Lithium ion battery ultrasonic image generation method and system based on generative adversarial network
By generating ultrasonic images of lithium-ion batteries using generative adversarial networks, the problems of limited longitudinal resolution and artifacts introduced by traditional interpolation algorithms in ultrasonic testing are solved, and high-precision three-dimensional reconstruction and defect detection of the internal structure of lithium-ion batteries are achieved.
Patent Information
- Application Number
- CN202610488796.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-14
- Publication Date
- 2026-07-10
AI Technical Summary
Existing ultrasonic testing and 3D reconstruction methods for lithium-ion batteries suffer from limited longitudinal resolution, interlayer breaks caused by sparse tomographic images, and the inability of traditional interpolation algorithms to accurately reconstruct the internal structure of the battery. Furthermore, traditional methods are prone to introducing artifacts in complex layered structures.
A generative adversarial network is used to generate intermediate layer images. Through feature extraction, deep location coding modulation and feature fusion modules, combined with adversarial training, intermediate layer images are generated, which solves the problems of tomographic image sequence expansion and 3D reconstruction.
It improves the quality of 3D reconstruction of the internal structure of lithium-ion batteries, reduces interlayer artifacts and fractures, enhances structural continuity and boundary clarity, and supports high-precision visualization and quantitative analysis of internal battery defects.
Smart Images
Figure CN122368326A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of battery detection and image processing technology, and more specifically, relates to a method and system for generating ultrasonic images of lithium-ion batteries based on generative adversarial networks. Background Technology
[0002] Lithium-ion batteries are widely used in electric vehicles and electronic products, and their safety and stability are crucial. During battery manufacturing and service, internal changes such as electrolyte wetting state, lithium metal deposition, gas generation, and electrode deformation are key indicators for assessing battery condition, but these often require battery disassembly for observation. While X-ray computed tomography (X-CT) can provide three-dimensional images, its high cost and long scanning time make rapid testing during production difficult. Currently, ultrasonic testing technology has become a popular non-destructive testing method for batteries due to its high sensitivity to mechanical modulus, dielectric density, and interface changes. This technology typically uses an ultrasonic transducer to scan the battery surface, acquiring a series of ultrasonic signals reflecting battery characteristics. Combined with reconstruction algorithms, it restores the acoustic parameter distribution inside the battery, thereby visualizing the three-dimensional structure of the battery's internal structure.
[0003] However, existing ultrasonic testing and 3D reconstruction methods still face many challenges in practical applications. On the one hand, since the ultrasonic attenuation rate increases with the increase of the emission frequency, the emission frequency of the ultrasonic transducer is constrained by the thickness of the battery being tested, resulting in limited longitudinal resolution of the transducer. Consequently, the number of tomographic images obtained is relatively sparse. Directly performing 3D reconstruction based on the limited slice images will cause obvious interlayer breaks, making it difficult to accurately present the structural evolution of the battery along the depth direction.
[0004] On the other hand, to alleviate the problem of interlayer sparsity, it is usually necessary to complete the image of the missing intermediate layers. Currently, the commonly used completion methods mostly employ traditional numerical methods such as linear interpolation or spline interpolation. However, the internal structure of a lithium-ion battery is a complex multilayer tightly stacked structure composed of positive and negative electrodes, separators, current collectors, etc. The propagation of sound waves between layers has highly nonlinear characteristics. Traditional interpolation algorithms lack prior understanding of the internal geometry of the battery and the laws of acoustic propagation. When the interlayer span is large, they cannot reproduce the subtle structural evolution between layers, which easily introduces severe interlayer artifacts into the completed image, resulting in a large difference between the completed image and the real tomographic image. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a method and system for generating ultrasonic images of lithium-ion batteries based on generative adversarial networks.
[0006] The present invention adopts the following technical solution.
[0007] A first aspect of the present invention provides a method for generating ultrasonic images of lithium-ion batteries based on generative adversarial networks, comprising: Two-dimensional ultrasonic tomographic images of lithium-ion batteries at multiple depth locations were acquired to form a tomographic image sequence; The two-dimensional ultrasound tomography images are preprocessed to obtain an enhanced tomography image sequence; The two adjacent images in the tomographic image sequence are input into a pre-trained generative adversarial network to generate an intermediate layer image located between the depths of the two adjacent images. The generated intermediate layer image is merged with the tomographic image sequence to obtain an extended tomographic image sequence; Three-dimensional reconstruction was performed based on extended tomographic image sequences to obtain three-dimensional volume data of the internal structure of lithium-ion batteries.
