A lithium battery pole piece surface defect detection method based on abnormal guidance

Abstract

This invention discloses a method for detecting surface defects of lithium battery electrodes based on anomaly guidance, belonging to the field of industrial vision inspection and intelligent manufacturing technology. The method first collects lithium battery electrode image data, constructs a dataset containing normal and defective samples, and then labels, enhances, and segments the data. Second, a feature reference model is established based on a large number of normal samples. Based on this, anomaly response calculation is performed on the input image to obtain anomaly guidance information reflecting potential defect areas. Then, a multi-branch feature reconstruction mechanism is introduced into the YOLO26n network to perform fine-grained enhancement of shallow features based on the anomaly response map. Simultaneously, edge information and anomaly responses are combined for synergistic enhancement, and an adaptive multi-scale fusion strategy is used to improve the representation ability of defects at different scales. Finally, the enhanced features are input into the improved YOLO26n model for training and inference, achieving accurate localization and classification of minute defects on the electrode surface. This invention can effectively improve the detection rate of minute defects, reduce false detections and false negatives in complex backgrounds, and has good robustness and engineering application value.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and industrial inspection technology, and in particular to a surface defect detection method and system based on anomaly guidance, applicable to the detection of minute defects on the surface of industrial materials such as lithium battery electrodes. Background Technology

[0002] With the rapid development of the new energy industry, lithium batteries, as core energy storage devices, directly affect product safety and lifespan due to their manufacturing quality. During lithium battery production, various defects such as scratches, cracks, pinholes, and foreign matter adhesion can easily appear on the electrode surface. These defects are typically small in size, irregular in shape, and randomly distributed, posing a significant challenge to automated inspection. Therefore, achieving efficient and accurate detection of surface defects on lithium battery electrodes has become an important research direction in the field of industrial vision inspection.

[0003] Existing surface defect detection methods mainly include those based on traditional image processing and those based on deep learning. Traditional methods rely on manually designed features, such as edge detection and threshold segmentation. While effective in specific scenarios, they exhibit poor robustness under complex backgrounds, varying lighting conditions, and diverse defect types, making them unsuitable for practical industrial applications. With the development of deep learning technology, convolutional neural network-based target detection algorithms (such as the YOLO series) have been widely applied in industrial defect detection. These methods can automatically learn feature representations, improving detection accuracy and efficiency to a certain extent.

[0004] However, existing target detection technologies still have the following shortcomings: First, for tiny defects in lithium battery electrodes, their proportion in the image is extremely low and their features are not obvious, making them easy to be submerged in the feature extraction process, resulting in serious missed detections; Second, in industrial scenarios, the number of normal samples far exceeds the number of defective samples, while existing methods usually only use labeled defective samples for supervised learning, failing to fully explore the structural information contained in normal samples, resulting in insufficient ability of the model to distinguish anomalies; In addition, the electrode surface is often accompanied by complex textures, reflections, and noise interference, which further increases the probability of false detection and affects the stability of the detection results.

[0005] Therefore, a new method for detecting surface defects in lithium battery electrodes is needed, which can effectively utilize information from normal samples to characterize abnormal areas and improve the feature representation ability of minute defects, thereby enhancing detection accuracy and robustness. Summary of the Invention

[0006] To address the problems of insufficient detection capability for minute defects, inadequate utilization of normal sample information, and susceptibility to missed and false detections in complex backgrounds in existing technologies, this invention provides an anomaly-guided method and system for detecting surface defects in lithium battery electrodes, thereby improving the detection rate of minute defects and the overall stability and robustness of the detection.

[0007] To achieve the above objectives, this invention employs an anomaly-guided method for detecting surface defects in lithium battery electrodes, comprising the following steps: Step 1: Acquire image data of lithium battery electrodes.

[0008] Step 2: Preprocess the lithium battery electrode image dataset, including labeling, augmentation, and segmentation of the data, and construct a training dataset.

[0009] Step 3: Construct a feature reference model based on normal samples in the dataset, and obtain the feature distribution representing the normal state through feature extraction and statistical modeling.

