A method and system for detecting defects of an electrode sheet for battery manufacturing by fusing multiple features

Abstract

This invention proposes a multi-feature fusion detection method and system for electrode defects in battery manufacturing, belonging to the field of battery manufacturing process optimization technology. It includes: acquiring and preprocessing surface images of lithium battery electrode coatings; constructing a defect response field based on the preprocessed image, and performing edge enhancement using a defect confidence map based on adaptive thresholds, combined with dual threshold processing to obtain a defect region image; inputting the defect region image into a defect perception network to obtain the defect category and classification confidence; wherein, during the training process of the network, an adaptive gradient optimization algorithm is used, with the optimization objective of maximizing classification accuracy and minimizing false negatives, to adaptively optimize the defect perception network and obtain the optimal combination of classification parameters including the optimal confidence threshold. This improves the defect recognition accuracy and the generalization performance of the defect recognition model.

Description

Technical Field

[0001] This invention relates to the field of battery manufacturing process optimization technology, and in particular to a method and system for multi-feature fusion detection of defects in electrode sheets used in battery manufacturing. Background Technology

[0002] The surface of existing lithium battery electrodes is subject to complex conditions such as coating texture and noise interference, which places stringent requirements on the accurate detection of weak defects such as micro-holes and minor scratches. The current mainstream detection methods are mainly divided into two categories: traditional machine vision and deep learning.

[0003] Traditional machine vision methods can only identify common defects under ideal conditions and are highly susceptible to interference from electrode texture and noise, leading to texture misjudgment and missed detection of weak defects. Although deep learning methods can automatically learn features, they still have shortcomings in practical applications: convolutional layers have difficulty distinguishing between texture and defects, edge detection is prone to producing false edges, and there is a contradiction between noise amplification and missed detection; activation functions are prone to gradient anomalies and cannot adapt to the extraction of weak defects; network parameters rely on manual adjustment, which does not meet the safety priority requirements of battery production and is prone to serious defect missed detection. Overall, the anti-interference ability and detection accuracy are difficult to meet the requirements of the production line. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a multi-feature fusion detection method and system for electrode defects in battery manufacturing, which improves defect prediction accuracy and the generalization performance of the identification network, thereby enhancing battery manufacturing efficiency and product safety and reliability.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for multi-feature fusion detection of defects in electrode sheets used in battery manufacturing, comprising: Collect and preprocess surface images of lithium battery electrode sheets after coating. A defect response field is constructed based on the preprocessed image, and edge enhancement is performed using a defect confidence map based on an adaptive threshold. The defect region image is obtained by combining dual threshold processing. Input the defect region image into the defect perception network to obtain the defect category and classification confidence. During the training process of the network, an adaptive gradient optimization algorithm is used to optimize the defect perception network with the highest classification accuracy and the lowest false negative rate as the optimization objective, so as to obtain the optimal combination of classification parameters including the optimal confidence threshold.

[0006] Secondly, the present invention provides a multi-feature fusion detection system for electrode defects in battery manufacturing, comprising: The preprocessing module is configured to acquire and preprocess surface images of lithium battery electrode sheets after coating. The feature extraction module is configured to construct a defect response field based on the preprocessed image, perform edge enhancement using a defect confidence map based on an adaptive threshold, and obtain a defect region image by combining dual threshold processing. The defect identification module is configured to input a defect region image into a defect perception network to obtain the defect category and classification confidence level. During the training process of the network, an adaptive gradient optimization algorithm is used to adaptively optimize the defect perception network with the goal of maximizing classification accuracy and minimizing false negative rate, thereby obtaining the optimal combination of classification parameters including the optimal confidence threshold.

[0007] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the multi-feature fusion detection method for electrode defects in battery manufacturing described in the first aspect.

[0008] Fourthly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the multi-feature fusion detection method for electrode defects in battery manufacturing described in the first aspect.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention achieves accurate edge extraction and region segmentation for weak contrast and minute defects through the Defect Response Field Adaptive Edge Segmentation (DRF-AES) algorithm. It combines end-to-end feature extraction using a lightweight defect-aware convolutional neural network (DFP-CNN), multi-dimensional defect feature fusion, and parameter adaptive optimization using the Defect Classification Response Gradient Optimization (DCR-GO) algorithm. First, it autonomously extracts the geometric, grayscale, and texture depth features of defects and completes accurate classification. Then, it establishes a comprehensive scoring model based on defect type weights, area proportions, and quantity statistics. With the collaborative guarantee of secondary confidence verification and manual re-inspection mechanisms, it obtains the optimal defect detection and conformity judgment results, achieving a balance between maximizing the defect detection rate and minimizing the risk of missed detection in the lithium battery electrode manufacturing process, thereby improving the efficiency of battery manufacturing and the safety and reliability of products.