[0008] Optionally, acquiring two-dimensional ultrasonic tomographic images of the lithium-ion battery at multiple depth locations includes: The surface of the lithium-ion battery is scanned in two dimensions using an ultrasonic transducer to obtain ultrasonic reflection time-domain signals at multiple scanning locations. The tomography step size is determined based on the longitudinal resolution of the ultrasonic transducer. Based on the ultrasonic reflection time-domain signal and the tomographic step, tomographic imaging calculations are performed to reconstruct two-dimensional ultrasonic tomographic images corresponding to different depths. The two-dimensional ultrasonic tomographic images at each depth are then combined in depth order to form a tomographic image sequence.
[0009] Optionally, preprocessing the two-dimensional ultrasound tomography image includes: Denoising the two-dimensional ultrasound tomography images; Contrast enhancement processing was performed on the denoised two-dimensional ultrasound tomography images.
[0010] Optionally, the generative adversarial network includes a generator and a discriminator; The generator is used to generate intermediate layer images, and the discriminator is used to distinguish between real ultrasound tomography images and intermediate layer images generated by the generator.
[0011] Optionally, the generator includes a feature extraction module, a depth location coding modulation module, a feature fusion module, and an image generation module, wherein: The feature extraction module is used to perform multi-scale feature extraction on the two input images respectively to obtain the first feature set and the second feature set; The depth position coding modulation module is used to introduce depth position coding during feature extraction to modulate the features; The feature fusion module is used to perform weighted fusion of the modulated first feature set and the second feature set to obtain intermediate layer features; The image generation module is used to reconstruct the image based on the intermediate layer features and generate an intermediate layer image.
[0012] Optionally, the depth position coding modulation module introduces a position modulation term in each feature extraction convolutional processing stage. The position modulation term is constructed by depth layer position coding, a first learnable parameter, and a second learnable parameter to change the response of the feature map to the attenuation characteristics of ultrasound signals at different depths.
[0013] Optionally, the feature fusion module uses a learnable weight map generated based on an attention mechanism to perform weighted fusion of the modulated first feature set and the second feature set at various scales to generate intermediate layer features, wherein the learnable weight map represents the degree of correlation between the generated intermediate layer structure and the features of two adjacent images.
[0014] Optionally, the discriminator adopts a block-based discriminant network structure to divide the input image to be discriminated into regions and output the authenticity discrimination results of local regions, and generate a global authenticity discrimination result based on the authenticity discrimination results of each local region.
[0015] Optionally, the loss function of the generative adversarial network during the pre-training process includes an adversarial loss term, a reconstruction loss term, and a gradient consistency loss term; The reconstruction loss term is used to characterize the difference in pixel grayscale between the intermediate layer image and the real ultrasound tomography image; the gradient consistency loss term is used to characterize the degree of gradient change at the structural edges in the intermediate layer image.
[0016] A second aspect of the present invention provides a lithium-ion battery ultrasonic image generation system based on generative adversarial networks (GANs), used to implement the lithium-ion battery ultrasonic image generation method based on GANs described in the first aspect of the present invention, comprising: The module comprises a data acquisition module, a preprocessing module, an inter-layer image generation module, and a 3D reconstruction module, among which: The data acquisition module is used to acquire two-dimensional ultrasonic tomographic images of lithium-ion batteries at multiple depth locations, forming a tomographic image sequence. The preprocessing module is used to preprocess the two-dimensional ultrasound tomography images to obtain an enhanced tomography image sequence; The interlayer image generation module is equipped with a pre-trained generative adversarial network, which is used to input two adjacent images in the tomographic image sequence into the generative adversarial network, generate an intermediate layer image located between the depths of the two adjacent images, and merge the generated intermediate layer image with the tomographic image sequence to obtain an extended tomographic image sequence. The 3D reconstruction module is used to perform 3D reconstruction based on the extended tomographic image sequence to obtain 3D volume data of the internal structure of the lithium-ion battery.
[0017] Compared with the prior art, the beneficial effects of the present invention include at least the following: 1. This invention effectively reduces noise interference in the acquired two-dimensional ultrasonic tomographic images by performing denoising and contrast enhancement preprocessing, and significantly improves the grayscale recognition of electrodes, separators and defect areas inside the battery, thus solving the problems of low signal-to-noise ratio and blurred structural details in the original ultrasonic tomographic images.