[0010] Step 4: Input the image to be detected into the feature extraction network, calculate the abnormal response map of the input image in combination with the feature reference model, and introduce the feature enhancement FEM mechanism based on the abnormal response into the YOLO26n model to enhance the shallow features in a fine-grained manner, and combine edge information to achieve synergistic enhancement of abnormal regions. At the same time, the adaptive multi-scale fusion FAM strategy is used to dynamically weight features of different scales to improve the representation ability of multi-scale defects.

[0011] Step 5: The enhanced features are then input into the improved YOLO26n model for training and inference to achieve the localization and classification output of defects on the surface of lithium battery electrodes.

[0012] Preferably, in steps one and two, each image corresponds to a TXT format annotation file, used to accurately annotate the category and location information of each defect in the image. The dataset consists of multiple 640×512 pixel lithium battery electrode images and their corresponding label files. The electrode dataset is expanded by image flipping, rotation, and brightness adjustment. Based on this, the enhanced dataset is randomly divided into a training set and a validation set in an 8:2 ratio.

[0013] Preferably, the abnormal response map serves as intermediate guiding information in the feature enhancement process. By weighted reconstruction of the features, the response of the defect region in the feature space is highlighted, thereby improving the distinguishability of minute defects.

[0014] Preferably, the improvements to the YOLO26n-based algorithm model include the following aspects: introducing an FEM mechanism after the second-layer feature extraction to compensate for the information of small targets in shallow features; adding an EESO structure after the C2PSA module to enhance the edge information of deep features; and replacing all Concat modules with FAM structures to improve the multi-scale feature fusion capability.

[0015] Preferably, the feature enhancement (FEM) mechanism includes fine-grained enhancement of shallow features while preserving their structure, and optimization of enhanced features by combining anomaly response maps to reduce information loss from minor defects during feature extraction.

[0016] Preferably, the EESO method enhances the representation of defect boundary regions by fusing edge information and abnormal responses, thereby improving the detection capability of irregular defects.

[0017] Preferably, the multi-scale fusion FAM structure adopts an adaptive weight allocation method, which dynamically adjusts the features at different scales according to spatial context information to improve the model's adaptability to defects of different sizes.

[0018] Compared with existing technologies, the present invention has the following beneficial effects: By introducing normal sample feature modeling and anomaly response guidance mechanism, the present invention changes the way that the traditional target detection model uniformly processes global features, enabling the model to focus on potential abnormal regions, thereby significantly improving the detection capability of small defects; at the same time, through multi-branch feature enhancement and multi-scale fusion strategies, the robustness of the model under complex background conditions is effectively enhanced, and the false detection and false detection rates are reduced, which has good engineering application value. Attached Figure Description

[0019] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and specific implementation methods; Figure 1 This is a flowchart illustrating the structural process of the anomaly-guided lithium battery electrode surface defect detection method of the present invention. Figure 2 This is a flowchart of the network structure based on the improved YOLO26n algorithm of this invention; Figure 3 The structure diagram of the shallow feature enhancement (FEM) module; Figure 4 This is a structural diagram of the Edge-Anomaly Collaboration EESO module. Figure 5 The structure diagram of the multi-scale feature fusion (FAM) module; Figure 6This is a diagram illustrating defects on the surface of lithium battery electrodes. Figure 7 This is a diagram showing the surface defects detected on a lithium battery electrode.

[0021] The specific implementation party was timid. The following detailed description, in conjunction with the accompanying drawings, provides a more comprehensive overview of the anomaly-guided method for detecting surface defects in lithium-ion battery electrodes. It is clear that the described embodiments are merely a portion of the specific implementations of this invention. Based on the technical solutions disclosed in this invention, those skilled in the art can make appropriate modifications or variations to the embodiments without departing from the inventive spirit of this invention, and all equivalent implementations obtained should be included within the protection scope of this invention.

[0022] In this embodiment, an anomaly-guided method for detecting surface defects in lithium battery electrodes is provided, the overall process of which is as follows: Figure 1 As shown, it includes the following steps: Step 1: Collect a dataset of lithium battery electrode images; for example, the resulting image size is 640×512 pixels, including multiple lithium battery electrode images and corresponding label files.