[0010] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0011] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0012] Figure 1The main flowchart of a multi-feature fusion detection method for electrode defects in battery manufacturing provided by an embodiment of the present invention is shown. Detailed Implementation

[0013] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0014] Lithium-ion batteries are the core power source for electric vehicles, energy storage systems, and portable electronic devices. The quality of their electrodes directly determines the battery's capacity consistency, cycle stability, and safety reliability. During the manufacturing processes of lithium-ion battery electrodes (positive and negative electrodes), such as coating, rolling, and slitting, five typical surface defects are easily generated: scratches, bubbles, metal leakage, particles, and decarburization. If these defects are not detected and assessed in a timely and accurate manner, they can lead to serious safety accidents such as internal micro-short circuits, lithium plating, capacity decay, and even thermal runaway. Therefore, high-precision and high-efficiency defect detection and conformity assessment of the electrode surface after coating is a crucial step in ensuring the quality and safety of battery manufacturing.

[0015] Existing methods for detecting defects in lithium battery electrodes mainly fall into the following two categories: The first category is based on traditional machine vision inspection methods, which have a certain ability to detect common defects under ideal lighting conditions. However, due to the influence of complex conditions such as electrode surface texture and uneven lighting, there are problems such as texture misjudgment and missed detection of weak defects, making it difficult to meet the requirements of industrial field inspection.

[0016] The second category is defect detection methods based on deep learning. These methods automatically learn defect features using models such as convolutional neural networks (CNNs), avoiding the incompleteness of manually designed features. They achieve high accuracy in academic research and small-batch testing. However, significant shortcomings remain in actual lithium battery electrode inspection scenarios: First, the regular coating texture of the electrode can easily mask weak defects such as micro-holes and minor scratches. Traditional convolutional layers cannot effectively distinguish between textures and defects, and edge detection is prone to noise interference, resulting in false edges. The contradiction between missing weak defects and noise amplification is prominent. Second, activation functions are prone to gradient vanishing or exploding problems, making it difficult to adapt to the extraction of weak defect features. Third, network parameters often rely on manual tuning, resulting in poor adaptability. Furthermore, parameter optimization does not prioritize the production safety requirements of low-missibility detection, easily leading to serious defect omissions and insufficient anti-interference and classification accuracy.

[0017] To address the shortcomings of existing methods, this invention proposes a multi-feature fusion detection method, system, medium, and equipment for electrode defects in battery manufacturing. It constructs a Defect Classification Response Gradient Optimization (DCR-GO) system for lithium battery electrode defect detection scenarios. This system utilizes a five-layer collaborative operation mechanism: Defect Response Field Adaptive Edge Segmentation (DRF-AES), defect clean region screening and normalization, DFP-CNN lightweight defect perception feature classification, DCR-GO global parameter adaptive optimization, and a multi-dimensional comprehensive quality scoring model for defects. This deeply integrates and unifies weak defect signal enhancement, accurate extraction of real defect edges, deep defect feature autonomous learning, intelligent network parameter optimization, and electrode finished product qualification judgment. It possesses the ability to accurately capture subtle defects and complex, irregularly shaped defects, while efficiently filtering out false defect interference from coating textures, stain reflections, and low-confidence suspected defect samples. This comprehensively improves the recognition accuracy, operational efficiency, and practicality of the entire detection solution for on-site engineering applications. This architecture ensures the accuracy of defect detection while strictly controlling the overall computational load to achieve lightweight deployment, and simultaneously achieves multiple advantages such as high defect detection rate, low false negative rate, high consistency of judgment results and high stability of algorithm operation.

[0018] Example 1 like Figure 1 As shown, this embodiment discloses a multi-feature fusion detection method for electrode defects in battery manufacturing, including the following steps: S1: Acquire surface images of lithium battery electrode sheets after coating and perform preprocessing; S2: Construct a defect response field based on the preprocessed image, and use the defect confidence map based on adaptive threshold for edge enhancement. Combine this with double threshold processing to obtain the defect region image. S3: Input the defect region image into the defect perception network to obtain the defect category and classification confidence. During the training process of the network, an adaptive gradient optimization algorithm is used to optimize the defect perception network with the highest classification accuracy and the lowest false negative rate as the optimization objective, so as to obtain the optimal combination of classification parameters including the optimal confidence threshold.

[0019] Next, combined Figure 1 This embodiment provides a detailed description of a multi-feature fusion detection method for electrode defects in battery manufacturing.

[0020] (a) Image preprocessing By using an industrial linear CCD camera with a linear light source, continuous image acquisition is performed on the surface of lithium battery electrode sheets after coating. The acquired raw images are preprocessed by grayscale conversion, contrast enhancement, and noise filtering to obtain clear and usable preprocessed images.