[0018] 2. This invention constructs a generative adversarial network that includes a feature extraction module, a depth position coding modulation module, a feature fusion module, and an image generation module. This network generates intermediate layer images between adjacent tomographic images, thereby expanding the tomographic image sequence and solving the problems of sparse interlayer sampling and interlayer breaks in 3D reconstruction caused by the limited longitudinal resolution of the ultrasonic transducer.
[0019] 3. This invention enhances the structural continuity and boundary clarity of the generated intermediate layer image at different depths by introducing a depth position coding modulation term during the feature extraction process and combining it with a learnable weight map based on an attention mechanism to perform weighted fusion of multi-scale features. This solves the problem that traditional interpolation methods are prone to artifacts and blurring in complex layered structures.
[0020] 4. This invention constructs three-dimensional volumetric data of the internal structure of lithium-ion batteries by spatially stacking and voxelizing based on extended tomographic image sequences. This enables high-precision visualization and quantitative analysis of the spatial distribution of defects such as cracks, bubbles, and lithium plating inside the battery, thereby improving the reliability and accuracy of non-destructive testing. Attached Figure Description
[0021] Figure 1 This is a schematic flowchart of a lithium-ion battery ultrasonic image generation method based on generative adversarial networks provided in accordance with an embodiment of the present invention. Figure 2 This is a schematic diagram of the ultrasound image acquisition process provided according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the adversarial training process of a generative adversarial network according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the generator structure provided in accordance with an embodiment of the present invention; Figure 5 This is a schematic diagram comparing the image generated by the method provided in the embodiments of the present invention with that generated by the traditional linear interpolation method; Figure 6 The images shown are physical photos of the test battery and schematic diagrams of the three-dimensional reconstruction results provided in accordance with embodiments of the present invention. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.
[0023] In Example 1, this invention provides a method for generating ultrasonic images of lithium-ion batteries based on generative adversarial networks, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain two-dimensional ultrasonic tomographic images of the lithium-ion battery at multiple depth locations.
[0024] Preferably, step 1 includes: The surface of the lithium-ion battery is scanned in two dimensions using an ultrasonic transducer to obtain ultrasonic reflection time-domain signals at multiple scanning locations. The tomography step size is determined based on the longitudinal resolution of the ultrasonic transducer. Based on the ultrasonic reflection time-domain signal and the tomographic step, tomographic imaging calculations are performed to reconstruct two-dimensional ultrasonic tomographic images corresponding to different depths. The two-dimensional ultrasonic tomographic images at each depth are then combined in depth order to form a tomographic image sequence.
[0025] More preferably, such as Figure 2 As shown, the two-dimensional scanning uses a focused ultrasonic transducer that moves along an "S"-shaped trajectory on the battery surface to perform two-dimensional scanning, acquiring ultrasonic reflection time-domain signals at M×N locations across the battery. , m represents the total number of sampling points for each reflected signal. The sequence of scanned ultrasound reflected signals, arranged according to spatial location, is as follows:
[0026] The formula for calculating the number of sampling points m is:
[0027] in, t represents the sampling frequency of the ultrasonic reflection time-domain signal; t represents the sampling duration.
[0028] More preferably, the longitudinal resolution is determined based on the propagation speed of the ultrasonic wave in the battery and the duration of the ultrasonic pulse or the center frequency of the ultrasonic transducer.
[0029] Specifically, the tomographic step distance is calculated using the scanning ultrasound reflection signals obtained above. Tomographic analysis was performed on the battery to obtain cross-sectional images at different depths. The selected tomographic step size was numerically equal to the longitudinal resolution of the ultrasonic transducer. Vertical resolution The calculation formula is:
[0030] Where is the speed at which ultrasound propagates in the battery. The duration of the ultrasound pulse. The center frequency of the ultrasonic transducer is 2.5 MHz. In this embodiment, the sound velocity in the cell is approximately 1500 m / s, and the calculated longitudinal resolution is 0.3 mm. Therefore, the tomographic step size is... Set to 0.3mm.
[0031] Tomographic computational imaging was performed on each battery, and the resulting images were processed according to the tomographic step size. Combined into a tomographic image set These images represent cross-sections of the battery at different depths, providing a wealth of fundamental information about the battery's internal structure, material distribution, and potential defects such as bubbles, lithium dendrites, and cracks.