[0023] Step 2: Preprocess the defect image samples of lithium battery electrodes based on defect type and location, including defect annotation, data augmentation, and dataset partitioning. Specifically, data preprocessing methods include labeling defect regions, data augmentation, and dataset partitioning. Data augmentation methods include, but are not limited to, image flipping, rotation, and brightness adjustment to improve the model's generalization ability. The dataset is divided into training and validation sets in an 8:2 ratio, and `train` and `val` files are generated to store the corresponding images and labels.

[0024] Step 3: Construction of a Normal Sample Feature Reference Model; Specifically, normal samples (i.e., defect-free samples) are selected from the dataset, and features are extracted from these normal samples using a feature extraction network. A normal feature reference model is then constructed based on the extracted features. Specifically, the backbone network of the YOLO26n model can be used to perform multi-layer feature extraction on the normal samples, and a feature distribution can be established through feature storage to characterize the feature patterns of lithium battery electrodes under normal conditions.

[0025] Step 4: Anomaly-guided feature enhancement and fusion; Specifically, an anomaly-based feature enhancement mechanism is introduced in the YOLO26n model to perform multi-branch processing on features, including: (1) Shallow feature enhancement: Fine-grained enhancement processing is performed on the shallow features in the feature extraction network to enhance the texture information of minor defects. Specifically, the shallow features are input into the RGB and grayscale images and processed through two branches. First, the feature map is downsampled and convolved to generate coarse-grained features to capture large-scale semantic information, and then dynamically upsampled to restore the original resolution to provide a reference for fine-grained information restoration. The structure calculates the residual between the original features and the upsampled coarse-grained features to accurately extract the detailed features lost during downsampling, and uses convolution to enhance the residual, strengthening the target edge and texture information. Then, the coarse-grained features and fine-grained residuals are fused to form a feature map that takes into account both contextual semantics and local details; at the same time, the outputs of the RGB and grayscale paths are channel-joined and fused by 1×1 convolution to further integrate color and structural information. Finally, the feature distance is calculated by combining the normal feature reference model constructed in step three to generate abnormal response information. Then, the abnormal response is used to perform weighted modulation on the fused features to reduce information loss during the feature extraction process. (2) Edge-Anomaly Co-enhancement: The feature distance is calculated by combining the normal feature reference model constructed in step three to generate abnormal response information. The abnormal response is then used to weight and modulate the features, guiding the structure to focus on small defect areas. On this basis, the edge information of the image is further extracted by deformable convolution, and the edge information is fused with the abnormal response map. The boundary features of the abnormal area are strengthened by weighting, thereby improving the ability to express irregular defects. (3) Multi-scale adaptive fusion: Feature maps of different scales are fused, and the weights of each scale are dynamically allocated according to spatial context information and abnormal responses. Specifically, deep features are upsampled to be consistent with shallow features, and then element-wise summed to perform initial feature fusion. Subsequently, 1x1 convolution is used to compress the channels of the fused features, and the sigmoid function is used to generate spatial attention weights. The generated weights are applied to the initial fused features to form new fused features, allowing the model to focus on the target region. In order to enhance the cross-channel feature correlation, the new features are divided into m groups along the channel dimension. Global mean normalized attention weights are calculated for each group, and channel attention masks are applied to each group of features. All groups are concatenated to obtain highly correlated aggregated features. Finally, the output features are obtained by fusing with the initial features to improve the model's adaptability to defects of different sizes.

[0026] Through the above multi-branch feature enhancement and fusion process, enhanced feature representations for detection are obtained, wherein the feature distance calculation formula is as follows: in, It is a feature of the input image at position x; is the normal feature reference model; c is the number of channels.

[0027] Step 4: Defect Detection and Result Output; Specifically, the enhanced features obtained in Step 4 are input into the improved YOLO26n model for training and inference. In this embodiment, the target detection model is a detection model improved based on the YOLO26n architecture. During the training phase, to train the YOLO26n network, this embodiment first configures a suitable operating environment, with Python version 3.8 and the deep learning framework PyTorch 1.13. Default parameter settings are used during training, with 64 images input to the network in each training batch, and pre-trained weights are used to initialize the model. After each training cycle, the loss of the entire network is calculated, and the model is iteratively optimized until the mAP metric on the validation set reaches a stable level. Once the mAP metric stabilizes, the training of the YOLO26 network can be stopped to obtain the final training model for lithium battery electrode defect detection.

