A lithium battery state of health estimation method and system based on end-cloud combination

Abstract

The application discloses a lithium battery health state estimation method and system based on end-cloud combination, which firstly uploads the battery data to the cloud for preprocessing after the battery data is collected, and prepares a data set for the battery data after preprocessing. Secondly, a cloud lithium battery health state estimation network model MSMixer is constructed. Finally, the data in the data set is input into the MSMixer network for training, and the data is used as a general initial model parameter for vehicle end training, the vehicle end fine-tunes the cloud model by using local real-time data, and the lithium battery health state is completed after fine-tuning, and the lithium battery health state estimation result is output. The application not only guarantees the global generalization ability of the model, but also improves the adaptability to the single-vehicle battery characteristics, has higher accuracy, and provides reliable technical support for electric vehicle battery health management.

Description

Technical Field

[0001] This invention relates to the field of power battery management technology, specifically to a method and system for estimating the health status of lithium batteries based on edge-cloud integration. Background Technology

[0002] With the rapid development of electric vehicles and energy storage power stations, the core power source—the high-voltage battery pack—inevitably undergoes performance degradation during long-term operation. This degradation manifests as gradual capacity decay, increasing internal resistance, and decreased charge / discharge efficiency. This degradation not only shortens battery life but may also lead to safety hazards such as thermal runaway, seriously threatening the reliability and safety of the system. Therefore, research on high-precision battery state of health (SOH) estimation methods is of great significance for ensuring the safe operation of battery packs, improving energy utilization efficiency, and extending service life.

[0003] Existing SOH estimation methods can be broadly categorized into two types: model-based methods and data-driven methods. Model-based methods rely on equivalent circuit models or electrochemical mechanism models to derive SOH. However, the internal reaction mechanisms of batteries are complex and operating conditions are highly variable, making it difficult to accurately obtain model parameters, thus limiting their application in real-world scenarios. In contrast, data-driven methods directly learn the mapping relationship between inputs and SOH using runtime data, avoiding complex modeling processes and making them more suitable for complex and variable operating environments. In recent years, deep learning methods such as convolutional neural networks (CNN), recurrent neural networks (RNN), and their variants (LSTM, GRU) have been widely applied to battery SOH estimation, achieving some success in feature extraction and time-series modeling. However, the following shortcomings remain: most studies have failed to fully utilize multi-source information, particularly neglecting battery internal resistance, a key parameter highly correlated with aging; complex deep models are typically deployed in the cloud, making it difficult to meet the real-time and personalized requirements of vehicle-side applications.

[0004] Therefore, a lithium battery state of health (SOH) estimation method based on edge-cloud integration is developed. In the cloud, a multi-scale fusion MSMixer network model is trained using large-scale historical operational data, fully leveraging its advantages in feature extraction, temporal modeling, and model generalization. On the vehicle side, the cloud model is fine-tuned based on real-time monitoring data, balancing high accuracy and real-time performance. Specifically, during the fine-tuning process, internal resistance is used as a constraint term in the loss function, thereby improving the adaptability of the SOH estimation to individual battery differences, avoiding overfitting due to data scarcity, and enhancing the physical rationality of the model. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a lithium battery health state estimation method and system based on edge-cloud integration. An MSMixer network model is pre-trained in the cloud and fine-tuned using real-time data on the vehicle. Simultaneously, battery internal resistance is introduced as a loss function constraint to achieve high-precision adaptation of the model to specific vehicle batteries. This method fully utilizes historical data from the cloud and real-time data from the vehicle, improving the accuracy and robustness of health state estimation, balancing real-time performance and computational efficiency, and reducing reliance on full cloud data.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] In one aspect, this invention provides a method for estimating the state of health of lithium batteries based on edge-cloud integration, comprising the following steps:

[0008] S1. After the battery data is collected, it is uploaded to the cloud for preprocessing, and the preprocessed battery data is used to create a dataset. The battery data includes the voltage, current, temperature and corresponding internal resistance and battery capacity data of the battery pack during its historical operation, which are stored in the cloud.

