Algorithm for fuzzy image recognition of battery cell pole group and pole piece

Abstract

The present invention provides an algorithm for identifying fuzzy images of pole pieces of a battery cell pole group, which is used to process image data obtained in the detection imaging of pole pieces of a battery cell pole group. The algorithm includes: determining the standard for the fuzzy imaging of the battery cell pole group and establishing a sample set of images with fuzzy imaging; processing the image data to produce a data set; using the image sample set as a training set, and performing computational processing on the data set based on a convolutional neural network to sort out battery cell pole groups with fuzzy imaging and normal ones. The algorithm for identifying fuzzy images of pole pieces of a battery cell pole group of the present invention pre-sets the standard for the fuzzy imaging of the battery cell pole group and establishes a sample set of fuzzy images, and uses a convolutional neural network model to sort out battery cell pole groups with fuzzy imaging, so as to facilitate manual secondary determination of whether the alignment of the pole pieces is compliant, thereby improving the situation in which the calculation accuracy of the pole piece alignment detection caused by the fuzzy imaging of the pole piece corners in the battery cell pole group detection is insufficient.

Description

Technical Field

[0001] The present invention relates to the technical field of battery cell detection, and in particular to an algorithm for recognizing fuzzy images of battery cell electrode groups and electrode sheets. Background Art

[0002] With the development of social economy and the improvement of people's living standards, people's environmental awareness has also increased. In order to reduce the environmental pollution caused by the use of traditional energy in vehicles, electric vehicles powered by lithium-ion batteries have begun to be widely used.

[0003] The power battery pack of electric vehicles is composed of a single battery, and a single battery is composed of several battery cell groups. During the production process, the battery cell group of each battery generally needs to be stacked. During the stacking process, it is necessary to detect whether the state of each electrode in the battery cell group meets the process standards. If it does not meet the specification requirements, it needs to be detected and scrapped.

[0004] At present, X-rays are mostly used to detect the folding of battery cell pole pieces. The four corners of the battery cell pole group are irradiated and imaged in a direction parallel to the pole piece plane to show the arrangement or folding of the corners of each pole piece in the battery cell. The pole piece image features are then analyzed through image analysis to determine whether they meet the process standards.

[0005] However, when the electrode group is delivered, the pole pieces in the electrode group cannot be guaranteed to be always perpendicular to the imaging device (CCD). If the tilt angle is too large, the projection of the tilted part will be blurred. Under this imaging state, the original alignment algorithm cannot accurately calculate the points. Therefore, the alignment calculation of the battery cell pole pieces is not reliable enough, making it impossible to effectively control the quality of the electrode group. Summary of the invention

[0006] In view of this, the present invention aims to propose an algorithm for recognizing fuzzy images of pole pieces of a battery cell pole group, so as to improve the insufficient accuracy of pole piece alignment detection calculation caused by fuzzy imaging of pole piece corners in battery cell pole group detection.

[0007] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0008] An algorithm for recognizing fuzzy images of pole pieces of a battery cell pole group is used to process image data obtained in the detection imaging of pole pieces of a battery cell pole group; the algorithm includes:

[0009] Determine the imaging blur standard of the battery cell electrode group and establish an image sample set with blurred imaging; process the image data to produce a data set; use the image sample set as a training set, and perform computational processing on the data set based on a convolutional neural network to sort the battery cell electrode groups with blurred imaging and those with normal imaging.

[0010] Furthermore, the image data is obtained by imaging the four corners of the battery cell electrode group respectively through the scheme of X-ray light + CCD + image intensifier.

[0011] Furthermore, the processing of the image data includes screening, labeling and cutting steps arranged in sequence; in the screening step, invalid images are eliminated and the images are rotated to make the directions of the images consistent; in the labeling step, the effective areas of the pole pieces and the ends of the pole pieces in the image are marked; in the cutting step, the original image is segmented to obtain the images of the effective areas and generate the data set.

[0012] Furthermore, in the labeling step, the LabelImg image labeling tool is used.

