status parameters, thereby improving the efficiency and accuracy of automatic identification and classification of battery electrode wrinkle defects and reducing the need for manual intervention.

[0051] Example 2, based on the same inventive concept as the automatic classification system for battery electrode wrinkling defects in the foregoing examples, such as... Figure 2As shown in the embodiments of this application, an automatic classification method for battery electrode wrinkling defects is provided, the method comprising:

[0052] Images of the battery electrode surface are acquired to obtain an electrode image dataset. Edge detection is performed, and wrinkle contour features are extracted to delineate multiple wrinkle regions. Defect identification and classification are performed by traversing the multiple wrinkle regions. The defect classification results are then subjected to multi-layer transfer learning to generate multiple wrinkle defect samples. Based on the multiple wrinkle defect samples, sorting and matching are performed, and a sorting strategy is formulated for automatic sorting, capturing sorting status parameters. Based on the sorting status parameters, real-time analysis is performed to obtain real-time sorting anomaly parameters, which are then used to update the sorting strategy to automatically classify battery electrode wrinkle defects and construct a sorting report.

[0053] Furthermore, the methods include:

[0054] The transmission path of the battery electrode is located, and the transmission start and end points are determined based on the transmission path. A first line scan camera is deployed based on the transmission start point, and a second line scan camera is deployed based on the transmission end point. The transmission speed parameters of the battery electrode are obtained by performing transmission calculations on the battery electrode based on the transmission path. The first and second line scan cameras are synchronously triggered to continuously acquire images of the battery electrode based on the transmission speed parameters, thereby obtaining multiple continuous surface images. The multiple continuous surface images are stitched and calibrated to obtain the electrode image dataset.

[0055] Furthermore, the methods include:

[0056] Grayscale processing is performed on the electrode image dataset to generate a grayscale image dataset of the electrode surface. Noise suppression is applied to the grayscale image dataset of the electrode surface to obtain a standard image dataset. Bidirectional calculation is performed on the dataset according to multiple sizes to obtain bidirectional gradient magnitudes. The bidirectional gradient magnitudes are nonlinearly fused to obtain an edge enhancement image set, which is then binarized to generate a binary edge image set for feature analysis to determine the wrinkle contour features. Contour tracking is performed by traversing the binary edge image set according to the wrinkle contour features to determine multiple connected edge contour information. It is then determined whether the multiple connected edge contour information is a closed contour. When the multiple connected edge contour information is a closed contour, the battery electrode is divided and identified to determine the multiple wrinkle regions. When the multiple connected edge contour information is an unclosed contour, the connection endpoints are extracted to perform a closing operation on the multiple connected edge contour information. Based on the closing result, the battery electrode is divided and identified to determine the multiple wrinkle regions.

[0057] Furthermore, the methods include:

[0058] Geometric feature analysis is performed on the multiple folded regions to obtain folded geometric features; texture feature analysis is performed on the multiple folded regions to obtain folded texture features; vector calculation is performed on the multiple folded regions based on the folded geometric features and the folded texture features to generate a region feature vector set; defect identification analysis is performed based on the region feature vector set to obtain defect classification results; confidence analysis is performed based on the defect classification results to determine confidence values ​​for multiple defect classes; multi-level transfer learning is performed on the defect classification results according to the confidence values ​​for multiple defect classes to generate the multiple folded defect samples.

[0059] Furthermore, the methods include:

[0060] The defect classification results are sorted in descending order according to the confidence values ​​of the multiple defect categories to generate a defect category sequence. Multiple confidence thresholds are set based on the defect classification results, and the confidence values ​​of the multiple defect categories are matched and compared according to the multiple confidence thresholds. The confidence values ​​of the multiple defect categories that are higher than the multiple confidence thresholds are extracted and marked as high confidence values. The high confidence values ​​are used as indexes to retrieve the defect classification results and extract high confidence samples. Multi-layer transfer learning is performed based on the high confidence samples to generate the multiple wrinkle defect samples.

[0061] Furthermore, the methods include:

[0062] Based on the high-confidence samples, high-dimensional features are extracted through random selection; the high-dimensional features are perturbed by Gaussian noise through a feature layer to generate a first type of sample; the high-dimensional features are linearly interpolated through a feature transformation layer to generate a second type of sample; incremental transfer learning is performed using the first type of sample and the second type of sample to generate the multiple wrinkle defect samples.

