Accurate lithium battery electrode positioning method based on multi-task two-dimensional point-line learning

Abstract

A precise lithium battery pole piece positioning method based on multi-task two-dimensional point-line learning first predicts the pole piece as a whole based on the point segmentation model, and then corrects the point prediction model through the line segmentation model. Among them, the line model guides the point model as input, and at the same time takes the low-level layer features in the encoder as input to ensure the effectiveness of edge extraction. In this way, the trend line level prediction of the positive and negative battery pole pieces of the lithium battery can be realized on the basis of the point-level prediction of the positive and negative battery pole pieces of the lithium battery, thereby improving the accuracy of the positioning of the lithium battery pole pieces, so as to improve the degree of automation and production efficiency of the lithium battery manufacturing process.

Description

Technical Field

[0001] The present application relates to the field of intelligent positioning technology, and more specifically, to a precise lithium battery pole piece positioning method based on multi-task two-dimensional point-line learning. Background Art

[0002] Lithium battery is a high-performance, high-energy-density battery widely used in electronic equipment, electric vehicles and other fields. In the production and quality inspection process of lithium batteries, the battery pole pieces need to be located and counted to ensure the quality and performance of the battery. The traditional lithium battery pole piece positioning method mainly relies on manual observation and manual division, which is inefficient and easily affected by subjective factors.

[0003] Therefore, an intelligent and precise lithium battery electrode positioning solution is desired. Summary of the invention

[0004] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a precise lithium battery pole piece positioning method based on multi-task two-dimensional point-line learning, which first predicts the pole piece as a whole based on the point segmentation model, and then corrects the point prediction model through the line segmentation model. Among them, the line model guides the point model as input, and at the same time takes the low-level layer features in the encoder as input, which ensures the effectiveness of edge extraction. In this way, on the basis of the point-level prediction of the positive and negative battery pole pieces of the lithium battery, the trend line level prediction of the positive and negative battery pole pieces of the lithium battery can be realized, thereby improving the accuracy of the positioning of the lithium battery pole pieces, so as to improve the degree of automation and production efficiency of the lithium battery manufacturing process.

[0005] In the first aspect, a method for accurately positioning lithium battery pole pieces based on multi-task two-dimensional point-line learning is provided, which includes:

[0006] Get X-ray images of lithium batteries;

[0007] Passing the lithium battery X-ray image through an encoder based on a deep residual network model to obtain a plurality of lithium battery X-ray feature maps, wherein the two lithium battery X-ray feature maps output by the first convolution layer and the second convolution layer of the encoder are a first shallow lithium battery X-ray feature map and a second shallow lithium battery X-ray feature map, wherein the output of the last convolution layer of the encoder is a deep lithium battery X-ray feature map;

[0008] Based on the skip-layer connection of the multiple lithium battery X-ray feature maps, the deep lithium battery X-ray feature map is input into a decoder including multiple deconvolution modules to obtain a decoding feature matrix;

[0009] Performing image semantic segmentation on the decoded feature matrix to obtain a positive prediction map and a negative prediction map; and

[0010] The positive electrode prediction map, the negative electrode prediction map, the first shallow lithium battery X-ray characteristic map and the second shallow lithium battery X-ray characteristic map are input into a line module to obtain an electrode line prediction result map.

[0011] In the above-mentioned precise lithium battery pole piece positioning method based on multi-task two-dimensional point and line learning, the positive electrode prediction map, the negative electrode prediction map, the first shallow lithium battery X-ray feature map and the second shallow lithium battery X-ray feature map are input into the line module to obtain the electrode line prediction result map, including: upsampling the second shallow lithium battery X-ray feature map to obtain the upsampled second shallow lithium battery X-ray feature map, wherein the upsampled second shallow lithium battery X-ray feature map has the same characteristic scale as the first shallow lithium battery X-ray feature map; fusing the upsampled second shallow lithium battery X-ray feature map with the first shallow lithium battery X-ray feature map to obtain a shallow fusion feature map. Figure; passing the shallow fusion feature map through the first point convolution layer to obtain a shallow feature matrix; fusing the positive prediction map and the shallow feature matrix to obtain a positive fusion prediction map, and fusing the negative prediction map and the shallow feature matrix to obtain a negative fusion prediction map; fusing the positive fusion prediction map and the shallow feature matrix to obtain a positive line fusion prediction map, and fusing the negative fusion prediction map and the shallow feature matrix to obtain a negative line fusion prediction map; passing the positive line fusion prediction map through the second point convolution layer and the Softmax classification function to obtain the positive line prediction result map; and, passing the negative line fusion prediction map through the third point convolution layer and the Softmax classification function to obtain the negative line prediction result map.

[0012] In the above-mentioned precise lithium battery pole piece positioning method based on multi-task two-dimensional point and line learning, the second shallow lithium battery X-ray feature map after upsampling is fused with the first shallow lithium battery X-ray feature map to obtain a shallow fusion feature map, including: fusing the second shallow lithium battery X-ray feature map after upsampling with the first shallow lithium battery X-ray feature map to obtain a shallow fusion feature map using the following feature fusion formula; wherein the feature fusion formula is:

[0013] F c =λF a +βF b

[0014] Among them, F c Represents the shallow fusion feature map, F a represents the X-ray characteristic image of the second shallow lithium battery after upsampling, Fb represents the first shallow lithium battery X-ray characteristic map, "+" represents the addition of the elements at the corresponding positions of the second shallow lithium battery X-ray characteristic map after upsampling and the first shallow lithium battery X-ray characteristic map, and λ and β represent weighted parameters for controlling the balance between the second shallow lithium battery X-ray characteristic map after upsampling and the first shallow lithium battery X-ray characteristic map.

[0015] In the above-mentioned precise lithium battery pole piece positioning method based on multi-task two-dimensional point and line learning, the encoder and the decoder have a symmetrical network structure.

[0016] In the above-mentioned precise lithium battery pole piece positioning method based on multi-task two-dimensional point and line learning, the encoder includes five convolutional layers, and the decoder includes five deconvolutional layers.

[0017] In the above-mentioned precise lithium battery pole piece positioning method based on multi-task two-dimensional point and line learning, the positive electrode prediction map and the shallow feature matrix are fused to obtain a positive electrode fusion prediction map, and the negative electrode prediction map and the shallow feature matrix are fused to obtain a negative electrode fusion prediction map, including: fusing the positive electrode prediction map and the shallow feature matrix with the following positive electrode fusion formula to obtain a positive electrode fusion prediction map; wherein the positive electrode fusion formula is:

[0018] F2=F1⊙M1

[0019] Wherein, F1 represents the positive electrode prediction map, M1 represents the shallow feature matrix, F2 represents the positive electrode fusion prediction map, ⊙ represents the point multiplication by position; and the negative electrode prediction map and the shallow feature matrix are fused with the following negative electrode fusion formula to obtain a negative electrode fusion prediction map; wherein, the negative electrode fusion formula is:

[0020] F4=F3⊙M1

[0021] Among them, F3 represents the negative electrode prediction map, M1 represents the shallow feature matrix, F4 represents the negative electrode fusion prediction map, and ⊙ represents point multiplication by position.

[0022] In the above-mentioned precise lithium battery pole piece positioning method based on multi-task two-dimensional point and line learning, the positive electrode fusion prediction map and the shallow feature matrix are fused to obtain a positive electrode line fusion prediction map, and the negative electrode fusion prediction map and the shallow feature matrix are fused to obtain a negative electrode line fusion prediction map, including: fusing the positive electrode fusion prediction map and the shallow feature matrix with the following positive electrode line fusion formula to obtain a positive electrode line fusion prediction map; wherein, the positive electrode line fusion formula is:

[0023] F5=αF2+γM1

[0024] Wherein, F5 represents the positive line fusion prediction map, F2 represents the positive fusion prediction map, M1 represents the shallow feature matrix, "+" represents the addition of the elements at the corresponding positions of the positive fusion prediction map and the shallow feature matrix, α and γ represent weighting parameters for controlling the balance between the positive fusion prediction map and the shallow feature matrix; and the negative line fusion prediction map and the shallow feature matrix are fused with the following negative line fusion formula to obtain the negative line fusion prediction map; wherein, the negative line fusion formula is:

[0025] F6=εF4+δM1

[0026] Among them, F6 represents the negative line fusion prediction map, F4 represents the negative fusion prediction map, M1 represents the shallow feature matrix, "+" represents the addition of the elements at corresponding positions of the negative fusion prediction map and the shallow feature matrix, and ε and δ represent weighted parameters for controlling the balance between the negative fusion prediction map and the shallow feature matrix.