[0021] The lithium battery electrode to be tested is obtained, and after a coating process, a coated surface is obtained. Then, continuous image acquisition is performed on the surface to obtain the original image.

[0022] As one implementation method, this embodiment uses an industrial linear CCD camera with a linear LED light source to acquire electrode images via transmission. Transmitted light effectively highlights penetrating defects such as pores and exposed metal, while reflected light is suitable for surface defects such as scratches and particles. The transmission or reflection illumination method can be selected according to actual inspection needs to acquire the original RGB image, denoted as... .

[0023] Next, the acquired three-channel raw RGB image is converted to grayscale using the standard weighted grayscale formula: ; Where R, G, and B are the pixel values ​​for the red, green, and blue channels, respectively. , , These correspond to the weighting coefficients. The human eye is most sensitive to green and least sensitive to blue, therefore green is assigned the highest weighting coefficient. For example, , , .

[0024] Considering that surface defects on electrodes often have low contrast with normal areas, direct edge detection can easily lead to missed detections. Therefore, this embodiment employs Limit Contrast Adaptive Histogram Equalization (CLAHE) to enhance the grayscale image. The advantage of this method is that it avoids the excessive noise amplification problem caused by traditional histogram equalization, specifically limiting the contrast amplification factor to 3.0. Subsequently, a 3×3 window median filter is used to remove salt-and-pepper noise; median filtering preserves edge information better than mean filtering. After the above processing, the preprocessed image is obtained. This serves as the input for subsequent edge detection.

[0025] (II) Feature Extraction A defect response field adaptive edge segmentation algorithm (DRF-AES) is constructed. Through defect response field construction, texture suppression filtering, adaptive edge enhancement, and defect morphology constraint connected component segmentation, the algorithm achieves accurate edge extraction and region segmentation of weak contrast and small defects on the electrode surface.

[0026] To address the issues of regular coating textures on lithium battery electrode surfaces, which easily obscure weak defect edges, and the generation of numerous false edges using traditional edge detection methods, this embodiment proposes a Defect Response Field (DRF) computational mechanism for preprocessed images. The response value is constructed pixel by pixel. This mechanism highlights abrupt defect changes by using the ratio of local gradient to background gradient, and suppresses periodic coating textures on the electrode surface by using the grayscale variance term. This achieves simultaneous processing of "preserving defects, suppressing textures, and removing noise," avoiding the pain point of traditional algorithms being unable to distinguish between textures and defects from the source. Specifically, the defect response field... Represented as: ; in, This represents the local gradient magnitude, used to capture abrupt changes at defect edges; This is the local gradient mean, used to characterize the intensity of background changes in the current region; The grayscale variance within the window is used to identify areas with coating texture. To prevent division by zero of extremely small positive numbers, for example, ω is the texture suppression coefficient, which maximizes the suppression of regular texture interference without losing the strength of defect edges. For example, ω=0.7.

[0027] Based on the defect response field, a defect confidence map based on DRF adaptive thresholding is introduced for edge enhancement, making the algorithm more sensitive to difficult-to-detect defects such as micro-holes, minor scratches, and weak-contrast exposed metal. This step enhances the edges of high-confidence defect regions through dynamic gain coefficients while maintaining smoothness in non-defect regions, resolving the contradiction in traditional algorithms where "over-enhancement leads to noise amplification, while insufficient enhancement leads to missed weak defects." (Defect confidence map) Calculated in the following way: ; Where k=5 is the gain coefficient. This is the global median of the DRF. Edge enhancement is then performed: ; in, For example, a defect enhancement factor. φ =1.5; The defect confidence map and the two work together to achieve an adaptive adjustment effect where the more obvious the defect, the greater the enhancement intensity; and the cleaner the background, the smaller the enhancement intensity, making subsequent segmentation more stable and reliable.

[0028] Furthermore, a dynamic dual-threshold calculation method based on global DRF distribution is adopted. High and low segmentation thresholds are automatically generated according to the defect response intensity of the entire electrode image. This eliminates the need for manual parameter tuning and is unaffected by uneven electrode brightness, thickness differences, or illumination drift, thus exhibiting stronger adaptability to field conditions. The threshold calculation formula is as follows: ; ; in, The mean of the DRF of the entire image. The standard deviation of the DRF for the entire image. High threshold. Used for accurately locating real and reliable defect edges, low threshold This adaptive thresholding mechanism is used to preserve weak edges and connect fracture defects, maximizing the retention of minute defect information while removing noisy edges. It dynamically adjusts according to the overall defect response distribution of the image, offering greater robustness compared to a fixed-ratio threshold.