[0063] Furthermore, the methods include:

[0064] Based on the analysis of the multiple wrinkle defect samples, multiple wrinkle defect category information is obtained; battery electrodes are dynamically matched according to the multiple wrinkle defect sample information to formulate a sorting strategy; the sorting strategy is decomposed and transformed to generate a sorting control instruction sequence; the sorting control instruction sequence is sent to the automatic sorting device to perform automatic sorting operation to track the position of the battery electrodes and generate electrode position stream data; the sorting action of the battery electrodes is recorded according to the electrode position stream data, and the sorting status parameters are captured.

[0065] Furthermore, the methods include:

[0066] Multiple state parameter threshold ranges are set for battery electrode sheets. It is determined whether the sorting state parameter is within the multiple state parameter threshold ranges. When the sorting state parameter is within the multiple state parameter threshold ranges, the battery electrode sheet is determined to be successfully sorted, and a first classification result is generated. Sorting anomaly detection is performed based on the first classification result, and a first real-time sorting anomaly parameter is generated. When the sorting state parameter is not within the multiple state parameter threshold ranges, the battery electrode sheet is determined to be sorted to be unsuccessful, and a second real-time sorting anomaly parameter is generated.

[0067] Furthermore, the methods include:

[0068] Based on the first real-time sorting anomaly parameter, sorting statistics are performed to generate a sorting success rate; sorting time is calculated according to the sorting success rate to obtain an average sorting time parameter; sorting priority analysis is performed based on the sorting success rate and the average sorting time parameter, and the sorting strategy is updated based on the analysis results to determine the sorting optimization result; battery electrode wrinkling defects are automatically classified a second time according to the sorting optimization result to construct the sorting report.

[0069] In summary, any step of the method described above can be stored as a computer instruction or program in an unrestricted computer memory, and can be called and identified by an unrestricted computer processor to implement any method in the embodiments of this application, without any additional restrictions.

[0070] Furthermore, the "first" or "second" mentioned above may not only represent a sequential relationship, but may also represent a specific concept, and / or refer to the individual or collective selection of multiple elements. Clearly, those skilled in the art can make various modifications and variations to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. An automatic classification system for battery electrode wrinkling defects, characterized in that, The system includes: The edge detection module is used to acquire images of the battery electrode surface, obtain an electrode image dataset for edge detection, extract wrinkle contour features, and delineate multiple wrinkle regions. The defect identification module is used to traverse the multiple folded regions to identify and classify defects, obtain the defect classification results, perform multi-level transfer learning, and generate multiple folded defect samples. The sorting and matching module is used to sort and match the multiple wrinkle defect samples, formulate sorting strategies for automatic sorting, and capture sorting status parameters. An automatic classification module is used to perform real-time analysis based on the sorting status parameters, obtain real-time sorting anomaly parameters to update the sorting strategy, automatically classify battery electrode wrinkle defects, and construct a sorting report. The defect identification module includes: Geometric feature analysis is performed on the multiple folded regions to obtain folded geometric features; texture feature analysis is performed on the multiple folded regions to obtain folded texture features; vector calculation is performed on the multiple folded regions based on the folded geometric features and the folded texture features to generate a region feature vector set; defect identification analysis is performed based on the region feature vector set to obtain defect classification results; confidence analysis is performed based on the defect classification results to determine confidence values ​​for multiple defect classes; multi-level transfer learning is performed on the defect classification results according to the confidence values ​​for multiple defect classes to generate the multiple folded defect samples. The sorting and matching module includes: Based on the analysis of the multiple wrinkle defect samples, multiple wrinkle defect category information is obtained; Based on the information of the multiple wrinkle defect samples, the battery electrode sheets are dynamically matched to formulate a sorting strategy. Based on the sorting strategy, a sorting control instruction sequence is generated by decomposition and transformation. The sorting control command sequence is sent to the automatic sorting device to perform automatic sorting operations, track the position of the battery electrode, and generate electrode position stream data; The sorting action of the battery electrode is recorded based on the electrode position stream data, and the sorting status parameters are captured.

2. The automatic classification system for battery electrode wrinkling defects as described in claim 1, characterized in that, The edge detection module includes: The transmission path of the battery electrode is located, the transmission start point and the transmission end point are determined according to the transmission path, a first line scan camera is arranged based on the transmission start point, and a second line scan camera is arranged based on the transmission end point. Based on the transmission path, the transmission speed parameters of the battery electrode are obtained by performing transmission calculations on the battery electrode. Based on the battery electrode transmission speed parameters, the first line array camera and the second line array camera are synchronously triggered to continuously acquire multiple continuous surface images of the battery electrode. The multiple consecutive surface images are stitched together and calibrated to obtain the electrode image dataset.

