A Real-Time State Estimation Method for Lithium-ion Batteries Based on Multivariate Sensing Data

Abstract

This invention discloses a real-time state estimation method for lithium-ion batteries based on multi-modal sensing data, belonging to the field of lithium-ion battery technology. The method includes: selecting important sensing points within the battery; deploying sensors within the battery; constructing a lithium-ion battery experimental dataset; performing three-dimensional interpolation reconstruction of the temperature field and three-dimensional slicing to obtain a temperature image dataset T; performing two-dimensional interpolation reconstruction of the pressure field and constructing a time-series pressure image dataset P; reconstructing EIS, voltage, and current data to obtain one-dimensional feature vectors; processing the dataset and feature vectors using 3D CNN, (2+1)D CNN, SFM, and ARD; extracting important global features using Self-Attention and Cross-Attention; and using MLP to estimate battery state parameters based on the transferred features. This invention has the advantages of facilitating real-time estimation and maintaining high accuracy even with insufficient training data. Because advanced sensing accurately reflects the internal field, it can better fit the real-world situation, and the fusion of multi-modal features improves estimation performance.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery technology, specifically relating to a real-time state estimation method for lithium-ion batteries based on multi-source sensing data. Background Technology

[0002] Lithium-ion batteries possess advantages such as high energy density, low self-discharge rate, and long cycle life, and are widely used in various products. Unlike the stable environment of laboratory testing, user application environments and battery usage are complex, often involving extreme conditions and battery abuse, which can lead to performance degradation or even failure. Therefore, it is necessary to monitor battery operation in a timely manner during use and scientifically plan the battery's usage limits to extend its service life. The state parameters of lithium-ion batteries reflect their operating state and performance information, providing parameter support for battery safety early warning, fault diagnosis, and equalization. Therefore, accurate and real-time estimation of various state parameters has significant practical implications.

[0003] Battery state parameters include State of Health (SOH), State of Charge (SOC), and Remaining Useful Life (RUL). Currently, several parameter estimation methods exist: 1. Ampere-hour integration method: This method is test-based, simple to operate, and highly accurate, but requires the battery to be fully charged and discharged in each cycle, which is difficult to meet under normal operating conditions. Furthermore, it only provides feedback at the end of the cycle, not in real-time. 2. Model-based methods: These methods struggle to balance generality, accuracy, and computational complexity. For example, electrochemical models describe the internal reaction processes of the battery but contain 10 nonlinear coupled partial differential equations, making them complex and difficult to solve. Low-order equivalent circuit models cannot accurately predict extreme conditions, while high-order models are only applicable to specific environments and have low generality. 3. Filtering-based methods: These methods cannot balance the computational complexity of the model with the richness of the electrochemical mechanism.

[0004] Data-driven methods avoid complex electrochemical process analysis and solution, reduce reliance on prior knowledge, are simple in principle, highly accurate, and facilitate fitting complex nonlinear relationships between battery state parameters and collected data. Currently, the characteristic quantities used in data-driven methods are mostly voltage and current, which have undergone various processing steps to predict battery state based on electrical performance parameters.

[0005] Building upon electrical performance parameters, combining high-information-density parameters that reflect more of the battery's internal conditions with network learning can better fit real-world scenarios, improving network accuracy and reducing errors. In recent years, with the rapid development of embedded sensors and flexible electronics, advanced sensing technologies have matured, leading to the emergence of various multi-physics, high-spatial-resolution micro-flexible sensor arrays that can be integrated into batteries, providing the technological possibility for real-time acquisition of internal battery pressure, strain, and temperature. The advent of portable, high-precision electrochemical impedance spectroscopy (EIS) acquisition devices also allows for real-time acquisition of electrochemical performance parameters for battery analysis, leveraging EIS's informational advantages regarding internal material properties, interface phenomena, and electrochemical reactions. Therefore, there is an urgent need to develop a real-time state estimation method for lithium-ion batteries that combines electrical performance parameters, electrochemical performance parameters, and physical performance parameters obtained from sensors. Compared to using electrical performance estimation alone, this method reduces the requirement for a large training set of data, achieving high accuracy even with insufficient training data, and significantly shortening dataset acquisition time and cost. Summary of the Invention

[0006] This invention addresses the shortcomings of existing technologies by proposing a real-time state estimation method for lithium-ion batteries based on multi-source sensing data.

