Abnormal electrode sheet detection method and apparatus, electronic device, and storage medium

Abstract

Embodiments of the present application relate to the technical field of computers. Provided are an abnormal electrode sheet detection method and apparatus, an electronic device, and a storage medium. The method comprises: determining a picture to be detected of a battery cell to be tested; and inputting said picture into an anomaly detection model to obtain an output target detection result, the anomaly detection model being obtained by training taking annotated battery cell pictures and unannotated battery cell pictures as training samples and taking annotation information corresponding to the annotated battery cell pictures and first results corresponding to the unannotated battery cell pictures as training labels, the first results being obtained in the following mode: inputting each unannotated battery cell picture into a trained comparison model, and obtaining a first result of each unannotated battery cell picture output by the comparison model, and the comparison model being trained by taking the annotated battery cell pictures as training samples and taking the annotation information corresponding to the annotated battery cell pictures as a training label. The embodiments of the present application improve the electrode sheet anomaly detection accuracy on the basis of a small number of annotated battery cell pictures.

Description

Method and device for detecting abnormal pole piece, electronic device and storage medium

[0001] The present application claims priority to Chinese Patent Application No. 202410832539.7 entitled "Method and device for detecting abnormal pole piece, electronic device and storage medium" filed on June 25, 2024, the disclosure of which is incorporated herein in its entirety as part or all of the present application. TECHNICAL FIELD

[0002] The present application relates to the technical field of computers, in particular, the present application relates to a method and device for detecting abnormal pole pieces, electronic devices and storage media. BACKGROUND

[0003] The pole piece is an important component of the battery, and its performance directly affects the overall performance and life of the battery. Therefore, quality control of the pole piece is crucial, especially for abnormal detection of the pole piece. In the manufacturing process of lithium batteries, if the production process is not proper, it may cause abnormal bonding of the pole piece, and such batteries often do not meet the production standards.

[0004] In recent years, with the development of deep learning technology, some research has begun to try to apply deep learning to the abnormal detection of the pole piece. However, a large amount of labeled data is needed to train the model, therefore, there is a problem that it is not possible to train a model with high accuracy for pole piece abnormal detection based on a small amount of labeled data. SUMMARY

[0005] The present application provides a method and device for detecting abnormal pole pieces, electronic devices and storage media, to solve the technical problem that it is not possible to train a model with high accuracy for pole piece abnormal detection based on a small amount of labeled data.

[0006] According to a first aspect of the present application, a method for detecting abnormal pole pieces is provided, the method comprising: determining a to-be-detected picture of a to-be-detected battery cell, the to-be-detected picture including a display area of a to-be-detected pole piece on the to-be-detected battery cell;

[0007] inputting the to-be-detected picture into an abnormality detection model to obtain a target detection result output by the abnormality detection model, the target detection result being used to represent whether the to-be-detected pole piece is normal;

[0008] The abnormality detection model is trained using the labeled battery cell pictures in the first training set and the unlabeled battery cell pictures in the second training set as training samples, and using the labeled information corresponding to the labeled battery cell pictures and the first result corresponding to the unlabeled battery cell pictures as training labels;

[0009] The first result is obtained by the following method:

[0010] inputting each unlabeled battery cell picture in the second training set into the trained contrast model to obtain a first result of each unlabeled battery cell picture output by the contrast model;

[0011] The contrast model is trained by taking the labeled battery cell pictures in the first training set as training samples and taking the label information corresponding to the labeled battery cell pictures as training labels;

[0012] Each labeled battery cell picture includes a display area of the pole piece on the corresponding first sample battery cell, and the label information of each labeled battery cell picture is used to represent whether the pole piece on the corresponding first sample battery cell is normal;

[0013] Each unlabeled battery cell picture includes a display area of the pole piece on the corresponding second sample battery cell, and the first result of each unlabeled battery cell picture is used to represent whether the pole piece on the corresponding first sample battery cell is normal.

[0014] In one possible implementation, the first training set is generated by the following method:

[0015] obtaining at least one labeled first battery cell picture and label information corresponding to the first battery cell picture; the first battery cell picture is a battery cell picture including a display area of a normal pole piece or a display area of an abnormal pole piece;

[0016] training the first model to convergence by taking the first battery cell picture as a training sample and taking the label information corresponding to the first battery cell picture as a training label, so as to serve as a screening model;

[0017] inputting the unlabeled battery cell picture into the screening model to obtain predicted label information of each unlabeled battery cell picture output by the screening model and a confidence corresponding to the predicted label information;

[0018] taking the unlabeled battery cell picture with a confidence lower than a confidence threshold as a second battery cell picture, labeling the second battery cell picture, and obtaining label information corresponding to each second battery cell picture;

[0019] taking both the first battery cell picture and the second battery cell picture as labeled battery cell pictures to constitute the first training set.

[0020] In another possible implementation, the training process of the contrast model includes:

[0021] training the second model to convergence based on the labeled battery cell pictures in the first training set and the label information corresponding to the labeled battery cell pictures, so as to serve as the contrast model;

[0022] wherein one round of iterative training includes:

[0023] inputting each labeled battery cell picture in the first training set into the second model to obtain a feature representation of each labeled battery cell picture;

[0024] For any two labeled battery cell pictures, a contrast loss between the two labeled battery cell pictures is determined according to the feature representations of the two labeled battery cell pictures;

[0025] A first contrast loss is determined according to the contrast losses between the two labeled battery cell pictures with the same annotation information;

[0026] A second contrast loss is determined according to the contrast losses between the two labeled battery cell pictures with different annotation information;

[0027] The parameters of the second model are updated according to the first contrast loss and the second contrast loss.

[0028] In yet another possible implementation, the anomaly detection model comprises a feature extraction layer, a feature fusion layer, and a detection layer; the to-be-detected picture is input into the feature extraction layer to obtain a plurality of target feature maps of different sizes output by the feature extraction layer;

[0029] The plurality of target feature maps of different sizes are input into the feature fusion layer for feature fusion to obtain a plurality of fusion feature maps output by the feature fusion layer;

[0030] The plurality of target fusion feature maps of different sizes are input into the detection layer to obtain a target detection result output by the detection layer.

[0031] In yet another possible implementation, the plurality of target feature maps of different sizes comprise a first target feature map, a second target feature map, a third target feature map, and a fourth target feature map;

[0032] The to-be-detected picture is input into the feature extraction layer to obtain a plurality of target feature maps of different sizes output by the feature extraction layer, comprising:

[0033] The to-be-detected picture is down-sampled, and the down-sampled feature map is segmented and spliced to obtain the first target feature map;

[0034] The first target feature map is down-sampled, and the down-sampled feature map is segmented and spliced to obtain the second target feature map;

[0035] The second target feature map is down-sampled, and the down-sampled feature map is segmented and spliced to obtain the third target feature map;

[0036] The third target feature map is down-sampled, and the down-sampled feature map is segmented and spliced to obtain the fourth target feature map.

[0037] In yet another possible implementation, the fourth target feature map is subjected to a pooling operation through a plurality of pooling layers connected in series to obtain a plurality of pooling feature maps output by the plurality of pooling layers;

[0038] The pooled feature maps output by each pooling layer are spliced to obtain a first spliced feature map;

[0039] The first spliced feature map is feature-extracted by a plurality of preset-size convolution kernels to obtain a plurality of first features of different sizes;

[0040] The first features of different sizes are spliced in the channel dimension to obtain a second spliced feature map;

[0041] The feature values on each channel of the second spliced feature map are average-pooled to obtain a first global description vector for the target feature map; the global description vector is used to represent the global information of each channel of the target feature map;

[0042] The first global description vector of the second spliced feature map is input into a first fully connected layer to obtain a second global description vector of a preset dimension output by the first fully connected layer; the preset dimension is smaller than the channel dimension of the target feature;

[0043] The second global description vector is input into a second fully connected layer to obtain a third global description vector; the number of neurons of the second fully connected layer is the same as the channel dimension of the fourth target feature map;

[0044] The third global description vector in the channel dimension is input into an activation function to obtain a weight value for each channel;

[0045] For each channel, the feature values of the channel are weighted based on the weight value of the channel to obtain weighted feature values;

[0046] Based on the weighted feature values in each channel, a fourth target feature map to be input into a feature fusion layer is obtained.

[0047] In yet another possible implementation, the fourth target feature map is upsampled, and the upsampled result is spliced with the third target feature map to obtain a first fusion feature map;

[0048] The first fusion feature map is segmented and spliced to obtain a spliced first fusion feature map;

[0049] The spliced first fusion feature map is upsampled, and the upsampled result is spliced with the second target feature map to obtain a second fusion feature map;

[0050] The second fusion feature map is segmented and spliced to obtain a spliced second fusion feature map;

[0051] The spliced second fusion feature map is upsampled, and the upsampled result is spliced with the first target feature map to obtain a third fusion feature map;

[0052] performing segmentation and splicing operations on the third fusion feature map to obtain a first target fusion feature map;

[0053] performing convolution operations on the first target fusion feature, splicing the convolution results with the spliced second fusion feature map, and performing segmentation and splicing operations on the spliced results to obtain a second target fusion feature map;

[0054] performing convolution operations on the second target fusion feature map, splicing the convolution results with the spliced first fusion feature map, and performing segmentation and splicing operations on the spliced results to obtain a third target fusion feature map;

[0055] performing convolution operations on the third target fusion feature map, splicing the convolution results with the fourth target feature map, and performing segmentation and splicing operations on the spliced results to obtain a fourth target fusion feature map;

[0056] obtaining the first target fusion feature, the second target fusion feature, the third target fusion feature, and the fourth target fusion feature output by the feature fusion layer.

[0057] In yet another possible implementation, the segmentation and splicing operations include:

[0058] For any feature map to be segmented and spliced, performing convolution operations on the feature map to obtain a convolutional feature map;

[0059] segmenting the convolutional feature map to obtain a first segmented feature image and a second segmented feature map;

[0060] inputting the first segmented feature image into a plurality of serial bottleneck blocks to perform convolution operations, and obtaining reference feature maps output by the bottleneck blocks;

[0061] splicing the reference feature maps output by the bottleneck blocks and the second segmented feature map to obtain a feature map corresponding to the feature map to be segmented and spliced, which has been subjected to segmentation and splicing operations.

[0062] In yet another possible implementation, the target detection result is a to-be-detected image or a to-be-detected image containing a detection frame, and the detection frame is used to represent an abnormal area of a to-be-detected pole on the to-be-detected image; the detection layer includes a detection module and a screening module;

[0063] When the target detection result is a to-be-detected image containing a detection frame, for any size of target fusion feature map, inputting the target fusion feature map into the detection module to obtain at least one candidate detection frame information output by the detection module; the candidate detection frame information includes the coordinates and confidence of the candidate detection frame;

[0064] According to the confidence of each candidate detection frame, the candidate detection frame information is placed in a candidate detection frame list in descending order of confidence.

[0065] According to the candidate bounding box information in the candidate bounding box list, a plurality of rounds of iteration screening are performed until the candidate bounding box list is empty or the maximum confidence in the candidate bounding box list is less than a preset confidence threshold, and a to-be-detected picture containing at least one detection bounding box is output;

[0066] One round of iteration screening includes:

[0067] The candidate bounding box with the maximum confidence is determined from the candidate bounding box list of the current round, and the candidate bounding box is taken as a detection bounding box, and the candidate bounding box is deleted from the candidate bounding box list;

[0068] For the detection bounding box and any candidate bounding box in the candidate bounding box list, according to the coordinates of the detection bounding box and the coordinates of the candidate bounding box, the overlapping area between the detection bounding box and the candidate bounding box is determined;

[0069] For any candidate bounding box, if the overlapping area between the candidate bounding box and the detection bounding box is greater than an overlapping threshold, the candidate bounding box is deleted from the candidate bounding box list, and a candidate bounding box list of the next round is obtained.