[0027] In the above-mentioned precise lithium battery pole piece positioning method based on multi-task two-dimensional point and line learning, a training step is also included: for training the encoder and the decoder; wherein the training step includes: obtaining training data, the training data including training lithium battery X-ray images; passing the training lithium battery X-ray images through the encoder based on the deep residual network model to obtain a plurality of training lithium battery X-ray feature maps, wherein the two training lithium battery X-ray feature maps output by the first convolutional layer and the second convolutional layer of the encoder are the first training shallow lithium battery X-ray feature map and the second training shallow lithium battery X-ray feature map, wherein the output of the last convolution layer of the encoder is a training deep lithium battery X-ray feature map; based on the skip-layer connection of the multiple training lithium battery X-ray feature maps, the training deep lithium battery X-ray feature map is input into the decoder comprising multiple deconvolution modules to obtain a training decoding feature matrix; the probability distribution shift information compensation loss function value of the training lithium battery X-ray image matrix and the training decoding feature matrix is ​​calculated; and the encoder and the decoder are trained based on the probability distribution shift information compensation loss function value and through back propagation of gradient descent.

[0028] In the above-mentioned precise lithium battery pole piece positioning method based on multi-task two-dimensional point and line learning, the probability distribution shift information compensation loss function value of the training lithium battery X-ray image matrix and the training decoding feature matrix is ​​calculated, including: calculating the probability distribution shift information compensation loss function value of the training lithium battery X-ray image matrix and the training decoding feature matrix using the following optimization formula; wherein the optimization formula is:

[0029]

[0030]

[0031]

[0032]

[0033] Wherein, M1 and M2 are the training lithium battery X-ray image matrix and the training decoding feature matrix, respectively, softmax(·) represents the softmax function, bool(·) represents the bool function, log represents the logarithmic function with base 2, τ1 and τ2 are shift compensation hyperparameters, and α is a weighted hyperparameter, Represents the probability distribution shift information compensation loss function value.

[0034] Compared with the prior art, the precise lithium battery pole piece positioning method based on multi-task two-dimensional point-line learning provided by the present application first predicts the pole piece as a whole based on the point segmentation model, and then corrects the point prediction model through the line segmentation model. Among them, the line model guides the point model as input, and at the same time uses the low-level layer features in the encoder as input, which ensures the effectiveness of edge extraction. In this way, on the basis of the point-level prediction of the positive and negative battery pole pieces of the lithium battery, the trend line level prediction of the positive and negative battery pole pieces of the lithium battery can be realized, thereby improving the accuracy of the positioning of the lithium battery pole pieces, so as to improve the degree of automation and production efficiency of the lithium battery manufacturing process. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0036] Figure 1 It is a multi-task point segmentation and line segmentation network diagram according to an embodiment of the present application.

[0037] Figure 2 is a structural diagram of a line segmentation module according to an embodiment of the present application.

[0038] Figure 3 A schematic diagram of a scenario of a method for accurately positioning lithium battery pole pieces based on multi-task two-dimensional point and line learning according to an embodiment of the present application.

[0039] Figure 4This is a flow chart of a method for accurately positioning lithium battery pole pieces based on multi-task two-dimensional point and line learning according to an embodiment of the present application.

[0040] Figure 5 Schematic diagram of the architecture of a precise lithium battery pole piece positioning method based on multi-task two-dimensional point and line learning according to an embodiment of the present application.

[0041] Figure 6 It is a flowchart of the sub-steps of step 150 in the precise lithium battery pole piece positioning method based on multi-task two-dimensional point and line learning according to an embodiment of the present application.

[0042] Figure 7 It is a flowchart of the sub-steps of step 160 in the precise lithium battery pole piece positioning method based on multi-task two-dimensional point and line learning according to an embodiment of the present application.

[0043] Figure 8 It is a block diagram of a precise lithium battery pole piece positioning system based on multi-task two-dimensional point and line learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0044] The following will describe the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0045] Unless otherwise specified, all technical and scientific terms used in the embodiments of the present application have the same meaning as those commonly understood by those skilled in the art of the present application. The terms used in this application are only for the purpose of describing specific embodiments and are not intended to limit the scope of this application.

[0046] In the embodiments of the present application, it should be noted that, unless otherwise specified and limited, the term "connection" should be understood in a broad sense. For example, it can be an electrical connection or a connection between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0047] It should be noted that the terms "first\second\third" involved in the embodiments of the present application are only used to distinguish similar objects, and do not represent a specific order for the objects. It is understandable that the specific order or sequence of "first\second\third" can be interchanged where permitted. It should be understood that the objects distinguished by "first\second\third" can be interchanged where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein.

[0048] Compared with fuel vehicles, when new energy vehicles first appeared, the most important issue for car owners was battery life. If the mobile phone is out of power, it can be turned on and used by charging it with a charging cable; if the new energy vehicle is out of power, it will end on the road. Therefore, at the beginning of the emergence of new energy vehicles, how to improve battery life became the primary problem faced by various car companies (at that time, due to the small base of new energy vehicles, battery safety issues had not yet fully emerged). From the 100-200km battery life represented by the small ant in the early stage, to the 300km-400km battery life represented by the BYD Qin / BAIC EU series recently, to the 500km-700km battery life represented by Weilai EC6 and Xiaopeng P7 in the past two years - in order to solve the "battery life anxiety" of new energy vehicle users, the R&D departments of major new energy vehicle companies are really working hard. Under this trend, the "high nickel" ternary lithium battery with the core orientation of improving energy density (i.e. battery life) has undoubtedly become the mainstream of the mainstream technology route. At the end of 2016, BYD, once the main advocate of lithium iron phosphate battery technology, announced that it would add 6GWh of ternary lithium battery production capacity in 2017. This was definitely explosive news for the lithium battery industry at the time, because BYD's move seemed to foreshadow the full retreat of the lithium iron phosphate battery route. In sharp contrast, CATL, the leader of "ternary lithium batteries", began to fully embrace "high-nickel" ternary lithium batteries in 2018 and became the first power battery manufacturer in the world to mass-produce high-nickel NCM 811 lithium batteries at that time. It was also from this year that ternary lithium batteries began to become the absolute overlord in the field of new energy vehicle power batteries with overwhelming advantages.

[0049] The quality inspection requirements of lithium batteries generally include the number and location statistics of pole pieces. The existing inspection method relies on the images collected by X-ray equipment, and relies on some unknown image processing enhancements, followed by human observation and manual segmentation. With the development of deep learning, this provides an AI-level solution for the pole piece detection of lithium batteries. This paper innovatively proposes a two-dimensional joint optimization segmentation model to complete the accurate prediction of the pole piece position. Based on the point segmentation model, the pole piece is first predicted as a whole, and the line segmentation model can correct the point prediction model. Among them, the line model guides the point model as input, and at the same time takes the low-level layer features in the encoder as input, which ensures the effectiveness of edge extraction.

[0050] The technical problems to be solved by this application include: 1. Providing an AI-level point segmentation model; 2. Providing an AI-level line segmentation model; 3. Realizing multi-task joint learning; 4. The algorithm model is robust and outputs predicted points and contour trend lines at the same time.

[0051] As mentioned above, during the production and quality inspection of lithium batteries, the battery pole pieces need to be located and counted to ensure the quality and performance of the battery. Traditional lithium battery pole piece positioning methods mainly rely on manual observation and manual division, which are inefficient and easily affected by subjective factors. Therefore, an intelligent and accurate lithium battery pole piece positioning solution is desired.