[0007] The purpose of this invention is to provide a real-time state estimation method for lithium-ion batteries based on multi-source sensing data. The specific steps are as follows:

[0008] Step 1: Select sensor placement locations. The locations should include key areas such as the positive and negative electrode tab areas and the center of the battery, selected in multiple layers.

[0009] Step 2: Design and deploy sensors for key locations on M lithium-ion batteries of the same model. Highly integrated, high-resolution flexible thin-film sensors are preferred, allowing for comprehensive, multi-dimensional monitoring of various states within the battery using fewer components, thus minimizing impact on the battery.

[0010] Step 3: Perform multiple complete charge-discharge cycles on M lithium-ion batteries of the same model equipped with sensors. Stop the experiment when the battery capacity decays to 80% of the nominal capacity and obtain the experimental data set. This includes temperature and pressure field data collected at multiple points during the complete charge-discharge process, dynamically acquired EIS, current, and voltage data, and the entire process's SOC, SOH, and RUL.

[0011] Step 4: Perform 3D interpolation reconstruction of temperature data based on spatial features to improve temperature resolution. Taking an orthogonal rectangular coordinate system as an example, the same applies to cylindrical and spherical coordinate systems. Using the X, Y, and Z coordinate axes as reference directions, perform three-dimensional multi-layer slicing on the reconstruction network, storing the temperature field images of each slice to form a temperature image dataset T, which includes a subset T of the three-dimensional slices. x T y T z Two-dimensional interpolation reconstruction was performed on the pressure data to obtain a high-resolution pressure field map. A time axis was constructed to obtain a video-like time-series pressure image dataset P. Both temperature and pressure images are single-channel images, which simplifies subsequent image processing and reduces computational overhead.

[0012] Step 5: Reconstruct the EIS data, voltage, and current at multiple frequencies to obtain a one-dimensional feature vector [real1, real2, ..., real2]. n imag1, imag2, ..., imag n ,V1,V2,……,V n ,I1,I2,……,I n ], where real i Let imag be the real part of the EIS at the i-th frequency point. i V is the imaginary part of EIS at the i-th frequency point. i with I i The voltage and current at this frequency point were collected and arranged in time sequence to form an electrochemical and electrical performance dataset E.

[0013] Step 6: Dataset partitioning and normalization. The datasets T, P, E and SOC, SOH, RUL described in Steps 4 and 5 are divided into training and test sets, and normalized using the Z-score normalization method.

[0014] Step 7: Process the temperature image dataset T. This involves processing the temperature image subset T along the x-axis under three-dimensional slices. x For example, taking T x T represents the current slice. x-1 T represents the previous adjacent slice in the x-direction. x+1 Represents the next adjacent slice in the x-direction, for T x-1 T x T x+1A 3D CNN with residual connections is used, employing i×j×3 3D convolutional kernels that slide across a 3D image sequence. Depth is progressively increased through convolutional computation to enrich semantics and extract local features. The features are then fused and enhanced using a Spatial Fusion Module (SFM). SFM first merges three consecutive images, transforming 3×H×W×D into H×W×3D, where H is the image height, W is the image width, and D is the number of image channels, denoted as T. SF The spatial information of the feature map is aggregated using average pooling and max pooling, and is represented as follows:

[0015] SFM(T SF ) = T SF ×σ(MLP(AvgPool(T SF ))+MLP(MaxPool(T SF )))

[0016] The spatial information is then input into a shared network, which consists of a multilayer perceptron (MLP) and hidden layers.