[0029] Finally, the edges obtained from the initial segmentation are subjected to electrode defect-specific morphological constraint tracking and breakpoint connection. Based on the geometric features of bubbles, exposed metal, scratches, particles, and decarburization, only edges conforming to the true defect structure are retained, while pseudo-edges caused by coating textures, dust, and reflections are completely filtered out. Subsequently, morphological closing operations are performed using 5×5 circular structuring elements to fill tiny breakpoints at the edges and connect adjacent broken edge segments, ultimately obtaining a continuous, complete, burr-free, and interference-free defect region image. This provides a highly accurate segmentation foundation for subsequent feature extraction and classification using a lightweight defect-aware convolutional neural network (DFP-CNN).

[0030] In this embodiment, a Defect Response Field Adaptive Edge Segmentation (DRF-AES) algorithm is employed to distinguish between areas of interference from regular coating textures on lithium battery electrodes and areas of actual defects formed during production, at the underlying mechanism level. By leveraging defect response field modeling, it achieves the effect of suppressing redundant textures and preserving complete defect features, fundamentally addressing the industry-wide problem of traditional edge segmentation algorithms easily misclassifying normal coating textures as defects. Simultaneously, combined with an adaptive enhancement strategy for defect confidence and a morphological constraint connected component filtering method, it effectively preserves the features of weak defects that are difficult to identify, such as tiny pinholes, minor scratches, and low-contrast metal leaks, while filtering out image noise and large-area coating texture interference. Facing complex industrial scenarios such as uneven electrode coating thickness and uneven illumination, compared to traditional conventional edge detection and segmentation algorithms, this embodiment significantly improves the integrity of defect edges, and the actual detection rate of ultra-small defects with an area of ​​less than 20 pixels is significantly improved. This enables the delivery of clean defect image samples with no redundant interference and high feature integrity to the backend intelligent classification network.

[0031] (III) Defect Identification The segmented defect region image is input into the defect-aware network (lightweight defect-aware convolutional neural network, DFP-CNN). This neural network performs defect depth feature extraction and defect type classification end-to-end in one go, and directly outputs the defect category and the corresponding classification confidence.

[0032] First, the defect area image Connected component labeling is performed, and the complete image of each defect region is extracted. Connected components with an area of ​​less than 5 pixels and isolated tiny edges with an area of ​​less than 20 pixels and a circularity greater than 0.9 (most of which are residual pseudo edges) are removed to avoid noise interference and ensure that the input of the lightweight defect-aware convolutional neural network (DFP-CNN) is a real defect region image.

[0033] Next, the extracted defect region image is normalized, mapping the image pixel values ​​to the [0,1] interval, and the input size is uniformly 224×224 pixels to ensure the consistency and stability of the lightweight defect-aware convolutional neural network (DFP-CNN) inference. Then, the normalized defect region image is input into the neural network to eliminate the interference caused by differences in image size and pixel values, allowing the network to focus on defect feature learning, improving the model's generalization ability, while reducing the amount of computation, speeding up the inference speed, and adapting to real-time industrial deployment.

[0034] The Defect-Aware Network (DFP-CNN) adopts a core structure of defect-aware convolution and lightweight feature fusion. It is an end-to-end architecture that does not require manual intervention in feature design and can automatically complete the extraction and classification of deep defect features, adapting to the real-time detection needs of industrial production lines.

[0035] The network structure includes: an input layer, three defect-aware convolutional layers (with an improved ReLU6 activation function), two adaptive pooling layers, one lightweight feature fusion layer, one fully connected layer, and an output layer. The defect-aware convolutional layers are responsible for accurately extracting defect features and suppressing texture interference, the adaptive pooling layers are responsible for dynamic dimensionality reduction, the lightweight feature fusion layer is responsible for fusing shallow and deep defect features, the fully connected layer is responsible for feature mapping, and the output layer is responsible for class output.

[0036] The core innovation of DFP-CNN lies in its defect-aware convolutional operations and lightweight feature fusion operations, and its core formula is as follows: (1) Defect-aware convolution differs from traditional convolution in that it has built-in texture suppression functionality: ; In the formula, The defect-aware convolutional layer outputs the pixel value of the feature map at (x,y). Weights for 3×3 defect-aware convolution kernels (core optimizable parameters); is the pixel value at the corresponding position in the input feature map; w is the texture suppression coefficient (fixed value, adapted to the texture characteristics of the electrode coating), for example, w=0.6; This is a texture confidence map, calculated from the local gray-level variance of the input image. It is used to characterize the intensity of texture interference at the current location; the stronger the texture, the higher the confidence level. The closer it is to 1, the stronger the texture suppression effect, achieving the exclusive function of "extracting defect features while suppressing texture interference", which is an effect that traditional convolutional layers do not have.