[0032] Step 2: Preprocess the two-dimensional ultrasound tomography image to obtain an enhanced tomography image sequence.
[0033] Preferably, the preprocessing includes: The two-dimensional ultrasound tomography image is denoised; the denoised image is then contrast-enhanced.
[0034] More preferably, the denoising process employs a Gaussian filter to denoise the image, thereby smoothing the image and suppressing noise.
[0035] Specifically, the Gaussian filter formula is as follows:
[0036] in, To input the original image, The image after noise reduction. This represents the Gaussian convolution kernel. This refers to the window size.
[0037] More preferably, the contrast enhancement process employs histogram equalization to enhance the image contrast of the battery region in the image, thereby highlighting the details of the battery's internal structure (such as electrodes, separators, defects, etc.).
[0038] Specifically, the grayscale values of different battery structure regions may be similar, leading to blurred details. Enhancing image contrast can help distinguish different areas of the battery, thereby improving image recognition.
[0039] Step 3: Input two adjacent images from the tomographic image sequence into a pre-trained generative adversarial network to generate an intermediate layer image located between the depths of the two adjacent images.
[0040] Preferably, the generative adversarial network includes a generator and a discriminator; The generator is used to generate intermediate layer images, and the discriminator is used to distinguish between real ultrasound tomography images and intermediate layer images generated by the generator.
[0041] More preferably, such as Figure 4 As shown, the generator includes: The feature extraction module is used to perform multi-scale feature extraction on the two input images respectively to obtain the first feature set and the second feature set; The depth position coding modulation module is used to introduce depth position coding during feature extraction to modulate the features; The feature fusion module is used to perform weighted fusion of the modulated first feature set and the second feature set to obtain intermediate layer features; The image generation module is used to reconstruct the image based on the intermediate layer features and generate an intermediate layer image.
[0042] Specifically, the feature extraction module employs a parameter-sharing convolutional neural network (CNN) to process two adjacent tomographic images. and Multi-scale feature extraction was performed separately to obtain the first feature set. Second feature set Features include local details (such as speckle) and global features (such as structural edges). Shared parameters avoid fusion errors caused by misalignment of feature distributions in different layers.
[0043] More preferably, the depth position coding modulation module introduces a position modulation term in each feature extraction convolutional processing stage. The position modulation term is constructed by depth layer position coding, a first learnable parameter, and a second learnable parameter to change the response of the feature map to the attenuation characteristics of ultrasound signals at different depths.
[0044] Specifically, the depth position encoding modulation module introduces a position modulation term in each convolutional block:
[0045] in, For depth-level bit encoding, As the first learnable parameter, As the second learnable parameter, this modulation term enables the feature extraction network to adapt to differences in ultrasonic attenuation at different depths.
[0046] More preferably, the feature fusion module uses a learnable weight map generated based on an attention mechanism to perform weighted fusion of the modulated first feature set and the second feature set at various scales to generate intermediate layer features, wherein the learnable weight map represents the degree of correlation between the generated intermediate layer structure and the features of two adjacent images.
[0047] Specifically, the feature fusion module performs weighted fusion of the modulated features at various scales using learnable weights to generate intermediate layer image features. The formula is:
[0048] Among them, the weighted graph Automatically generated by attention network This indicates the degree of similarity between the generated feature structure and the image features of the upper and lower layers, thereby improving the interlayer continuity of the structural features; Generate intermediate layer image features Perform deconvolution to generate intermediate layer images. .
[0049] More preferably, the generative adversarial network is optimized through adversarial training.
[0050] Specifically, such as Figure 3 As shown, in the adversarial training, the discriminator is used to distinguish between real and generated ultrasound tomography images. The input is a single real ultrasound tomography image or a generated inter-layer image. A Markov discriminator (PatchGAN) network structure is used to first divide the input image into regions, then perform local real / fake detection, and finally output an average global real / fake probability. .
[0051] The specific adversarial training process is as follows: (1) Update the discriminator parameters With fixed generator parameters, the generated image is The discriminator loss function is:
[0052] Backpropagation And only update parameters .
[0053] (2) Update generator parameters With the discriminator parameters fixed, calculate backwards again. And minimize the adversarial loss function.
[0054] The overall generation loss function is:
[0055] in, To reconstruct the loss term, ensure that the generated image is closer to the real image in grayscale; This is a gradient consistency loss term, used to maintain sharp image edges and avoid overly smooth transitions. , , These are the weighting coefficients.