[0070] According to a second aspect of an embodiment of the present application, an abnormal pole piece detection device is provided, and the device includes:

[0071] A determination module is configured to determine a to-be-detected picture of a to-be-detected battery cell, and the to-be-detected picture includes a display area of a to-be-detected pole piece on the to-be-detected battery cell;

[0072] An input module is configured to input the to-be-detected picture into an abnormality detection model, and obtain a target detection result output by the abnormality detection model, wherein the target detection result is used to represent whether the to-be-detected pole piece is normal;

[0073] The abnormality detection model is trained by taking the labeled battery cell pictures in the first training set and the unlabeled battery cell pictures in the second training set as training samples, and taking the labeling information corresponding to the labeled battery cell pictures and the first result corresponding to the unlabeled battery cell pictures as training labels;

[0074] The first result is obtained by the following method:

[0075] Each unlabeled battery cell picture in the second training set is input into the trained comparison model, and the first result of each unlabeled battery cell picture output by the comparison model is obtained;

[0076] The comparison model is trained by taking the labeled battery cell pictures in the first training set as training samples, and taking the labeling information corresponding to the labeled battery cell pictures as training labels;

[0077] Each labeled cell picture includes a display area of the pole piece on the corresponding first sample cell, and the labeling information of each labeled cell picture is used to represent whether the pole piece on the corresponding first sample cell is normal.

[0078] Each unlabeled cell picture includes a display area of the pole piece on the corresponding second sample cell, and the first result of each unlabeled cell picture is used to represent whether the pole piece on the corresponding first sample cell is normal.

[0079] According to a third aspect of the embodiments of the present application, an electronic device is provided, which includes a memory, a processor and a computer program stored in the memory, and the processor implements the steps of the method provided in the first aspect when executing the program.

[0080] According to a fourth aspect of the embodiments of the present application, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the steps of the method provided in the first aspect.

[0081] According to a fifth aspect of the embodiments of the present application, a computer program product is provided, which includes computer instructions stored in a computer readable storage medium, and when a processor of a computer device reads the computer instructions from the computer readable storage medium, the processor executes the computer instructions to make the computer device execute the steps of the method provided in the first aspect.

[0082] The technical scheme provided by the embodiments of the present application has the following beneficial effects:

[0083] The detection method of the abnormal pole piece provided by the embodiments of the present application determines the detection picture of the cell to be detected, and inputs the detection picture into the abnormal detection model. Since the detection picture includes the display area of the pole piece to be detected of the cell to be detected, the output of the abnormal detection model representing whether the pole piece to be detected is normal can be obtained. Since the abnormal detection model is trained with the labeled cell pictures and the unlabeled cell pictures as training samples, with the labeling information corresponding to the labeled cell pictures and the first result corresponding to the unlabeled cell pictures as training labels, and the first result used to train the abnormal detection model is obtained by the contrast model based on the labeled cell pictures, the labeled cell pictures labeled by a small amount of manual work can also obtain a large amount of training samples for training the abnormal detection model, so that the characteristics of the normal pole piece and the abnormal pole piece can be effectively learned in the process of training the abnormal detection model, the pole piece abnormal detection accuracy of the abnormal detection model can be greatly improved without increasing the number of labeled cell pictures labeled by manual work, and the labor cost is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0084] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the description of the embodiments of the present application will be briefly introduced.

[0085] Fig. 1 is a schematic diagram of a system architecture for implementing the method for detecting abnormal pole pieces provided by the embodiments of the present application;

[0086] Fig. 2 is a schematic diagram of a flow of a method for detecting abnormal pole pieces provided by the embodiments of the present application;

[0087] Fig. 3 is a schematic diagram of a flow of a method for generating a first training set in the method for detecting abnormal pole pieces provided by the embodiments of the present application;

[0088] Fig. 4 is a schematic diagram of a flow of a method for one round of iterative training of a comparative model in the method for detecting abnormal pole pieces provided by the embodiments of the present application;

[0089] Fig. 5 is a schematic diagram of a flow of a method for obtaining a target detection result in the method for detecting abnormal pole pieces provided by the embodiments of the present application;

[0090] Fig. 6 is a schematic diagram of a network structure of a comparative model in the method for detecting abnormal pole pieces provided by the embodiments of the present application;

[0091] Fig. 7 is a schematic diagram of a flow of a method for obtaining a plurality of target feature maps of different sizes in the method for detecting abnormal pole pieces provided by the embodiments of the present application;

[0092] Fig. 8 is a schematic diagram of a flow of a method for obtaining a fourth target feature map to be input to a feature fusion layer in the method for detecting abnormal pole pieces provided by the embodiments of the present application;

[0093] Fig. 9 is a schematic diagram of a flow of a method for obtaining a plurality of target fusion features in the method for detecting abnormal pole pieces provided by the embodiments of the present application;

[0094] Fig. 10 is a schematic diagram of a flow of a method for performing segmentation and splicing operations on a feature map in the method for detecting abnormal pole pieces provided by the embodiments of the present application;

[0095] Fig. 11 is a schematic diagram of a flow of a method for obtaining a target detection result in another method for detecting abnormal pole pieces provided by the embodiments of the present application;

[0096] Fig. 12 is a schematic diagram of a flow of a method for one round of iterative screening in another method for detecting abnormal pole pieces provided by the embodiments of the present application;

[0097] Fig. 13 is a schematic diagram of a network structure of an abnormality detection model in the method for detecting abnormal pole pieces provided by the embodiments of the present application;

[0098] Fig. 14 is a schematic diagram of a flow of a method for detecting abnormal pole pieces provided by the embodiments of the present application;

[0099] FIG. 15 is a structural schematic diagram of an abnormal pole piece detection device provided by an embodiment of the present application;

[0100] FIG. 16 is a structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0101] Embodiments of the present application will be described below in conjunction with the accompanying drawings. It should be understood that the embodiments described below in conjunction with the accompanying drawings are exemplary descriptions of the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions of the embodiments of the present application.

[0102] Those skilled in the art can understand that, unless specifically stated otherwise, the singular forms "a", "an" and "the" used herein can also include the plural forms. It should be further understood that the terms "comprise" and "include" used in the embodiments of the present application mean that the corresponding features can be implemented as the presented features, information, data, steps, operations, elements and / or components, but do not exclude other features, information, data, steps, operations, elements, components and / or combinations thereof supported by the present technology. It should be understood that when we say that an element is "connected" or "coupled" to another element, the element can be directly connected or coupled to the other element, or can mean that the element and the other element establish a connection relationship through an intermediate element. In addition, the "connection" or "coupling" used herein can include wireless connection or wireless coupling. The term "and / or" used herein indicates that at least one of the items defined by the term, for example, "A and / or B" can be implemented as "A", or as "B", or as "A and B".

[0103] To make the purposes, technical solutions and advantages of the present application clearer, the embodiments of the present application will be described in further detail below in conjunction with the accompanying drawings.

[0104] The related art will be described below:

[0105] Currently, the detection of abnormality in pole pieces mainly relies on manual visual inspection or traditional machine vision techniques. However, these methods have some problems: on the one hand, manual visual inspection is low in efficiency, easy to fatigue and subjective; on the other hand, traditional machine vision techniques can usually only handle simple and regular image features, and have limited ability to recognize complex and nonlinear abnormal patterns. In recent years, with the development of deep learning technology, some researches have begun to try to apply deep learning to the detection of abnormality in pole pieces. For example, by using convolutional neural networks (CNN) to extract features from pole piece images, and then using traditional machine learning models such as support vector machines (SVM) or random forests (Random Forest) for classification to determine whether the pole piece is normal. However, although deep learning has shown potential in the detection of abnormality in pole pieces, there are still the following problems:

[0106] 1. Existing deep learning-based detection methods usually require a large amount of labeled data to train the model. However, in the actual application of pole piece abnormality detection, the number of abnormal samples is often small, making it difficult to implement this method. The abnormality of the pole piece is mostly due to the similar color and is not easy to find, which brings certain difficulty to the model recognition;

[0107] 2. Traditional machine vision techniques or shallow learning models can only handle simple and regular image features, and have limited ability to recognize complex and nonlinear abnormal patterns. Deep learning models designed for specific types and characteristics of pole pieces may not generalize well to other types of pole pieces;

[0108] 3. Some existing deep learning models have high computational complexity, which limits their application in real-time online abnormality detection systems;

[0109] 4. Current methods for detecting abnormality in pole pieces still rely heavily on manual visual inspection, which is not only low in efficiency, but also prone to human error.

[0110] In view of at least one of the above technical problems or aspects in need of improvement in the related art, the present application provides an abnormal pole piece detection method. The abnormal pole piece detection method includes determining a to-be-detected picture of a to-be-detected battery cell and inputting the to-be-detected picture into an abnormality detection model. Since the to-be-detected picture includes a display area of a to-be-detected pole piece of the to-be-detected battery cell, the abnormality detection model can output a result indicating whether the to-be-detected pole piece is normal. The abnormality detection model is trained using labeled battery cell pictures and unlabeled battery cell pictures as training samples, using labeled information corresponding to the labeled battery cell pictures and a first result corresponding to the unlabeled battery cell pictures as training labels, and the first result used to train the abnormality detection model is obtained by a contrast model based on the labeled battery cell pictures. Therefore, the labeled battery cell pictures labeled by a small amount of manual work can also be used to obtain a large amount of training samples for training the abnormality detection model, so that the abnormality detection model can effectively learn the features of normal pole pieces and abnormal pole pieces during training, and the pole piece abnormality detection accuracy of the abnormality detection model can be greatly improved without increasing the number of labeled battery cell pictures labeled by manual work, and the labor cost is reduced.

[0111] The technical solutions of the embodiments of the present application and the technical effects of the technical solutions of the present application will be described below through the description of several exemplary embodiments. It should be noted that the following embodiments can be mutually referenced, borrowed or combined. For the same terms, similar features and similar implementation steps in different embodiments, they will not be described repeatedly.

[0112] FIG. 1 is a schematic diagram of a system architecture for implementing the abnormal pole piece detection method according to an embodiment of the present application. The system architecture includes a terminal 120 and a server 140.

[0113] The terminal 120 installs and runs an application program of the abnormal pole piece detection method. The terminal 120 is configured to determine whether a to-be-detected pole piece in a to-be-detected picture is normal according to an abnormality detection model.

[0114] The terminal 120 is connected to the server 140 through a wireless network or a wired network.

[0115] The server 140 includes at least one of a server, multiple servers, a cloud computing platform, and a virtualization center. Illustratively, the server 140 includes a processor 144 and a memory 142, and the memory 142 includes a display module 1421, a control module 1422, and a receiving module 1423. The server 140 is configured to provide background services for an application program of the method for detecting abnormal pole pieces. Optionally, the server 140 undertakes the main computing work, and the terminal 120 undertakes the secondary computing work; or the server 140 undertakes the secondary computing work, and the terminal 120 undertakes the main computing work; or the server 140 and the terminal 120 adopt a distributed computing architecture to perform collaborative computing.