[0017] In the formula, SFM represents the spatial fusion module, and T SF Let σ represent three consecutive images in a certain direction in the temperature image dataset T used for the spatial fusion module, where σ is the sigmoid function, MLP represents multilayer perceptron, AvgPool represents average pooling, and MaxPool represents max pooling.

[0018] For T x T y T z All convolutional layers are processed as described above, and batch normalization is performed on all convolutional layers.

[0019] Step 8: Process the stress image dataset P. Use a (2+1)D convolution with residual connections, factorizing the convolution to handle the spatial and temporal dimensions separately, reducing the size of the parameter and weight matrices. The convolution kernel of this method consists of a kernel of size N. i-1 A two-dimensional convolution kernel of size M is defined as ×1×d×d. i It consists of a temporal convolution kernel of size ×t×1×1, where t represents the time range and d represents the spatial width and height. Since it is a single-channel image, N... i-1 =1. Hyperparameter M i The intermediate subspace dimension of the signal projection between spatial and temporal convolutions is determined, and the choice is... In the formula, M i N represents the number of output channels of the temporal convolution kernel, t represents the time range, d represents the spatial width and height, and N represents the spatial width and height. i-1N represents the number of output channels of the previous convolutional layer. i-1 =1, N i This indicates the number of output channels in the current convolutional layer. Batch normalization is applied to all convolutional layers to accelerate convergence and simplify the parameter tuning process, while global spatiotemporal average pooling is applied to the final convolutional tensor.

[0020] Step 9: Extract high contribution rates from EIS. Add an Automatic Relevance Determination (ARD) covariance function kernel to the Gaussian Process Regression (GPR) and optimize the parameters using the maximum likelihood method. After identifying important features, remove redundant columns from the 4n columns of real, imaginary, voltage, and current data to simplify the network.

[0021] Step 10: After processing the temperature, pressure, electrochemical and electrical performance modules, add a self-attention mechanism to assign weights to each feature and extract important global features.

[0022] Step 11: Use fully connected layers to map the globally important features of temperature and pressure to one-dimensional vectors, using the number of important features as the vector length. After concatenating the two vectors, use a fully connected layer to adjust the dimension to be the same as the dimension of the feature vectors in dataset E after Step 9, forming the query vector Q.

[0023] Step 12: Using the high-contribution-rate electrochemical and electrical performance feature vectors obtained in Step 9 as the bond vector K and value vector V, establish cross-correlation relationships between different modalities using a cross-attention mechanism. Then, fuse the output of the cross-attention mechanism with other features.

[0024] Step 13: Pass the fused features to a multilayer perceptron (MLP) to estimate the state parameters SOH, SOC, and RUL of the lithium-ion battery, and then inverse normalize them to obtain the final result.

[0025] The beneficial effects of this application are as follows: This invention, based on an implantable battery sensor and a portable electrochemical impedance spectroscopy (EIS) monitoring method, combines multi-physics fields, EIS spectral information, and electrical performance information within the battery, fusing multiple modalities to leverage the high information density of direct state parameters within the battery. This alleviates the training difficulties caused by insufficient training data while improving estimation accuracy. Spatial and temporal slices are applied to the internal physical fields, and 3D CNN and (2+1)D CNN methods are used to directly process the sliced ​​images, facilitating real-time estimation. The ARD method is used to filter important frequencies, reducing EIS acquisition time. Finally, cross-correlation relationships are established, and the fusion of various important global features enhances estimation performance. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart of the real-time state estimation method for lithium-ion batteries based on multi-source sensing data according to the present invention.

[0028] Figure 2 This is a model structure diagram of the real-time state estimation method for lithium-ion batteries based on multivariate sensing data according to the present invention.

[0029] Figure 3 This is a schematic diagram of the real-time state estimation method for lithium-ion batteries based on multi-source sensing data according to the present invention.

[0030] Figure 4 This is a schematic diagram illustrating the temperature and pressure data processing method of the real-time state estimation method for lithium-ion batteries based on multi-source sensing data according to the present invention.