[0013] Furthermore, the convolutional neural network adopts the VGG16 network model and uses the pytorch framework for network construction.

[0014] Furthermore, the VGG16 network model only uses 13 convolutional layers of VGG16.

[0015] Furthermore, the classifier of the VGG16 network model is set to two fully connected layers: nn.Linear(512*7*7,256) and nn.Linear(256,2).

[0016] Furthermore, the convolutional neural network is provided with a model training and parameter adjustment link; the convolutional neural network is trained based on a training set, a test set composed of the data set, and a generated verification set, an optimizer is constructed, and parameters are updated according to the calculated gradients.

[0017] Furthermore, the algorithm also includes an evaluation step; in the evaluation step, the overkill rate and underkill rate indicators are set, and the convolutional neural network is optimized through training to achieve the set indicator requirements; wherein:

[0018] Overkill rate = (NG number determined by the algorithm - actual NG number) / total number of detections; underkill rate = 1 - (NG number determined correctly by the verification algorithm / total number of NG by manual visual inspection); NG refers to the battery cell groups with blurred images sorted out by the algorithm.

[0019] Furthermore, the set indicators are: the under-kill rate and the over-kill rate are both no more than 0.1%.

[0020] Compared with the prior art, the present invention has the following advantages:

[0021] The algorithm for recognizing blurred images of pole pieces of battery cell pole groups of the present invention pre-sets the standard for blurred imaging of battery cell pole groups, establishes a sample set of blurred images, and uses a convolutional neural network model to sort out battery cell pole groups with blurred imaging, so as to facilitate manual secondary determination of whether the alignment of the pole pieces is compliant. For battery cell pole groups with normal imaging, the algorithm can be used to directly determine whether the alignment of the pole pieces is compliant, thereby improving the situation in which the calculation accuracy of the pole piece alignment detection caused by blurred imaging of the pole piece corners in battery cell pole group detection is insufficient.

[0022] At the same time, the four corners of the battery cell electrode group are imaged separately using the X-ray light + CCD + image intensifier solution to obtain high-quality images, which is conducive to accurately judging the alignment and folding of the electrode pieces. The technology is stable and is conducive to the judgment and sorting of fuzzy images by convolutional neural networks.

[0023] In addition, through the processing of screening, labeling and cutting, the image specifications of the battery cell electrode group generated by imaging can be standardized, thereby forming a data set to be judged by the algorithm, which is conducive to quickly and accurately sorting out blurred and normal images, and facilitating the accurate judgment of the electrode alignment of normal images. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings, which constitute part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention. The directional terms such as front and back, up and down, etc. involved therein are only used to indicate relative positional relationships and do not constitute improper limitations on the present invention. In the accompanying drawings:

[0025] Figure 1 It is a schematic diagram of the overall flow of the algorithm for recognizing fuzzy images of electrode groups and electrode pieces of a battery cell according to an embodiment of the present invention;

[0026] Figure 2 This is an example of an image showing normal imaging of the electrode group and electrode piece of the battery cell according to the embodiment of the present invention;

[0027] Figure 3 This is an example of an image in which the electrode sheet of the electrode group of the battery cell according to the embodiment of the present invention is blurred;

[0028] Figure 4 An example of an image generated by annotation according to an embodiment of the present invention;

[0029] Figure 5 Another example of an image generated by annotation according to an embodiment of the present invention;

[0030] Figure 6 This is an example of the storage path and format of the training set and the validation set described in the embodiment of the present invention;

[0031] Figure 7This is an example of a normal image sorted out according to an embodiment of the present invention;

[0032] Figure 8 This is an example of a blurred image sorted out according to an embodiment of the present invention. DETAILED DESCRIPTION

[0033] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0034] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0035] The present invention relates to an algorithm for recognizing fuzzy images of pole pieces of a battery cell pole group, which is used for calculating and evaluating imaging data of pole piece corners in detection of pole piece alignment of a battery cell pole group, and is helpful for improving the situation where the calculation accuracy of pole piece alignment detection is insufficient due to fuzzy imaging of pole piece corners.