[0056] (3) Backpropagation Iterative Optimization In the discriminator's output, a large gradient is generated with a low probability. This gradient is backpropagated to the generator, guiding it to correct image features and make the generated image closer to the real image. This parameter update phase is iterated repeatedly to achieve the optimization goal.
[0057] Step 4: Merge the generated intermediate layer image with the tomographic image sequence to obtain an extended tomographic image sequence.
[0058] Specifically, the original dataset containing N real ultrasonic tomographic images of lithium batteries... Step 3 expands it to include Extended dataset of images :
[0059] Step 5: Perform three-dimensional reconstruction based on the extended tomographic image sequence to obtain three-dimensional volume data of the battery's internal structure.
[0060] Preferably, step 5 includes: creating volume data based on image depth information, and smoothing the volume data.
[0061] Specifically, the formula for creating volume data V is:
[0062] in, .
[0063] To eliminate potential subtle brightness fluctuations or texture discontinuities between adjacent layers, a three-dimensional Gaussian filter is used for the three-dimensional volume. The three directions are convolved with Gaussian kernels for smoothing while preserving the overall structural shape. The formula is as follows:
[0064] in, The standard deviation is The three-dimensional Gaussian kernel.
[0065] This invention significantly improves the 3D reconstruction quality of ultrasonic tomographic images of lithium-ion batteries by constructing an intermediate layer image generation framework based on generative adversarial networks (GANs). During the tomographic image sequence expansion process, an interpolation strategy based on features from adjacent layers is employed to generate intermediate layers, effectively increasing the sampling density along the battery thickness direction and alleviating interlayer breaks and structural discontinuities caused by the limited number of layers. The generator incorporates an attention-weighted fusion mechanism and depth position encoding. Depth position encoding enhances the network's adaptability to the attenuation characteristics of ultrasonic signals at different depths by injecting layer information during feature extraction and modulating it using learnable parameters. Attention-weighted fusion dynamically generates fusion weights based on the correlation of input image features, thereby improving the coherence and boundary clarity of interlayer structural transitions. At the adversarial training level, a least-squares GAN is combined with a block-based discriminant network. Based on adversarial loss, reconstruction loss terms and gradient consistency loss terms are introduced to jointly constrain the generation process, improving the realism of the generated intermediate layer images while suppressing edge blurring caused by excessive smoothing, ensuring the sharpness of structural details in the generated images. Based on the extended tomographic image sequence, spatial stacking and voxelization are performed according to the depth information corresponding to each image to construct volume data characterizing the three-dimensional structure inside the battery, thereby supporting rapid detection and quantitative analysis of defects such as crack connectivity, bubble volume fraction, lithium plating or dendrite spatial distribution.
[0066] From a technical perspective, this invention firstly reduces noise interference in ultrasound images through Gaussian filtering denoising and contrast enhancement preprocessing, improving the grayscale recognition of electrodes, diaphragms, and defect areas, thus providing higher-quality input data for the subsequent generative network. Secondly, the combination of intermediate layer images generated by the generative adversarial network and the attention-weighted fusion mechanism makes the expanded tomographic sequence more structurally continuous, significantly reducing interlayer artifacts and discontinuities, and enhancing the overall structural fidelity of the image. Thirdly, even under unfavorable conditions such as low signal-to-noise ratio or large tomographic step size, this invention can still effectively improve the accuracy and completeness of 3D volumetric data through the reasonable generation of intermediate layer images. Finally, the reliability of defect detection and quality assessment based on high-quality 3D volumetric data is significantly improved, enabling more accurate identification and localization of potential defects inside the battery.
[0067] At the application level, the technical solution of this invention can be widely deployed in the production quality inspection, usage status assessment, and failure analysis of power batteries for new energy vehicles and lithium-ion batteries for consumer electronics. For example, in production line quality inspection scenarios, batteries after liquid injection or formation can be quickly sampled for inspection to identify manufacturing defects such as internal bubbles and electrode wrinkles; in retired battery screening scenarios, the deformation and aging degree of the battery's internal structure can be observed using 3D reconstruction results, providing a basis for the battery's secondary use and safety assessment.