[0031] In the diagram: 1. Sensor insertion location; 2. Flexible thin-film sensor; 3. Temperature sensor; 4. Pressure sensor. Detailed Implementation

[0032] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0033] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0034] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0035] Combination Figure 1 , Figure 2 The real-time state estimation method for lithium-ion batteries based on multi-source sensing data of the present invention includes the following specific steps in the embodiments of this application:

[0036] Step 1: Select key sensor locations. These locations should include critical areas such as the positive and negative electrode tabs and the battery center, selected in multiple layers. To minimize impact on the battery's internal structure, it is preferable to select key locations on 2-3 planes.

[0037] Step Two, Refer to Figure 3 Sensors are designed and deployed at key locations. Implantable sensors are preferably highly integrated, high-resolution, multi-physics coupled flexible thin-film sensors, which facilitate comprehensive and multi-dimensional monitoring of multiple areas within the battery using fewer components. Their thinner size and flexible nature allow for conformal integration with the battery's internal structure, reducing impact on battery performance. The number of sensors is the same as the number of planes selected in step one, and each sensor monitoring unit is arranged on the plane according to the key locations.

[0038] Step 3, Refer to Figure 3 A multi-cycle complete charge-discharge experiment was conducted on M identical square soft-pack lithium-ion batteries equipped with sensors. A high-rate discharge of 1C was selected to accelerate aging, and the experimental time was shortened. The experiment was stopped when the battery capacity decayed to 80% of its nominal capacity, and the experimental data dataset was obtained. This includes temperature and pressure field data collected by sensors during the complete charge-discharge process, EIS, current, and voltage data dynamically acquired by the EIS acquisition system, and SOC, SOH, and RUL data collected by the battery performance acquisition system throughout the process. The consistency of all acquired data over time was ensured.

[0039] Step 4, Refer to Figure 4 Based on spatial features, three-dimensional interpolation reconstruction of temperature data is performed to improve temperature resolution. The reconstruction network is sliced ​​in three directions using the X, Y, and Z coordinate axes as reference directions, where X represents the battery width, Y represents the battery length, and Z represents the battery thickness. Preferably, there are at least five slices in each direction. The temperature image dataset T consists of temperature field images of the slice cross-sections, containing subsets T of the three-directional slices. x T y T z Since pressure is a vector quantity, a high-resolution pressure field map is obtained by reconstructing it in two dimensions along the X and Y directions based solely on the pressure data magnitude. A time axis is then constructed to obtain a video-like time-series pressure image dataset P. Both temperature and pressure images are single-channel images, with color intensity representing numerical values ​​to simplify subsequent image processing and reduce computational overhead.

[0040] Step 5: Reconstruct the EIS data, voltage, and current at multiple frequencies to obtain a one-dimensional feature vector [real1,real2,……,realn,imag1,imag2,……,imagn,V1,V2,……,Vn,I1,I2,……,In], where real1 is the real part of the EIS at the i-th frequency point, imag1 is the imaginary part of the EIS at the i-th frequency point, and imag1 is the imaginary part of the EIS at the i-th frequency point. i with I i The voltage and current data collected at this frequency point using EIS data are arranged in time sequence to form an electrochemical and electrical performance dataset E. That is, the first feature vector is collected after n seconds, the second feature vector is collected after 2n seconds, and so on.

[0041] Step 6: Dataset partitioning and normalization. The datasets T, P, E and SOC, SOH, RUL described in Steps 4 and 5 are divided into training set and test set, with 70% as training set and 30% as test set. Z-score normalization is then used for normalization.