[0116] Optionally, the device types of the terminal include at least one of a smartphone, a tablet computer, an e-book reader, a Moving Picture Experts Group Audio Layer III (MP3) player, a Moving Picture Experts Group Audio Layer IV (MP4) player, a laptop computer, and a desktop computer.

[0117] It can be appreciated by those skilled in the art that the number of the terminals can be more or less. For example, the terminals can be only one, or tens or hundreds, or more. The number and the device types of the terminals are not limited in the embodiments of the present application.

[0118] In the embodiments of the present application, a method for detecting abnormal pole pieces is provided, as shown in FIG. 2, which includes the following steps.

[0119] S101, determining a to-be-detected picture of a to-be-detected battery, the to-be-detected picture including a display region of a to-be-detected pole piece on the to-be-detected battery.

[0120] In the embodiments of the present application, the to-be-detected battery is composed of key materials such as pole pieces, electrolytes, and separators. In the manufacturing process of the battery, abnormal pole pieces will cause the battery to fail to meet the detection standard. Therefore, after the production of the battery is completed, the to-be-detected pole pieces of the to-be-detected battery need to be detected. The to-be-detected picture of the to-be-detected battery includes the display region of the pole piece of the to-be-detected battery, which is used to determine whether the pole piece of the to-be-detected battery is abnormal according to the to-be-detected picture.

[0121] S102, inputting the to-be-detected picture into an abnormality detection model to obtain a target detection result output by the abnormality detection model, the target detection result being used to represent whether the to-be-detected pole piece is normal.

[0122] The abnormality detection model is trained by taking the labeled battery cell pictures in the first training set and the unlabeled battery cell pictures in the second training set as training samples, and taking the label information corresponding to the labeled battery cell pictures and the first result corresponding to the unlabeled battery cell pictures as training labels.

[0123] In the embodiment of the present application, the to-be-detected picture is input into the trained abnormality detection model to obtain a target detection result output by the abnormality detection model, which is used to represent whether the to-be-detected pole piece is normal. That is, the target detection result can be displayed as that the to-be-detected pole piece of the to-be-detected battery cell is normal, or that the to-be-detected pole piece of the to-be-detected battery cell is abnormal.

[0124] In the embodiment of the present application, the labeled battery cell picture in the first training set is a picture whose corresponding label information is labeled after artificial confirmation of whether the pole piece is abnormal. The above battery cell picture can be labeled by Labelimg. The unlabeled battery cell picture in the second training set is a picture that is not labeled by artificial labeling, and the pole piece abnormality is detected by the trained comparison model to obtain the corresponding first result.

[0125] In the embodiment of the present application, each labeled battery cell picture includes a display area of the pole piece on the corresponding first sample battery cell, and the label information of each labeled battery cell picture is used to represent whether the pole piece on the corresponding first sample battery cell is normal.

[0126] In the embodiment of the present application, the labeled battery cell picture has a display area of the pole piece of the corresponding first sample battery cell, and the label information of the labeled battery cell picture is used to represent whether the pole piece on the corresponding first sample battery cell is normal. That is, the label information can be displayed as that the pole piece is abnormal, or that the pole piece is normal.

[0127] In the embodiment of the present application, each unlabeled battery cell picture includes a display area of the pole piece on the corresponding second sample battery cell, and the first result of each unlabeled battery cell picture is used to represent whether the pole piece on the corresponding first sample battery cell is normal.

[0128] In the embodiment of the present application, the unlabeled battery cell picture also has a display area of the pole piece on the corresponding second sample battery cell, and the first result of the unlabeled battery cell picture is used to represent whether the pole piece on the corresponding second sample battery cell is abnormal. That is, the first result obtained by the comparison model can be displayed as that the pole piece is abnormal, or that the pole piece is normal.

[0129] In the embodiment of the present application, the number of labeled battery cell pictures in the first training set is much smaller than the number of unlabeled battery cell pictures in the second training set. The labeled battery cell pictures in the first training set and the unlabeled battery cell pictures in the second training set are used as training samples, and the label information corresponding to the labeled battery cell pictures and the first result corresponding to the unlabeled battery cell pictures are used as training labels to train the anomaly detection model. Without providing a large number of manually labeled labeled battery cell pictures, the anomaly detection model can effectively learn the features of the abnormal pole piece and the normal pole piece during the training process, greatly reducing the time spent on manual labeling of battery cell pictures, reducing labor costs, and also improving the robustness of the anomaly detection model.

[0130] In the embodiment of the present application, the first result is obtained by inputting each unlabeled battery cell picture in the second training set into the trained contrast model to obtain the first result of each unlabeled battery cell picture output by the contrast model. The contrast model is trained with the labeled battery cell pictures in the first training set as training samples and the label information corresponding to the labeled battery cell pictures as training labels.

[0131] In the embodiment of the present application, all unlabeled battery cell pictures in the second training set are input into the trained contrast model to obtain the first result predicted by the contrast model for each unlabeled battery cell picture, and the contrast model is also trained with the labeled battery cell pictures in the first training set as training samples and the label information of the labeled battery cell pictures as training labels. That is, the anomaly detection model and the contrast model are trained using the same first training set, so that more training labels available for training can be provided for training the anomaly detection model by the trained contrast model without adding more manually labeled labeled battery cell pictures.

[0132] In the embodiment of the present application, the generation method of the first training set is shown in FIG. 3, and the specific content is as follows:

[0133] S201, obtaining at least one labeled first battery cell picture and label information corresponding to the first battery cell picture; the first battery cell picture is a battery cell picture containing a display area of a normal pole piece or a display area of an abnormal pole piece;

[0134] S202, training the first model to convergence with the first battery cell picture as a training sample and the label information corresponding to the first battery cell picture as a training label, to serve as a screening model;

[0135] S203, inputting an unlabeled battery cell picture into the screening model to obtain predicted label information of each unlabeled battery cell picture output by the screening model and a confidence corresponding to the predicted label information;

[0136] S204, the unlabeled battery cell picture with a confidence lower than the confidence threshold is taken as a second battery cell picture, the second battery cell picture is labeled, and labeled information corresponding to each second battery cell picture is obtained;

[0137] S205, the first battery cell picture and the second battery cell picture are both taken as labeled battery cell pictures to constitute a first training set.

[0138] In S201 of the embodiment of the present application, the obtained first battery cell picture is a picture artificially labeled with corresponding labeled information, the obtained first battery cell picture is a battery cell picture containing a display area of a normal pole piece or containing an abnormal pole piece, and the normal pole piece feature of the display area of the normal pole piece is representative, and the abnormal pole piece feature of the display area of the abnormal pole piece is representative.

[0139] In S202 of the embodiment of the present application, the first model is trained by taking the first battery cell picture as a training sample and taking the labeled information of the first battery cell picture as a training label for multiple rounds of iteration until convergence, and the trained first model is taken as a screening model.

[0140] In S203 of the embodiment of the present application, the unlabeled battery cell picture is input to the screening model, and the predicted labeled information output by the screening model for each unlabeled battery cell picture, that is, the confidence corresponding to the predicted labeled information, is obtained, that is, when obtaining the predicted labeled information representing the normal or abnormality of the pole piece output by the screening model for each unlabeled battery cell picture, a confidence is also output to represent the accuracy of the current prediction result.

[0141] In S204 of the embodiment of the present application, since a lower confidence represents a lower accuracy of the current prediction result, it is indicated that many features in the corresponding unlabeled battery cell picture have not been effectively learned by the model. Therefore, the unlabeled battery cell picture with a confidence lower than a preset threshold is selected as a second battery cell picture, and the second battery cell picture is labeled to obtain labeled information corresponding to each second battery cell picture, so that the most informative and representative battery cell pictures are selected for labeling, and then used for training the abnormality detection model and the comparison model.

[0142] In S205 of the embodiment of the present application, the first battery cell picture and the second battery cell picture are both taken as labeled battery cell pictures to constitute a first training set, that is, on the basis of the labeled first battery cell picture, the most informative unlabeled battery cell picture is selected by the screening model through selective labeling to expand the labeled battery cell picture, so as to obtain the first training set for training the abnormality detection model and the comparison model.

[0143] In the embodiment of the present application, the most informative and representative unlabeled battery cell pictures are selected by the screening model for labeling, which reduces the number of samples that need to be labeled, helps the trained model to learn the features of abnormal and normal pole pieces more effectively, and greatly reduces the labor cost by intelligently selecting the unlabeled battery cell pictures that need to be labeled.

[0144] In the embodiment of the present application, after obtaining the labeled battery cell pictures or unlabeled battery cell pictures, the battery cell pictures need to be preprocessed to make them suitable for deep learning training and detection. The preprocessing steps include adjusting the size of the image, normalization, enhancement (such as adding noise, rotating, cropping, etc.), and data augmentation (such as random sampling, flipping, and mirroring, etc.), which are used to improve the generalization ability and robustness of the model.

[0145] In the embodiment of the present application, the second model is trained based on the labeled battery cell pictures in the first training set and the corresponding labeling information of the labeled battery cell pictures until it converges, serving as a comparison model.

[0146] In the embodiment of the present application, during the training of the comparison model, the labeled battery cell pictures in the first training set and the corresponding labeling information are used to iteratively train the second model for multiple rounds until it converges, so that the trained second model serves as the comparison model.

[0147] In the embodiment of the present application, the method for one round of iterative training of the comparison model is shown in FIG. 4, and the specific content includes:

[0148] S301, input each labeled battery cell picture in the first training set into the second model to obtain the feature representation of each labeled battery cell picture;

[0149] S302, for any two labeled battery cell pictures, determine the contrast loss between the two labeled battery cell pictures according to the feature representation of the two labeled battery cell pictures;

[0150] S303, determine the first contrast loss according to the contrast loss between the two labeled battery cell pictures with the same labeling information;

[0151] S304, determine the second contrast loss according to the contrast loss between the two labeled battery cell pictures with different labeling information;

[0152] S305, update the parameters of the second model according to the first contrast loss and the second contrast loss.

[0153] In S301 of the embodiment of the present application, each labeled battery cell picture in the first training set is input into the second model to obtain a feature representation of each labeled battery cell picture in a low-dimensional space vector. The feature representation can represent whether the detected labeled battery cell picture is abnormal and the region of the abnormal region on the labeled battery cell picture.

[0154] In S302 of the embodiment of the present application, according to the feature representation of each labeled battery cell picture, a contrast loss between each pair of battery cell pictures is calculated. The contrast loss refers to the distance between the feature representations corresponding to each pair of samples.

[0155] In the embodiment of the present application, each pair of labeled battery cell pictures with the same labeling information is a positive sample pair, and each pair of labeled battery cell pictures with different labeling information is a negative sample pair. For example, the labeling information of labeled battery cell pictures A, B, C and D is represented as normal pole piece, normal pole piece, abnormal pole piece and abnormal pole piece, respectively. The positive sample pairs formed by the above four unlabeled battery cell pictures are: labeled battery cell pictures A and B, and labeled battery cell pictures C and D. The negative sample pairs are: labeled battery cell pictures A and C, labeled battery cell pictures A and D, labeled battery cell pictures B and C, and labeled battery cell pictures B and D.

[0156] In S303 of the embodiment of the present application, each pair of labeled battery cell pictures with the same labeling information is a positive sample pair. The first contrast loss is determined according to the contrast losses between all positive sample pairs.