[0036] Overall, referring to Figure 1 As shown in the figure, for the image data obtained in the electrode detection imaging of the battery cell electrode group, the algorithm includes the following main steps:

[0037] 1. Determine the standard of image blur of the battery cell group and establish an image sample set with image blur;

[0038] 2. Process the image data and create a data set;

[0039] 3. Using the image sample set as the training set, the data set is computationally processed based on the convolutional neural network to sort out the battery cell groups with blurred imaging and normal imaging.

[0040] Based on the above general principles, the specific implementation content of each of the above steps is flexibly adjusted according to the process detection and quality control requirements. Below, this algorithm is explained in a preferred specific implementation method.

[0041] When performing image recognition of the battery cell pole group through this algorithm, the scanning imaging of the pole piece in the battery cell pole group should be completed first. The imaging of the pole piece can be achieved by using X-rays and supporting imaging equipment. In this embodiment, the image data of the pole piece is preferably obtained by imaging the four corners of the battery cell pole group respectively through the scheme of X-ray light + CCD + image intensifier. By using the scheme of X-ray light + CCD + image intensifier to image the four corners of the battery cell pole group respectively, high-quality images can be obtained, which is conducive to accurately judging the alignment, folding, etc. of the pole piece. The technology is stable and is conducive to the judgment and sorting of fuzzy images by convolutional neural networks.

[0042] In the specific implementation, we first need to clarify the requirements for electrode blur detection and the collection of sample sets to define the electrode imaging blur. When imaging the battery electrode group, if the internal positive and negative electrodes and the imaging device are tilted at an angle, it is very easy to cause local imaging blur and the imaging between the electrodes will blend, making it impossible to distinguish the electrodes. Examples of common normal images and blurred images are as follows: Figure 2 and Figure 3 shown.

[0043] After completing the collection of the above-mentioned image data, necessary preprocessing should be performed on the image data. In this embodiment, the processing steps include screening, marking and image cutting steps which are arranged in sequence.

[0044] Among them, the screening step is used to eliminate invalid images and rotate the image to make the direction of the image consistent. The annotation step marks the effective area of ​​the pole piece end in the image. In the image cutting step, the original image is first segmented to obtain the image of the effective area, and then a data set is generated. Through the processing of screening, annotation and image cutting, the image specification of the battery cell pole group generated by imaging can be standardized to form a data set to be determined by the algorithm, which is conducive to quickly and accurately sorting out blurred and normal images, and is conducive to accurately determining the alignment of the pole pieces in normal images.

[0045] In the above-mentioned labeling process, the LabelImg image labeling tool is preferably used. LabelImg image labeling tool is used to label the effective area of ​​the end of the pole piece in the image of the battery cell pole group, which is not only conducive to generating standard image data to be identified, but also has the characteristics of good operation stability and high processing efficiency.

[0046] In the specific operation, the selected valid images can be rotated using cv2.getRotationMatrix2D to ensure the consistent imaging direction. Then use the image annotation tool labelimg to annotate the images. Each image is annotated with the corresponding polar region, annotated as abnormal / normal, and the annotation text corresponding to the image is generated (jpg image and xml text). The annotated image is as follows Figure 4 and Figure 5 As shown in the figure, the effective area for fuzzy recognition is concentrated at the end of the pole piece, and the effective area is marked during marking.

[0047] Based on the situation that the fuzzy area in this technical solution is concentrated at the end of the pole piece, the image is cut corresponding to the effective area when the data set is produced, and saved in two folders of abnormal / normal by category. After that, glob.glob is used to obtain all XML text path information in the folder, ET.parse opens the corresponding XML document and parses it to obtain the xmin, xmax, ymin, and ymax data under the corresponding abnormal / normal node, and img[ymin:ymax,xmin:xmax] is further used to segment the original image and save it in the corresponding abnormal / normal folder.

