Lithium ion battery health prediction system and method based on microscopic image multi-scale and region adaptive fusion

Abstract

The invention discloses a lithium ion battery health prediction system and method based on microscopic image multi-scale and region adaptive fusion. The prediction system comprises a data acquisition module, a data preprocessing module, a feature extraction module, a multi-scale feature fusion module, a health state prediction module and a data enhancement and optimization module. The method overcomes the limitation based on traditional electric signal analysis, is suitable for various microscopic imaging technologies, enables the battery health state prediction to be directly evaluated based on the change of the internal microstructure of the electrode, and improves the prediction precision, applicability and physical interpretation. The method can be widely applied to battery life prediction, battery management system optimization and intelligent monitoring systems, and provides reliable technical support for battery safety management, performance optimization and industrial intelligence.

Description

Technical Field

[0001] The present invention relates to lithium-ion battery health management and prediction technology, and in particular to a lithium-ion battery health prediction system and method based on multi-scale and regional adaptive fusion of microscopic images. Background Art

[0002] With the development of electric vehicles, renewable energy storage systems, and high-performance portable devices, the life prediction and health management of lithium-ion batteries have become important research directions. Existing health status prediction methods mainly rely on external data such as voltage, current, internal resistance, and temperature, and use equivalent circuit modeling or data-driven machine learning methods for analysis. However, these data can only reflect changes in the external macroscopic state of the battery and cannot directly reflect the microstructural evolution of the internal battery materials, resulting in limited prediction accuracy under complex operating conditions. Especially under high temperature, fast charging, or long-term cycle aging conditions, the degradation mechanism of the battery is complex, and it is difficult to accurately predict the health status based solely on external signals. Therefore, finding a method that can deeply analyze the degradation mechanism of battery materials has become the key to improving the accuracy of health prediction.

[0003] In recent years, microscopic imaging techniques, such as scanning electron microscopy (SEM), transmission electron microscopy (TEM), scanning tunneling microscopy (STM), atomic force microscopy (AFM), and super-depth-of-field optical microscopy (SDFM), have provided high-resolution information for studying the microstructural evolution of battery materials. These imaging techniques can reveal key degradation features of electrode materials, such as crack propagation, particle fragmentation, active material shedding, interface evolution, and solid electrolyte interphase (SEI) thickening, making them important tools for studying the state of health of batteries. However, traditional microscopic image analysis methods rely primarily on manual observation or simple statistical analysis, such as edge detection and fractal analysis, which cannot form an automated and intelligent prediction framework, limiting their scalability in industrial applications. Furthermore, single-magnification image analysis has certain limitations. While low magnification can observe overall structural changes, it is difficult to resolve subtle local degradation features. While high magnification can reveal material details, it lacks global information. Therefore, single-magnification images cannot fully characterize the degradation process of electrodes, affecting the accuracy and stability of health prediction.

[0004] The contribution of the health status of different regions to the overall degradation of the battery is also uneven. There are significant differences in the aging rates and patterns of the core, transition and edge regions of the electrode, but existing methods find it difficult to dynamically optimize the influence weights of different regions, resulting in low prediction stability. In addition, the acquisition cost of microscopic imaging data is high, and there are few samples of certain battery degradation states, making it difficult for existing deep learning models to fully learn the health changes of the battery throughout its life cycle under data constraints, which in turn affects the generalization ability of the model. In addition, most prediction methods use a fixed model training method and lack an adaptive learning mechanism, making it difficult to adapt to different working environments and battery types, further affecting the prediction accuracy. At the same time, the high cost of data acquisition and privacy protection issues limit the widespread application of deep learning models. There is an urgent need for an efficient, accurate, and applicable health prediction method for different battery types and working environments. Summary of the Invention

[0005] The present invention proposes a lithium-ion battery health prediction system and method based on multi-scale and regional adaptive fusion of microscopic images, aiming to address the shortcomings of existing health prediction technologies in terms of accuracy, stability and interpretability.

[0006] 1. Technical Solution Overview

[0007] This paper provides a lithium-ion battery health prediction system and method based on multi-scale and regionally adaptive fusion of microscopic images. This system utilizes a variety of advanced technologies to improve the accuracy, stability, and interpretability of battery state-of-health (SoH) predictions. Based on microscopic image data, this method combines computer vision, deep learning, multi-scale feature fusion, regionally adaptive weighting, data augmentation, and federated learning to accurately model the battery aging process.

[0008] The system of the present invention consists of modules such as data acquisition, data processing, multi-scale feature extraction, multi-scale information fusion, regional adaptive weighting, health status prediction, data enhancement and intelligent generation, model optimization and dynamic learning. The data acquisition module uses a variety of microscopic imaging devices to obtain image data of different magnifications and different regions, and standardizes them through the data processing module to improve the consistency and stability of feature extraction. The multi-scale feature extraction module extracts key degradation features of battery materials through computer vision and deep learning technology, and combines physical modeling methods to enhance the interpretability of health prediction. The pyramid attention mechanism is used to fuse image information of different magnifications, so that macroscopic structure and microscopic detail features are effectively integrated. At the same time, a regional adaptive weighting strategy is adopted to optimize the contribution of different electrode regions in health prediction and improve the stability of prediction. In addition, the combination of deep learning and knowledge graph makes health status prediction not only data-driven but also physically reasonable. To solve the problem of data scarcity, the present invention uses methods such as CutMix, DiffusionModel, and Generative Adversarial Network (GAN) to generate synthetic data to improve the generalization ability of the model. Federated learning enables distributed model training, improving cross-device prediction capabilities while ensuring data privacy. This invention overcomes the limitations of traditional health prediction methods and builds an efficient, accurate, and widely applicable battery health prediction system that can be applied to new energy vehicles, grid energy storage, battery management systems, and other fields.