[0157] In S304 of the embodiment of the present application, each pair of labeled battery cell pictures with different labeling information is a negative sample pair. The second contrast loss is determined according to the contrast losses between all negative sample pairs.

[0158] In the embodiment of the present application, the contrast loss is calculated in the following manner:

[0159] wherein q represents the feature representation of the current labeled battery cell picture, k + represents the feature vector of the positive sample related to the current labeled battery cell picture, k - represents a set of negative sample feature vectors unrelated to the current labeled battery cell picture, and τ is a temperature hyperparameter for scaling the inner product, which can control the softening of the probability distribution.

[0160] In S305 of the embodiment of the present application, the parameters of the second model are updated using a gradient descent algorithm or other optimization algorithm based on the first contrast loss and the second contrast loss, so as to narrow the distance between the positive sample pairs in the feature space and to push away the distance between the negative sample pairs in the feature space, thereby prompting the second model to learn more discriminative feature representations.

[0161] In the embodiment of the present application, the anomaly detection model comprises a feature extraction layer, a feature fusion layer and a detection layer, and the method for obtaining the target detection result is shown in FIG. 5, and the specific content comprises:

[0162] S401, input the picture to be detected into the feature extraction layer to obtain a plurality of target feature maps of different sizes output by the feature extraction layer;

[0163] S402, input the plurality of target feature maps of different sizes into the feature fusion layer for feature fusion to obtain a plurality of fusion feature maps output by the feature fusion layer;

[0164] S403, input the plurality of target fusion feature maps of different sizes into the detection layer to obtain a target detection result output by the detection layer.

[0165] In the embodiment of the present application, the anomaly detection model is an improved YOLOV8 model, which comprises a feature extraction layer, a feature fusion layer and a detection layer.

[0166] In S401 of the embodiment of the present application, the picture to be detected is input into the feature extraction layer, a plurality of target feature maps of different sizes are obtained from the feature extraction layer, different multiple down-sampling processing is used to extract target feature maps of different sizes of the picture to be detected, representative features of the picture to be detected are extracted by the feature extraction layer, and with the increase of the down-sampling multiple, more discriminative information is contained in the extracted features.

[0167] In S402 of the embodiment of the present application, the plurality of target feature maps of different sizes are input into the feature fusion layer, the feature fusion layer fuses the target feature maps of different sizes together to obtain a plurality of fusion feature maps of different sizes, and by fusing the plurality of fusion features of different sizes, the anomaly detection model can better classify the features of the pole piece, and by multi-layer feature fusion, the advantages of rich semantic information of high-level feature maps and rich spatial detail information of low-level feature maps can be combined at the same time, and the anomaly detection efficiency is improved.

[0168] In S403 of the embodiment of the present application, the fusion feature maps of different scales output by the feature fusion layer are detected respectively, and according to the features of the display area of the pole piece of the battery cell to be detected, the target detection result of the picture to be detected is output.

[0169] In the embodiment of the present application, the network structure of the comparison model is shown in FIG. 6, and the specific content is as follows:

[0170] The comparison model comprises a data enhancement module Data Augmentation, an encoder Encode and a contrastive loss module.

[0171] In the process of using the contrast model, the unlabeled data set is input into the data enhancement model; in the process of training the contrast model, the labeled data set is input into the data enhancement module, the original sample Anchor output by the data enhancement module, the positive sample belonging to the same class as the original sample, and the negative sample belonging to different classes are obtained, the original sample, the positive sample and the negative sample are input into the encoder, the corresponding feature representation output by the encoder is obtained, the positive sample pair Positive Pairs and the negative sample pair Negative Pairs are input into the contrast loss module, the contrast loss is calculated, and the encoder is optimized according to the contrast loss.

[0172] In the embodiment of the present application, the plurality of target feature maps of different sizes includes a first target feature map, a second target feature map, a third target feature map and a fourth target feature map, and the method for obtaining a plurality of target feature maps of different sizes is shown in FIG. 7, and the specific content includes:

[0173] S501, the to-be-detected picture is down-sampled, the down-sampled feature map is segmented and spliced, and a first target feature map is obtained;

[0174] S502, the first target feature map is down-sampled, the down-sampled feature map is segmented and spliced, and a second target feature map is obtained;

[0175] S503, the second target feature map is down-sampled, the down-sampled feature map is segmented and spliced, and a third target feature map is obtained;

[0176] S504, the third target feature map is down-sampled, the down-sampled feature map is segmented and spliced, and a fourth target feature map is obtained.

[0177] In the embodiment of the present application, the target feature map includes: a first target feature map obtained by down-sampling the to-be-detected picture by four times, a second target feature map obtained by down-sampling the to-be-detected picture by eight times, a third target feature map obtained by down-sampling the to-be-detected picture by 16 times, and a fourth target feature map obtained by down-sampling the to-be-detected picture by 32 times.

[0178] In S501 of the embodiment of the present application, the to-be-detected picture is down-sampled by two convolution modules to obtain a down-sampled feature map, and the down-sampled feature map is segmented and spliced by a C2F module to obtain a first target feature map.

[0179] In S502 of the embodiment of the present application, the first target feature map is down-sampled by one convolution module to obtain a down-sampled feature map, and the down-sampled feature map is segmented and spliced by a C2F module to obtain a second target feature map.

[0180] In S503 of the embodiment of the present application, the second target feature map is down-sampled through a convolution module to obtain a down-sampled feature map, and the down-sampled feature map is segmented and spliced through a C2F module to obtain a third target feature map.

[0181] In S504 of the embodiment of the present application, the third target feature map is down-sampled through a convolution module to obtain a down-sampled feature map, and the down-sampled feature map is segmented and spliced through a C2F module to obtain a fourth target feature map.

[0182] By repeatedly stacking the convolution module, the target feature maps under different down-sampling multiples are extracted, which helps the anomaly detection model to more comprehensively understand the image and improves the accuracy of model recognition.

[0183] In the embodiment of the present application, the method for obtaining the fourth target feature map to be input to the feature fusion layer is shown in FIG. 8, and the specific content includes:

[0184] S601, the fourth target feature map is pooled through a plurality of pooling layers connected in series to obtain a pooled feature map output by each pooling layer;

[0185] S602, the pooled feature maps output by each pooling layer are spliced to obtain a first spliced feature map;

[0186] S603, a plurality of convolution kernels of a preset size are used to respectively extract features from the first spliced feature map to obtain a plurality of first features of different sizes;

[0187] S604, the first features of different sizes are spliced in the channel dimension to obtain a second spliced feature map;

[0188] S605, the feature values on each channel of the second spliced feature map are averaged-pooled to obtain a first global description vector for the target feature map; the global description vector is used to represent the global information of each channel of the target feature map;

[0189] S606, the first global description vector of the second spliced feature map is input to a first fully connected layer to obtain a second global description vector of a preset dimension output by the first fully connected layer; the preset dimension is smaller than the channel dimension of the target feature;

[0190] S607, the second global description vector is input to a second fully connected layer to obtain a third global description vector; the number of neurons of the second fully connected layer is the same as the channel dimension of the fourth target feature map;

[0191] S608, the channel dimension third global description vector is input to an activation function to obtain a weight value of each channel;

[0192] S609, for each channel, weighting the feature value of the channel based on the weight of the channel to obtain a weighted feature value;

[0193] S610, based on the weighted feature value in each channel, obtaining a fourth target feature map to be input to a feature fusion layer.

[0194] In S601 of the embodiment of the present application, the fourth target feature map is subjected to a pooling operation by a plurality of maximum pooling layers connected in series in the SPPF module to obtain a plurality of pooling feature maps output by the pooling layers. For example, there are a first pooling layer, a second pooling layer and a third pooling layer connected in series. The fourth target feature is input to the first pooling layer to obtain a first pooling feature map output by the first pooling layer. The first pooling feature map is input to the second pooling layer to obtain a second pooling feature map. The second pooling feature map is input to the third pooling layer to obtain a third pooling feature map.

[0195] In S602 of the embodiment of the present application, the first pooling feature map, the second pooling feature map and the third pooling feature map obtained above are spliced into one feature map to obtain a first spliced feature map.

[0196] In S603 of the embodiment of the present application, the first spliced feature map is subjected to feature extraction by a fusion module (Inception module). The fusion module (Inception module) is composed of a plurality of branches. Each branch extracts features of the first spliced feature map by a convolution kernel of different size to obtain first features of different sizes output by the branches.

[0197] In S604 of the embodiment of the present application, the first features of different sizes are spliced in the channel dimension, i.e. stacked in the depth direction to obtain a second spliced feature map.

[0198] In S605 of the embodiment of the present application, the feature values on each channel of the second spliced feature map are subjected to average pooling to compress the feature map of each channel into a numerical value, capture global information of each channel and obtain a first global description vector for the target feature map.

[0199] In S606 of the embodiment of the present application, the first global description vector of the second spliced feature map is input to a first fully connected layer. The first global description vector is mapped to a low-dimensional space to obtain a second global vector of a preset dimension. The preset dimension is smaller than the channel dimension of the fourth target feature map.

[0200] In S607 of the embodiment of the present application, the second global description vector is input into a second full connection layer, the second full connection layer restores the dimension of the second global description vector to the dimension of the fourth target feature map, and a third global description vector is obtained, so as to realize modeling of the mutual dependence relationship between channels.

[0201] In S608 of the embodiment of the present application, the third global description vector of the channel dimension is input into a Sigmoid activation function, and the third global description vector is normalized to the range of [0, 1] to obtain the importance weight of each channel, that is, the weight of each channel is obtained.

[0202] In S609 of the embodiment of the present application, for each channel of the fourth target feature map, the feature value of the channel is weighted based on the weight of the channel, and a weighted feature value of each channel is obtained. By weighting the feature value, the channel intensity of the fourth target feature map is adjusted, the importance of each channel is recalibrated, the representation of important features is further strengthened, and the representation of unimportant features is weakened, so that the anomaly detection model pays more attention to important feature information.

[0203] In S610 of the embodiment of the present application, after the feature value of each channel in the fourth target feature map is weighted, a new fourth target feature map is reconstructed.

[0204] In the embodiment of the present application, S603-S610 are completed through an SENet module. The SENet module is composed of a fusion module (Inception module), an average pooling module, two full connection layers, an activation function and a Scale module.

[0205] By learning the importance of each channel, the ability of feature representation is enhanced, important features are emphasized and unimportant features are suppressed, so that the network is more sensitive in capturing key information. On the basis of not significantly increasing the calculation complexity, the extreme piece detection accuracy of the anomaly detection model is significantly improved.