[0052] Accordingly, it is considered that in the actual quality inspection process of lithium batteries, the number statistics of pole pieces and the position statistics of pole pieces are generally included. The existing detection method relies on X-ray equipment to collect X-ray images of lithium batteries, and after image enhancement processing, the enhanced images are observed by human eyes and manually divided. With the development of deep learning, an intelligent solution is provided for the pole piece detection of lithium batteries. Based on this, in the technical solution of the present application, a precise lithium battery pole piece positioning method based on multi-task two-dimensional point and line learning is proposed, which completes the precise pole piece position prediction by adopting a two-dimensional joint optimization segmentation model. Specifically, based on the point segmentation model, the pole piece is first predicted as a whole, and then the point prediction model is corrected by the line segmentation model. Among them, the line model guides the point model as input, and at the same time takes the low-level layer features in the encoder as input, which ensures the effectiveness of edge extraction.

[0053] The multi-task learning mechanism has been proven to be able to promote and improve the performance of two types of tasks at the same time. Therefore, in the technical solution of this application, the two related tasks designed are point segmentation and line segmentation. Figure 1 As shown, the entire network structure is composed of an encoder-decoder, where the encoder adopts ResNet-50, which has 5 layers. In the decoder, the original image resolution is gradually restored by progressively fusing the corresponding features in the encoder, so as to output the positive and negative pole predictions of the first task-pole points at the end of the decoder. Then, the positive and negative pole prediction maps are combined with the high-level features of the encoder and input into the counting module to obtain the counting statistics. In particular, here, the low-level has low-dimensional detail features, which can provide edge information about the poles in the image. As shown Figure 2 As shown in the figure, in the line segmentation module structure, the point segmentation map of the positive and negative poles is multiplied with the low-level feature segmentation of the encoder as an attention map guide to obtain the low-level enhanced features of the positive and negative pole regions, which are combined with the original low-level features as residual branches to obtain the final line prediction result map.

[0054] The technical effects produced by the present application include: 1. being able to complete point-level predictions of lithium battery pole pieces; 2. being able to complete trend-line-level predictions of lithium battery pole pieces.

[0055] Specifically, in the technical solution of the present application, first, a lithium battery X-ray image is obtained. Then, a convolutional neural network model with excellent performance in implicit feature extraction is used to perform feature mining on the lithium battery X-ray image. In particular, considering that when extracting the features of the lithium battery X-ray image, it is necessary not only to focus on the deep semantic features of the lithium battery pole piece in the lithium battery X-ray image, but also to focus on the shallow basic details and edge feature information of the lithium battery pole piece, which is conducive to the subsequent positioning of the lithium battery pole piece. Therefore, in the technical solution of the present application, the lithium battery X-ray image is passed through an encoder based on a deep residual network model to obtain a plurality of lithium battery X-ray feature maps, wherein the two lithium battery X-ray feature maps output by the first convolutional layer and the second convolutional layer of the encoder are the first shallow lithium battery X-ray feature map and the second shallow lithium battery X-ray feature map, and the output of the last convolutional layer of the encoder is a deep lithium battery X-ray feature map. In particular, here, the encoder includes five convolutional layers, and the first to fifth convolutional layers of the encoder are used to respectively extract the implicit feature information about the lithium battery pole piece at different depths in the lithium battery X-ray image. That is to say, when performing feature mining on the lithium battery X-ray image, while extracting the deep and rich semantic features about the lithium battery pole piece in the lithium battery X-ray image, the basic details and contours in the shallow layer are retained, which is conducive to the subsequent positioning of the lithium battery pole piece.

[0056] Then, in the decoding stage, based on the skip-layer connection of the multiple lithium battery X-ray feature maps, the deep lithium battery X-ray feature map is input into a decoder including multiple deconvolution modules to obtain a decoding feature matrix. In particular, here, the encoder and the decoder have a symmetrical structure. That is to say, a symmetrical U-shape form is adopted in the decoder, and the decoder includes five deconvolution layers, which gradually restore the original image resolution by progressively fusing the corresponding features in the encoder. In this way, the first task - the positive and negative pole prediction of the pole point can be output at the end of the decoder.

[0057] That is, specifically, after obtaining the decoding feature matrix, in order to locate the positive and negative electrodes of the lithium battery, it is necessary to perform image semantic segmentation on the decoding feature matrix, so as to perform corresponding mask operations after detecting the lithium battery electrode position area, so as to obtain the positive electrode prediction map and the negative electrode prediction map. In particular, in the technical solution of the present application, when locating the lithium battery electrode, in order to effectively distinguish the positive and negative electrodes, the output channels are designed to be 2, but all are completed in one decoder, which implies a contrast learning mechanism, which increases the discrimination power of the network. In this way, compared with the dual decoder, not only more parameters are saved, but also the accuracy is higher.

[0058] Furthermore, after obtaining the positive electrode prediction map and the negative electrode prediction map of the lithium battery, the positive and negative electrode prediction maps are jointly input into the line module with the low-level features of the encoder to obtain a trend line prediction. It should be understood that the shallow lithium battery X-ray feature map extracted by the encoder has low-dimensional detail features and can provide information about the basic edges of the lithium battery pole piece. In this way, the pole piece can be first predicted as a whole through the point segmentation model, and then the line model can be used to correct the output features of the point prediction model. In other words, the line model guides the point model as input, and at the same time uses the low-level layer features in the encoder as input, thereby ensuring the effectiveness of edge extraction.

[0059] Specifically, the positive electrode prediction map, the negative electrode prediction map, the first shallow lithium battery X-ray feature map and the second shallow lithium battery X-ray feature map are input into the line module to obtain the electrode line prediction result map. That is, the point segmentation map of the positive and negative electrodes, that is, the positive electrode prediction map and the negative electrode prediction map are respectively multiplied with the low-level feature segmentation of the encoder as the attention map guide, and the low-level enhanced features of the positive and negative electrode regions are obtained as residual branches combined with the original low-level features to obtain the final line prediction result map. In this way, on the basis of completing the point-level prediction of the positive and negative battery poles of the lithium battery, the trend line level prediction of the positive and negative battery poles of the lithium battery can be achieved, thereby improving the accuracy of the positioning of the lithium battery poles.

[0060] In particular, in the technical solution of the present application, here, for the decoding feature matrix obtained by the image codec based on the deep residual network model, it is expected that it can be consistent with the source image, that is, the lithium battery X-ray image, in the probability distribution expression of image semantics. In other words, it is expected that when the lithium battery X-ray image is encoded and decoded by the image codec based on the deep residual network model, the probability distribution expression of its image semantics will not encounter the degradation problem of the respective feature probability distribution expression caused by the shift of the probability distribution when it is back-propagated in the parameter space of the model, thereby enhancing the expression accuracy of the decoding feature matrix.

[0061] Based on this, the applicant of the present application further introduces a probability distribution shift information compensation loss function for the lithium battery X-ray image matrix, for example, recorded as M1, and the decoding feature matrix, for example, recorded as M2, which is expressed as:

[0062]

[0063]

[0064]

[0065]

[0066] Among them, τ1 and τ2 are shift compensation hyperparameters, and α is a weighting hyperparameter.

[0067] Here, the probability distribution expression values ​​obtained from the lithium battery X-ray image matrix M1 and the decoding feature matrix M2 respectively based on the Softmax function will follow the characteristics of the probability distribution for their respective feature distributions. The probability distribution shift information compensation loss function is used to compensate for the shift of the probability distribution of the feature representation of the lithium battery X-ray image matrix M1 and the decoding feature matrix M2, and the cross information entropy brought by the compensation is maximized through the bool function. In this way, the decoding feature matrix M2 can restore the probability distribution expression information of the lithium battery X-ray image matrix M1 to the greatest extent, thereby improving the expression accuracy of the decoding feature matrix. In this way, the prediction of the trend line level of the positive and negative battery poles of the lithium battery can be realized on the basis of the point-level prediction of the positive and negative battery poles of the lithium battery, thereby improving the accuracy of the positioning of the lithium battery poles, so as to improve the automation and production efficiency of the lithium battery manufacturing process.