[0037] (2) Lightweight feature fusion calculation formula (to achieve adaptive fusion of deep and shallow features, balancing lightweight and classification accuracy): ; In the formula, This is the fused defect feature map; , The fusion weights are fixed values, verified through numerous experiments, and adapted to the distribution of electrode defect characteristics. The basic features (edges, grayscale features) output by the shallow defect-aware convolutional layer. Complex features (texture and morphological features) output by convolutional layers that perceive deep defects. This formula enables efficient fusion of shallow and deep features through element-wise weighted fusion operations. It preserves the integrity of basic defect features, enhances the classification discrimination of deep features, and avoids redundant calculations in traditional feature fusion, thus balancing lightweight design with classification accuracy.

[0038] (3) Improved ReLU6 activation function formula (adapts to defect feature distribution, avoids gradient vanishing, and improves network training stability): ; In the formula: This is the output value of the improved ReLU6 activation function. Based on the traditional ReLU6, this function introduces a small offset of 0.01 and a logarithmic correction term to solve the gradient vanishing problem caused by the gradient being 0 when x≤0 in the traditional ReLU6. At the same time, it limits the maximum output value to 6 to avoid gradient explosion, adapts to the weak response characteristics of the electrode defect features, and improves the stability of network training and the accuracy of feature extraction.

[0039] In this embodiment, the original lightweight defect-aware convolutional neural network (DFP-CNN) operates in an end-to-end manner. Defect-aware convolutional layers and adaptive pooling layers work alternately to automatically extract deep features of defects. The shallow defect-aware convolutional layer extracts basic features such as defect edges and grayscale using the aforementioned defect-aware convolution formula, while simultaneously suppressing interference from polarimetric coating textures through texture suppression terms. The deep defect-aware convolutional layer extracts complex features such as defect texture, shape, and boundary contours using the same convolution formula, eliminating the need for manually designing any shape, grayscale, or texture-related features. The lightweight feature fusion layer combines the shallow and deep features using a fusion formula. Adaptive fusion enhances the discriminative power of defect features. The fully connected layer maps the fused features, and the output layer outputs the category probabilities and corresponding classification confidence scores of five types of defects (scratches, bubbles, metal leakage, particles, and decarburization) through the softmax activation function. The category with the highest probability is the preliminary defect category. This eliminates the cumbersome process of traditional manual feature design and avoids the problems of incomplete and poor adaptability of manual feature design. End-to-end integrated feature extraction and classification are completed to improve detection efficiency. At the same time, through the synergistic effect of each core module, texture suppression, weak defect extraction, and accurate classification are completed simultaneously, balancing detection accuracy and speed.

[0040] (iv) Online training During the training of the defect-aware network, an adaptive gradient optimization algorithm is used to construct a parameter optimization model. With the highest classification accuracy and lowest false negative rate as the optimization objectives, the lightweight defect-aware convolutional neural network is adaptively optimized to obtain the optimal combination of classification parameters, including the optimal confidence threshold.

[0041] First, a multi-parameter adaptive weight optimization model is constructed, targeting the core network parameters (defect-aware convolutional kernel weights, adaptive pooling layer parameters, lightweight feature fusion layer weights, and fully connected layer weights) and classification confidence thresholds in the Lightweight Defect-Aware Convolutional Neural Network (DFP-CNN). A total of 128 key parameters (a fixed number of parameters, requiring no fine-tuning) are adaptively optimized.

[0042] This embodiment proposes a Defect Classification Response Gradient Optimization (DCR-GO) algorithm. The core parameters of the neural network are combined with a confidence threshold to form a parameter vector to be optimized. ,in The core parameters of the original lightweight defect-aware convolutional neural network (DFP-CNN) are as follows. The threshold for determining the confidence level of defect classification is set. The improvement lies in the targeted optimization of the core parameters of DFP-CNN, addressing the problems of low efficiency and poor adaptability associated with traditional manual parameter adjustment. Through adaptive optimization, the network parameters are precisely matched to the electrode defect detection scenario, further enhancing classification accuracy and anti-interference capabilities.

[0043] A fitness function prioritizing safety in defect classification is constructed to comprehensively evaluate the merits of parameter combinations. This fitness function balances classification accuracy and production safety, with the core objective of reducing the false negative rate. Its expression is: ; In the formula, Acc represents the overall accuracy of defect classification, used to characterize the overall discrimination performance of the model; FNR is the false negative rate, representing the probability that a real defect is misclassified as a non-defect. The coefficient settings reflect the "safety first" principle in battery production, aiming to improve... The optimization direction is shifted towards lower false negatives and higher reliability, avoiding safety hazards caused by missed detection of serious defects. This improvement aligns with the core requirement of "safety first" in battery production, ensuring the safety and reliability of battery products while also maintaining classification accuracy, achieving a dual optimization of "safety and precision."