[0009] 2. System module design

[0010] The lithium-ion battery health prediction system, based on multi-scale and regional adaptive fusion of microscopic images, consists of multiple modules: a data acquisition module, a data processing module, a multi-scale feature extraction module, a multi-scale information fusion module, a regional adaptive weighting module, a health status prediction module, a data enhancement and intelligent generation module, and a model optimization and dynamic learning module. These modules work together to achieve efficient prediction and management of battery SoH.

[0011] 1) The data acquisition module is used to acquire microscopic imaging data of lithium-ion battery electrode materials, supporting imaging equipment such as SEM, TEM, AFM, STM, and SDFM. The acquired data covers images of different magnifications (e.g., 200×, 500×, 1000×, and 2000×) and different regions (core, transition, and edge), ensuring multi-scale and regional diversity of imaging data to comprehensively characterize the aging process of electrode materials.

[0012] 2) The data processing module standardizes the acquired microscopic images, including preprocessing operations such as denoising, normalization, contrast enhancement, adaptive filtering, and gamma correction, to optimize image quality and mitigate the impact of experimental conditions. Furthermore, this module employs regional adaptive filtering, optimized for different imaging regions, to enhance the stability of subsequent feature extraction.

[0013] 3) The multi-scale feature extraction module combines computer vision (CV) and deep learning (DL) to extract microscopic features of electrode materials. Computer vision methods are used to extract morphological features such as crack distribution, particle morphology, porosity, and oxide layer thickness. Deep learning methods employ convolutional neural networks (CNNs), visual transformers (ViTs), and graph neural networks (GNNs) for high-dimensional feature learning. Furthermore, in conjunction with electrochemical kinetic models, a correlation is established between microscopic image features and battery aging mechanisms, improving the model's interpretability.

[0014] 4) The multi-scale information fusion module employs a pyramid attention mechanism (PAM) to fuse features from images at different magnifications, effectively combining macroscopic structural information with microscopic detail information. Furthermore, an adaptive feature selection mechanism is employed to filter redundant information and optimize computational efficiency. Furthermore, the module incorporates spectroscopic data (such as XRD spectra and EDS spectra) to enhance the characterization capabilities of microscopic image data, enabling the prediction model to more comprehensively capture material degradation trends.

[0015] 5) The regional adaptive weighting module dynamically optimizes the weighting of different regions of the battery electrode (core, transition, and edge). Because different regions have different aging rates and patterns, a regional weighting strategy is adopted, using deep learning to calculate the weights of different regional characteristics to ensure the stability and accuracy of health prediction results.

[0016] 6) The health status prediction module uses deep learning-based regression methods, including multi-layer perceptrons (MLPs), long short-term memory networks (LSTMs), and transformers, to predict health status and optimize the loss function to minimize prediction error. Furthermore, this module incorporates knowledge graphs, combining material degradation mechanisms with data-driven prediction models, making prediction results more interpretable.

[0017] 7) Data enhancement and intelligent generation module, to solve the problem of high cost of obtaining microscopic image data, adopts CutMix, Diffusion Model,

[0018] GAN generates synthetic data to improve model training results. In addition, this module uses adversarial data augmentation (generating more challenging samples) and combines contrastive learning based on physical priors to optimize the robustness of deep learning feature extraction by comparing positive and negative samples.

[0019] 8) The Model Optimization and Dynamic Learning module utilizes federated learning to collaboratively train prediction models across diverse data sources, leveraging large-scale distributed data to optimize models while ensuring data privacy and security. Furthermore, this module employs Bayesian hyperparameter optimization and genetic algorithms to automatically adjust model parameters, improving prediction accuracy and model generalization. It also incorporates an adaptive learning rate adjustment strategy to optimize the training process and accelerate model convergence.

[0020] 3. Implementation steps

[0021] The present invention also provides a method for predicting the health status of a lithium-ion battery based on the above system, comprising the following steps:

[0022] Step 1: Data collection and preprocessing

[0023] The present invention first uses microscopic imaging equipment to collect image data of lithium-ion battery electrode materials to comprehensively characterize the microstructural changes of the materials at different aging stages. The imaging equipment used includes SEM, TEM, AFM, STM, and SDFM, capable of high-resolution imaging of the electrode materials' micromorphology, particle structure, crack evolution, and surface chemical composition. The acquired data covers different magnifications and electrode regions to ensure multi-scale coverage and regional integrity. In terms of magnification, low-magnification images (e.g., 200×) reveal the overall electrode structure, medium-magnification images (e.g., 500× and 1000×) analyze the distribution of particles, and high-magnification images (e.g., 2000×) reveal nanoscale cracks, pores, and interface features. In terms of spatial distribution, the acquired images cover the core, transition, and edge regions to capture differences in aging across different regions of the material. All acquired images are categorized and stored by magnification and region, and imaging parameters such as accelerating voltage, working distance, and contrast adjustment are recorded for subsequent analysis and data standardization.