[0068] Figure 3 Schematic diagram of a scenario of a method for accurately positioning lithium battery pole pieces based on multi-task two-dimensional point-line learning according to an embodiment of the present application. Figure 3 As shown, in this application scenario, first, obtain a lithium battery (for example, Figure 3 An X-ray image of M) (e.g., Figure 3 Then, the obtained lithium battery X-ray image is input to a server (for example, Figure 3 In S) shown in , the server is capable of processing the lithium battery X-ray image based on a precise lithium battery pole piece positioning algorithm based on multi-task two-dimensional point and line learning, so as to input the positive electrode prediction map, the negative electrode prediction map, the first shallow lithium battery X-ray feature map and the second shallow lithium battery X-ray feature map into a line module to obtain an electrode line prediction result map.

[0069] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0070] In one embodiment of the present application, Figure 4 Flow chart of a method for accurately positioning lithium battery pole pieces based on multi-task two-dimensional point-line learning according to an embodiment of the present application. Figure 4As shown, according to an embodiment of the present application, a precise lithium battery pole piece positioning method 100 based on multi-task two-dimensional point-line learning includes: 110, obtaining a lithium battery X-ray image; 120, passing the lithium battery X-ray image through an encoder based on a deep residual network model to obtain a plurality of lithium battery X-ray feature maps, wherein the two lithium battery X-ray feature maps output by the first convolution layer and the second convolution layer of the encoder are a first shallow lithium battery X-ray feature map and a second shallow lithium battery X-ray feature map, wherein the output of the last convolution layer of the encoder is The output is a deep lithium battery X-ray feature map; 130, based on the skip-layer connection of the multiple lithium battery X-ray feature maps, the deep lithium battery X-ray feature map is input into a decoder including multiple deconvolution modules to obtain a decoding feature matrix; 140, image semantic segmentation is performed on the decoding feature matrix to obtain a positive electrode prediction map and a negative electrode prediction map; and, 150, the positive electrode prediction map, the negative electrode prediction map, the first shallow lithium battery X-ray feature map and the second shallow lithium battery X-ray feature map are input into a line module to obtain an electrode line prediction result map.

[0071] Figure 5 Schematic diagram of the architecture of a precise lithium battery pole piece positioning method based on multi-task two-dimensional point-line learning according to an embodiment of the present application. Figure 5 As shown, in the network architecture, first, a lithium battery X-ray image is obtained; then, the lithium battery X-ray image is passed through an encoder based on a deep residual network model to obtain a plurality of lithium battery X-ray feature maps, wherein the two lithium battery X-ray feature maps output by the first convolutional layer and the second convolutional layer of the encoder are a first shallow lithium battery X-ray feature map and a second shallow lithium battery X-ray feature map, wherein the output of the last convolutional layer of the encoder is a deep lithium battery X-ray feature map; then, based on the skip-layer connection of the plurality of lithium battery X-ray feature maps, the deep lithium battery X-ray feature map is input into a decoder including a plurality of deconvolution modules to obtain a decoding feature matrix; then, image semantic segmentation is performed on the decoding feature matrix to obtain a positive electrode prediction map and a negative electrode prediction map; and, finally, the positive electrode prediction map, the negative electrode prediction map, the first shallow lithium battery X-ray feature map and the second shallow lithium battery X-ray feature map are input into a line module to obtain an electrode line prediction result map.

[0072] Specifically, in step 110, an X-ray image of a lithium battery is obtained. As described above, during the production and quality inspection of lithium batteries, the pole pieces of the battery need to be located and counted to ensure the quality and performance of the battery. Traditional lithium battery pole piece positioning methods mainly rely on manual observation and manual division, which are inefficient and easily affected by subjective factors. Therefore, an intelligent and precise lithium battery pole piece positioning solution is desired.

[0073] Accordingly, it is considered that in the actual quality inspection process of lithium batteries, the number statistics of pole pieces and the position statistics of pole pieces are generally included. The existing detection method relies on X-ray equipment to collect X-ray images of lithium batteries, and after image enhancement processing, the enhanced images are observed by human eyes and manually divided. With the development of deep learning, an intelligent solution is provided for the pole piece detection of lithium batteries. Based on this, in the technical solution of the present application, a precise lithium battery pole piece positioning method based on multi-task two-dimensional point and line learning is proposed, which completes the precise pole piece position prediction by adopting a two-dimensional joint optimization segmentation model. Specifically, based on the point segmentation model, the pole piece is first predicted as a whole, and then the point prediction model is corrected by the line segmentation model. Among them, the line model guides the point model as input, and at the same time takes the low-level layer features in the encoder as input, which ensures the effectiveness of edge extraction.

[0074] Specifically, in the technical solution of the present application, first, an X-ray image of a lithium battery is acquired.

[0075] Specifically, in step 120, the lithium battery X-ray image is passed through an encoder based on a deep residual network model to obtain a plurality of lithium battery X-ray feature maps, wherein the two lithium battery X-ray feature maps output by the first convolutional layer and the second convolutional layer of the encoder are a first shallow lithium battery X-ray feature map and a second shallow lithium battery X-ray feature map, wherein the output of the last convolutional layer of the encoder is a deep lithium battery X-ray feature map.

[0076] Next, a convolutional neural network model with excellent performance in implicit feature extraction is used to perform feature mining on the lithium battery X-ray image. In particular, considering that when extracting the features of the lithium battery X-ray image, it is necessary to focus not only on the deep semantic features of the lithium battery pole pieces in the lithium battery X-ray image, but also on the shallow basic details and edge feature information of the lithium battery pole pieces, which is conducive to the subsequent positioning of the lithium battery pole pieces.

[0077] Therefore, in the technical solution of the present application, the lithium battery X-ray image is passed through an encoder based on a deep residual network model to obtain multiple lithium battery X-ray feature maps, wherein the two lithium battery X-ray feature maps output by the first convolution layer and the second convolution layer of the encoder are the first shallow lithium battery X-ray feature map and the second shallow lithium battery X-ray feature map, and the output of the last convolution layer of the encoder is a deep lithium battery X-ray feature map. In particular, here, the encoder includes five convolution layers, and the first to fifth convolution layers of the encoder are used to respectively extract the implicit feature information about the lithium battery pole piece at different depths in the lithium battery X-ray image.

[0078] That is to say, when performing feature mining on the lithium battery X-ray image, while extracting the deep and rich semantic features of the lithium battery pole pieces in the lithium battery X-ray image, its basic details and contours at the shallow level are retained, which is beneficial for the subsequent positioning of the lithium battery pole pieces.

[0079] Convolutional Neural Network (CNN) is an artificial neural network that is widely used in image recognition and other fields. Convolutional neural networks can include input layers, hidden layers, and output layers, where the hidden layers can include convolutional layers, pooling layers, activation layers, and fully connected layers. The previous layer performs corresponding operations based on the input data and outputs the operation results to the next layer. The initial input data is processed through multiple layers of operations to obtain a final result.

[0080] The convolutional neural network model uses the convolution kernel as a feature filtering factor and has excellent performance in extracting local features of images. Compared with traditional image feature extraction algorithms based on statistics or feature engineering, the convolutional neural network model has stronger feature extraction generalization and fitting capabilities.

[0081] Specifically, in step 130, based on the skip-layer connection of the multiple lithium battery X-ray feature maps, the deep lithium battery X-ray feature map is input into a decoder including multiple deconvolution modules to obtain a decoding feature matrix. Then, in the decoding stage, based on the skip-layer connection of the multiple lithium battery X-ray feature maps, the deep lithium battery X-ray feature map is input into a decoder including multiple deconvolution modules to obtain a decoding feature matrix. In particular, here, the encoder and the decoder have a symmetrical structure. That is to say, a symmetrical U-shape form is adopted in the decoder, and the decoder includes five deconvolution layers, which gradually restore the original image resolution by progressively fusing the corresponding features in the encoder. In this way, the first task - the positive and negative pole prediction of the pole point can be output at the end of the decoder.

[0082] The automatic codec includes an encoder and a decoder, and the encoder contains two convolutional layers. In an example, the first convolution layer has 1 input channel, 2 output channels, 10 convolution kernel size, 10 sliding step, 10 zero padding width, and then a normalization layer and a ReLU nonlinear activation layer are set; the second convolution layer has 25 input channels, 50 output channels, 3 convolution kernel size, 3 sliding step, 0 zero padding width, and then a normalization layer and a ReLU nonlinear activation layer are set; the encoder has a fully connected layer at the tail end with 10 neurons; the decoder has a fully connected layer at the head end with 850 neurons, and then two deconvolution layers are connected. The first deconvolution layer has 50 input channels, 25 output channels, 4 convolution kernel size, 3 sliding step, 1 zero padding width, and then a normalization layer and a ReLU nonlinear activation layer are set; the second deconvolution layer has 25 input channels, 1 output channel, 1 convolution kernel size, 10 sliding step, 1 zero padding width, and then a Sigmoid nonlinear activation layer is set.