[0009] S2. Construct a cloud-based lithium battery health state estimation network model, MSMixer. The MSMixer network model includes a multi-scale feature extraction network, a parallel MS temporal modeling network (comprising a sparse attention module and a Mamba module), and a spatiotemporal MLP network. Different data correspond to their respective branches of the feature extraction network and temporal modeling network. The feature vectors from the four branches are concatenated and then fused and uniformly calculated by the spatiotemporal MLP networks of two branches.

[0010] S3. Input historical data from the cloud into the MSMixer network for training, save the trained model, and use it as the general initial model parameters for vehicle-side training. The vehicle-side uses local real-time data to fine-tune the cloud-based model, and after fine-tuning, completes the lithium battery health status and outputs the lithium battery health status estimation result.

[0011] Preferably, in step S1, the method for collecting raw data of the electric vehicle battery pack is as follows: the actual vehicle operation data is uploaded, and the data uploaded to the cloud is normalized to obtain voltage, current, temperature, and corresponding battery capacity and internal resistance data, which are used as the dataset for network training.

[0012] Preferably, the method for preprocessing the original data of the electric vehicle battery pack in step S1 is to perform normalization processing using the min-max method.

[0013] 1. Preferably, in step S1, the original unprocessed voltage, current, and temperature data are... , , After performing the above preprocessing operations, the input data for the corresponding network branch is obtained. , , Where L represents the number of sampling points for voltage, current, and temperature data in one cycle of charging data.

[0014] Preferably, in step S2, the method for building the MSMixer network architecture is as follows:

[0015] First, the multi-scale feature extraction networks of different branches extract rich semantic information features from the normalized voltage, current and temperature data respectively;

[0016] Secondly, the extracted semantic information features are modeled temporally through the MS temporal modeling network. The Mamba module can effectively address the problem of uneven feature distribution in the input data, while the sparse attention module can provide direct modeling capability for long-range dependencies. The combination of the two outputs a more reliable temporal modeling feature vector.

[0017] Finally, the two-branch spatiotemporal MLP networks concatenate and fuse the time-series modeling vectors of the voltage, current, and temperature branches, and reduce the dimensionality by fusing the information of the time dimension and the feature dimension, and output the final predicted internal resistance and capacity results respectively.

[0018] Preferably, step S2 includes the following sub-steps:

[0019] Multi-scale feature extraction network:

[0020] 1. First, voltage data The initial features are obtained by passing the data through a one-dimensional convolutional neural network with a kernel size of 1 and a stride of 1, followed by average pooling. ;

[0021] 2. The initial features will go through a multi-scale feature extraction stage consisting of three one-dimensional convolutions of different receptive field sizes arranged side by side, to obtain multi-scale feature vectors extracted by convolution kernels of three different receptive field sizes;

[0022] 3. Concatenate and convolve the three multi-scale feature vectors to obtain the final output vector of the multi-scale feature extraction module. ;

[0023] MS time series modeling network:

[0024] 1. The feature vectors obtained after the multi-scale feature extraction module are then fed into the sparse attention module and the Mamba module in the MS time series modeling module, respectively; the calculation of the sparse attention module branch is as follows: After processing using a sparse attention mechanism, and then connecting it to the input via residual connections... The features are obtained by adding elements one by one and then normalizing. ; final characteristics After passing through the feedforward network, and then through the residual connection and input... The features are obtained by adding elements one by one and then normalizing them. ;

[0025] 2. The calculation for the Mamba module is as follows: After passing through a one-dimensional convolutional neural network with a kernel size of 1, features are obtained using the Silu activation function. ; final characteristics The state-space model SSM is processed by a fully connected layer to output features. ;

[0026] 3. Final output of the MS time series modeling network Depend on and It is obtained by adding each element together.