[0042] Step 7: Process the temperature image dataset T. This involves processing the temperature image subset T along the x-axis under three-dimensional slices. x For example, taking T x T represents the current slice. x-1 T represents the previous adjacent slice in the x-direction. x+1 Represents the next adjacent slice in the x-direction, for T x-1 T x T x+1 A 3D CNN with residual connections is used, employing i×j×3 3D convolutional kernels that slide across a 3D image sequence. The kernels i and j are preferably chosen to match the aspect ratio of the image, and zero padding is preferred to preserve edge information. Depth is progressively increased through convolutional computation to enrich semantics and extract local features. Features are then fused and enhanced using a Spatial Fusion Module (SFM). SFM first merges three consecutive images, transforming 3×H×W×D into H×W×3D, where H is the image height, W is the image width, and D is the number of image channels, denoted as T. SF The spatial information of the feature map is aggregated using average pooling and max pooling, and is represented as follows:

[0043] SFM(T SF ) = T SF ×σ(MLP(AvgPool(T SF ))+MLP(MaxPool(T SF )))

[0044] The spatial information is then input into a shared network, which consists of a multilayer perceptron (MLP) and hidden layers.

[0045] In the formula, SFM represents the spatial fusion module, and T SF Let σ represent three consecutive images in a certain direction in the temperature image dataset T used for the spatial fusion module, where σ is the sigmoid function, MLP represents multilayer perceptron, AvgPool represents average pooling, and MaxPool represents max pooling.

[0046] Step 8: Process the stress image dataset P. Use a (2+1)D convolution with residual connections, factorizing the convolution to handle the spatial and temporal dimensions separately, reducing the size of the parameters and weight matrices. The ReLU activation function is preferred. The convolution kernel of this method consists of a kernel of size N. i-1 A two-dimensional convolution kernel of size M is defined as ×1×d×d. i It consists of a temporal convolution kernel of size ×t×1×1, where t represents the time range and d represents the spatial width and height. Since it is a single-channel image, N... i-1 =1. Hyperparameter M i The intermediate subspace dimension of the signal projection between spatial and temporal convolutions is determined, and the choice is... In the formula, M i N represents the number of output channels of the temporal convolution kernel, t represents the time range, d represents the spatial width and height, and N represents the spatial width and height. i-1 N represents the number of output channels of the previous convolutional layer. i-1 =1, N i This indicates the number of output channels in the current convolutional layer. Batch normalization is applied to all convolutional layers to accelerate convergence and simplify the parameter tuning process, while global spatiotemporal average pooling is applied to the final convolutional tensor.

[0047] Step 9: Extract high contribution rates from EIS. Add an Automatic Relevance Determination (ARD) covariance function kernel to the Gaussian Process Regression (GPR) dataset, and optimize the parameters using the maximum likelihood method. After identifying important features, remove redundant columns from the 4n columns of real, imaginary, voltage, and current data to simplify the network. To maintain maximum estimation accuracy, retain at least 20 columns.

[0048] Step 10: After processing the temperature, pressure, electrochemical and electrical performance modules, add a self-attention mechanism to assign attention weights to each feature, extract important global features, and remove irrelevant and redundant features to accelerate subsequent operations.

[0049] Step 11: Use fully connected layers to map the globally important features of temperature and pressure to one-dimensional vectors, using the number of important features as the vector length. After concatenating the two vectors, use a fully connected layer to adjust the dimension to match the dimension of the feature vectors in dataset E after Step 9, forming the query vector Q.

[0050] Step 12: Using the high-contribution-rate electrochemical and electrical performance feature vectors obtained in Step 9 as the bond vector K and value vector V, establish cross-correlation relationships between different modalities using a cross-attention mechanism. Then, fuse the output of the cross-attention mechanism with other features.

[0051] Step 13: Pass the fused features to a multilayer perceptron (MLP) to estimate the state parameters SOH, SOC, and RUL of the lithium-ion battery. The preferred hidden layer of the perceptron is 4 layers, and the preferred activation function is the ReLU activation function. The final result is obtained by inverse normalization.

[0052] The embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. It should be noted that implementations not illustrated or described in the drawings or the main text of the specification are forms known to those skilled in the art and have not been described in detail.