3. The automatic classification system for battery electrode wrinkling defects as described in claim 1, characterized in that, The edge detection module includes: Grayscale processing is performed on the electrode image dataset to generate a grayscale image dataset of the electrode surface. Noise suppression is performed on the grayscale image dataset of the electrode surface to obtain a standard image dataset. Bidirectional calculation is then performed on multiple sizes to obtain bidirectional gradient magnitudes. The bidirectional gradient magnitudes are nonlinearly fused to obtain an edge-enhanced image set, which is then binarized to generate a binarized edge image set for feature analysis to determine the wrinkle contour features. Contour tracking is performed by traversing the binary edge image set according to the wrinkle contour features to determine the contour information of multiple connected edges. The process involves iterating through and determining whether the multiple connected edge contours are closed contours. If the multiple connected edge contours are closed contours, the battery electrode is divided and identified to determine the multiple wrinkled regions. When the multiple connected edge contour information is an unclosed contour, the connection endpoints are extracted to perform a closing operation on the multiple connected edge contour information, and the battery electrode is divided and identified according to the closing result to determine the multiple wrinkled regions.

4. The automatic classification system for battery electrode wrinkling defects as described in claim 1, characterized in that, The defect identification module includes: The defect classification results are sorted in descending order according to the confidence values ​​of the multiple defect types to generate a defect category sequence. Based on the defect classification results, multiple confidence thresholds are set, and the confidence values ​​of the multiple types of defects are matched and compared according to the multiple confidence thresholds. The confidence values ​​of the multiple types of defects that are higher than the multiple confidence thresholds are extracted and marked as high confidence values. The high confidence value is used as an index to retrieve the defect classification results and extract high confidence samples; Multi-layer transfer learning is performed based on the high-confidence samples to generate the multiple wrinkle defect samples.

5. The automatic classification system for battery electrode wrinkling defects as described in claim 4, characterized in that, The defect identification module includes: Based on the high-confidence samples, high-dimensional features are extracted through random selection. The high-dimensional features are perturbed by Gaussian noise through the feature layer to generate the first type of sample. The high-dimensional features are linearly interpolated using a feature transformation layer to generate a second type of sample. Incremental transfer learning is performed using the first type of samples and the second type of samples to generate the multiple wrinkle defect samples.

6. The automatic classification system for battery electrode wrinkling defects as described in claim 1, characterized in that, The automatic classification module includes: Set multiple state parameter threshold ranges for battery electrodes, and determine whether the sorting state parameters are within the multiple state parameter threshold ranges; When the sorting status parameters are within the threshold range of the plurality of status parameters, the battery electrode is determined to be successfully classified, and a first classification result is generated. Based on the first classification result, sorting anomaly detection is performed to generate the first real-time sorting anomaly parameter; When the sorting status parameter is not within the threshold range of the multiple status parameters, the battery electrode classification is determined to be failed, and a second real-time sorting abnormal parameter is generated.

7. The automatic classification system for battery electrode wrinkling defects as described in claim 6, characterized in that, The automatic classification module includes: Based on the first real-time sorting anomaly parameters, sorting statistics are performed to generate a sorting success rate. Calculate the sorting time based on the sorting success rate to obtain the average sorting time parameter; Based on the sorting success rate and the average sorting time parameter, a sorting priority analysis is performed, and the sorting strategy is updated based on the analysis results to determine the sorting optimization result. Based on the sorting optimization results, the battery electrode wrinkling defects are automatically classified a second time, and the sorting report is constructed.

8. An automatic classification method for battery electrode wrinkling defects, characterized in that, The method is performed by the automatic classification system for battery electrode wrinkling defects according to any one of claims 1 to 7, and the method includes: Images of the battery electrode surface are acquired to obtain an electrode image dataset for edge detection, and wrinkle contour features are extracted to delineate multiple wrinkle regions. The multiple folded regions are traversed to identify and classify defects. The defect classification results are then subjected to multi-level transfer learning to generate multiple folded defect samples. Based on the multiple wrinkle defect samples, sorting and matching are performed, a sorting strategy is formulated for automatic sorting, and sorting status parameters are captured. Based on the sorting status parameters, real-time analysis is performed to obtain real-time sorting anomaly parameters, and the sorting strategy is updated to automatically classify battery electrode wrinkling defects and construct a sorting report.

Images