[0024] Since the quality of the collected images is affected by the experimental conditions, data preprocessing is required before entering the feature extraction link to improve the consistency and usability of the images. The present invention adopts a series of standardization methods, including denoising, normalization, contrast enhancement and brightness adjustment, to eliminate noise interference in the imaging process and optimize the resolvability of the image. In the denoising link, adaptive filtering technology is used to adaptively adjust the denoising intensity by analyzing the pixel gradient distribution, remove high-frequency noise, and maintain the detailed information of the microstructure. In order to ensure the comparability of image data between different batches, the present invention performs normalization processing on all images and adopts the Min-Max normalization method to standardize the pixel values to the [0,1] interval to eliminate the influence of brightness fluctuations. The normalization formula is as follows:

[0025]

[0026] Where I(x, y) represents the grayscale value of the original image, min(I) and max(I) are the minimum and maximum pixel values of the image, respectively.

[0027] In addition, to enhance the structural information of the image, the present invention uses the contrast-limited adaptive histogram equalization (CLAHE) method to improve the recognizability of low-contrast areas, making microstructures such as cracks, grain boundaries, and pores clearer. CLAHE calculates the grayscale histogram through a local window and equalizes it to avoid over-enhancement or information loss that may be caused by global histogram equalization. At the same time, the present invention uses gamma correction technology to optimize the image brightness. The mathematical expression of gamma transform is as follows:

[0028] I gamma (x, y) = I norm (x, y) γ

[0029] γ is an adjustment parameter. A value greater than 1 enhances detail in dark areas of the image, while a value less than 1 reduces overexposure in bright areas. By enhancing contrast and optimizing brightness, the image achieves greater structural clarity and stability during the subsequent feature extraction phase.

[0030] In data preprocessing, the present invention also takes into account the noise distribution characteristics of the image and adopts an adaptive denoising strategy for different imaging conditions. When the image is acquired at a high acceleration voltage (>10kV), local noise may increase due to electron beam scattering. Therefore, in this case, the intensity of the Gaussian filter is appropriately increased to reduce the impact of noise. For images collected at a low acceleration voltage (<5kV), bilateral filtering is used to maintain the edge clarity of the microstructure while denoising. In addition, when the working distance is large, the image may be slightly out of focus. The present invention detects the out-of-focus area through Laplace transform and uses sharpening filtering to compensate to restore the detailed information of the microstructure.

[0031] Step 2: Multi-scale feature extraction

[0032] After completing image data acquisition and preprocessing, the present invention performs multi-scale feature extraction on the image data to fully analyze the micromorphological changes of lithium-ion battery electrode materials at different aging stages. The present invention uses a combined CV and DL approach to extract multiple types of features from the image data, including morphological features, texture features, high-dimensional deep learning features, and physical constraint features, to comprehensively describe the material's degradation pattern.

[0033] First, in the computer vision feature extraction phase, the present invention uses image segmentation to identify key structures such as cracks, pores, and oxide layers, and calculates their proportions, morphology, and distribution characteristics. Crack feature extraction is based on an edge detection algorithm, which uses the Canny algorithm to enhance image edges and fills the crack region with morphological closing operations to ensure complete crack identification. The crack proportion calculation formula is as follows:

[0034]

[0035] Among them, A crack is the pixel area of the crack area, A total is the total area of the image. In addition, the present invention detects the pore structure by Otsu threshold segmentation method and calculates the porosity:

[0036]

[0037] It is used to measure the pore distribution of electrode materials. This indicator is crucial for evaluating active material loss and electrolyte wettability. In order to further analyze the material microstructure, the present invention calculates the particle diameter distribution, uses the watershed segmentation algorithm to separate adjacent particles, and calculates the average particle diameter and variance:

[0038]

[0039] Among them, d i Represents the diameter of a single particle, and N is the total number of particles. This indicator is used to evaluate the uniformity and degradation trend of electrode material particles.

[0040] In terms of texture feature analysis, the present invention uses the gray level co-occurrence matrix (GLCM) to calculate image texture parameters such as contrast, correlation, uniformity, and energy to evaluate the roughness and structural changes of the material surface. The specific calculation formula is as follows:

[0041]

[0042] Here, P(i, j) is the joint distribution probability of adjacent pixels in the image, contrast (C) measures the local variation in pixel values, and entropy (H) describes texture complexity. These texture features can capture morphological changes in material surfaces, providing more dimensional information for health status prediction.

[0043] In addition to computer vision features, the present invention further uses deep learning methods to extract high-dimensional features. First, the pre-processed image is input into a pre-trained CNN or ViT model to extract deep representation features. The calculation formula for deep learning features is as follows:

[0044] F DL =f CNN(I) or F DL =f ViT (I)

[0045] Among them, f CNN and f ViT (I) represents the feature extraction network, F DL The 2048-dimensional high-dimensional feature vector is capable of capturing the complex texture information and global structural characteristics of electrode materials. Compared to traditional computer vision methods, deep learning features can automatically learn multi-level information from images, avoiding the limitations of manually set feature extraction rules.