[0206] In the embodiment of the present application, the method for obtaining a plurality of target fusion features is shown in FIG. 9, and the specific content is as follows:

[0207] S701, upsampling the fourth target feature map, and splicing the upsampling result with the third target feature map to obtain a first fusion feature map;

[0208] S702, segmenting and splicing the first fusion feature map to obtain a spliced first fusion feature map;

[0209] S703, upsampling the spliced first fusion feature map, and splicing the upsampling result with the second target feature map to obtain a second fusion feature map;

[0210] S704, segmentation and splicing operations are performed on the second fusion feature map to obtain a spliced second fusion feature map;

[0211] S705, upsampling is performed on the spliced second fusion feature map, and the upsampling result is spliced with the first target feature map to obtain a third fusion feature map;

[0212] S706, segmentation and splicing operations are performed on the third fusion feature map to obtain a first target fusion feature map;

[0213] S707, convolution is performed on the first target fusion feature, the convolution result is spliced with the spliced second fusion feature map, and segmentation and splicing operations are performed on the spliced result to obtain a second target fusion feature map;

[0214] S708, convolution is performed on the second target fusion feature map, the convolution result is spliced with the spliced first fusion feature map, and segmentation and splicing operations are performed on the spliced result to obtain a third target fusion feature map;

[0215] S709, convolution is performed on the third target fusion feature map, the convolution result is spliced with the fourth target feature map, and segmentation and splicing operations are performed on the spliced result to obtain a fourth target fusion feature map;

[0216] S710, the first target fusion feature, the second target fusion feature, the third target fusion feature and the fourth target fusion feature output by the feature fusion layer are obtained.

[0217] In S701 of the embodiment of the present application, the fourth target feature is upsampled by an upsampling module (Upsample module), and at this time, the size of the upsampling result of the fourth target feature map after upsampling is the same as that of the third target feature map. The upsampling result is spliced with the third target feature to obtain a first fusion feature map.

[0218] In S702 of the embodiment of the present application, the first fusion feature map is segmented by a C2F module, that is, the first fusion feature map is split into two parts, one part of the feature map directly participates in the splicing operation, and the other part of the feature map is spliced with the feature map directly participating in the splicing operation after a series of convolution, normalization and activation operations, thereby obtaining a spliced first fusion feature map.

[0219] In S703 of the embodiment of the present application, the spliced first fusion feature map is upsampled, and the size of the upsampling result obtained after the first fusion feature map is upsampled is the same as that of the second target feature map. The upsampling result is spliced with the second target feature map to obtain a second fusion feature map.

[0220] In S704 of the embodiment of the present application, the second fusion feature map is segmented by the C2F module, that is, the second fusion feature map is split into two parts, one part of the feature map directly participates in the splicing operation, and the other part of the feature map is spliced with the feature map directly participating in the splicing operation after a series of convolution, normalization and activation operations, so as to obtain the spliced second fusion feature map.

[0221] In S705 of the embodiment of the present application, the spliced second fusion feature map is up-sampled, and the up-sampling result obtained after the spliced first fusion feature map is up-sampled is the same in size as the first target feature map. The up-sampling result is spliced with the first target feature map to obtain a third fusion feature map.

[0222] In S706 of the embodiment of the present application, the third fusion feature map is segmented by the C2F module, and the third fusion feature map is split into two parts, one part of the feature map directly participates in the splicing operation, and the other part of the feature map is spliced with the feature map directly participating in the splicing operation after a series of convolution, normalization and activation operations, so as to obtain the first target fusion feature to be input to the detection layer.

[0223] In S707 of the embodiment of the present application, the first target fusion feature is subjected to a convolution operation (down-sampling operation) by the convolution module. After the convolution operation of the first target fusion feature map, the convolution result obtained is the same in size as the spliced second fusion feature map. The convolution result is spliced with the second fusion feature map to obtain a splicing result. The splicing result is segmented into two feature maps by the C2F module, one feature map directly participates in the splicing operation, and the other feature map is spliced with the feature map directly participating in the splicing operation after a series of convolution, normalization and activation operations, so as to obtain the second target fusion feature to be input to the detection layer.

[0224] In S708 of the embodiment of the present application, the second target fusion feature map is subjected to a convolution operation by the convolution module. After the convolution operation of the second target fusion feature, the convolution result obtained is the same in size as the spliced first fusion feature map. The convolution result is spliced with the first fusion feature map to obtain a splicing result. The splicing result is segmented into two feature maps by the C2F module, one feature map directly participates in the splicing operation, and the other feature map is spliced with the feature map directly participating in the splicing operation after a series of convolution, normalization and activation operations, so as to obtain the third target fusion feature to be input to the detection layer.

[0225] In S709 of the embodiment of the present application, the third target fusion feature map is subjected to convolution operation by a convolution module, and the third target fusion feature after the convolution operation is obtained. The size of the convolution result is the same as that of the spliced fourth target feature map. The convolution result is spliced with the fourth target feature map to obtain a splicing result. The splicing result is segmented by a C2F module, and the splicing result is divided into two feature maps. One feature map directly participates in the splicing operation, and the other feature map is spliced with the feature map directly participating in the splicing operation after a series of convolution, normalization and activation operations, to obtain a fourth target fusion feature to be input to a detection layer.

[0226] In S710 of the embodiment of the present application, the first target fusion feature, the second target fusion feature, the third target fusion feature and the fourth target fusion feature output by the feature fusion layer are obtained.

[0227] Through the feature fusion of the feature maps of different sizes, the expression ability of the network can be enhanced, and the detection effect can be improved. After the feature fusion, each size of the target fusion feature map contains rich feature information from the feature map of another size, so that the anomaly detection model is more accurate when detecting the wafer, and is more conducive to the positioning and classification of the abnormal wafer.

[0228] In the embodiment of the present application, the method of performing segmentation and splicing operation on the feature map is shown in FIG. 10, and the specific content is as follows:

[0229] S801, for any feature map to be segmented and spliced, performing convolution operation on the feature map to obtain a feature map after convolution;

[0230] S802, segmenting the feature map after convolution to obtain a first segmented feature picture and a second segmented feature map;

[0231] S803, inputting the first segmented feature picture into a plurality of serial bottleneck blocks to perform convolution operation, and obtaining reference feature maps output by the bottleneck blocks;

[0232] S804, splicing the reference feature maps output by the bottleneck blocks and the second segmented feature map to obtain a feature map corresponding to the feature map to be segmented and spliced, which is subjected to segmentation and splicing operation.

[0233] In S801 of the embodiment of the present application, the feature map to be subjected to segmentation and splicing operation is subjected to convolution operation by a convolution module in the C2F module to obtain a feature map after convolution.

[0234] In S802 of the embodiment of the present application, the feature map after convolution is segmented by a Split module to divide the feature map after convolution into two parts to obtain a first segmented feature map and a second segmented feature map.

[0235] In S803 of the embodiment of the present application, the first segmentation feature map is input into a plurality of serial bottleneck blocks for processing to obtain reference feature maps output by the bottleneck blocks. The bottleneck block is usually composed of two convolution layers and a skip connection. The first convolution layer is used to reduce the channel number of the feature map, the second convolution is used to extract the features of the feature map and increase the receptive field, and the skip connection is used to add a skip connection between the convolution layers to connect the input of the bottleneck module directly to the output. This connection mode can alleviate the gradient vanishing problem and help the model learn better.

[0236] In S804 of the embodiment of the present application, the reference feature maps output by the bottleneck blocks are fused with the second segmentation feature map to obtain a feature map corresponding to the feature map to be segmented and spliced and subjected to segmentation and splicing operations.

[0237] The segmentation and splicing operations on the feature map by the C2F module reduce the channel number of the feature map, reduce the computational complexity, can extract more features, improve the expression ability of the model, and use the skip connection to alleviate the gradient vanishing problem.

[0238] In the embodiment of the present application, the target detection result is a to-be-detected picture or a to-be-detected picture containing a detection frame. The detection frame is used to represent an abnormal area of the to-be-detected tab in the to-be-detected picture.

[0239] In the embodiment of the present application, when the output target detection result is a to-be-detected picture, it indicates that the tab of the to-be-detected battery in the to-be-detected picture is normal. When the output target detection result is a to-be-detected picture containing a detection frame, it indicates that the tab of the to-be-detected battery in the to-be-detected picture is abnormal, and the area of the tab abnormality is displayed through the area defined by the detection frame.

[0240] In the embodiment of the present application, the detection layer includes a detection module and a screening module. When the target detection result is a to-be-detected picture containing a detection frame, the method for obtaining the target detection result is shown in FIG. 11, and the specific content is as follows:

[0241] S901, for any size of target fusion feature map, input the target fusion feature map into the detection module to obtain at least one candidate detection frame information output by the detection module; the candidate detection frame information includes the coordinates and confidence of the candidate detection frame;

[0242] S902, according to the confidence of each candidate detection frame, the candidate detection frame information is put into the candidate detection frame list in the order of confidence from large to small;

[0243] S903, performing multi-round iteration screening according to each candidate bounding box information in the candidate bounding box list until the candidate bounding box list is empty or the maximum confidence in the candidate bounding box list is less than a preset confidence threshold, and outputting the to-be-detected picture containing at least one detection frame.

[0244] In the embodiment of the present application, the detection layer includes a detection module and a screening module. The detection module is composed of multiple detectors. Different sizes of target fusion feature maps are input into corresponding detectors for anomaly detection. Each detector is composed of a group of convolution kernel fully connected layers, which are used to predict the category of the extreme sheet on the target fusion feature map of the current size, i.e. abnormal or normal.

[0245] In S901 of the embodiment of the present application, each size of target fusion feature map is input into the corresponding detector in the detection module, and at least one candidate detection frame information for the target fusion feature map of the corresponding size output by each detector is obtained. The candidate detection information includes the coordinate region of the candidate detection block on the target feature map and the accuracy degree (confidence) of the existence of the abnormal region in the candidate detection block.

[0246] In S902 of the embodiment of the present application, the candidate detection frame is sequentially placed in the candidate detection frame list according to the order from large to small of the confidence, so that the candidate detection frame with the highest confidence can be preferentially processed in subsequent processing of the candidate detection frame.

[0247] In S903 of the embodiment of the present application, multi-round iteration screening is performed according to the confidence degree and coordinates of each candidate detection frame until the candidate detection frame list is empty or the maximum confidence in the candidate detection frame list is less than a preset confidence threshold. The candidate detection frame list is empty, which means that all candidate frames have been processed, so that the detection frame for outputting the target detection result can be obtained according to the current screening result. Since the candidate detection frame information in the candidate detection frame list is placed in the order from large to small according to the confidence, when the maximum confidence in the candidate detection frame list is less than the preset confidence threshold, it means that the confidence of the candidate detection frame with the highest confidence in the current candidate detection frame list is already low. Therefore, all candidate detection frames in the current candidate detection frame list cannot be screened out as the detection frame for outputting the target detection result, and the screening is stopped.

[0248] In the embodiment of the present application, the method of one-round iteration screening is shown in FIG. 12, and the specific content includes:

[0249] S1001, determining the candidate detection frame with the maximum confidence from the candidate frame list of the current round, taking the candidate detection frame as the detection frame, and deleting the candidate detection frame from the candidate frame list;

[0250] S1002, for any candidate detection frame in the list of candidate detection frames, according to the coordinates of the detection frame and the coordinates of the candidate detection frame, determine the overlapping area between the detection frame and the candidate detection frame;

[0251] S1003, for any candidate detection frame, if the overlapping area between the candidate detection frame and the detection frame is greater than the overlap threshold, the candidate detection frame is deleted from the candidate frame list, and the next round of candidate detection frame list is obtained.

[0252] In S1001 of the embodiment of the present application, the candidate detection frame with the maximum confidence is taken out from the candidate detection frame list of the current round, that is, the confidence of all candidate detection frames is compared first, and the candidate detection frame with the maximum confidence is determined as the detection frame, which is displayed on the picture to be detected to represent the abnormal area of the pole piece.