[0027] Sparse attention module:

[0028] 1. Sparse attention mechanism operation The first step requires first determining the input feature vector. The query is derived by projecting three different parameter matrices. ), Key ( ) and Value ( )matrix;

[0029] 2. Subsequently, the obtained projection matrix needs to be processed. Perform random sampling, the number of samples is , This indicates the length of the input data. A new matrix will be obtained after sampling. ;

[0030] 3. Utilize the sampled Matrix and Matrix to calculate sparsity measure :

[0031]

[0032] in , They are respectively represented as Matrix and Eigenvectors in a matrix The feature dimension of the attention matrix;

[0033] 4. The matrix is ​​then modified to retain only... highest indivual Vectors in As a custom hyperparameter, the remaining positions in the matrix are filled with 0, and the final new matrix is ​​defined as follows: ;

[0034] 5. Utilizing sparse matrices Replace the original matrix Perform correlation matrix operations to obtain a hollow matrix containing several attention scores and a value of 0. ;

[0035] 6. The position of the value 0 in the matrix is ​​used Mean of all values ​​in the matrix To fill in the gaps. This is how the sparse attention mechanism is calculated. The final output result;

[0036] Mamba modules in two parallel branches:

[0037] 1. State-space model computation The first step requires processing the input feature matrix. Performing linear projection and further splitting yields three different mapping matrices. ;

[0038] 2. It is necessary to base it on the existing three parameter matrices. Calculate the new discretization Matrix and The matrix is ​​calculated using the following steps:

[0039]

[0040] in To perform exponentiation on the values ​​within the matrix, It is the identity matrix. This represents the generated static parameter matrix;

[0041] 3. Perform state-space model operations The output result is ,So The value of each vector in the vector can be obtained by The vector values ​​in the matrix are calculated accordingly, and the specific calculation steps are as follows:

[0042]

[0043] State-space model computation The output result is Each vector in the equation can be obtained using the above calculation formula, thus obtaining... The final output result, where This represents the static parameter matrix generated by the initialization system;

[0044] Spatiotemporal MLP Networks:

[0045] 1. The first step in the operation of a spatiotemporal MLP network is to compute the temporal modeling vectors in the three branches. By concatenating the matrices, a concatenated matrix is ​​obtained.

[0046] 2. Use a fully connected layer to stitch the matrix together. Dimensionality reduction is performed to obtain the fused moments after dimensionality reduction. ;

[0047] 3. For the fusion matrix To perform spatiotemporal MLP operations, the information in the time dimension is first fused: After transposition, the result is passed through a linear layer and an activation function to obtain the output processed in the time dimension. ;

[0048] 4. Then perform information fusion along the feature dimensions: After transposition, the result is passed through a linear layer and an activation function to obtain the output processed in the time dimension. ;

[0049] 5. The final output of the network requires dimensionality reduction of the feature matrix output. This is achieved by using two fully connected layers to reduce the dimensionality of each dimension of the feature matrix, resulting in the final capacity prediction result. , which represents the estimated SOH value for the current charging cycle;

[0050] 6. The network for predicting internal resistance branches has the same structure as the MSMixe network model. The final output of the network is the predicted internal resistance using... express.

[0051] In a second aspect, the present invention also provides a lithium battery health status estimation system based on edge-cloud integration, comprising the following modules:

[0052] The cloud data module is used to collect battery data, upload it to the cloud for preprocessing, and create a dataset from the preprocessed battery data.

[0053] The health status estimation module is used to build the cloud-based lithium battery health status estimation network model MSMixer, which includes a sequentially cascaded multi-scale feature extraction network, a parallel MS time-series modeling network with sparse attention and Mamba modules, and a spatiotemporal MLP network.

[0054] The vehicle-side prediction module is used to input the centralized data of the dataset into the MSMixer network for training, and uses it as the general initial model parameters for vehicle-side training. The vehicle-side uses local real-time data to fine-tune the cloud model, and finally outputs the cloud-based lithium battery health status estimation result.