[0046] To enhance the interpretability of the model, the present invention combines physical modeling methods to establish a connection between microscopic image features and electrochemical kinetic properties. For example, the present invention analyzes the relationship between the change in oxide layer thickness and the increase in battery internal resistance, and uses an oxide layer boundary recognition algorithm in the image to calculate the oxide layer thickness:

[0047]

[0048] Among them, h i is the oxide layer thickness at different pixel locations, and N is the total number of pixels. Subsequently, combined with Nyquist impedance spectroscopy data, a mathematical model was established between oxide layer thickness and battery health status to enhance the physical significance of microscopic image features.

[0049] Step 3: Multi-scale information fusion

[0050] After completing multiscale feature extraction, the present invention fuses the information contained in images of different magnifications to ensure that both macrostructural features and microscopic details are captured. The present invention employs PAM for multiscale feature fusion, a mechanism that dynamically adjusts the feature weights of images of different magnifications, enhancing the stability and accuracy of health status prediction. Furthermore, the present invention incorporates an adaptive feature selection mechanism to eliminate redundant information, improve computational efficiency, and enhance the physical characterization capabilities of microscopic image features by combining spectroscopic data (such as XRD spectra and EDS spectra).

[0051] In the method of the present invention, each battery sample contains images of different magnifications, such as 200×, 500×, 1000× and 2000×, and images of different magnifications correspond to information of different scales. Low-magnification images (200×) can provide overall structural information, such as crack extension, particle aggregation and macroscopic deformation, while high-magnification images (2000×) can reveal details such as microscopic cracks, pore structure and interface degradation. The present invention first standardizes the features of images of different magnifications so that they are calculated in the same feature space, and then uses the attention mechanism for weighted fusion. The attention calculation formula is as follows:

[0052]

[0053] Among them, Q, K, and V are query matrix, key value matrix, and value matrix respectively. k is the scaling factor used to stabilize gradient updates. By calculating the attention weights between images of different magnifications, we ensure that the most representative features are given higher weights during the fusion process. Finally, the multi-scale fusion feature is expressed as follows:

[0054]

[0055] Among them, ω m is the feature weight of different magnifications, F scale，m Through this mechanism, the present invention can achieve cross-scale information integration, so that the global information provided by low-magnification images and the local detail information provided by high-magnification images complement each other, thereby improving the performance of the prediction model.

[0056] In addition, the present invention uses an adaptive feature selection mechanism to screen important features based on their information contribution while eliminating redundant information to optimize computational efficiency. The specific method is to calculate the variance and information entropy of each feature and perform dimensionality reduction on features with low information contribution. For a given feature set F, the variance information of the features is first calculated:

[0057]

[0058] Among them, μ F is the mean of the feature, Reflects the distribution range of the feature. Then, calculate the entropy value of the feature:

[0059]

[0060] Among them, P(F i ) is the probability distribution of feature values. If the entropy value of a feature is too low, it indicates that its contribution to the overall prediction is limited and can be considered for removal. Ultimately, only high-information features are retained and entered into the health status prediction model.

[0061] In order to enhance the physical interpretability of microscopic image features, the present invention further combines spectroscopic data, such as XRD spectrum and EDS spectrum, to combine chemical composition information with morphological features. For example, the XRD spectrum can be used to determine the crystal structure changes of the material, and can be combined with the crack distribution in the image to evaluate the degree of structural damage. Assuming that the XRD peak intensity I XRD To reflect the crystallinity of the material, the following correlation model can be defined:

[0062] F combined =λ1F multi-scale +λ2IXRD

[0063] Among them, λ1 and λ2 are fusion weight parameters. In this way, health prediction not only relies on the morphological characteristics of microscopic images, but also combines the information of the intrinsic structural changes of the material, improving the scientific rationality of the prediction.

[0064] Step 4: Region Adaptive Weighting

[0065] After completing the multi-scale information fusion, the present invention further optimizes the regional contribution of the imaging data, and enhances the role of different electrode regions in health prediction through a regional adaptive weighting mechanism to ensure the stability and accuracy of the prediction. Since different regions of the electrode (core area, transition area, edge area) show different degradation patterns during the aging process, traditional methods usually use fixed weights to process all regions and cannot flexibly respond to regional differences. The present invention adopts a regional dynamic weighting strategy based on deep learning to adaptively adjust the feature contribution of different regions, making the health status prediction more accurate and stable.

[0066] In the method of the present invention, features are first extracted from the core region, transition region, and edge region, and regional normalization is performed to ensure that data from different regions are calculated at the same scale. Then, deep learning is used to calculate the weights of the features of each region, so that the prediction model can automatically adjust the importance of the region according to the specific conditions of different battery samples. For each region r, the regional feature representation is defined as follows:

[0067]

[0068] Among them, F s is the feature vector of a single sampling point in the region, ω s is the adaptive weight of the sampling point, n r is the total number of sampling points in the area.