[0083] Specifically, in step 140, the decoding feature matrix is ​​subjected to image semantic segmentation to obtain a positive pole prediction map and a negative pole prediction map. That is, specifically, after obtaining the decoding feature matrix, in order to be able to locate the positive and negative poles of the lithium battery, it is necessary to perform image semantic segmentation on the decoding feature matrix, so as to perform corresponding mask operations after detecting the lithium battery pole position area, so as to obtain a positive pole prediction map and a negative pole prediction map. In particular, in the technical solution of the present application, when locating the lithium battery pole, in order to be able to effectively distinguish the positive and negative poles, the output channels are designed to be 2, but all are completed in one decoder, which implies a contrast learning mechanism, which increases the discrimination power of the network. In this way, compared with the dual decoder, not only more parameters are saved, but also the accuracy is higher.

[0084] Specifically, in step 150, the positive electrode prediction map, the negative electrode prediction map, the first shallow lithium battery X-ray feature map and the second shallow lithium battery X-ray feature map are input into the line module to obtain the electrode line prediction result map. Furthermore, after obtaining the positive electrode prediction map and the negative electrode prediction map of the lithium battery, the positive and negative electrode prediction maps are jointly input into the line module with the low-level features of the encoder to obtain the trend line prediction. It should be understood that the shallow lithium battery X-ray feature map extracted by the encoder has low-dimensional detail features and can provide information about the basic edges of the lithium battery pole piece. In this way, the pole piece can be first predicted as a whole through the point segmentation model, and then the line model can be used to correct the output features of the point prediction model. In other words, the line model guides the point model as input, and at the same time uses the low-level layer features in the encoder as input, thereby ensuring the effectiveness of edge extraction.

[0085] Specifically, the positive electrode prediction map, the negative electrode prediction map, the first shallow lithium battery X-ray feature map and the second shallow lithium battery X-ray feature map are input into the line module to obtain the electrode line prediction result map. That is, the point segmentation map of the positive and negative electrodes, that is, the positive electrode prediction map and the negative electrode prediction map are respectively multiplied with the low-level feature segmentation of the encoder as the attention map guide, and the low-level enhanced features of the positive and negative electrode regions are obtained as residual branches combined with the original low-level features to obtain the final line prediction result map. In this way, on the basis of completing the point-level prediction of the positive and negative battery poles of the lithium battery, the trend line level prediction of the positive and negative battery poles of the lithium battery can be achieved, thereby improving the accuracy of the positioning of the lithium battery poles.

[0086] Figure 6 Flow chart of the sub-steps of step 150 in the method for accurate lithium battery pole piece positioning based on multi-task two-dimensional point and line learning according to an embodiment of the present application, such as Figure 6As shown, the positive electrode prediction map, the negative electrode prediction map, the first shallow lithium battery X-ray feature map and the second shallow lithium battery X-ray feature map are input into the line module to obtain the electrode line prediction result map, including: 151, upsampling the second shallow lithium battery X-ray feature map to obtain the upsampled second shallow lithium battery X-ray feature map, wherein the upsampled second shallow lithium battery X-ray feature map and the first shallow lithium battery X-ray feature map have the same characteristic scale; 152, fusing the upsampled second shallow lithium battery X-ray feature map and the first shallow lithium battery X-ray feature map to obtain a shallow fusion feature map; 153, the shallow fusion feature map is passed through the A point convolution layer is used to obtain a shallow feature matrix; 154, the positive electrode prediction map and the shallow feature matrix are fused to obtain a positive electrode fusion prediction map, and the negative electrode prediction map and the shallow feature matrix are fused to obtain a negative electrode fusion prediction map; 155, the positive electrode fusion prediction map and the shallow feature matrix are fused to obtain a positive electrode line fusion prediction map, and the negative electrode fusion prediction map and the shallow feature matrix are fused to obtain a negative electrode line fusion prediction map; 156, the positive electrode line fusion prediction map is passed through a second point convolution layer and a Softmax classification function to obtain the positive electrode line prediction result map; and, 157, the negative electrode line fusion prediction map is passed through a third point convolution layer and the Softmax classification function to obtain the negative electrode line prediction result map.

[0087] Wherein, fusing the upsampled second shallow lithium battery X-ray feature map with the first shallow lithium battery X-ray feature map to obtain a shallow fusion feature map, including: fusing the upsampled second shallow lithium battery X-ray feature map with the first shallow lithium battery X-ray feature map to obtain a shallow fusion feature map using the following feature fusion formula; wherein the feature fusion formula is:

[0088] F c =λF a +βF b

[0089] Among them, F c Represents the shallow fusion feature map, F a represents the X-ray characteristic image of the second shallow lithium battery after upsampling, F b represents the first shallow lithium battery X-ray characteristic map, "+" represents the addition of the elements at the corresponding positions of the second shallow lithium battery X-ray characteristic map after upsampling and the first shallow lithium battery X-ray characteristic map, and λ and β represent weighted parameters for controlling the balance between the second shallow lithium battery X-ray characteristic map after upsampling and the first shallow lithium battery X-ray characteristic map.

[0090] Further, fusing the positive electrode prediction map and the shallow feature matrix to obtain a positive electrode fusion prediction map, and fusing the negative electrode prediction map and the shallow feature matrix to obtain a negative electrode fusion prediction map, including: fusing the positive electrode prediction map and the shallow feature matrix to obtain a positive electrode fusion prediction map using the following positive electrode fusion formula; wherein the positive electrode fusion formula is:

[0091] F2=F1⊙M1

[0092] Wherein, F1 represents the positive electrode prediction map, M1 represents the shallow feature matrix, F2 represents the positive electrode fusion prediction map, ⊙ represents the point multiplication by position; and the negative electrode prediction map and the shallow feature matrix are fused with the following negative electrode fusion formula to obtain a negative electrode fusion prediction map; wherein, the negative electrode fusion formula is:

[0093] F4=F3⊙M1

[0094] Among them, F3 represents the negative electrode prediction map, M1 represents the shallow feature matrix, F4 represents the negative electrode fusion prediction map, and ⊙ represents point multiplication by position.

[0095] Furthermore, the positive electrode fusion prediction map and the shallow feature matrix are fused to obtain a positive electrode line fusion prediction map, and the negative electrode fusion prediction map and the shallow feature matrix are fused to obtain a negative electrode line fusion prediction map, including: fusing the positive electrode fusion prediction map and the shallow feature matrix to obtain a positive electrode line fusion prediction map using the following positive electrode line fusion formula; wherein the positive electrode line fusion formula is:

[0096] F5=αF2+γM1

[0097] Wherein, F5 represents the positive line fusion prediction map, F2 represents the positive fusion prediction map, M1 represents the shallow feature matrix, "+" represents the addition of the elements at the corresponding positions of the positive fusion prediction map and the shallow feature matrix, α and γ represent weighting parameters for controlling the balance between the positive fusion prediction map and the shallow feature matrix; and the negative line fusion prediction map and the shallow feature matrix are fused with the following negative line fusion formula to obtain the negative line fusion prediction map; wherein, the negative line fusion formula is:

[0098] F6=εF4+δM1

[0099] Among them, F6 represents the negative line fusion prediction map, F4 represents the negative fusion prediction map, M1 represents the shallow feature matrix, "+" represents the addition of the elements at corresponding positions of the negative fusion prediction map and the shallow feature matrix, and ε and δ represent weighted parameters for controlling the balance between the negative fusion prediction map and the shallow feature matrix.