[0055] This invention has the following characteristics and beneficial effects:

[0056] By combining edge and cloud computing, this method fully leverages large-scale historical data in the cloud and real-time data from the vehicle, ensuring both the model's global generalization ability and its adaptability to individual vehicle battery characteristics. The MSMixer network built in the cloud efficiently extracts multi-scale features, enhancing the accuracy and robustness of health status estimation. Introducing actually measured internal resistance parameters as a loss function constraint during fine-tuning on the vehicle not only avoids overfitting due to insufficient data but also ensures consistency between the model's estimation results and the battery's physical characteristics. This method possesses high accuracy and provides reliable technical support for electric vehicle battery health management. Attached Figure Description

[0057] Figure 1 This is a framework diagram of an embodiment of the present invention;

[0058] Figure 2 This is a diagram of the MSMixer neural network according to an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of the feature extraction module according to an embodiment of the present invention;

[0060] Figure 4 This is a schematic diagram of the runaway MLP module according to an embodiment of the present invention. Detailed Implementation

[0061] This invention provides a method for estimating the state of health of lithium batteries based on edge-cloud integration, such as... Figure 1 and Figure 2 As shown, it includes the following steps:

[0062] In a first aspect, this invention provides a method for estimating the state of health of a lithium battery based on edge-cloud integration, comprising the following steps:

[0063] S1. After the battery data is collected, it is uploaded to the cloud for preprocessing, and the preprocessed battery data is used to create a dataset. The battery data includes the voltage, current, temperature and corresponding internal resistance and battery capacity data of the battery pack during its historical operation, which are stored in the cloud.

[0064] S2. Construct a cloud-based lithium battery health status estimation network model—MSMixer. The MSMixer network model includes a multi-scale feature extraction network, a parallel MS temporal modeling network with sparse attention and Mamba modules, and a spatiotemporal MLP network. Different data correspond to their respective branches of the feature extraction network and temporal modeling network. The feature vectors from the four branches are concatenated and then fused and uniformly calculated by the spatiotemporal MLP networks of two branches.

[0065] S3. Input historical cloud data into the MSMixer network for training, save the trained model, and use it as the general initial model parameters for vehicle-side training. The vehicle-side uses local real-time data to fine-tune the cloud model, and after fine-tuning, completes the lithium battery health status and outputs the lithium battery health status estimation result. During this process, the vehicle-side directly measures the battery internal resistance parameter and introduces it as a constraint into the fine-tuning stage as a loss function. By adding an internal resistance loss term, the model is updated, thereby improving the model's adaptability to the specific battery characteristics of the vehicle.

[0066] In step S1, the battery data is collected by uploading actual vehicle operating data once every 3 seconds, and normalizing the data uploaded to the cloud to obtain voltage, current, temperature and corresponding battery capacity and internal resistance data, which are used as the dataset for network training.

[0067] The method for preprocessing the battery data collected in step S1 is as follows:

[0068] Normalization is performed using the min-max method. The normalized time series data expression is as follows:

[0069]

[0070] in, It is the original input sequence, namely the timing information of voltage, current, and temperature; and These are the maximum and minimum values ​​of the input sequence, respectively.

[0071] In step S1, the original unprocessed voltage, current, and temperature data are... , , After performing the above preprocessing operations, the input data for the corresponding network branch is obtained. , , Where L represents the number of sampling points for voltage, current, and temperature data in one cycle of charging data. All input data will be divided into training, validation, and test sets in a 5:1:1 ratio.

[0072] In step S2, the specific steps for processing the input data are as follows:

[0073] S2-1. The multi-scale feature extraction network with different branches extracts rich semantic information features from the normalized voltage, current and temperature data respectively.

[0074] S2-2. The extracted semantic information features are modeled temporally using the MS temporal modeling network. The Mamba module can effectively address the problem of uneven feature distribution in the input data, while the sparse attention module can provide direct modeling capability for long-range dependencies. The combination of the two outputs a more reliable temporal modeling feature vector.