[0068] It should be noted that the specific network structures and parameter settings described in the above embodiments are merely preferred embodiments of the present invention and are not the only limitation on the scope of protection. For example, in the selection of the feature extraction network, in addition to parameter-sharing convolutional neural networks, residual networks, U-Net structures, or Transformer architectures can also be used for feature extraction; in terms of feature fusion methods, depth position encoding can be linear or sinusoidal or cosine position encoding; in addition to attention-weighted fusion, fusion methods can also be achieved by convolution after channel concatenation, or even by using an average fusion strategy, although the latter may lead to image blurring, it is still a way to generate intermediate layer images; in terms of loss functions, in addition to least squares generative adversarial loss, the original generative adversarial network cross-entropy loss, Wasserstein generative adversarial loss, or hinge loss can all be used as alternatives to adversarial loss, and gradient consistency loss can also be replaced by perceptual loss or structural similarity loss to maintain the detail quality of the generated image. Any non-substantial modifications or equivalent substitutions made based on the spirit of the present invention should be covered within the scope of protection of the present invention.
[0069] To verify the effectiveness of the method of the present invention, the method described in Embodiment 1 of the present invention and the traditional linear interpolation method were used to generate intermediate layer images on the same ultrasonic tomography image sequence of a lithium-ion battery, respectively. The results were compared, for example... Figure 5 As shown. Among them, Figure 5 (a) is the intermediate layer image generated by the method of the present invention. Figure 5 (b) is a real ultrasound tomography image acquired at the same depth location. Figure 5 (c) is the corresponding intermediate layer image generated using the traditional linear interpolation method. The comparison shows that the intermediate layer image generated by the method of this invention is superior to the traditional linear interpolation method in terms of structural coherence and edge sharpness, and can more realistically reproduce the subtle evolution of the internal interlayer structure of the battery, effectively suppressing the generation of interlayer artifacts.
[0070] Figure 6 The image shows a physical photograph of the test battery and the rendered result of the 3D volume data obtained after 3D reconstruction using the method of this invention. Figure 6 (a) is a photograph of the actual lithium-ion battery to be tested. Figure 6(b) shows the three-dimensional visualization result of the battery's internal structure reconstructed based on the extended tomographic image sequence. The reconstruction results clearly distinguish the electrode stacking structure and the spatial distribution of potential defect areas inside the battery, verifying the beneficial effect of the method of the present invention in improving the accuracy of three-dimensional reconstruction and the reliability of defect detection.
[0071] In Embodiment 2, this invention provides a lithium-ion battery ultrasonic image generation system based on generative adversarial networks (GANs), used to implement the lithium-ion battery ultrasonic image generation method based on GANs described in Embodiment 1, comprising: The module comprises a data acquisition module, a preprocessing module, an inter-layer image generation module, and a 3D reconstruction module, among which: The data acquisition module is used to acquire two-dimensional ultrasonic tomographic images of lithium-ion batteries at multiple depth locations, forming a tomographic image sequence. The preprocessing module is used to preprocess the two-dimensional ultrasound tomography images to obtain an enhanced tomography image sequence; The interlayer image generation module is equipped with a pre-trained generative adversarial network, which is used to input two adjacent images in the tomographic image sequence into the generative adversarial network, generate an intermediate layer image located between the depths of the two adjacent images, and merge the generated intermediate layer image with the tomographic image sequence to obtain an extended tomographic image sequence. The 3D reconstruction module is used to perform 3D reconstruction based on the extended tomographic image sequence to obtain 3D volume data of the internal structure of the lithium-ion battery.
[0072] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.
[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.
Claims
1. A method for generating ultrasonic images of lithium-ion batteries based on generative adversarial networks, characterized in that, include: Two-dimensional ultrasonic tomographic images of lithium-ion batteries at multiple depth locations were acquired to form a tomographic image sequence; The two-dimensional ultrasound tomography images are preprocessed to obtain an enhanced tomography image sequence; The two adjacent images in the tomographic image sequence are input into a pre-trained generative adversarial network to generate an intermediate layer image located between the depths of the two adjacent images. The generated intermediate layer image is merged with the tomographic image sequence to obtain an extended tomographic image sequence; Three-dimensional reconstruction was performed based on extended tomographic image sequences to obtain three-dimensional volume data of the internal structure of lithium-ion batteries.