[0069] The present invention further adopts the attention mechanism to calculate the global contribution of different regions. First, the features of each region are mapped to a shared low-dimensional feature space, and then the attention weights of different regions are calculated:

[0070]

[0071] Among them, ω r Represents the learning weight of region r. MLP is a multi-layer perceptron, which is used to calculate the importance of different regions and normalize them through Softmax so that the weights of all regions meet the following requirements:

[0072]

[0073] This method can automatically adjust the weight of each region according to the regional characteristics of different battery samples, thereby ensuring that the most representative regional information is used first in the health prediction process, while regions with less information or greater noise contribute less.

[0074] In addition, the present invention optimizes the regional weighting strategy based on the physical significance of different regions. The core area usually contains relatively stable structural information, and its decay rate is relatively slow in long-term cycles, so it often provides more reliable feature information in health prediction. The present invention assigns a higher default weight to the core area in the initial stage, and dynamically adjusts the importance of this area during the model training process. On the contrary, the edge area is affected by factors such as the current collector and electrolyte infiltration, and its microscopic morphology changes greatly, which may lead to greater prediction fluctuations. Therefore, in the regional weighting process, the present invention adopts an outlier detection method to eliminate data points that may be greatly affected by environmental factors.

[0075] In the specific implementation process, the present invention further combines physical modeling methods to optimize the health status contribution of different areas. For example, the edge area often has phenomena such as oxide layer thickening and crack expansion. The present invention calculates the oxide layer thickness:

[0076]

[0077] Among them, h i is the oxide layer thickness at different pixel locations, and N is the total number of pixels. Then, combined with electrochemical impedance spectroscopy data, the regional weighting parameters are optimized so that the regional adaptive weights are not only based on deep learning calculations but also constrained by physical parameters, improving the scientific rationality of the prediction results.

[0078] Ultimately, the method of this invention dynamically adjusts the characteristic contributions of different electrode regions, enabling the health prediction model to adaptively adjust the influence weight of each region based on the specific conditions of different battery samples. By optimizing the utilization of regional information through the attention mechanism and combining it with physical modeling to enhance interpretability, the regional adaptive weighting strategy of this invention effectively improves the accuracy and stability of predictions, providing more reasonable and efficient input data for subsequent health status predictions.

[0079] Step 5: Health status prediction

[0080] After completing multi-scale information fusion and regional adaptive weighting, the present invention enters the health status prediction stage. The goal of this stage is to build a deep learning regression model based on the fused image features to achieve high-precision prediction of lithium-ion battery SoH. This invention uses regression models such as MLP, LSTM, and Transformer for prediction, and combines knowledge graphs and physical modeling to improve the interpretability and reliability of the predictions.

[0081] First, the present invention uses the fused multi-scale and multi-region features as input data to build a deep learning regression model. For a single battery sample, the fused features are expressed as:

[0082]

[0083] Among them, ω r is the regional adaptive weighting coefficient, is the feature vector of the rth region. Subsequently, this global feature vector is input into the deep learning regression model for health status prediction:

[0084] y SoH =f MLP / LSTM / Transformer (F global )

[0085] Among them, f is the regression model, the input is the image fusion feature, and the output is the SoH of the battery.

[0086] The regression model of the present invention mainly includes the following three structures:

[0087] 1) MLP: Used to process high-dimensional feature data, it uses a fully connected layer for mapping and combines it with BatchNormalization for normalization to improve training stability.

[0088] 2) LSTM: Suitable for processing time series data. It is used to model the long-term dependencies between different image features in battery health status prediction, improving the adaptability of the prediction model to historical samples.

[0089] 3) Transformer: Used to capture the attention relationship between different feature dimensions and improve the stability and generalization ability of prediction through the multi-head self-attention mechanism.

[0090] In order to improve the prediction accuracy, the present invention uses the mean square error (MSE) loss function for optimization:

[0091]

[0092] Among them, y i is the real SoH value, To predict the SoH value, N is the total number of samples. By minimizing the MSE error, the adaptability of the prediction model to different battery samples is improved.

[0093] In addition, to enhance the interpretability of the prediction model, the present invention combines the knowledge graph to integrate the degradation mechanism of battery materials with the data-driven prediction model. For example, the present invention establishes association rules between image features such as crack ratio, oxide layer thickness, and porosity and health status:

[0094] P(y|F)∝P(F|M)P(M)

[0095] Here, P(F|M) represents the relationship between the material degradation mechanism and image features, and P(M) represents the prior probability of material aging. By optimizing the regression model training process through knowledge graphs, the prediction results are not only data-driven but also conform to the laws of physics and materials science, improving the model's interpretability.

[0096] Step 6: Model optimization and training

[0097] To ensure the accuracy and stability of the health status prediction model, the present invention introduces a series of optimization strategies during the training process, including hyperparameter optimization, federated learning, dynamic learning rate adjustment, and regularization methods, to enhance the generalization ability of the model and reduce overfitting problems.