[0075] S2-3. The two-branch spatiotemporal MLP network concatenates and fuses the temporal modeling vectors of the three branches, and reduces the dimensionality by fusing the information of the time dimension and the feature dimension, and outputs the final predicted internal resistance and capacity results respectively.

[0076] In step S2-1, one of the methods for extracting voltage data features from a multi-scale feature extraction branch is as follows:

[0077] S2-1-1, First, voltage data The initial features are obtained by passing the data through a one-dimensional convolutional neural network with a kernel size of 1 and a stride of 1, followed by average pooling.

[0078]

[0079] in The initial feature vector, Indicates the number of feature channels. This indicates the average pooling operation. Represents the ReLU activation function. Indicates the kernel size as Convolution operation;

[0080] S5-2, the initial features will undergo a multi-scale feature extraction process consisting of three parallel one-dimensional convolutions with different receptive field sizes, such as... Figure 3 As shown, the specific calculation steps are as follows:

[0081]

[0082] in Multi-scale feature vectors extracted from three convolutional kernels with different receptive field sizes;

[0083] S2-1-2. The final output of the multi-scale feature extraction module requires concatenation and convolution fusion of feature vectors at multiple different scales. The specific calculation steps are as follows:

[0084]

[0085] in This represents the final output vector of the multi-scale feature extraction module. This indicates a vector concatenation operation;

[0086] In step S2-2, the method for modeling the feature vector by one of the MS time-series modeling branches is as follows:

[0087] S2-2-1. The feature vectors obtained after the multi-scale feature extraction module will enter the sparse attention module and the Mamba module in the MS time series modeling module, respectively. The calculation steps of the sparse attention module branch in the two parallel branches are as follows:

[0088]

[0089] in This represents the final output value of the sparse attention module branch. This indicates a normalization operation. This represents a single-layer feedforward network. This represents the sparse attention mechanism operation;

[0090] S2-2-2, The calculation steps of the Mamba module in the two parallel branches are as follows:

[0091]

[0092] in This is the final output value of the Mamba module branch. This represents a fully connected layer. This represents the state-space model computation. This represents the Silu activation function;

[0093] S2-2-3. The final output of the MS temporal modeling network needs to combine the temporal modeling features of the sparse attention module branch and the Mamba module branch. The specific combination method is as follows:

[0094]

[0095] Step S2-2-1 includes the following sub-steps:

[0096] S2-2-1-1, Sparse Attention Mechanism Operation The first step requires first determining the input feature vector. Different projections yield the Query ( ), Key ( ) and Value ( The matrix is ​​calculated using the following steps:

[0097]

[0098] in These represent three different parameter matrices;

[0099] S2-2-1-2, Next, the obtained projection matrix needs to be processed. Perform random sampling, the number of samples is A new matrix will be obtained after sampling. ;

[0100] S2-2-1-3, Utilizing the sampled... Matrix and Matrix to calculate sparsity measure , The calculation formula is as follows:

[0101]

[0102] in, This represents the feature dimension of the attention matrix.

[0103] S2-2-1-4, To The matrix is ​​then modified to retain only... highest indivual Vectors in For custom hyperparameters, the remaining positions in the matrix are filled with 0;

[0104] S2-2-1-4, Utilizing Sparse Matrices Replace the original matrix Perform correlation matrix operations to obtain a hollow matrix containing several attention scores and a value of 0. The specific calculation formula is as follows:

[0105]

[0106] S2-2-1-5, will The position of the value 0 in the matrix is ​​used Mean of all values ​​in the matrix To fill in the gaps. This is how the sparse attention mechanism is calculated. The final output result;

[0107] Step S2-2-2 includes the following sub-steps:

[0108] S2-2-2-1, State-space model computation The first step requires processing the input feature matrix. After projection and further splitting, three different mapping matrices are obtained. The specific calculation steps are as follows:

[0109]

[0110] in This indicates that the matrix is ​​being split. These represent the parameter matrix and the bias parameters, respectively.