[0053] In summary, this disclosure provides a flexible sensing array capable of monitoring multiple parameters at multiple points within a lithium battery. The array, designed and fabricated using flexible electronics technology, comprises several sensing units, each integrating two or more sensors. Compared to existing battery internal monitoring technologies, the flexible sensing array provided in this disclosure can simultaneously monitor multiple parameters in multiple regions within a lithium battery, exhibiting higher spatial resolution.

[0054] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for real-time estimation of the state of a lithium-ion battery based on multi-source sensing data, characterized in that, The specific steps are as follows: S1: Select important locations for sensor placement. These locations should include the positive and negative tab areas and the center area of ​​the battery. Select multiple locations in layers. S2: Design implantable sensors for key locations and place implantable sensors on M lithium-ion batteries of the same model. S3: Perform multiple cycles of complete charge and discharge on M lithium-ion batteries of the same model equipped with sensors. Stop the experiment when the battery capacity decays to 80% of the nominal capacity and obtain the experimental data set. The experimental data set includes temperature field and pressure field data collected at multiple points during the complete charge and discharge process, as well as dynamically collected EIS, current and voltage data, and the full process SOC, SOH and RUL. S4: Process temperature field data and pressure field data to construct a three-dimensional slice image dataset T based on three-dimensional reconstruction and a video-like sampling time-series pressure image dataset P based on two-dimensional reconstruction, respectively. S5: Reconstruct the EIS, voltage, and current data at multiple frequencies and arrange them in time sequence to form an electrochemical and electrical performance dataset E; S6: Dataset partitioning and normalization: The datasets T, P, E and SOC, SOH, RUL described in S4 and S5 are divided into training set and test set, and normalization is performed. S7: Process the temperature image dataset T, use a 3D CNN with residual connections for every three adjacent slices, input the results into the spatial fusion module to fuse and enhance the features, and perform batch normalization on all convolutional layers in this process; S8: Process the stress image dataset P, use (2+1)D convolution with residual connections, factorize the convolution to process the spatial and temporal dimensions respectively, apply batch normalization to all convolutional layers to accelerate convergence and simplify the parameter tuning process, and apply global spatiotemporal average pooling to the final convolutional tensor. S9: Add a covariance function kernel with automatic correlation determination to Gaussian process regression, extract high contribution rate frequencies of EIS, and remove redundant columns from the 4n columns of real part, imaginary part, voltage and current data after important feature identification to simplify the network; S10: Add a self-attention mechanism to extract globally important features based on attention weights; S11: Use a fully connected layer to map the globally important features of temperature and pressure to a one-dimensional vector. Use the number of important features as the vector length, concatenate the two vectors, and then use a fully connected layer to adjust the dimension to be the same as the dimension of the feature vector in dataset E after S9, thus forming the query vector Q. S12: Using the high contribution rate feature vector obtained in S9 as the key vector K and value vector V, the cross-attention mechanism is used to establish the cross-correlation relationship between different modalities, and the output of the cross-attention mechanism is fused with other features; S13: The fused features are passed to the multilayer perceptron to estimate the state parameters SOH, SOC, and RUL of the lithium-ion battery, and then inversely normalized to obtain the final result.

2. The real-time state estimation method for lithium-ion batteries based on multi-source sensing data according to claim 1, characterized in that, The implantable sensor is a highly integrated, high-resolution, multi-physics field coupled flexible thin-film sensor.

3. The real-time state estimation method for lithium-ion batteries based on multi-source sensing data according to claim 2, characterized in that, The number of flexible thin-film sensors is 2-3, and each sensor monitoring unit is arranged on a plane according to important points.

4. The real-time state estimation method for lithium-ion batteries based on multi-source sensing data according to claim 1, characterized in that, Step S4 processes the temperature field data and pressure field data. The specific steps are as follows: S401: Perform three-dimensional interpolation reconstruction on the temperature data. Taking the orthogonal rectangular coordinate system as an example, the cylindrical coordinate system and the spherical coordinate system are similar. Using the X, Y, and Z coordinate axes as the reference directions, perform three-dimensional multi-layer slicing on the reconstruction network, and store the temperature field map of the slice cross section to form a temperature image dataset T, which includes a subset T of the three-dimensional slices. x T y T z； S402: Perform two-dimensional interpolation reconstruction on the pressure data to obtain a high-resolution pressure field map, construct a time axis, and obtain a video-like sampling time-series pressure image dataset P.