[0044] A metric for evaluating the inter-class discriminative power of defect features is calculated, serving as the core guide for parameter iteration and updates. This metric reflects the ability of each layer of the original lightweight defect-aware convolutional neural network (DFP-CNN) to distinguish defect types based on extracted features. The stronger the discriminative power, the larger the metric value, and the higher the priority of network parameter updates, achieving adaptive matching between network parameters and defect types. The evaluation metric expression is: ; in, , Let be the mean values ​​of the features extracted by the i-th layer of the Lightweight Defect-Aware Convolutional Neural Network (DFP-CNN) for the k-th and j-th defect classes, respectively. , The corresponding standard deviation is K, where K is the total number of defect categories, and ε = To prevent division by zero of extremely small positive numbers. The benefit of this improvement is that it makes parameter optimization more targeted, avoids blind iteration, prioritizes optimizing network parameters that contribute significantly to defect classification, improves optimization efficiency, and ensures that the optimized network has stronger feature extraction capabilities, higher discrimination between different defect categories, reduces classification confusion, and improves classification accuracy.

[0045] Furthermore, an adaptive step-size update mechanism based on discriminability feedback is introduced for iterative optimization. The step size is dynamically adjusted according to the discriminability evaluation index of each feature in the current iteration. When the discriminability changes drastically, a large step size is used for rapid exploration; when the discriminability tends to stabilize, a small step size is used for fine-grained search. The step-size update formula is: ; In the formula, Let be the step size of the i-th dimension feature in the t-th iteration. , These represent the maximum and minimum step sizes, respectively, and λ is the decay coefficient (2.0 is recommended). The current discrimination index, The maximum discriminant metric among all current features. Update the parameter vector along the direction of discriminant improvement: ; ; This mechanism can effectively avoid problems such as local optima, slow convergence, and oscillation divergence that are prone to occur in traditional algorithms, and significantly improve the optimization efficiency and result stability.

[0046] The maximum number of iterations is set to 100. Each iteration recalculates the discriminative gradient between defect classes, the fitness function value, and the parameter update amount, while simultaneously retaining historically optimal parameter combinations to prevent degradation. After iteration, the optimal network parameters and optimal confidence threshold that maximize the fitness function are output, used to improve the discrimination accuracy, anti-interference ability, and industrial applicability of the lightweight defect-aware convolutional neural network (DFP-CNN). The benefits of this improvement are that it avoids performance degradation during parameter optimization, ensuring that the final output parameter combination is optimal, further improving the classification accuracy and anti-interference ability of DFP-CNN, making it adaptable to electrode detection scenarios with different brightness and coating textures, and enhancing its industrial applicability.

[0047] After multiple iterations, a well-trained defect perception network is obtained based on the optimal combination of classification parameters. The image of the defect region to be identified is input into the defect perception network to obtain the defect category and classification confidence score.

[0048] The optimal confidence threshold obtained through iterative optimization A two-level confidence verification mechanism is introduced to achieve accurate defect classification. Specifically, the classification confidence score output by DFP-CNN is compared with... Compare, if the classification confidence is ≥ If the classification confidence level is < 0, then the defect category is confirmed as the final classification result; if the classification confidence level is < 0, then the defect category is confirmed as the final classification result. If the defect is identified as a "suspected defect," a manual re-inspection prompt will be generated, allowing for manual confirmation of the defect category to avoid misjudgment.

[0049] In this embodiment, the defect-aware convolutional operator based on the defect-aware network simultaneously performs feature extraction and texture suppression. An adaptive pooling structure preserves minute defect details, and an improved activation function enhances the activation capability of weak defect features. Furthermore, a pre-training strategy specific to high-quality wafers and a defect-specific data augmentation strategy optimize the training process, achieving excellent defect category discrimination capabilities without requiring large-scale labeling of defect samples. Combined with a two-level confidence verification mechanism, low-confidence identification results are automatically classified as suspected defects and transferred to the manual review process, preventing quality inspection errors caused by forced judgment of uncertain samples. The entire network structure is streamlined, with few parameters and fast inference speed, and can be directly installed on ordinary industrial control equipment for stable operation, perfectly adapting to the high-speed online inspection work rhythm of production lines.