[0098] First, the present invention uses Bayesian optimization to adjust hyperparameters to automatically search for the optimal network structure, learning rate, and regularization parameters. The goal of Bayesian optimization is to find the optimal hyperparameter combination that minimizes the model error:

[0099] θ * =arg min θ L(y,f θ (F))

[0100] Among them, θ * is the optimized model parameter, L(y, f θ (F) represents the loss function. The present invention adopts Gaussian process (GP) to model the distribution of hyperparameters and continuously adjusts them during the training process to improve the adaptability of the model.

[0101] In addition, the present invention introduces an adaptive learning rate adjustment mechanism, which uses a larger learning rate in the early stage of training to speed up convergence, and reduces the learning rate in the later stage of training to prevent oscillation and overfitting:

[0102]

[0103] Among them, η t is the learning rate of the tth round of training, η0 is the initial learning rate, and λ is the decay coefficient. Through adaptive learning rate adjustment, the present invention can find the optimal solution more quickly and improve training efficiency.

[0104] In order to improve the privacy protection capability and cross-device applicability of the model, this paper adopts a federated learning mechanism to enable multiple devices or laboratories to conduct joint training without sharing original data. Assume there are K devices, each device i maintains local model parameters θ i , then the global model parameters are updated as follows

[0105]

[0106] Among them, ω i Represents the data weight of device i. Through federated learning, the present invention can share prediction models between different laboratories and battery manufacturers without leaking sensitive data, thereby increasing the application range of the algorithm. In addition, to prevent overfitting, the present invention adopts regularization methods (such as L2 regularization) to constrain the complexity of the model:

[0107] L=L MSE +λ||θ|| 2

[0108] Among them, L MSE is the mean square error loss, λ||θ|| 2 Represents the regularization term, which effectively suppresses the overfitting of the model to the training data and improves its generalization ability.

[0109] Step 7: Data Enhancement and Intelligent Generation

[0110] Since the acquisition cost of microscopic imaging data is high and there are fewer samples in some health states (such as extreme aging states), the present invention adopts data enhancement and intelligent data generation methods to improve the generalization ability of the model.

[0111] First, the present invention uses region-adaptive CutMix data enhancement to perform random splicing between different microscopic image regions to increase sample diversity. The CutMix calculation formula is as follows:

[0112] I aug =M(x,y)·I1(x,y)+(1-M(x,y))·I2(x,y)

[0113] Where M(x, y) is a binary mask matrix that controls how image regions are stitched together. Through CutMix, the present invention can simulate transitions between different health states and improve the model's robustness to subtle morphological changes.

[0114] In addition, the present invention uses Diffusion Model and GAN for data synthesis to supplement the limited experimental data. The training process of the diffusion model is as follows:

[0115]

[0116] Among them, x t is the generated image at time step t, α t and β t is the noise parameter of the diffusion process. Through reverse diffusion, the present invention can generate high-quality synthetic images with physical consistency and expand the scale of the data set.

[0117] To further improve data diversity, this paper uses adversarial data augmentation to generate more challenging samples through adversarial training:

[0118]

[0119] Among them, G is the generator, D is the discriminator, and P data represents the real data distribution, P z By generating data through GAN, the present invention effectively improves the model's adaptability to complex morphological changes.

[0120] 4.Technological advantages

[0121] Compared with the existing technology, the present invention realizes the accurate prediction of battery health status based on microscopic imaging data, without relying on traditional electrical signal analysis, avoiding complex voltage and current feature extraction, and improving the physical intuitiveness and interpretability of the prediction. The present invention optimizes the feature contribution of different magnifications and different electrode areas through multi-scale feature fusion and regional adaptive weighting, so as to reduce the health prediction error and control the root mean square error (RMSE) within 0.0132. The deep learning regression model is combined with physical modeling to make the prediction results more stable and applicable to different aging modes. The present invention further combines data enhancement and intelligent data generation technology to improve the generalization ability under small sample conditions, and optimizes model training through federated learning to enhance cross-device adaptability. This method shows strong predictive ability under different battery materials and operating environments, and can be applied to new energy vehicles, grid energy storage and battery management systems and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0122] In order to more clearly describe the technical solution of the present invention, the following figures are provided to illustrate the system architecture, main functional modules and key method processes of the present invention. The figures of the present invention are for illustrative purposes only and do not constitute a limitation on the scope of protection of the present invention.

[0123] Figure 1 : Schematic diagram of the overall architecture of the lithium-ion battery health prediction system of the present invention, showing key modules such as data acquisition, preprocessing, multi-scale feature extraction, information fusion, regional adaptive weighting, health status prediction and model optimization;

[0124] Figure 2 : Schematic diagram of the multi-scale feature extraction of the present invention, illustrating how image features at different magnifications are processed within a computer vision and deep learning framework to extract key features such as cracks, pores, and particle distribution;

[0125] Figure 3: Schematic diagram of the multi-scale information fusion and regional weighting mechanism of the present invention, showing how to use the pyramid attention mechanism to optimize the feature contribution of images with different magnifications and enhance the health prediction ability of different electrode regions through the regional weighting strategy;

[0126] Figure 4 : A schematic diagram of the health status prediction process of the present invention illustrates the deep learning modeling process from image feature input to final battery SoH prediction, including feature fusion, regression calculation and optimization strategy. DETAILED DESCRIPTION

[0127] The technical solution of the present invention is further described below with reference to specific embodiments. The present invention is not limited to the following examples, and can be adjusted according to different battery types, data acquisition equipment, and health prediction requirements during implementation.