[0111] S2-2-2-2, It is necessary to base it on the existing three parameter matrices. Calculate the new discretization Matrix and The matrix is ​​calculated using the following steps:

[0112]

[0113] in This indicates that the values ​​within the matrix are subjected to exponentiation. Represents the identity matrix. This represents the generated static parameter matrix;

[0114] S2-2-2-3, Let the state-space model be operated The output result is ,So The value of each vector in the vector can be obtained by The vector values ​​in the matrix are calculated accordingly, and the specific calculation steps are as follows:

[0115]

[0116] State-space model computation The output result is Each vector in the equation can be obtained using the above calculation formula, thus obtaining... The final output result, where This represents the static parameter matrix generated by the initialization system;

[0117] In steps S2-3, such as Figure 4 As shown, the specific working method of the spatiotemporal MLP network branch used for capacity prediction is as follows:

[0118] S2-3-1, The first step in the operation of the spatiotemporal MLP network is to compute the temporal modeling vectors in the three branches. By concatenating the matrices, a concatenated matrix is ​​obtained. ;

[0119] S2-3-2, Concatenation matrix through a fully connected layer The specific computational steps for dimensionality reduction are as follows:

[0120]

[0121] in This is the fusion matrix after dimensionality reduction of the spliced ​​matrix;

[0122] S2-3-3, Regarding the fusion matrix To perform spatiotemporal MLP operations, the information in the time dimension must first be fused. The specific calculation steps are as follows:

[0123]

[0124] in The output results are processed for the time dimension. This indicates a transpose operation on the matrix;

[0125] S2-3-4 Output results of time dimension processing Next, information fusion is performed along the feature dimensions. The specific calculation steps are as follows:

[0126]

[0127] in The output result of processing the feature dimensions;

[0128] S2-3-5. The final output of the network needs to undergo dimensionality reduction processing on the output of the feature dimension processing. This is done by using two fully connected layers to reduce the dimensionality of each of the two dimensions of the feature matrix. The specific calculation steps are as follows:

[0129]

[0130] in The final capacity prediction result of the network output is represented by the estimated SOH value for the current charging cycle;

[0131] S2-3-6, The network steps for predicting the internal resistance branch are the same as those for predicting the capacity branch. The final output of the network is the predicted internal resistance result. express.

[0132] The hyperparameters of the MSMixer network model constructed in step S2 are determined by cross-validation, specifically including the sliding learning rate, number of iterations, and batch size.

[0133] The loss function used in step S3 is as follows:

[0134]

[0135] in, , These are the weighting coefficients. Used to constrain the prediction accuracy of SOH The actual internal resistance measured at the vehicle end is used as a "strong supervision signal" to force the model to maintain consistency with the physical measurement in internal resistance estimation. and Representing the estimated and actual values ​​of SOH, and This represents the estimated and actual values ​​of the measured internal resistance.

[0136] During the vehicle-side fine-tuning phase, the vehicle utilizes locally collected real-time data to personalize the cloud-based pre-trained model, enhancing its adaptability to specific vehicle battery characteristics. Specifically, the vehicle directly measures the battery's internal resistance parameters and incorporates this as a "strong supervision signal" into the model fine-tuning process. In the loss function design, in addition to the conventional error term constraining the accuracy of SOH prediction, an additional internal resistance loss term is added to force the model to maintain consistency with physical measurements in internal resistance estimation.

[0137] In a second aspect, the present invention also provides a lithium battery health status estimation system based on edge-cloud integration, comprising the following modules:

[0138] The cloud data module is used to collect battery data, upload it to the cloud for preprocessing, and create a dataset from the preprocessed battery data.

[0139] The health status estimation module is used to build the cloud-based lithium battery health status estimation network model MSMixer, which includes a sequentially cascaded multi-scale feature extraction network, a parallel MS time-series modeling network with sparse attention and Mamba modules, and a spatiotemporal MLP network.