[0100] In one embodiment of the present application, the accurate lithium battery pole piece positioning method based on multi-task two-dimensional point-line learning further includes a training step: for training the encoder and the decoder; Figure 7 Flow chart of the sub-steps of step 160 in the method for accurate lithium battery pole piece positioning based on multi-task two-dimensional point and line learning according to an embodiment of the present application, such as Figure 7 As shown, the training step 160 includes: 161, obtaining training data, the training data including training lithium battery X-ray images; 162, passing the training lithium battery X-ray images through the encoder based on the deep residual network model to obtain multiple training lithium battery X-ray feature maps, wherein the two training lithium battery X-ray feature maps output by the first convolution layer and the second convolution layer of the encoder are the first training shallow lithium battery X-ray feature map and the second training shallow lithium battery X-ray feature map, wherein the output of the last convolution layer of the encoder is Output is a training deep lithium battery X-ray feature map; 163, based on the skip-layer connection of the multiple training lithium battery X-ray feature maps, the training deep lithium battery X-ray feature map is input into the decoder including multiple deconvolution modules to obtain a training decoding feature matrix; 164, calculate the probability distribution shift information compensation loss function value of the training lithium battery X-ray image matrix and the training decoding feature matrix; and, 165, based on the probability distribution shift information compensation loss function value and through gradient descent back propagation to train the encoder and the decoder.

[0101] In particular, in the technical solution of the present application, here, for the decoding feature matrix obtained by the image codec based on the deep residual network model, it is expected that it can be consistent with the source image, that is, the lithium battery X-ray image, in the probability distribution expression of image semantics. In other words, it is expected that when the lithium battery X-ray image is encoded and decoded by the image codec based on the deep residual network model, the probability distribution expression of its image semantics will not encounter the degradation problem of the respective feature probability distribution expression caused by the shift of the probability distribution when it is back-propagated in the parameter space of the model, thereby enhancing the expression accuracy of the decoding feature matrix.

[0102] Based on this, the applicant of the present application further introduces a probability distribution shift information compensation loss function for the lithium battery X-ray image matrix, for example, denoted as M1 and the decoding feature matrix, for example, denoted as M2, which is expressed as: the probability distribution shift information compensation loss function value of the training lithium battery X-ray image matrix and the training decoding feature matrix is ​​calculated by the following optimization formula; wherein the optimization formula is:

[0103]

[0104]

[0105]

[0106]

[0107] Wherein, M1 and M2 are the training lithium battery X-ray image matrix and the training decoding feature matrix, respectively, softmax(·) represents the softmax function, bool(·) represents the bool function, log represents the logarithmic function with base 2, τ1 and τ2 are shift compensation hyperparameters, and α is a weighted hyperparameter, Represents the probability distribution shift information compensation loss function value.

[0108] Here, the probability distribution expression values ​​obtained from the lithium battery X-ray image matrix M1 and the decoding feature matrix M2 respectively based on the Softmax function will follow the characteristics of the probability distribution for their respective feature distributions. The probability distribution shift information compensation loss function is used to compensate for the shift of the probability distribution of the feature representation of the lithium battery X-ray image matrix M1 and the decoding feature matrix M2, and the cross information entropy brought by the compensation is maximized through the bool function. In this way, the decoding feature matrix M2 can restore the probability distribution expression information of the lithium battery X-ray image matrix M1 to the greatest extent, thereby improving the expression accuracy of the decoding feature matrix. In this way, the prediction of the trend line level of the positive and negative battery poles of the lithium battery can be realized on the basis of the point-level prediction of the positive and negative battery poles of the lithium battery, thereby improving the accuracy of the positioning of the lithium battery poles, so as to improve the automation and production efficiency of the lithium battery manufacturing process.

[0109] In summary, a precise lithium battery pole piece positioning method 100 based on multi-task two-dimensional point and line learning based on the embodiment of the present application is explained, which first predicts the pole piece as a whole based on the point segmentation model, and then corrects the point prediction model through the line segmentation model. Among them, the line model guides the point model as input, and at the same time takes the low-level layer features in the encoder as input, which ensures the effectiveness of edge extraction. In this way, on the basis of the point-level prediction of the positive and negative battery pole pieces of the lithium battery, the trend line level prediction of the positive and negative battery pole pieces of the lithium battery can be realized, thereby improving the accuracy of the positioning of the lithium battery pole pieces, so as to improve the degree of automation and production efficiency of the lithium battery manufacturing process.

[0110] In one embodiment of the present application, Figure 8 : is a block diagram of a precise lithium battery pole piece positioning system based on multi-task two-dimensional point-line learning according to an embodiment of the present application. Figure 8As shown, according to the embodiment of the present application, the precise lithium battery pole piece positioning system 200 based on multi-task two-dimensional point and line learning includes: an image acquisition module 210, which is used to acquire a lithium battery X-ray image; an encoding module 220, which is used to pass the lithium battery X-ray image through an encoder based on a deep residual network model to obtain a plurality of lithium battery X-ray feature maps, wherein the two lithium battery X-ray feature maps output by the first convolution layer and the second convolution layer of the encoder are a first shallow lithium battery X-ray feature map and a second shallow lithium battery X-ray feature map, wherein the output of the last convolution layer of the encoder is a deep A deep lithium battery X-ray feature map; a decoding module 230, which is used to input the deep lithium battery X-ray feature map into a decoder including multiple deconvolution modules based on the skip-layer connection of the multiple lithium battery X-ray feature maps to obtain a decoding feature matrix; an image semantic segmentation module 240, which is used to perform image semantic segmentation on the decoding feature matrix to obtain a positive electrode prediction map and a negative electrode prediction map; and a line module 250, which is used to input the positive electrode prediction map, the negative electrode prediction map, the first shallow lithium battery X-ray feature map and the second shallow lithium battery X-ray feature map into a line module to obtain an electrode line prediction result map.

[0111] In a specific example, in the above-mentioned precise lithium battery pole piece positioning system based on multi-task two-dimensional point-line learning, the line module includes: a sampling unit, which is used to upsample the second shallow lithium battery X-ray feature map to obtain the upsampled second shallow lithium battery X-ray feature map, wherein the upsampled second shallow lithium battery X-ray feature map and the first shallow lithium battery X-ray feature map have the same characteristic scale; a shallow fusion unit, which is used to fuse the upsampled second shallow lithium battery X-ray feature map with the first shallow lithium battery X-ray feature map to obtain a shallow fusion feature map; a first convolution unit, which is used to pass the shallow fusion feature map through a first point convolution layer to obtain a shallow feature matrix; positive and negative electrode fusion A fusion unit is used to fuse the positive electrode prediction map and the shallow feature matrix to obtain a positive electrode fusion prediction map, and to fuse the negative electrode prediction map and the shallow feature matrix to obtain a negative electrode fusion prediction map; a positive and negative electrode line fusion unit is used to fuse the positive electrode fusion prediction map and the shallow feature matrix to obtain a positive electrode line fusion prediction map, and to fuse the negative electrode fusion prediction map and the shallow feature matrix to obtain a negative electrode line fusion prediction map; a second convolution unit is used to pass the positive electrode line fusion prediction map through a second point convolution layer and a Softmax classification function to obtain the positive electrode line prediction result map; and a third convolution unit is used to pass the negative electrode line fusion prediction map through a third point convolution layer and the Softmax classification function to obtain the negative electrode line prediction result map.

[0112] In a specific example, in the above-mentioned precise lithium battery pole piece positioning system based on multi-task two-dimensional point and line learning, the shallow fusion unit is used to: fuse the upsampled second shallow lithium battery X-ray feature map with the first shallow lithium battery X-ray feature map to obtain a shallow fusion feature map using the following feature fusion formula; wherein the feature fusion formula is:

[0113] F c =λF a +βF b

[0114] Among them, F c Represents the shallow fusion feature map, F a represents the X-ray characteristic image of the second shallow lithium battery after upsampling, F b represents the first shallow lithium battery X-ray characteristic map, "+" represents the addition of the elements at the corresponding positions of the second shallow lithium battery X-ray characteristic map after upsampling and the first shallow lithium battery X-ray characteristic map, and λ and β represent weighted parameters for controlling the balance between the second shallow lithium battery X-ray characteristic map after upsampling and the first shallow lithium battery X-ray characteristic map.