[0048] After completing the above data set preparation, the following calculation and judgment based on the convolutional neural network model can be performed. Preferably, the convolutional neural network adopts the VGG16 network model, and the network is built using the pytorch framework. Among them, the VGG16 network model only uses the 13 convolutional layers of VGG16 (i.e., the feature extraction part); and the classifier of the VGG16 network model is set to two fully connected layers of nn.Linear(512*7*7,256) and nn.Linear(256,2).

[0049] Based on the above model algorithm, the battery cell pole groups with fuzzy molding and difficult to automatically determine the pole piece alignment can be accurately sorted out. Afterwards, by manually determining whether the pole piece alignment is compliant, the accuracy of the battery cell pole group alignment inspection and the purpose of improving product quality can be achieved.

[0050] To further improve the accuracy of the algorithm. In this embodiment, the convolutional neural network is provided with a model training and parameter adjustment link; the convolutional neural network is trained based on the training set, the test set composed of the data set, and the generated verification set, an optimizer is constructed, and the parameters are updated according to the calculated gradient. By setting the training and parameter adjustment and update links in the convolutional neural network model, the performance of the algorithm can be continuously optimized, thereby improving the accuracy of the algorithm's sorting judgment.

[0051] The specific settings in the algorithm can be referred to the following scheme:

[0052] First, classify the fuzzy and normal images into two types: abnormal and normal. Save the segmented images into abnormal and normal folders according to their categories. The training set and validation set formats are as follows: Figure 6 shown.

[0053] After that, set the image training size to 204*204; and load the data, using datasets.ImageFolder and torch.utils.data.DataLoader to load the training set and validation set respectively. At the same time, construct the optimizer object Optimizer, optimizer = optim.Adam (model.parameters(), lr = learning_rate), which is used to save the current state and update the parameters according to the calculated gradient.

[0054] Among them, the cross entropy loss function loss_func = nn.CrossEntropyLoss() is used to determine the closeness between the actual output and the expected output. The training model parameter setting can use the following function:

[0055] model(learning_rate=0.0002

[0056] batch_size=4

[0057] epochs=100 )

[0059] Among them, learning_rate is the learning rate, which determines whether the objective function can converge to the local minimum and when it converges to the minimum. Batch_size is the number of samples contained in each batch. Epoch is the process of training all training samples once.

[0060] Through the above model optimization, the algorithm can make the judgment more accurate. When predicting and outputting the algorithm, you can first set the detection target category to abnormal and normal; and use img.resize to convert the loaded image to 224*224 without distortion. After that, pass the image to the trained model; after obtaining the model's prediction result, use torch.max(out,1)[1] to obtain the corresponding index of the classification result, 0 for abnormal and 1 for normal. At this time, according to the index value, if it is 0, it is recorded as NG, and 1 is OK, and use cv2.putText to display it on the image.

[0061] Finally, the image classification results are saved locally (for data recall), and the result signal is transmitted to the receiving end through communication (such as the PLC / MES system in the equipment, which performs a selection function in the downstream channel and discharges the fuzzy pole group for manual secondary inspection). The normal image and fuzzy image finally sorted are as follows: Figure 7 and Figure 8 shown.

[0062] In addition, the algorithm of the present invention also includes an evaluation step. In the evaluation step, the overkill rate and underkill rate indicators are set, and the convolutional neural network is optimized through training to achieve the set indicator requirements; wherein:

[0063] Overkill rate = (number of NG determined by the algorithm - actual number of NG) / total number of tests;

[0064] Missed kill rate = 1-(number of NGs correctly determined by the verification algorithm / total number of NGs detected by manual visual inspection);

[0065] NG refers to the battery cell groups with blurred images sorted out by the algorithm.

[0066] Evaluation steps are set for the algorithm, and the overkill rate and underkill rate indicators required for the detection process are set. With the help of the algorithm's own settings and optimization strategies, a high-quality image recognition and judgment algorithm model can be formed, thereby achieving the expected goals of the algorithm, which can well meet the process detection requirements of battery cell pole group fuzzy image recognition and pole piece alignment judgment.