[0253] In S1002 of the embodiment of the present application, since there may be multiple abnormal areas of the pole piece on the picture to be detected, it is necessary to determine the candidate detection frame which is different from the coordinate area of the currently identified detection frame. Therefore, first, the overlapping area between the detection frame and each candidate detection frame is calculated according to the coordinates, and according to the size of the overlapping area, it can be determined which candidate detection frame in the list of candidate detection frames represents the same abnormal area.

[0254] In S1003 of the embodiment of the present application, if the overlapping area between the detection frame and the candidate detection frame is greater than the overlap threshold, it means that the current candidate detection frame and the detection frame represent the same abnormal area, and all candidate detection frames representing the same abnormal area are deleted from the candidate detection frame list, so as to obtain the candidate detection frame list for the next round of iteration screening. Then, the candidate detection frame list participating in the next round of iteration screening only contains candidate detection frames representing different abnormal areas from the current detection frame.

[0255] In the embodiment of the present application, when the candidate detection frame of the next round of iteration screening satisfies the iteration condition, the above S1001-S1003 is continued to be executed, so as to select the detection frame representing the abnormal area different from the detection frame of the last round according to the confidence of the candidate frame. Through multiple rounds of iteration screening, all areas of the pole piece with abnormality on the picture to be detected can be effectively screened out.

[0256] In the embodiment of the present application, a network structure diagram of an abnormal detection model is provided as shown in FIG. 13.

[0257] The feature extraction module (i.e., the feature extraction layer) is composed of a convolution module for downsampling operation, a C2F module for segmentation and splicing operation, an SPPF module for pooling operation, and an SE-Net module for compression and excitation operation; wherein the convolution module is composed of Conv2d+BatchNom2d+SiLU, and the SE-Net module is composed of inception+Global average pooling+FC+Sigmoid+Scale.

[0258] The feature fusion module (i.e., the feature fusion layer) is composed of an Upsample module for upsampling, a concat module for connecting feature maps, a C2F module for segmentation and splicing operation, and a convolution module for pooling operation.

[0259] The detection head (i.e., the detection layer) is composed of four detection heads P2, P3, P4, and P5 for performing anomaly classification on target fusion feature maps of corresponding sizes, respectively, and outputting target detection results based on the classification results of each detection head.

[0260] In the embodiment of the present application, as shown in FIG. 14, a flowchart of an abnormality detection method of a pole piece is provided.

[0261] S1101, inputting a to-be-detected picture of a to-be-detected battery cell into an abnormality detection model;

[0262] S1102, obtaining target feature maps of different sizes by a feature extraction layer of the abnormality detection model;

[0263] S1103, performing feature fusion on the target feature maps of different sizes by a feature fusion layer of the abnormality detection model to obtain a plurality of target fusion feature maps of different sizes;

[0264] S1104, performing detection on the target fusion feature maps of different sizes by a detection layer of the abnormality detection model, respectively, to obtain a plurality of candidate bounding box information;

[0265] S1105, determining whether there is an outputable bounding box and the number of outputable bounding boxes from the plurality of candidate bounding boxes according to the candidate bounding box coordinates and the confidence in each candidate bounding box information, and determining a target detection result;

[0266] S1106, according to the target detection result, if the classification result is a normal pole piece, displaying the to-be-detected picture on a visualization interface; if the classification result is an abnormal pole piece, displaying the to-be-detected picture and at least one bounding box on the visualization interface.

[0267] In the embodiments of the present application, the official YOLOv8n model, YOLOv8n-p2 and the anomaly detection model provided in the present application in which the SE-Net is added to the small size channel in the feature fusion layer part are respectively trained for 500 times, and then the trained YOLOv8n model, YOLOv8n-p2 and the anomaly detection model provided in the present application YOLOv8n-p2+SENet are compared, and the comparison results are shown in the following table:

[0268] Table 1

[0269] From the above comparison results, it can be seen that the precision P, recall rate R and average mean precision (mAP50) of the anomaly detection model provided in the embodiments of the present application are all higher than the values of the official YOLOv8, and therefore, the detection effect of the anomaly detection model provided in the embodiments of the present application is higher than that of the official method.

[0270] In the embodiments of the present application, the present disclosure can be widely applied to the field of pole piece anomaly detection, improve the quality control level of the pole piece, reduce the production cost of the battery, and improve the overall performance and life of the battery.

[0271] The anomaly pole piece detection method provided in the embodiments of the present application determines the to-be-detected picture of the to-be-detected battery cell, and inputs the to-be-detected picture into the anomaly detection model. Since the to-be-detected picture includes the display area of the to-be-detected pole piece of the to-be-detected battery cell, the anomaly detection model can output a result representing whether the to-be-detected pole piece is normal. Since the anomaly detection model is trained using the labeled battery cell picture and the unlabeled battery cell picture as training samples, and using the labeled information corresponding to the labeled battery cell picture and the first result corresponding to the unlabeled battery cell picture as training labels, and the first result used for training the anomaly detection model is obtained based on the labeled battery cell picture and the comparison model, the labeled battery cell picture labeled by a small amount of manual work can also obtain a large amount of training samples for training the anomaly detection model, so that the anomaly detection model can effectively learn the features of the normal pole piece and the abnormal pole piece during the training process, and the pole piece anomaly detection accuracy of the anomaly detection model can be greatly improved without increasing the number of labeled battery cell pictures labeled by manual work, while the labor cost is reduced. In addition, the improved YOLOv8 model accelerates the recognition rate, realizes a real-time monitoring model with superior performance, can adapt to complex pole piece characteristics, and is suitable for real-time online anomaly detection systems.

[0272] The anomaly pole piece detection device provided in the embodiments of the present application can include a determination module 1501 and an input module 1502, and the specific contents are as follows:

[0273] The determining module 1501 is configured to determine a to-be-detected picture of the to-be-detected battery cell, and the to-be-detected picture includes a display region of a to-be-detected pole piece on the to-be-detected battery cell.

[0274] The input module 1502 is configured to input the to-be-detected picture into the abnormality detection model to obtain a target detection result output by the abnormality detection model, and the target detection result is used to represent whether the to-be-detected pole piece is normal.

[0275] The abnormality detection model is trained by taking the labeled battery cell pictures in the first training set and the unlabeled battery cell pictures in the second training set as training samples, and taking the label information corresponding to the labeled battery cell pictures and the first result corresponding to the unlabeled battery cell pictures as training labels.

[0276] The first result is obtained by the following method:

[0277] The second training set is input into the trained comparison model to obtain the first result of each unlabeled battery cell picture output by the comparison model.

[0278] The comparison model is trained by taking the labeled battery cell pictures in the first training set as training samples and taking the label information corresponding to the labeled battery cell pictures as training labels.

[0279] Each labeled battery cell picture includes a display region of a pole piece on a corresponding first sample battery cell, and the label information of each labeled battery cell picture is used to represent whether the pole piece on the corresponding first sample battery cell is normal.

[0280] Each unlabeled battery cell picture includes a display region of a pole piece on a corresponding second sample battery cell, and the first result of each unlabeled battery cell picture is used to represent whether the pole piece on the corresponding first sample battery cell is normal.

[0281] The device provided in the embodiments of the present application can execute the method provided in the embodiments of the present application, and the implementation principles are similar. The actions performed by each module in the device of the embodiments of the present application are corresponding to the steps in the method of the embodiments of the present application. For the detailed function description of each module of the device, refer to the description of the corresponding method in the foregoing description, and details are not described herein.

[0282] Optionally, in one possible implementation, the first training set is generated by the following method:

[0283] At least one labeled first battery cell picture and label information corresponding to the first battery cell picture are obtained. The first battery cell picture is a battery cell picture including a display region of a normal pole piece or a display region of an abnormal pole piece.

[0284] The first battery cell picture is taken as a training sample, and the first battery cell picture corresponding label information is taken as a training label to train the first model to convergence, so as to be taken as a screening model;

[0285] The un-labeled battery cell picture is input into the screening model, the prediction label information of each un-labeled battery cell picture output by the screening model and the confidence corresponding to the prediction label information are obtained;

[0286] The un-labeled battery cell picture with a confidence lower than the confidence threshold is taken as a second battery cell picture, the second battery cell picture is labeled, and the label information corresponding to each second battery cell picture is obtained;

[0287] The first battery cell picture and the second battery cell picture are both taken as labeled battery cell pictures to constitute a first training set.

[0288] In another possible implementation, the training process of the comparison model includes:

[0289] The second model is trained based on the labeled battery cell pictures in the first training set and the label information corresponding to the labeled battery cell pictures to convergence, so as to be taken as a comparison model;

[0290] Wherein, one round of iterative training includes:

[0291] Each labeled battery cell picture in the first training set is input into the second model to obtain the feature representation of each labeled battery cell picture;

[0292] For any two labeled battery cell pictures, the comparison loss between the two labeled battery cell pictures is determined according to the feature representation of the two labeled battery cell pictures;

[0293] The first comparison loss is determined according to the comparison loss between the two labeled battery cell pictures with the same label information;

[0294] The second comparison loss is determined according to the comparison loss between the two labeled battery cell pictures with different label information;

[0295] The parameters of the second model are updated according to the first comparison loss and the second comparison loss.

[0296] In still another possible implementation, the anomaly detection model includes: a feature extraction layer, a feature fusion layer and a detection layer; the picture to be detected is input into the feature extraction layer to obtain a plurality of target feature maps of different sizes output by the feature extraction layer;

[0297] The plurality of target feature maps of different sizes are input into the feature fusion layer for feature fusion to obtain a plurality of fusion feature maps output by the feature fusion layer;

[0298] The plurality of target fusion feature maps of different sizes are input into the detection layer to obtain a target detection result output by the detection layer.

[0299] In yet another possible implementation, the plurality of target feature maps of different sizes includes a first target feature map, a second target feature map, a third target feature map, and a fourth target feature map;

[0300] The to-be-detected picture is input into the feature extraction layer to obtain a plurality of target feature maps of different sizes output by the feature extraction layer, including:

[0301] The to-be-detected picture is down-sampled, and the down-sampled feature map is segmented and spliced to obtain a first target feature map;

[0302] The first target feature map is down-sampled, and the down-sampled feature map is segmented and spliced to obtain a second target feature map;

[0303] The second target feature map is down-sampled, and the down-sampled feature map is segmented and spliced to obtain a third target feature map;

[0304] The third target feature map is down-sampled, and the down-sampled feature map is segmented and spliced to obtain a fourth target feature map.

[0305] In yet another possible implementation, the fourth target feature map is subjected to a pooling operation through a plurality of pooling layers connected in series to obtain a plurality of pooled feature maps output by the pooling layers;

[0306] The pooled feature maps output by the pooling layers are spliced to obtain a first spliced feature map;

[0307] A plurality of preset-size convolution kernels are used to respectively extract features from the first spliced feature map to obtain a plurality of first features of different sizes;

[0308] The first features of different sizes are spliced in the channel dimension to obtain a second spliced feature map;

[0309] The feature values on each channel of the second spliced feature map are subjected to average pooling to obtain a first global description vector for the target feature map; the global description vector is used to represent the global information of each channel of the target feature map;

[0310] The first global description vector of the second spliced feature map is input into a first fully connected layer to obtain a second global description vector of a preset dimension output by the first fully connected layer; the preset dimension is smaller than the channel dimension of the target feature;

[0311] The second global description vector is input into a second fully connected layer to obtain a third global description vector; the number of neurons of the second fully connected layer is the same as the channel dimension of the fourth target feature map;

[0312] input the third global description vector of the channel dimension into an activation function to obtain a weight value of each channel;

[0313] For each channel, the characteristic value of the channel is weighted based on the weight value of the channel to obtain a weighted characteristic value;

[0314] Based on the weighted characteristic value in each channel, a fourth target feature map to be input to the feature fusion layer is obtained.