[0115] In a specific example, in the above-mentioned precise lithium battery pole piece positioning system based on multi-task two-dimensional point and line learning, the encoder and the decoder have a symmetrical network structure.

[0116] In a specific example, in the above-mentioned precise lithium battery pole piece positioning system based on multi-task two-dimensional point and line learning, the encoder includes five convolutional layers, and the decoder includes five deconvolutional layers.

[0117] In a specific example, in the above-mentioned precise lithium battery pole piece positioning system based on multi-task two-dimensional point and line learning, the positive and negative electrode fusion unit includes: a positive electrode fusion subunit, which is used to fuse the positive electrode prediction map and the shallow feature matrix with the following positive electrode fusion formula to obtain a positive electrode fusion prediction map; wherein, the positive electrode fusion formula is:

[0118] F2=F1⊙M1

[0119] Wherein, F1 represents the positive electrode prediction map, M1 represents the shallow feature matrix, F2 represents the positive electrode fusion prediction map, ⊙ represents the point multiplication by position; and, a negative electrode fusion subunit is used to fuse the negative electrode prediction map and the shallow feature matrix with the following negative electrode fusion formula to obtain a negative electrode fusion prediction map; wherein, the negative electrode fusion formula is:

[0120] F4=F3⊙M1

[0121] Among them, F3 represents the negative electrode prediction map, M1 represents the shallow feature matrix, F4 represents the negative electrode fusion prediction map, and ⊙ represents point multiplication by position.

[0122] In a specific example, in the above-mentioned precise lithium battery pole piece positioning system based on multi-task two-dimensional point and line learning, the positive and negative electrode line fusion unit includes: a positive electrode line fusion unit subunit, which is used to fuse the positive electrode fusion prediction map and the shallow feature matrix with the following positive electrode line fusion formula to obtain a positive electrode line fusion prediction map; wherein, the positive electrode line fusion formula is:

[0123] F5=αF2+γM1

[0124] Wherein, F5 represents the positive line fusion prediction map, F2 represents the positive fusion prediction map, M1 represents the shallow feature matrix, "+" represents the addition of the elements at the corresponding positions of the positive fusion prediction map and the shallow feature matrix, α and γ represent weighting parameters for controlling the balance between the positive fusion prediction map and the shallow feature matrix; and, a negative line fusion unit subunit, which is used to fuse the negative fusion prediction map and the shallow feature matrix with the following negative line fusion formula to obtain a negative line fusion prediction map; wherein, the negative line fusion formula is:

[0125] F6=εF4+δM1

[0126] Among them, F6 represents the negative line fusion prediction map, F4 represents the negative fusion prediction map, M1 represents the shallow feature matrix, "+" represents the addition of the elements at corresponding positions of the negative fusion prediction map and the shallow feature matrix, and ε and δ represent weighted parameters for controlling the balance between the negative fusion prediction map and the shallow feature matrix.

[0127] In a specific example, in the above-mentioned precise lithium battery pole piece positioning system based on multi-task two-dimensional point and line learning, a training module for training the encoder and the decoder is also included; wherein the training module includes: a training image acquisition unit for acquiring training data, wherein the training data includes a training lithium battery X-ray image; a training encoding unit for passing the training lithium battery X-ray image through the encoder based on the deep residual network model to obtain a plurality of training lithium battery X-ray feature maps, wherein the two training lithium battery X-ray feature maps output by the first convolution layer and the second convolution layer of the encoder are a first training shallow lithium battery X-ray feature map and a second training shallow lithium battery X -ray feature map, wherein the output of the last convolutional layer of the encoder is a training deep lithium battery X-ray feature map; a training decoding unit, which is used to input the training deep lithium battery X-ray feature map into the decoder including multiple deconvolution modules based on the skip-layer connection of the multiple training lithium battery X-ray feature maps to obtain a training decoding feature matrix; a compensation loss function value unit, which is used to calculate the probability distribution shift information compensation loss function value of the training lithium battery X-ray image matrix and the training decoding feature matrix; and a training unit, which is used to train the encoder and the decoder based on the probability distribution shift information compensation loss function value and through back propagation of gradient descent.

[0128] In a specific example, in the above-mentioned precise lithium battery pole piece positioning system based on multi-task two-dimensional point and line learning, the compensation loss function value unit is used to: calculate the probability distribution shift information compensation loss function value of the training lithium battery X-ray image matrix and the training decoding feature matrix using the following optimization formula; wherein the optimization formula is:

[0129]

[0130]

[0131]

[0132]

[0133] Wherein, M1 and M2 are the training lithium battery X-ray image matrix and the training decoding feature matrix, respectively, softmax(·) represents the softmax function, bool(·) represents the bool function, log represents the logarithmic function with base 2, τ1 and τ2 are shift compensation hyperparameters, and α is a weighted hyperparameter, Represents the probability distribution shift information compensation loss function value.

[0134] Here, those skilled in the art can understand that the specific functions and operations of each unit and module in the above-mentioned precise lithium battery pole piece positioning system based on multi-task two-dimensional point and line learning have been referred to above. Figures 1 to 7 It has been introduced in detail in the description of the accurate lithium battery pole piece positioning method based on multi-task two-dimensional point and line learning, and therefore, its repeated description will be omitted.

[0135] As described above, the precise lithium battery pole piece positioning system 200 based on multi-task two-dimensional point and line learning according to the embodiment of the present application can be implemented in various terminal devices, such as a server for precise lithium battery pole piece positioning based on multi-task two-dimensional point and line learning. In one example, the precise lithium battery pole piece positioning system 200 based on multi-task two-dimensional point and line learning according to the embodiment of the present application can be integrated into the terminal device as a software module and / or a hardware module. For example, the precise lithium battery pole piece positioning system 200 based on multi-task two-dimensional point and line learning can be a software module in the operating system of the terminal device, or it can be an application developed for the terminal device; of course, the precise lithium battery pole piece positioning system 200 based on multi-task two-dimensional point and line learning can also be one of the many hardware modules of the terminal device.

[0136] Alternatively, in another example, the precise lithium battery pole piece positioning system 200 based on multi-task two-dimensional point and line learning and the terminal device may also be separate devices, and the precise lithium battery pole piece positioning system 200 based on multi-task two-dimensional point and line learning may be connected to the terminal device via a wired and / or wireless network, and transmit interactive information in accordance with an agreed data format.

[0137] The present application also provides a computer program product, which includes instructions. When the instructions are executed, the device performs operations corresponding to the above method.

[0138] In one embodiment of the present application, a computer-readable storage medium is further provided, wherein the computer-readable storage medium stores a computer program for executing the above-mentioned method.

[0139] It should be understood that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware can be adopted. Moreover, the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code can be adopted.

[0140] The method, system, and computer program product of the embodiments of the present application are described by flowcharts and / or block diagrams. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 one or more processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0141] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 one or more processes and / or boxes Figure 1 A function specified in one or more boxes.

[0142] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 one or more processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0143] The basic principles of the present application are described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, effects, etc. mentioned in the present application are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. are required by each embodiment of the present application. In addition, the specific details disclosed above are only for the purpose of illustration and ease of understanding, not for limitation, and the above details do not limit the present application to being implemented by adopting the above specific details.

[0144] The block diagrams of the devices, apparatuses, equipment, and systems involved in this application are only illustrative examples and are not intended to require or imply that they must be connected, arranged, and configured in the manner shown in the block diagram. As will be appreciated by those skilled in the art, these devices, apparatuses, equipment, and systems can be connected, arranged, and configured in any manner. Words such as "including", "comprising", "having", etc. are open words, referring to "including but not limited to", and can be used interchangeably with them. The words "or" and "and" used here refer to the words "and / or" and can be used interchangeably with them, unless the context clearly indicates otherwise. The words "such as" used here refer to the phrase "such as but not limited to", and can be used interchangeably with them.

[0145] It should also be noted that in the apparatus, device and method of the present application, each component or each step can be decomposed and / or recombined. Such decomposition and / or recombination should be regarded as equivalent solutions of the present application.

[0146] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the present application. Therefore, the present application is not intended to be limited to the aspects shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

[0147] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements.

[0148] The above description has been given for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.