[0067] The specific setting indexes of the miss-kill rate and the over-kill rate can be set according to the process quality control requirements. Preferably, the setting index is: the miss-kill rate and the over-kill rate are both no greater than 0.1%. For example, the miss-kill rate is 0, and the over-kill rate is ≤ 0.1%.

[0068] To sum up, the algorithm for recognizing blurred images of pole pieces of battery cell pole groups in this embodiment pre-sets the standard for blurred imaging of battery cell pole groups and establishes a sample set of blurred images. It uses a convolutional neural network model to sort out battery cell pole groups with blurred imaging, so as to facilitate manual secondary determination of whether the alignment of the pole pieces is compliant. For battery cell pole groups with normal imaging, the algorithm can be used to directly determine whether the alignment of the pole pieces is compliant, thereby improving the insufficient accuracy of pole piece alignment detection calculation caused by blurred imaging of the pole piece corners in battery cell pole group detection.

[0069] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An algorithm for fuzzy image recognition of a cell electrode group pole piece, used for processing image data obtained in the detection imaging of a cell electrode group pole piece; characterized in that: The algorithm includes: Determine the imaging blur standard of the battery cell electrode group and establish an image sample set of imaging blur; the imaging blur standard is: when imaging the battery cell electrode group, due to the inclination of the internal positive and negative electrode sheets and the imaging device, the local imaging blur or the imaging blending between the electrode sheets causes the inability to distinguish the electrode sheets; Processing the image data to create a data set; Using the image sample set as a training set, the data set is computationally processed based on a convolutional neural network to sort out the battery cell electrode groups with blurred imaging and those with normal imaging; The convolutional neural network is provided with a model training and parameter adjustment link; the convolutional neural network is trained based on the training set, the test set composed of the data set, and the generated verification set, an optimizer is constructed, and parameters are updated according to the calculated gradient; The algorithm also includes an evaluation step, in which indicators of overkill rate and underkill rate are set, and the convolutional neural network is optimized through training to achieve the set indicator requirements.

2. The algorithm for recognizing fuzzy images of battery cell electrode groups and electrode pieces according to claim 1 is characterized by: The image data is obtained by imaging the four corners of the battery cell electrode group respectively through the scheme of X-ray light + CCD + image intensifier.

3. The algorithm for recognizing fuzzy images of battery cell electrode groups and electrode pieces according to claim 1 is characterized by: The processing of the image data includes screening, marking and cutting steps arranged in sequence; In the screening process, invalid images are eliminated, and the images are rotated to make the directions of the images consistent; In the marking step, the effective area of ​​the pole piece end in the image is marked; In the image cutting step, the original image is segmented to obtain an image of a valid area and generate the data set.

4. The algorithm for recognizing fuzzy images of battery cell electrode groups and electrode pieces according to claim 3 is characterized by: In the labeling process, the LabelImg image labeling tool is used.

5. The algorithm for recognizing fuzzy images of battery cell electrode groups and electrode pieces according to claim 1 is characterized by: The convolutional neural network adopts the VGG16 network model and uses the pytorch framework for network construction.

6. The algorithm for recognizing fuzzy images of battery cell electrode groups and electrode pieces according to claim 5 is characterized by: The VGG16 network model only uses 13 convolutional layers of VGG16.

7. The algorithm for recognizing fuzzy images of battery cell electrode groups and electrode pieces according to claim 5 is characterized by: The classifier of the VGG16 network model is set to two fully connected layers: nn.Linear(512*7*7, 256) and nn.Linear(256, 2).

8. The algorithm for recognizing fuzzy images of electrode groups and electrode pieces of a battery cell according to any one of claims 1 to 7, characterized in that: In the evaluation step: Overkill rate = (number of NG determined by the algorithm - actual number of NG) / total number of tests; Missed kill rate = 1-(number of NGs correctly determined by the verification algorithm / total number of NGs detected by manual visual inspection); NG refers to the battery cell groups with blurred images sorted out by the algorithm.

9. The algorithm for recognizing fuzzy images of battery cell electrode groups and electrode pieces according to claim 8, characterized in that: The set indicators are: the under-kill rate and over-kill rate are no more than 0.1%.

Images