[0315] In yet another possible implementation, the fourth target feature map is upsampled, and the upsampled result is spliced with the third target feature map to obtain a first fusion feature map;

[0316] The first fusion feature map is segmented and spliced to obtain a spliced first fusion feature map;

[0317] The spliced first fusion feature map is upsampled, and the upsampled result is spliced with the second target feature map to obtain a second fusion feature map;

[0318] The second fusion feature map is segmented and spliced to obtain a spliced second fusion feature map;

[0319] The spliced second fusion feature map is upsampled, and the upsampled result is spliced with the first target feature map to obtain a third fusion feature map;

[0320] The third fusion feature map is segmented and spliced to obtain a first target fusion feature map;

[0321] The first target fusion feature is convoluted, and the convolution result is spliced with the spliced second fusion feature map, and the spliced result is segmented and spliced to obtain a second target fusion feature map;

[0322] The second target fusion feature map is convoluted, and the convolution result is spliced with the spliced first fusion feature map, and the spliced result is segmented and spliced to obtain a third target fusion feature map;

[0323] The third target fusion feature map is convoluted, and the convolution result is spliced with the fourth target feature map, and the spliced result is segmented and spliced to obtain a fourth target fusion feature map;

[0324] The first target fusion feature, the second target fusion feature, the third target fusion feature and the fourth target fusion feature output by the feature fusion layer are obtained.

[0325] In yet another possible implementation, the segmentation and splicing operation includes:

[0326] For any feature map to be segmented and spliced, a convolution operation is performed on the feature map to obtain a feature map after convolution;

[0327] The feature map after convolution is segmented to obtain a first segmented feature picture and a second segmented feature map;

[0328] The first segmented feature picture is input into a plurality of serial bottleneck blocks for convolution operation to obtain reference feature maps output by each bottleneck block;

[0329] The reference feature maps output by each bottleneck block and the second segmented feature map are spliced to obtain a feature map corresponding to the feature map to be segmented and spliced, which is subjected to segmentation and splicing operations.

[0330] In yet another possible implementation, the target detection result is a detection picture or a detection picture containing a detection frame, the detection frame being used to represent an abnormal area of a detection target on the detection picture; the detection layer includes a detection module and a screening module;

[0331] When the target detection result is a detection picture containing a detection frame, for any size of target fusion feature map, the target fusion feature map is input into the detection module to obtain at least one candidate detection frame information output by the detection module; the candidate detection frame information includes coordinates of the candidate detection frame and a confidence;

[0332] According to the confidence of each candidate detection frame, the candidate detection frame information is placed in a candidate detection frame list in descending order of confidence;

[0333] According to the candidate detection frame information in the candidate detection frame list, a plurality of rounds of iterative screening are performed until the candidate detection frame list is empty or the maximum confidence in the candidate detection frame list is less than a preset confidence threshold, and a detection picture containing at least one detection frame is output;

[0334] One round of iterative screening includes:

[0335] The candidate detection frame with the maximum confidence is determined from the candidate frame list of this round, and the candidate detection frame is taken as the detection frame, and the candidate detection frame is deleted from the candidate frame list;

[0336] For any candidate detection frame in the candidate frame list, according to the coordinates of the detection frame and the coordinates of the candidate detection frame, the overlapping area between the detection frame and the candidate detection frame is determined;

[0337] For any candidate detection frame, if the overlapping area between the candidate detection frame and the detection frame is greater than an overlapping threshold, the candidate detection frame is deleted from the candidate frame list to obtain a candidate detection frame list of the next round.

[0338] The embodiment of the present application provides an electronic device (computer device / equipment / system), which comprises a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to realize the steps of the detection method of the abnormal pole piece. Compared with the related art, the following advantages can be achieved: by determining a to-be-detected picture of a to-be-detected battery and inputting the to-be-detected picture into an abnormality detection model, since the to-be-detected picture comprises a display area of a to-be-detected pole piece of the to-be-detected battery, the abnormality detection model can output a result for representing whether the to-be-detected pole piece is normal. Since the abnormality detection model is trained by taking the labeled battery picture and the unlabeled battery picture as training samples, taking the labeled information corresponding to the labeled battery picture and the first result corresponding to the unlabeled battery picture as training labels, and the first result used for training the abnormality detection model is obtained by the contrast model based on the labeled battery picture, the labeled battery picture labeled by a small amount of manual work can also be used to obtain a large amount of training samples for training the abnormality detection model, so that the characteristics of the normal pole piece and the abnormal pole piece can be effectively learned in the process of training the abnormality detection model, the pole piece abnormality detection accuracy of the abnormality detection model can be greatly improved without increasing the number of the labeled battery pictures labeled by manual work, and the manual cost is reduced. In an optional embodiment, an electronic device is provided, as shown in FIG. 16. The electronic device 4000 shown in FIG. 16 comprises a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, for example, through a bus 4002. Optionally, the electronic device 4000 can further comprise a transceiver 4004, which can be used for data interaction, such as data sending and / or data receiving, between the electronic device and other electronic devices. It should be noted that the transceiver 4004 is not limited to one in actual application, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present application.

[0339] The processor 4001 can be a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), a FPGA (Field Programmable Gate Array) or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It can implement or execute various exemplary logical blocks, modules and circuits described in connection with the disclosure. The processor 4001 can also be a combination of computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0340] The bus 4002 can include a path for transmitting information between the above-mentioned components. The bus 4002 can be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The bus 4002 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is shown in FIG. 16, but it does not mean that there is only one bus or only one type of bus.

[0341] The memory 4003 can be a ROM (Read Only Memory) or other type of static storage device that can store static information and instructions, a RAM (Random Access Memory) or other type of dynamic storage device that can store information and instructions, an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, an optical disk storage (including a compact disk, a laser disk, an optical disk, a digital versatile disk, a Blu-ray disk, etc.), a magnetic disk storage medium, other magnetic storage device, or any other medium capable of carrying or storing computer programs and capable of being read by a computer, without limitation.

[0342] The memory 4003 is configured to store a computer program for implementing the embodiments of the present application, and the processor 4001 is configured to control the execution of the computer program stored in the memory 4003. The processor 4001 is configured to execute the computer program stored in the memory 4003 to implement the steps shown in the foregoing method embodiments.

[0343] The electronic device can include, but is not limited to, a mobile terminal such as a mobile phone, a notebook computer, a digital broadcast receiver, a PDA (Personal Digital Assistant), a PAD (Tablet Personal Computer), a PMP (Portable Multimedia Player), a car terminal (for example, a car navigation terminal), and the like, and a fixed terminal such as a digital TV, a desktop computer, and the like. The electronic device shown in FIG. 16 is merely an example, and should not impose any limitation on the functions and use range of the embodiments of the present disclosure.

[0344] The embodiments of the present application provide a computer readable storage medium, and the computer readable storage medium stores a computer program. When the computer program is executed by a processor, the steps and corresponding contents of the foregoing method embodiments can be implemented. Compared with the prior art, the following can be achieved:

[0345] By determining the to-be-detected picture of the to-be-detected battery cell and inputting the to-be-detected picture into the anomaly detection model, since the to-be-detected picture includes the display area of the to-be-detected pole piece of the to-be-detected battery cell, the anomaly detection model output used to represent whether the to-be-detected pole piece is normal can be obtained. Since the anomaly detection model is trained by taking the labeled battery cell picture and the unlabeled battery cell picture as training samples, taking the labeled information corresponding to the labeled battery cell picture and the first result corresponding to the unlabeled battery cell picture as training labels, and the first result used to train the anomaly detection model is obtained based on the labeled battery cell picture and the contrast model detection, the implementation of using a small amount of manual labeling of the labeled battery cell picture can also obtain a large amount of training samples for training the anomaly detection model, so that the anomaly detection model can effectively learn the features of the normal pole piece and the abnormal pole piece in the process of training the anomaly detection model, and the pole piece anomaly detection accuracy of the anomaly detection model can be greatly improved without increasing the number of manually labeled labeled battery cell pictures, and the artificial cost is reduced.

[0346] It should be noted that the computer readable medium described above in the present disclosure can be a computer readable signal medium or a computer readable storage medium or any combination of the two. The computer readable storage medium may, for example, be, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination of the above. More specific examples of the computer readable storage medium can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present disclosure, the computer readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus or device. In the present disclosure, the computer readable signal medium can include a data signal carried in a baseband or as a part of a carrier wave, which carries computer readable program code. Such a propagated data signal can take many forms, including but not limited to an electromagnetic signal, an optical signal or any suitable combination of the above. The computer readable signal medium can also be any computer readable medium other than the computer readable storage medium that can send, propagate or transmit the program for use by or in connection with an instruction execution system, apparatus or device. The program code contained on the computer readable medium can be transmitted by any suitable medium, including but not limited to a wire, an optical fiber, an RF (radio frequency) or the like, or any suitable combination of the above.

[0347] The embodiments of the present application also provide a computer program product comprising a computer program, which, when executed by a processor, can implement the steps and corresponding contents of the foregoing method embodiments. Compared with the prior art, the following can be achieved:

[0348] By determining the to-be-detected picture of the to-be-detected battery cell and inputting the to-be-detected picture into the anomaly detection model, since the to-be-detected picture includes the display area of the to-be-detected pole piece of the to-be-detected battery cell, the anomaly detection model can output a result for characterizing whether the to-be-detected pole piece is normal. Since the anomaly detection model is trained by using the labeled battery cell pictures and the unlabeled battery cell pictures as training samples, using the labeled information corresponding to the labeled battery cell pictures and the first result corresponding to the unlabeled battery cell pictures as training labels, and the first result used for training the anomaly detection model is obtained by the contrast model based on the labeled battery cell pictures, the labeled battery cell pictures labeled by a small amount of manual work can also obtain a large amount of training samples for training the anomaly detection model, so that the anomaly detection model can effectively learn the features of the normal pole piece and the abnormal pole piece in the process of training the anomaly detection model, and the pole piece anomaly detection accuracy of the anomaly detection model can be greatly improved without increasing the number of labeled battery cell pictures labeled by manual work, and the labor cost is reduced.

[0349] The terms "first", "second", "third", "fourth", "1", "2", and the like (if any) in the description, claims, and drawings of the present application are used to distinguish similar objects, and do not necessarily have to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described.

[0350] It should be understood that, although the flowcharts of the embodiments of the present application indicate the implementation order of each operation step by arrows, the implementation order of these steps is not limited to the order indicated by the arrows. Unless otherwise specified herein, in some implementation scenarios of the embodiments of the present application, the implementation steps in each flowchart can be executed in other orders as required. In addition, part or all of the steps in each flowchart can include multiple sub-steps or multiple stages based on the actual implementation scenario. Part or all of these sub-steps or stages can be executed at the same time, and each of these sub-steps or stages can also be executed at different times. In the scenario where the execution times are different, the execution order of these sub-steps or stages can be flexibly configured as required, and the embodiments of the present application do not limit this.