Claims

1. A precise lithium battery pole piece positioning method based on multi-task two-dimensional point-line learning, characterized in that: include: Get X-ray images of lithium batteries; Passing the lithium battery X-ray image through an encoder based on a deep residual network model to obtain a plurality of lithium battery X-ray feature maps, wherein the two lithium battery X-ray feature maps output by the first convolution layer and the second convolution layer of the encoder are a first shallow lithium battery X-ray feature map and a second shallow lithium battery X-ray feature map, wherein the output of the last convolution layer of the encoder is a deep lithium battery X-ray feature map; Based on the skip-layer connection of the multiple lithium battery X-ray feature maps, the deep lithium battery X-ray feature map is input into a decoder including multiple deconvolution modules to obtain a decoding feature matrix; Performing image semantic segmentation on the decoded feature matrix to obtain a positive prediction map and a negative prediction map; and Input the positive electrode prediction map, the negative electrode prediction map, the first shallow lithium battery X-ray characteristic map and the second shallow lithium battery X-ray characteristic map into a line module to obtain an electrode line prediction result map; The positive electrode prediction map, the negative electrode prediction map, the first shallow lithium battery X-ray characteristic map and the second shallow lithium battery X-ray characteristic map are input into the line module to obtain the electrode line prediction result map, including: Upsampling the second shallow lithium battery X-ray characteristic map to obtain an upsampled second shallow lithium battery X-ray characteristic map, wherein the upsampled second shallow lithium battery X-ray characteristic map and the first shallow lithium battery X-ray characteristic map have the same characteristic scale; Fusion of the upsampled second shallow lithium battery X-ray feature map with the first shallow lithium battery X-ray feature map to obtain a shallow fusion feature map; Passing the shallow fusion feature map through a first point convolution layer to obtain a shallow feature matrix; fusing the positive prediction map and the shallow feature matrix to obtain a positive fusion prediction map, and fusing the negative prediction map and the shallow feature matrix to obtain a negative fusion prediction map; fusing the positive electrode fusion prediction map and the shallow feature matrix to obtain a positive electrode line fusion prediction map, and fusing the negative electrode fusion prediction map and the shallow feature matrix to obtain a negative electrode line fusion prediction map; Passing the positive line fusion prediction map through a second point convolution layer and a Softmax classification function to obtain a positive line prediction result map; and The negative line fusion prediction map is passed through the third point convolution layer and the Softmax classification function to obtain a negative line prediction result map.

2. The accurate lithium battery pole piece positioning method based on multi-task two-dimensional point-line learning according to claim 1 is characterized in that: Fusion of the upsampled second shallow lithium battery X-ray feature map and the first shallow lithium battery X-ray feature map to obtain a shallow fusion feature map, including: fusing the upsampled second shallow lithium battery X-ray feature map and the first shallow lithium battery X-ray feature map to obtain a shallow fusion feature map using the following feature fusion formula; Among them, the feature fusion formula is: F c =λF a +βF b Among them, F c Represents the shallow fusion feature map, F a represents the X-ray characteristic image of the second shallow lithium battery after upsampling, F b represents the first shallow lithium battery X-ray characteristic map, "+" represents the addition of the elements at the corresponding positions of the second shallow lithium battery X-ray characteristic map after upsampling and the first shallow lithium battery X-ray characteristic map, and λ and β represent weighted parameters for controlling the balance between the second shallow lithium battery X-ray characteristic map after upsampling and the first shallow lithium battery X-ray characteristic map.

3. The accurate lithium battery pole piece positioning method based on multi-task two-dimensional point-line learning according to claim 2 is characterized in that: The encoder and the decoder have a symmetrical network structure.

4. The accurate lithium battery pole piece positioning method based on multi-task two-dimensional point-line learning according to claim 3 is characterized in that: The encoder includes five convolutional layers and the decoder includes five deconvolutional layers.

5. The accurate lithium battery pole piece positioning method based on multi-task two-dimensional point-line learning according to claim 4 is characterized in that: The positive prediction map and the shallow feature matrix are fused to obtain a positive fusion prediction map, and the negative prediction map and the shallow feature matrix are fused to obtain a negative fusion prediction map, including: The positive electrode prediction map and the shallow feature matrix are fused with the following positive electrode fusion formula to obtain a positive electrode fusion prediction map; Among them, the positive electrode fusion formula is: F2=F1⊙M1 Wherein, F1 represents the positive electrode prediction map, M1 represents the shallow feature matrix, F2 represents the positive electrode fusion prediction map, and ⊙ represents point multiplication by position; and The negative electrode prediction map and the shallow feature matrix are fused using the following negative electrode fusion formula to obtain a negative electrode fusion prediction map; Among them, the negative electrode fusion formula is: F4=F3⊙M1 Among them, F3 represents the negative electrode prediction map, M1 represents the shallow feature matrix, F4 represents the negative electrode fusion prediction map, and ⊙ represents point multiplication by position.

6. The accurate lithium battery pole piece positioning method based on multi-task two-dimensional point-line learning according to claim 5 is characterized in that: The positive electrode fusion prediction map and the shallow feature matrix are fused to obtain a positive electrode line fusion prediction map, and the negative electrode fusion prediction map and the shallow feature matrix are fused to obtain a negative electrode line fusion prediction map, including: The positive line fusion prediction map and the shallow feature matrix are fused using the following positive line fusion formula to obtain a positive line fusion prediction map; Among them, the positive line fusion formula is: F5=αF2+γM1 Wherein, F5 represents the positive line fusion prediction map, F2 represents the positive electrode fusion prediction map, M1 represents the shallow feature matrix, "+" represents the addition of the elements at the corresponding positions of the positive electrode fusion prediction map and the shallow feature matrix, and α and γ represent weighting parameters for controlling the balance between the positive electrode fusion prediction map and the shallow feature matrix; and The negative electrode fusion prediction map and the shallow feature matrix are fused using the following negative electrode line fusion formula to obtain a negative electrode line fusion prediction map; Among them, the negative line fusion formula is: F6=εF4+δM1 Among them, F6 represents the negative line fusion prediction map, F4 represents the negative fusion prediction map, M1 represents the shallow feature matrix, "+" represents the addition of the elements at corresponding positions of the negative fusion prediction map and the shallow feature matrix, and ε and δ represent weighted parameters for controlling the balance between the negative fusion prediction map and the shallow feature matrix.

7. The accurate lithium battery pole piece positioning method based on multi-task two-dimensional point-line learning according to claim 6 is characterized in that: Also included is a training step: used to train the encoder and the decoder; Wherein, the training step includes: Acquire training data, wherein the training data includes training lithium battery X-ray images; Passing the training lithium battery X-ray image through the encoder based on the deep residual network model to obtain a plurality of training lithium battery X-ray feature maps, wherein the two training lithium battery X-ray feature maps output by the first convolution layer and the second convolution layer of the encoder are the first training shallow lithium battery X-ray feature map and the second training shallow lithium battery X-ray feature map, wherein the output of the last convolution layer of the encoder is the training deep lithium battery X-ray feature map; Based on the skip-layer connection of the plurality of training lithium battery X-ray feature maps, the training deep lithium battery X-ray feature map is input into the decoder comprising a plurality of deconvolution modules to obtain a training decoding feature matrix; Calculating the probability distribution shift information compensation loss function value of the training lithium battery X-ray image matrix and the training decoding feature matrix; and The encoder and the decoder are trained by compensating the loss function value based on the probability distribution shift information and by back propagation of gradient descent.

8. The accurate lithium battery pole piece positioning method based on multi-task two-dimensional point-line learning according to claim 7 is characterized in that: Calculating the probability distribution shift information compensation loss function value of the training lithium battery X-ray image matrix and the training decoding feature matrix, including: The probability distribution shift information compensation loss function value of the training lithium battery X-ray image matrix and the training decoding feature matrix is ​​calculated using the following optimization formula; Wherein, the optimization formula is: Wherein, M1 and M2 are the training lithium battery X-ray image matrix and the training decoding feature matrix, respectively, softmax(·) represents the softmax function, bool(·) represents the bool function, log represents the logarithmic function with base 2, τ1 and τ2 are shift compensation hyperparameters, and α is a weighted hyperparameter, Represents the probability distribution shift information compensation loss function value.

Images