[0128] Example 1: Lithium-ion battery health prediction based on microscopic imaging data

[0129] This example uses a button-type lithium-ion battery (Panasonic ML2020) in a spinel lithium manganese oxide (LMO) system as the research object, uses SEM to collect microscopic image data of the electrode material, and uses the method of the present invention to predict the battery SoH. Figures 1 to 4 , describe in detail the system architecture, data processing flow and prediction method of the present invention.

[0130] like Figure 1 As shown, the health prediction system of the present invention includes key modules such as data acquisition, preprocessing, feature extraction, information fusion, regional weighting, health prediction, and model optimization. This example uses a JEOL JCM-6000 SEM device to obtain microscopic images of the electrode surface at 200×, 500×, 1000×, and 2000× magnifications, sampling the core region, transition region, and edge region to fully cover the electrode degradation characteristics.

[0131] The collected image data has been standardized, including denoising, normalization, contrast enhancement and brightness optimization, to reduce the impact of the experimental environment on the data. Figure 2 As shown in the figure, denoising uses adaptive filtering technology to remove high-frequency noise and maintain microstructural details; contrast enhancement is based on CLAHE to improve the recognizability of low-contrast areas and make cracks, particle boundaries and pore features clearer.

[0132] like Figure 3As shown, after data preprocessing, it enters the multi-scale feature extraction stage. This embodiment combines computer vision and deep learning methods to extract key features, including morphological features such as crack ratio, porosity, and particle distribution, and uses the ResNet50 pre-training model to extract high-dimensional depth features. In order to make full use of multi-scale information, this embodiment uses a pyramid attention mechanism for multi-magnification feature fusion. Among them, low-magnification images (200×) provide overall structural information, and high-magnification images (2000×) analyze microscopic cracks, pores, and interface degradation features. During the fusion process, different magnification information is dynamically weighted so that high-information features occupy a greater weight, while eliminating redundant information to improve computational efficiency.

[0133] The present invention further calculates the health contribution of the core area, transition area and edge area through a regional adaptive weighting mechanism. Figure 3 As shown, the core area has more representative characteristics due to its higher structural stability, while the edge area is more affected by oxidation and crack propagation, and the prediction error may be higher. To this end, this embodiment adopts a dynamic weighting strategy based on deep learning to optimize the contribution of different regional characteristics to health prediction and improve prediction stability.

[0134] Finally, the fused image features are input into the MLP+Transformer prediction model, and the interpretability of the results is optimized by combining the knowledge graph. Figure 4 As shown, the prediction process includes feature fusion, regression calculation, and optimization strategy. Test results of this embodiment on an experimental dataset show that the root mean square error (RMSE) of single magnification image prediction is approximately 0.0524. After adopting multi-scale fusion and regional weighting, the RMSE is reduced to 0.0132, a 75% reduction in prediction error. Furthermore, under extreme aging conditions with SoH ≤ 75%, the method of the present invention can still maintain stable predictions, with an error reduction of 40.8%, demonstrating its efficiency and robustness.

[0135] Example 2: Health prediction based on different battery systems

[0136] The difference from Example 1 is that this example uses a nickel-cobalt-manganese (NCM) ternary material lithium-ion battery and uses TEM to collect data to verify the applicability of the present invention in different battery systems.

[0137] An 18650 cylindrical battery of the NCM system was selected, and TEM was used to collect the microstructural information of the electrode material at magnifications of 500×, 1000×, and 5000×, respectively, and combined with XRD spectrum data as auxiliary information. The data preprocessing method is the same as that in Example 1, and the same feature extraction and fusion method is adopted. The TEM image and XRD spectrum features are integrated through the spectral fusion mechanism to improve the ability to characterize the aging state of the material. The experimental results show that the RMSE of this method in the NCM system is controlled within 0.0154, which is 68% lower than the prediction error of the single magnification image. It shows strong generalization ability under different working conditions, proving its wide applicability.

[0138] Example 3: Health prediction based on different operating conditions

[0139] The difference from Example 1 is that this example performs health prediction on lithium iron phosphate (LFP) batteries under different temperatures and charge and discharge rate conditions to verify the adaptability of the present invention to different operating conditions.

[0140] LFP square aluminum shell batteries were selected for accelerated aging tests at three temperatures: 22°C, 30°C, and 40°C, and SEM was used to collect image data. The charge and discharge rates were set to 0.2C, 0.5C, and 1.0C to simulate different workloads. This method combines multi-scale feature fusion, regional weighting, and Transformer prediction model, and takes temperature and rate as input variables to adapt to different working conditions. Experimental results show that the RMSE of this method under different environmental conditions is controlled within 0.0175, and the prediction error is reduced by 55.3% under high temperature and high rate conditions, indicating that it is not only suitable for different battery materials, but also can adapt to complex operating environments.