[0140] The vehicle-side prediction module is used to input the centralized data of the dataset into the MSMixer network for training, and uses it as the general initial model parameters for vehicle-side training. The vehicle-side uses local real-time data to fine-tune the cloud model, and finally outputs the cloud-based lithium battery health status estimation result.

[0141] This invention conducts a comparative experiment on the battery pack SOH estimation performance using actual electric vehicle operating data, employing MAE and RMSE as evaluation metrics. The comparison methods utilize widely used prediction methods such as Convolutional Network-Long Short-Term Memory Network (CNN-LSTM), Temporal Convolutional Network (TCN), Transformer, and Recurrent Network (GRU). The test results are shown in Table 1.

[0142] Table 1

[0143]

[0144] Where MAE and RMSE are the mean absolute error and root mean square error loss functions, respectively, and SOH Error is the error in the online estimation of the battery pack's state of health (SOH). As shown in Table 1, the edge-cloud combined lithium battery health estimation method proposed in this invention can effectively achieve the SOH estimation task for electric vehicle battery packs while ensuring high accuracy.

[0145] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments, including components, without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A method for estimating the state of health of lithium batteries based on edge-cloud integration, characterized in that, Includes the following steps: S1. After battery data is collected, it is uploaded to the cloud for preprocessing. The preprocessed battery data is used to create a dataset. S2. Construct a cloud-based lithium battery health status estimation network model, MSMixer; S3. Input the collected data into the MSMixer network for training, and use it as the general initial model parameters for vehicle-side training. The vehicle-side uses local real-time data to fine-tune the cloud model, and outputs the lithium battery health status estimation result.

2. The lithium battery health status estimation method based on edge-cloud integration according to claim 1, characterized in that, In step S1, the battery data includes voltage, current, temperature, and corresponding internal resistance and battery capacity data of the battery pack during its historical operation, which are stored in the cloud. The preprocessing specifically involves normalizing the original, unprocessed voltage, current, and temperature data to obtain the input data for the corresponding network branches. , , The number of sampling points for voltage, current, and temperature data in one charging cycle is: .

3. The lithium battery health status estimation method based on edge-cloud integration according to claim 2, characterized in that, The MSMixer network model includes a multi-scale feature extraction network, a parallel MS temporal modeling network with sparse attention and Mamba modules, and a spatiotemporal MLP network. The specific implementation process is as follows: First, the multi-scale feature extraction networks of different branches extract semantic information features from the normalized voltage, current and temperature data respectively; Secondly, the extracted semantic information features are temporally modeled using the MS temporal modeling network; Finally, the two-branch spatiotemporal MLP networks concatenate and fuse the time-series modeling vectors of the voltage, current, and temperature branches, and reduce the dimensionality by fusing the information of the time dimension and the feature dimension, and output the final predicted internal resistance and capacity results respectively.

4. The lithium battery health status estimation method based on edge-cloud integration according to claim 3, characterized in that, The multi-scale feature extraction network is implemented as follows: First, the voltage data. The initial features are obtained by using a one-dimensional convolutional neural network with a kernel of 1 and then applying average pooling. ; The initial features are processed through a multi-scale feature extraction step consisting of three parallel one-dimensional convolutions of different receptive field sizes, resulting in multi-scale feature vectors extracted by convolution kernels of three different receptive field sizes. The three multi-scale feature vectors are concatenated and fused using convolution to obtain the final output vector of the multi-scale feature extraction module. .