[0351] The above is only an optional implementation of some implementation scenarios of the present application. It should be pointed out that, for those skilled in the art, other similar implementation means based on the technical idea of the present application can also be used without departing from the technical concept of the present application, and such implementation also falls within the protection scope of the embodiments of the present application. Industrial applicability:

[0352] The scheme provided in the embodiments of the present application can be applied to the field of computer technology. In the embodiments of the present application, a to-be-detected picture of a to-be-detected battery is determined; the to-be-detected picture is input into an anomaly detection model to obtain an output target detection result; the anomaly detection model is trained by taking labeled battery pictures and unlabeled battery pictures as training samples, and taking label information corresponding to the labeled battery pictures and a first result corresponding to the unlabeled battery pictures as training labels; the first result is obtained by the following manner: inputting each unlabeled battery picture into a trained comparison model to obtain the first result of each unlabeled battery picture output by the comparison model; the comparison model is trained by taking the labeled battery pictures as training samples and taking the label information corresponding to the labeled battery pictures as training labels, thereby realizing improving the anomaly detection accuracy of the pole piece on the basis of a small amount of labeled battery pictures.

Claims

1. A method for detecting abnormal electrodes, comprising: A test image of the battery cell to be tested is determined, wherein the test image includes the display area of ​​the electrode to be tested on the battery cell; The image to be detected is input into the anomaly detection model to obtain the target detection result output by the anomaly detection model. The target detection result is used to characterize whether the electrode to be detected is normal. The anomaly detection model is trained using labeled battery cell images from the first training set and unlabeled battery cell images from the second training set as training samples, and using the labeling information corresponding to the labeled battery cell images and the first result corresponding to the unlabeled battery cell images as training labels. The first result was obtained in the following way: Input each unlabeled cell image from the second training set into the trained comparison model to obtain the first result of each unlabeled cell image output by the comparison model; The comparison model is trained using the labeled battery cell images in the first training set as training samples and the labeling information corresponding to the labeled battery cell images as training labels. Each labeled cell image includes a display area of ​​the electrode on the corresponding first sample cell, and the labeling information of each labeled cell image is used to characterize whether the electrode on the corresponding first sample cell is normal. Each unlabeled cell image includes a display area of ​​the electrode on the corresponding second sample cell, and the first result of each unlabeled cell image is used to characterize whether the electrode on the corresponding first sample cell is normal.

2. The method for detecting abnormal electrodes according to claim 1, wherein, The first training set was generated in the following way: Obtain at least one labeled first cell image and the corresponding labeling information for the first cell image; the first cell image is a cell image containing a display area of ​​normal electrodes or a display area of ​​abnormal electrodes; The first battery cell image is used as a training sample, and the annotation information corresponding to the first battery cell image is used as a training label to train the first model until convergence, which is then used as a screening model. Input unlabeled cell images into the screening model, and obtain the predicted labeling information of each unlabeled cell image output by the screening model and the confidence level corresponding to the predicted labeling information; Unlabeled cell images with confidence levels below the confidence threshold are used as second cell images. The second cell images are labeled, and the labeling information corresponding to each second cell image is obtained. Both the first cell image and the second cell image are used as labeled cell images to form the first training set.

3. The method for detecting abnormal electrodes according to claim 1, wherein, The training process of the comparison model includes: The second model is trained until convergence based on the labeled battery cell images in the first training set and the corresponding labeling information of the labeled battery cell images, and is used as the comparison model. One round of iterative training includes: Input each labeled cell image from the first training set into the second model to obtain the feature representation of each labeled cell image; For any two labeled cell images, determine the contrast loss between the two labeled cell images based on the feature representation of each pair of labeled cell images; The first comparison loss is determined based on the comparison loss between pairs of labeled cell images with identical labeling information. The second comparison loss is determined based on the comparison loss between pairs of labeled cell images with different labeling information. The parameters of the second model are updated based on the first contrast loss and the second contrast loss.

4. The method for detecting abnormal electrodes according to claim 1, wherein, The anomaly detection model includes: a feature extraction layer, a feature fusion layer, and a detection layer; The step of inputting the image to be detected into the anomaly detection model and obtaining the target detection result output by the anomaly detection model includes: The image to be detected is input into the feature extraction layer to obtain multiple target feature maps of different sizes output by the feature extraction layer; The multiple target feature maps of different sizes are input into the feature fusion layer for feature fusion to obtain multiple fused feature maps output by the feature fusion layer; The multiple target fusion feature maps of different sizes are input into the detection layer to obtain the target detection results output by the detection layer.

5. The method for detecting abnormal electrodes according to claim 4, wherein, The plurality of target feature maps of different sizes include a first target feature map, a second target feature map, a third target feature map, and a fourth target feature map; The step of inputting the image to be detected into the feature extraction layer to obtain multiple target feature maps of different sizes output by the feature extraction layer includes: The image to be detected is downsampled, and the downsampled feature map is segmented and stitched together to obtain the first target feature map. The first target feature map is downsampled, and the downsampled feature map is segmented and stitched together to obtain the second target feature map. The second target feature map is downsampled, and the downsampled feature map is segmented and stitched together to obtain the third target feature map. The third target feature map is downsampled, and the downsampled feature map is segmented and stitched together to obtain the fourth target feature map.

6. The method for detecting abnormal electrodes according to claim 5, wherein, After obtaining the fourth target feature map, the process further includes: The fourth target feature map is pooled by multiple pooling layers in series to obtain the pooled feature map output by each pooling layer. The pooling feature maps output by each pooling layer are concatenated to obtain the first concatenated feature map; By using multiple convolution kernels of preset sizes, features are extracted from the first spliced ​​feature map to obtain multiple first features of different sizes; The first features of different sizes are stitched together along the channel dimension to obtain the second stitched feature map; Average pooling is performed on the feature values ​​of each channel of the second concatenated feature map to obtain a first global description vector for the target feature map; the global description vector is used to characterize the global information of each channel of the target feature map; The first global description vector of the second concatenated feature map is input into the first fully connected layer to obtain a second global description vector of a preset dimension output by the first fully connected layer; the preset dimension is smaller than the channel dimension of the target feature. The second global description vector is input into the second fully connected layer to obtain the third global description vector; the number of neurons in the second fully connected layer is the same as the channel dimension of the fourth target feature map; The third global description vector of the channel dimension is input into the activation function to obtain the weights of each channel; For each channel, the feature values ​​of the channel are weighted based on the channel's weight to obtain weighted feature values; Based on the weighted feature values ​​in each channel, the fourth target feature map to be input into the feature fusion layer is obtained.

7. The method for detecting abnormal electrodes according to claim 5, wherein, The step of inputting the multiple target feature maps of different sizes into the feature fusion layer for feature fusion to obtain multiple target fused feature maps output by the feature fusion layer includes: The fourth target feature map is upsampled, and the upsampled result is concatenated with the third target feature map to obtain the first fused feature map; The first fused feature map is segmented and stitched together to obtain the stitched first fused feature map. The first fused feature map after splicing is upsampled, and the upsampled result is spliced ​​with the second target feature map to obtain the second fused feature map; The second fused feature map is segmented and stitched together to obtain the stitched second fused feature map. The concatenated second fused feature map is upsampled, and the upsampled result is concatenated with the first target feature map to obtain a third fused feature map; The third fused feature map is segmented and stitched together to obtain the first target fused feature map; Perform a convolution operation on the first target fusion feature, and then concatenate the convolution result with the concatenated second fusion feature map. Perform segmentation and concatenation operations on the concatenated result to obtain the second target fusion feature map. A convolution operation is performed on the second target fusion feature map, and the convolution result is concatenated with the first fusion feature map after concatenation. The concatenation result is then segmented and concatenated to obtain the third target fusion feature map. A convolution operation is performed on the third target fusion feature map, and the convolution result is concatenated with the fourth target feature map. The concatenated result is then segmented and concatenated to obtain the fourth target fusion feature map. Obtain the first target fusion feature, the second target fusion feature, the third target fusion feature, and the fourth target fusion feature output by the feature fusion layer.

8. The method for detecting abnormal electrodes according to claim 5 or 7, wherein, The segmentation and splicing operations include: For any feature map to be segmented and stitched, perform a convolution operation on the feature map to obtain the convolutional feature map; The convolutional feature map is segmented to obtain a first segmented feature image and a second segmented feature map. The first segmentation feature image is input into multiple concatenated bottleneck blocks for convolution operation to obtain the reference feature map output by each bottleneck block; The reference feature maps output by each bottleneck block and the second segmentation feature map are concatenated to obtain a feature map that has undergone segmentation and concatenation operations, corresponding to the feature map to be segmented and concatenated.

9. The method for detecting abnormal electrodes according to claim 4, wherein, The target detection result is the image to be detected or the image to be detected containing a detection box, wherein the detection box is used to characterize the abnormal area of ​​the electrode to be detected on the image to be detected; the detection layer includes: a detection module and a filtering module; When the target detection result is a detection image containing a detection box, the step of inputting the multiple target fusion feature maps of different sizes into the detection layer to obtain the target detection result output by the detection layer includes: For a target fusion feature map of any size, the target fusion feature map is input into the detection module to obtain at least one candidate detection box information output by the detection module; the candidate detection box information includes: the coordinates and confidence score of the candidate detection box; Based on the confidence level of each candidate detection box, the information of each candidate detection box is placed into the candidate detection box list in descending order of confidence level; Multiple rounds of iterative filtering are performed based on the information of each candidate detection box in the candidate detection box list until the candidate detection box list is empty or the maximum confidence in the candidate detection box list is less than a preset confidence threshold, and the image to be detected containing at least one detection box is output. One round of iterative screening includes: The candidate detection box with the highest confidence is determined from the candidate box list in this round, and the candidate detection box is used as the detection box and then deleted from the candidate detection box list. For the detection box and any candidate detection box in the candidate detection box list, the overlap area between the detection box and the candidate detection box is determined based on the coordinates of the detection box and the coordinates of the candidate detection box. For any candidate detection box, if the overlap area between the candidate detection box and the detection box is greater than the overlap threshold, the candidate detection box is deleted from the candidate box list, and the next round of candidate detection box list is obtained.

10. A detection device for abnormal electrodes, comprising: The determination module is configured to determine the image to be tested of the battery cell to be tested, wherein the image to be tested includes the display area of ​​the electrode to be tested on the battery cell to be tested; The input module is configured to input the image to be detected into the anomaly detection model and obtain the target detection result output by the anomaly detection model. The target detection result is used to characterize whether the electrode to be detected is normal. The anomaly detection model is trained using labeled battery cell images from the first training set and unlabeled battery cell images from the second training set as training samples, and using the labeling information corresponding to the labeled battery cell images and the first result corresponding to the unlabeled battery cell images as training labels. The first result was obtained in the following way: Input each unlabeled cell image from the second training set into the trained comparison model to obtain the first result of each unlabeled cell image output by the comparison model; The comparison model is trained using the labeled battery cell images in the first training set as training samples and the labeling information corresponding to the labeled battery cell images as training labels. Each labeled cell image includes a display area of ​​the electrode on the corresponding first sample cell, and the labeling information of each labeled cell image is used to characterize whether the electrode on the corresponding first sample cell is normal. Each unlabeled cell image includes a display area of ​​the electrode on the corresponding second sample cell, and the first result of each unlabeled cell image is used to characterize whether the electrode on the corresponding first sample cell is normal.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory, the processor executing the computer program to implement the steps of the method according to any one of claims 1-9.

12. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method according to any one of claims 1-9.

13. A computer program product comprising a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-9.

Images