5. The real-time state estimation method for lithium-ion batteries based on multi-source sensing data according to claim 4, characterized in that, Both the temperature and pressure images are single-channel images, which simplifies subsequent image processing and reduces computational overhead.

6. The real-time state estimation method for lithium-ion batteries based on multi-source sensing data according to claim 1, characterized in that, Step S5 processes the EIS data and current and voltage data, specifically: reconstructing the EIS, voltage, and current data at multiple frequencies to obtain a one-dimensional feature vector [real1, real2, ..., real2]. n imag1, imag2, ..., imag n ,V1,V2,……,V n ,I1,I2,……,I n ], where real i Let imag be the real part of the EIS at the i-th frequency point. i V is the imaginary part of EIS at the i-th frequency point. i with I i The voltage and current at this frequency point were collected and arranged in time sequence to form an electrochemical and electrical performance dataset E.

7. The real-time state estimation method for lithium-ion batteries based on multi-source sensing data according to claim 1, characterized in that, In step S6, the Z-score normalization method is used for normalization.

8. The real-time state estimation method for lithium-ion batteries based on multi-source sensing data according to claim 1, characterized in that, Step S7 processes the temperature image dataset T, and the specific steps are as follows: S701: For the temperature image subset T in the x-direction under three-dimensional slices x T x T represents the current slice. x-1 T represents the previous adjacent slice in the x-direction. x+1 Represents the next adjacent slice in the x-direction, for T x-1 T x T x+1 Using a 3D CNN with residual connections, 3D convolutional kernels of size i×j×3 slide across a 3D image sequence. Batch normalization is performed on all convolutional layers, and depth is progressively increased through convolutional computation to enrich semantics, extract local features, and apply T... y T z All should be processed as described above; S702: Features are fused and enhanced using the Spatial Fusion Module (SFM). SFM first merges three consecutive images, converting 3×H×W×D into H×W×3D, where H is the image height, W is the image width, and D is the number of image channels, denoted as T. SF The spatial information of the feature map is aggregated using average pooling and max pooling, and is represented as follows: SFM(T SF )=T SF ×σ(MLP(AvgPool(T SF ))+MLP(MaxPool(T SF ))) The spatial information is then input into a shared network, which consists of a multilayer perceptron (MLP) and hidden layers. In the formula, SFM represents the spatial fusion module, and T SF Let σ represent three consecutive images in a certain direction in the temperature image dataset T used for the spatial fusion module, where σ is the sigmoid function, MLP represents multilayer perceptron, AvgPool represents average pooling, and MaxPool represents max pooling.

9. The real-time state estimation method for lithium-ion batteries based on multi-source sensing data according to claim 1, characterized in that, Step S8 processes the stress image dataset P, and the specific method is as follows: Using (2+1)D convolutions with residual connections, the spatial and temporal dimensions are factored separately. The convolution kernel of this method consists of a kernel of size N. i-1 A two-dimensional convolution kernel of size M is defined as ×1×d×d. i It consists of a temporal convolution kernel of ×t×1×1, where t represents the time range and d represents the spatial width and height; Since it is a single-channel image, N i-1 =1, hyperparameter M i The intermediate subspace dimension of the signal projection between spatial and temporal convolutions is determined, and the choice is... Batch normalization is applied to all convolutional layers to accelerate convergence and simplify the parameter tuning process, while global spatiotemporal average pooling is applied to the final convolutional tensor. In the formula, M i N represents the number of output channels of the temporal convolution kernel, t represents the time range, d represents the spatial width and height, and N represents the spatial width and height. i-1 N represents the number of output channels of the previous convolutional layer. i-1 =1, N i This indicates the number of output channels of the current convolutional layer.

Images