5. The lithium battery health status estimation method based on edge-cloud integration according to claim 4, characterized in that, The MS temporal modeling network, which operates in parallel with the sparse attention module and the Mamba module, is implemented as follows: Feature vector In the MS time series modeling module, enter the sparse attention module and the Mamba module respectively; The calculation of the sparse attention module branches is as follows: After processing using a sparse attention mechanism, and then connecting it to the input via residual connections... The features are obtained by adding elements one by one and then normalizing. ; final characteristics After passing through the feedforward network, and then through the residual connection and input... The features are obtained by adding elements one by one and then normalizing them. ; The calculation in the Mamba module is as follows: After passing through a one-dimensional convolutional neural network with a kernel size of 1, features are obtained using the Silu activation function. ; final characteristics The state-space model SSM is processed by a fully connected layer to output features. ; The final output of the MS time series modeling network Depend on and It is obtained by adding each element together.

6. The lithium battery health state estimation method based on edge-cloud integration according to claim 5, characterized in that, The sparse attention mechanism operation first calculates based on the input feature vector. The result is obtained by projecting from three different parameter matrices. , and matrix; Then the projection matrix Perform random sampling, the number of samples is ,use This indicates that a new matrix will be obtained after sampling. ; Using the sampled Matrix and Matrix computation of sparsity metrics ,right The matrix is ​​then modified to retain only... The highest value indivual Vectors in As a custom hyperparameter, the remaining positions in the matrix are filled with 0, and the resulting new matrix is ​​defined as follows: ; Using sparse matrices Replace the original matrix Performing correlation matrix operations yields a hollow matrix containing several attention scores and a value of 0. ,Will The position where the median value is 0 is used The mean of all values ​​in the matrix is ​​filled to obtain the final output of the sparse attention mechanism operation.

7. The lithium battery health state estimation method based on edge-cloud integration according to claim 6, characterized in that, The state-space model SSM operation performs linear projection and further decomposition on the input features to obtain three different mapping matrices. ; According to the matrix Calculate the new discretization Matrix and matrix, Depend on and static parameter matrix The result is obtained by performing an exponential operation on the values ​​within both matrices. Depend on Subtract from the identity matrix element by element and then multiply by Finally multiplied by get; Let the output of the state-space model operation be ,So The value of each vector in the vector is determined by... The vector values ​​in the new matrix , , and static parameter matrix The corresponding state space is calculated.

8. The lithium battery health state estimation method based on edge-cloud integration according to claim 7, characterized in that, The specific implementation process of the spatiotemporal MLP network is as follows: The first step in working a spatiotemporal MLP network is to compute the temporal modeling vectors in the three branches. By concatenating the matrices, a concatenated matrix is ​​obtained. ; splicing matrix through a fully connected layer Dimensionality reduction is performed to obtain the fused moments after dimensionality reduction. ; For fusion matrix To perform spatiotemporal MLP operations, the information in the time dimension is first fused: After transposition, the result is passed through a linear layer and an activation function to obtain the output processed in the time dimension. ; right Then perform information fusion along the feature dimensions: After transposition, the result is passed through a linear layer and an activation function to obtain the output processed in the time dimension. ; The final output is the result of processing the feature dimensions. The feature matrix is ​​dimensionality reduced in both dimensions by a fully connected layer, resulting in the final capacity prediction result. , which represents the estimated SOH value for the current charging cycle; The network for predicting internal resistance branches has the same structure as the MSMixe network model. The final output of the network predicts the internal resistance using... express.

9. A lithium battery health state estimation system based on edge-cloud integration, used to implement the lithium battery health state estimation method according to any one of claims 1 to 8, characterized in that, Includes the following modules: The cloud data module is used to upload battery data to the cloud for preprocessing after collection, and to create a dataset from the preprocessed battery data. The health status estimation module is used to build the cloud-based lithium battery health status estimation network model MSMixer, which includes a sequentially cascaded multi-scale feature extraction network, a parallel MS time-series modeling network with sparse attention and Mamba modules, and a spatiotemporal MLP network. The vehicle-side prediction module is used to input the centralized data of the dataset into the MSMixer network for training, and uses it as the general initial model parameters for vehicle-side training. The vehicle-side uses local real-time data to fine-tune the cloud model, and finally outputs the cloud-based lithium battery health status estimation result.

Images