[0050] Furthermore, the proposed Defect Classification Response Gradient Optimization Algorithm (DCR-GO) and multi-factor comprehensive quality scoring model overcome the design flaw of traditional detection technologies where the identification process and quality judgment are separated. It also completely differs from existing publicly available optimization algorithms such as general gradient descent and swarm intelligence optimization. Based on an original iterative update logic, it achieves integrated collaborative adaptive optimization of the core weight parameters of the lightweight defect perception network and the defect classification confidence threshold. With high classification accuracy and low high-risk defect false negative rate as dual optimization objectives, coupled with an adaptation function design prioritizing industrial production safety, the parameter iterative optimization process prioritizes minimizing false negative risks, highly aligning with the core production principle of prioritizing safety in lithium battery manufacturing.

[0051] In this embodiment, the optimization algorithm adopts a safety-first fitness function, which takes into account both classification accuracy and the requirement of low false negatives. Furthermore, it uses the inter-class discrimination of defect features as the iterative guide to accurately optimize network parameters and improve optimization efficiency. By setting a discrimination feedback adaptive step size, it avoids the problems of local optima and slow convergence in traditional algorithms, and realizes the coordinated and synchronous optimization of network parameters and confidence thresholds, which is deeply integrated with the confidence verification mechanism.

[0052] (v) Result Judgment A comprehensive defect scoring model is established. The comprehensive score is obtained by weighting the defect type, defect area ratio, and defect quantity statistics. The comprehensive score is compared with the preset pass threshold, and the judgment result is automatically output.

[0053] First, a comprehensive scoring model is constructed, and the scoring formula is as follows: ; in, Where A is the defect type and A is the defect area (pixels). The area threshold is used; a value of 1 is taken when the ratio exceeds 1. N is the number of defects. The threshold is the number of values ​​(10 recommended), and a value of 1 is used when the ratio exceeds 1. α, β, and γ are weighting coefficients that satisfy α + β + γ = 1.

[0054] The physical meaning of this formula is that the defect type is the most important evaluation factor, followed by the defect area, and finally the defect quantity. This is because a single large-area metal leakage defect poses a much greater threat to the battery safety than multiple tiny particle defects.

[0055] After that, a qualified threshold T is set, preferably with a value of 0.5, for qualified determination. The determination rules are as follows: (1) If Score < T, it is determined as qualified; (2) If Score ≥ T, it is determined as unqualified.

[0056] For the situation determined as "suspected defect" in S4, "manual re-inspection" is output regardless of the Score.

[0057] The following information is output: defect type, defect center coordinates, defect area, comprehensive score Score, and determination conclusion (qualified / unqualified / manual re-inspection). Thus, accurate identification and interpretability of the defects on the electrode sheet for battery manufacturing are achieved, ensuring the safety and stability of the battery production process.

[0058] In this embodiment, the defect comprehensive scoring model combines the defect type, the actual area ratio of the defect, and the defect distribution quantity to complete the quantitative scoring, sets a fixed qualified determination threshold to complete the automated rating, and at the same time reserves a manual re-inspection channel for suspected defects. The whole set of determination logic conforms to the actual production quality inspection standard of the lithium battery electrode sheet, and has extremely strong on-site practicability and feasibility.

[0059] This specific embodiment builds an integrated framework for intelligent detection and determination of lithium battery electrode sheet defects by integrating defect response field adaptive edge segmentation, lightweight defect perception neural network classification, and adaptive parameter collaborative optimization, effectively overcoming the actual industry problems existing in the existing detection technologies, such as serious interference of coating texture, easy omission of defects with weak contrast, large number of pseudo-edges, low classification accuracy in small sample scenarios, weak network generalization ability, cumbersome debugging of model parameters, and disconnection of the detection and determination process. By introducing defect response field modeling, differential texture suppression, and adaptive edge constraint segmentation means in the image segmentation stage, a dedicated lightweight defect perception convolutional neural network is constructed in the intelligent classification stage to complete end-to-end feature learning and defect recognition.

[0060] This specific embodiment is an intelligent detection method that can achieve high-precision defect detection under small sample conditions, has good interpretability, is lightweight and can be deployed on industrial production lines, and can organically integrate defect detection and qualification determination, significantly improving the defect detection rate in the manufacturing process of lithium battery electrode sheets, reducing the risk of missed detection, and achieving the balance between battery safety guarantee and manufacturing cost control.

[0061] Embodiment 2 This embodiment provides a multi-feature fusion detection system for electrode defects in battery manufacturing, including: The preprocessing module is configured to acquire and preprocess surface images of lithium battery electrode sheets after coating. The feature extraction module is configured to construct a defect response field based on the preprocessed image, perform edge enhancement using a defect confidence map based on an adaptive threshold, and obtain a defect region image by combining dual threshold processing. The defect identification module is configured to input a defect region image into a defect perception network to obtain the defect category and classification confidence level. During the training process of the network, an adaptive gradient optimization algorithm is used to adaptively optimize the defect perception network with the goal of maximizing classification accuracy and minimizing false negative rate, thereby obtaining the optimal combination of classification parameters including the optimal confidence threshold.