[0028] During the inference phase, improvements to the YOLO26n model resulted in robust detection of lithium-ion battery electrode defects, achieving higher detection accuracy and faster inference speed, while also facilitating deployment in real-world environments. Using an NVIDIA GeForce RTX 3090 GPU and maintaining consistent basic parameter settings, different algorithm models were trained on the same dataset to locate and classify surface defects in lithium-ion battery electrodes. The experimental results are as follows: Model Average accuracy (%) Parameters GFLOPs FPS Faster-RCNN 82.3 8344763 12.4 57.8 YOLOv8n 85.7 3011027 8.2 107.40 YOLO11n 92.2 2590019 6.4 83.88 YOLO26n 93.4 2504190 5.8 84.89 The improved YOLO26n of this invention 96.8 2580518 6.1 102.08 The algorithm employed in this invention exhibits significant advantages over the unmodified YOLO26n, YOLOv8, YOLO11, and the two-stage object detection algorithm Faster-RCNN in terms of average accuracy, number of parameters, computational cost (GLOPs), and inference speed (FPS). All the aforementioned calculation formulas are performed by removing dimensions and using numerical values. The formulas are derived from extensive data collection and software simulation to closely approximate real-world conditions. The preset parameters and thresholds in the formulas are set appropriately by those skilled in the art based on specific application requirements.

[0029] It should be noted that the above description is merely a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Those skilled in the art can make appropriate modifications, substitutions, or improvements to the embodiments without departing from the spirit and innovative principles of the present invention, and all equivalent embodiments should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the contents of the appended claims.

Claims

1. A method for detecting surface defects of lithium battery electrode sheet based on abnormal guidance, characterized in that, Includes the following steps: Step 1: Acquire images of lithium battery electrodes and construct an image dataset; Step 2: Preprocess the lithium battery electrode image dataset, including labeling defective targets, data augmentation, and dividing the training and validation sets; Step 3: Construct a feature reference model based on normal samples from the lithium battery electrode image dataset; Step 4: Construct an improved YOLO26n model, use the feature reference model to extract features and calculate differences in the input image, generate an anomaly response map to characterize the degree of anomaly, introduce a feature enhancement mechanism based on the anomaly response map into the model, reconstruct the features of the input image with weights, and complete the model training. Step 5: Input the image of the lithium battery electrode to be detected into the trained model, use the abnormal response map to guide feature expression, and output the detection results of surface defects of the lithium battery electrode.

2. The defect detection method according to claim 1, wherein The method for constructing the feature reference model in step three is as follows: using normal samples to extract multi-layer features through a feature extraction network, and establishing a feature distribution under normal conditions based on the features.

3. The defect detection method according to claim 2, wherein The abnormal response map is obtained by calculating the difference between the input image features and the normal feature reference model. The difference is referred to as the feature distance, and the feature distance calculation formula is as follows: ； wherein, is a feature of the input image at position x; is a normal feature reference model; c is the number of channels.

4. The defect detection method according to claim 3, characterized by, The abnormal response map serves as intermediate guiding information in the feature enhancement process, rather than being directly output as a defect detection result.

5. The defect detection method of claim 1, wherein The feature enhancement FEM mechanism in step four includes fine-grained enhancement of shallow features and optimization of enhanced features in combination with anomaly response maps to improve the feature response capability of small defect areas.

6. The defect detection method of claim 4, wherein The feature enhancement EESO mechanism further includes a joint enhancement method based on edge information and anomaly response map, which improves the expressive power of defect boundary regions through weighted fusion of edge information and anomaly response.

7. The defect detection method of claim 5, wherein The model includes a multi-scale feature fusion (FAM) structure, which adaptively weights features at different scales to achieve effective detection of defects of different sizes.

Images