2. The method for generating ultrasonic images of lithium-ion batteries based on generative adversarial networks according to claim 1, characterized in that: The acquisition of two-dimensional ultrasonic tomographic images of the lithium-ion battery at multiple depth locations includes: The surface of the lithium-ion battery is scanned in two dimensions using an ultrasonic transducer to obtain ultrasonic reflection time-domain signals at multiple scanning locations. The tomography step size is determined based on the longitudinal resolution of the ultrasonic transducer. Based on the ultrasonic reflection time-domain signal and the tomographic step, tomographic imaging calculations are performed to reconstruct two-dimensional ultrasonic tomographic images corresponding to different depths. The two-dimensional ultrasonic tomographic images at each depth are then combined in depth order to form a tomographic image sequence.
3. The method for generating ultrasonic images of lithium-ion batteries based on generative adversarial networks according to claim 1, characterized in that: Preprocessing of the two-dimensional ultrasound tomography image includes: Denoising the two-dimensional ultrasound tomography images; Contrast enhancement processing was performed on the denoised two-dimensional ultrasound tomography images.
4. The method for generating ultrasonic images of lithium-ion batteries based on generative adversarial networks according to claim 1, characterized in that: The generative adversarial network includes a generator and a discriminator; The generator is used to generate intermediate layer images, and the discriminator is used to distinguish between real ultrasound tomography images and intermediate layer images generated by the generator.
5. The method for generating ultrasonic images of lithium-ion batteries based on generative adversarial networks according to claim 4, characterized in that: The generator includes a feature extraction module, a depth location coding modulation module, a feature fusion module, and an image generation module, wherein: The feature extraction module is used to perform multi-scale feature extraction on the two input images respectively to obtain the first feature set and the second feature set; The depth position coding modulation module is used to introduce depth position coding during feature extraction to modulate the features; The feature fusion module is used to perform weighted fusion of the modulated first feature set and the second feature set to obtain intermediate layer features; The image generation module is used to reconstruct the image based on the intermediate layer features and generate an intermediate layer image.
6. The method for generating ultrasonic images of lithium-ion batteries based on generative adversarial networks according to claim 5, characterized in that: The depth position coding modulation module introduces a position modulation term in each feature extraction convolutional processing stage. The position modulation term is constructed by depth layer position coding, a first learnable parameter, and a second learnable parameter to change the response of the feature map to the attenuation characteristics of ultrasound signals at different depths.
7. The method for generating ultrasonic images of lithium-ion batteries based on generative adversarial networks according to claim 5, characterized in that: The feature fusion module uses a learnable weight map generated based on an attention mechanism to perform weighted fusion of the modulated first feature set and the second feature set at various scales to generate intermediate layer features. The learnable weight map represents the degree of correlation between the generated intermediate layer structure and the features of two adjacent images.
8. The method for generating ultrasonic images of lithium-ion batteries based on generative adversarial networks according to claim 4, characterized in that: The discriminator adopts a block-based discriminant network structure, which is used to divide the input image to be discriminated into regions and output the authenticity discrimination results of local regions, and generate a global authenticity discrimination result based on the authenticity discrimination results of each local region.
9. The method for generating ultrasonic images of lithium-ion batteries based on generative adversarial networks according to claim 1, characterized in that: The loss function of the generative adversarial network during the pre-training process includes an adversarial loss term, a reconstruction loss term, and a gradient consistency loss term; The reconstruction loss term is used to characterize the difference in pixel grayscale between the intermediate layer image and the real ultrasound tomography image; the gradient consistency loss term is used to characterize the degree of gradient change at the structural edges in the intermediate layer image.
10. A lithium-ion battery ultrasonic image generation system based on generative adversarial networks, used to implement the lithium-ion battery ultrasonic image generation method based on generative adversarial networks as described in any one of claims 1-9, characterized in that, include: The module comprises a data acquisition module, a preprocessing module, an inter-layer image generation module, and a 3D reconstruction module, among which: The data acquisition module is used to acquire two-dimensional ultrasonic tomographic images of lithium-ion batteries at multiple depth locations, forming a tomographic image sequence. The preprocessing module is used to preprocess the two-dimensional ultrasound tomography images to obtain an enhanced tomography image sequence; The interlayer image generation module is equipped with a pre-trained generative adversarial network, which is used to input two adjacent images in the tomographic image sequence into the generative adversarial network, generate an intermediate layer image located between the depths of the two adjacent images, and merge the generated intermediate layer image with the tomographic image sequence to obtain an extended tomographic image sequence. The 3D reconstruction module is used to perform 3D reconstruction based on the extended tomographic image sequence to obtain 3D volume data of the internal structure of the lithium-ion battery.