[0062] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the multi-feature fusion detection method for electrode defects in battery manufacturing as described in Embodiment 1 above.

[0063] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the multi-feature fusion detection method for electrode defects in battery manufacturing as described in Embodiment 1 above.

[0064] The steps or modules involved in Embodiments 2 to 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting defects of an electrode sheet for battery manufacturing, characterized by comprising: include: Collect and preprocess surface images of lithium battery electrode sheets after coating. A defect response field is constructed based on the preprocessed image, and edge enhancement is performed using a defect confidence map based on an adaptive threshold. The defect region image is obtained by combining dual threshold processing. Input the defect region image into the defect perception network to obtain the defect category and classification confidence. During the training process of the network, an adaptive gradient optimization algorithm is used to optimize the defect perception network with the highest classification accuracy and the lowest false negative rate as the optimization objective, so as to obtain the optimal combination of classification parameters including the optimal confidence threshold.

2. The method for detecting defects of an electrode sheet for manufacturing a battery according to claim 1, wherein The construction of the defect response field based on the preprocessed image specifically includes: constructing the defect response field based on the ratio of the local gradient to the background gradient of the preprocessed image and the grayscale variance, which is used to identify the coated texture region and suppress regular texture interference.

3. The method for detecting defects of an electrode sheet for manufacturing a battery according to claim 1, wherein The edge enhancement using a defect confidence map based on an adaptive threshold specifically includes: Based on the Sigmoid function, combined with the gain coefficient to amplify the difference in the defect response field and the global median to correct the overall deviation, the pixel values ​​of the defect response field are converted into confidence levels in the range of 0-1. Edge enhancement is performed based on the confidence level, defect response field, and defect enhancement factor.

4. The method for multi-feature fusion detection of electrode defects for battery manufacturing as described in claim 1, characterized in that, In the dual threshold method, the high threshold is used to locate the true defect edges of the image after edge enhancement, and the low threshold is used to retain the weak edges of the image after edge enhancement and connect the broken defects.

5. The method for multi-feature fusion detection of electrode defects for battery manufacturing as described in claim 1, characterized in that, The defect-aware network comprises an input layer, three defect-aware convolutional layers, two adaptive pooling layers, one lightweight feature fusion layer, one fully connected layer, and an output layer; the defect-aware convolutional layers are used to extract multi-scale defect features and suppress texture interference. The lightweight feature fusion layer is used to fuse shallow and deep defect features in multi-scale defect features.

6. The method for multi-feature fusion detection of electrode defects for battery manufacturing as described in claim 1, characterized in that, The adaptive gradient optimization algorithm is specifically as follows: iterative optimization is performed based on an adaptive step size update mechanism with discriminability feedback; the step size is dynamically adjusted according to the discriminability evaluation index of each feature in the current iteration, and when the rate of change of discriminability is greater than a preset threshold, the first step size is selected to perform exploration. When the rate of change of the discrimination is less than or equal to a preset threshold, a second step size search is selected, wherein the first step size is greater than the second step size, and the parameter vector is updated along the direction of discrimination improvement.

7. The method for multi-feature fusion detection of electrode defects for battery manufacturing as described in claim 1, characterized in that, The step of inputting the defect region image into the defect perception network to obtain the defect category and classification confidence score also includes: The classification confidence score is compared with the optimal confidence score threshold. If the classification confidence score is greater than or equal to the optimal confidence score threshold, the defect category is confirmed as the final classification result; otherwise, it is judged as a suspected defect and a manual re-inspection prompt is generated. A comprehensive score is calculated based on the defect type, defect area ratio, and defect quantity. The comprehensive score is then compared with a preset pass threshold. If the comprehensive score is less than the preset pass threshold, the electrode is deemed to be qualified; otherwise, it is deemed unqualified.

8. A multi-feature fusion detection system for electrode defects in battery manufacturing, characterized in that, include: The preprocessing module is configured to acquire and preprocess surface images of lithium battery electrode sheets after coating. The feature extraction module is configured to construct a defect response field based on the preprocessed image, perform edge enhancement using a defect confidence map based on an adaptive threshold, and obtain a defect region image by combining dual threshold processing. The defect identification module is configured to input a defect region image into a defect perception network to obtain the defect category and classification confidence level. During the training process of the network, an adaptive gradient optimization algorithm is used to adaptively optimize the defect perception network with the goal of maximizing classification accuracy and minimizing false negative rate, thereby obtaining the optimal combination of classification parameters including the optimal confidence threshold.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the multi-feature fusion detection method for electrode defects in battery manufacturing as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the multi-feature fusion detection method for electrode defects in battery manufacturing as described in any one of claims 1-7.

Images