Claims

1. A lithium-ion battery health prediction system based on multi-scale and regional adaptive fusion of microscopic images, characterized by: The system includes: 1) Data acquisition module, used to obtain microscopic image data of battery electrodes and classify and store image data of different magnifications, imaging modes, and sampling areas; 2) Data processing module, used to standardize the collected microscopic image data, including contrast enhancement, adaptive denoising, gamma correction and regional feature optimization to improve feature separability; 3) A multi-scale feature extraction module for extracting material degradation features from image data of different magnifications. The feature extraction includes: a) Computer vision-based morphological analysis, including crack detection, porosity calculation, particle distribution analysis, oxide layer thickness assessment, and edge roughness measurement; b) Feature learning based on deep learning, using a hybrid neural network architecture including convolutional neural network (CNN), visual transformer (ViT), and graph neural network (GNN); c) Feature modeling based on physical constraints, combined with electrochemical kinetic models, to establish the correlation between imaging features and battery material aging mechanisms; 4) A multi-scale information fusion module is used to optimize the feature fusion of image data of different magnifications. The fusion method includes: a) Pyramid attention mechanism: While extracting macroscopic structural information from low-magnification images, it also uses high-magnification images to extract microscopic degradation features, and optimizes the multi-scale fusion strategy through dynamic feature weighting; b) Adaptive feature selection mechanism to screen features of images at different magnifications and remove low-relevance or redundant information to optimize computational efficiency; c) Combined with spectral data (XRD spectrum, EDS spectrum, etc.) to enhance the characterization capability of microscopic imaging data; 5) Regional adaptive weighting module, which is used to dynamically adjust the weight according to the changes in the microscopic characteristics of different regions of the electrode material (core region, transition region, edge region), thereby improving the stability of health prediction; 6) A health status prediction module, which is used to predict the battery state of health (SoH) based on the fused multi-scale and multi-region features, using a deep learning regression model, including but not limited to a multi-layer perceptron (MLP), a long short-term memory network (LSTM), a transformer, or an adaptive recurrent neural network, and using a minimum mean square error (MSE) to optimize the loss function; combined with a knowledge graph, the material degradation mechanism is combined with the deep learning prediction model to improve the interpretability of the prediction results; 7) Model optimization and dynamic learning module, used to optimize the training process of health prediction models, including: a) Adopting an adaptive learning rate adjustment strategy to dynamically optimize the learning rate at different training stages and improve the model convergence speed; b) Combining Bayesian hyperparameter optimization and genetic algorithm search to adjust the hyperparameters of deep learning models and improve model generalization capabilities; c) Adopting a federated learning mechanism to collaboratively train models across different battery data sources to improve privacy protection capabilities; 8) Data enhancement and intelligent generation module, used to improve the model generalization ability when data is limited. The data enhancement method includes: a) Regionally adaptive CutMix data augmentation, which performs random stitching between different microscopic image regions to improve the model's adaptability to regional changes; b) Data synthesis based on diffusion models or generative adversarial networks (GANs) to supplement limited experimental data; c) Adversarial data enhancement: generating more challenging samples through adversarial training to improve model robustness; d) Contrastive learning based on physical priors optimizes the robustness of deep learning feature extraction by comparing positive and negative samples.

2. The system according to claim 1, wherein: The microscopic image data comes from one or more microscopic imaging devices, including but not limited to scanning electron microscope (SEM), transmission electron microscope (TEM), scanning tunneling microscope (STM), atomic force microscope (AFM), super depth of field optical microscope (SDFM) and other electronic or optical microscopic imaging technologies.

3. The system according to claim 1, wherein: The health status prediction model is further combined with a federated learning mechanism to optimize data privacy protection and improve cross-device prediction capabilities.

4. The system according to claim 1, wherein: The data processing module adopts a self-supervised learning strategy to optimize feature representation through unlabeled sample learning, thereby improving the adaptability of the model under data-limited conditions.

5. The system according to claim 1, wherein: The system is applicable to the health prediction of spinel lithium manganese oxide, lithium iron phosphate, lithium cobalt oxide, nickel cobalt manganese ternary, and nickel cobalt aluminum ternary lithium-ion batteries.

6. A lithium-ion battery health prediction method based on multi-scale and regional adaptive fusion of microscopic images, characterized in that: The following steps are involved: a) Collect a variety of microscopic image data (including SEM, TEM, STM, AFM, SFDFM, etc.), and classify and store images of different magnifications and different areas; b) Standardize the image data, including contrast enhancement, denoising, and region optimization, to improve the stability of feature extraction; c) Extract multi-scale microscopic image features, including computer vision features, deep learning features, and physical information constraint features; d) Using the pyramid attention mechanism to optimize multi-rate image feature fusion, eliminate inefficient feature information, and improve computational efficiency; e) Optimize the contribution weights of different electrode region characteristics based on regional adaptive weighting strategy to improve the stability of health prediction; f) Build a deep learning regression model using Transformer, LSTM, or MLP for SoH prediction, and combine it with knowledge graphs to optimize the interpretability of the results; g) Using CutMix and diffusion models for data augmentation to improve model generalization capabilities; h) Bayesian optimization and federated learning mechanisms are used to dynamically adjust model parameters to adapt the prediction model to the long-term battery degradation process.

Images