A lithium ion battery state of health estimation method considering temperature and charging current
By constructing a TCN-Transformer parallel model and fusing local and global features, the accuracy problem of lithium-ion battery health state estimation is solved, achieving efficient and robust SOH estimation that adapts to battery aging behavior under different conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING UNIV OF TECH
- Filing Date
- 2025-04-09
- Publication Date
- 2026-06-19
AI Technical Summary
Existing technologies make it difficult to accurately estimate the health status of lithium-ion batteries, which affects the safety and reliability of electric vehicles.
A parallel neural network TCN-Transformer model is constructed, which integrates local and global features and incorporates temperature and charging current variations. The model parameters are optimized through meta-learning and gradient descent to achieve efficient and fast SOH estimation.
It improves the accuracy and robustness of lithium-ion battery health state estimation, reduces estimation errors, and enhances adaptability and computational efficiency under different conditions.
Smart Images
Figure CN120352774B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of lithium-ion battery health status estimation technology, specifically a lithium-ion battery health status estimation method that takes into account temperature and charging current. Background Technology
[0002] Electric vehicles (EVs) offer advantages such as high energy density, long cycle life, and low self-discharge rate. However, with increasing charge-discharge cycles, irreversible chemical reactions within the battery lead to a gradual decline in performance. To quantify this degradation, a state of health (SOH) metric is typically used, defined as the ratio of the current maximum discharge capacity to the initial capacity. Accurate SOH estimation is crucial for ensuring the safety and reliability of EVs and protecting the safety of users and property. Summary of the Invention
[0003] To address the aforementioned problems in the prior art, this invention provides a method for estimating the health status of lithium-ion batteries that considers temperature and charging current. This method possesses robust generalization ability, high accuracy, and strong potential for practical applications, enabling efficient and rapid assessment of the health status of lithium-ion batteries.
[0004] To achieve the above objectives, this invention proposes a method for estimating the state of health of lithium-ion batteries considering temperature and charging current. This method includes the following steps:
[0005] S1. Obtain data on different charging segments of lithium-ion batteries: Conduct cyclic charge and discharge tests on the battery through experiments until the terminal voltage reaches 4.2V and the current drops to the cutoff current of 0.02C, and record the battery charging voltage and capacity data in real time.
[0006] S2, Data Processing: The voltage obtained in S1 is sliced to obtain the charging capacity sequence corresponding to different voltage sequences, and the capacity increment is used as a health feature.
[0007] S3. Construct a parallel model of Temporal Convolutional Network (TCN) and Transformer, and integrate local and global features: integrate the parallel neural network TCN-Transformer with meta-learning, while taking into account the changes in temperature and charging current.
[0008] S4. Initialize parameters and perform model pre-training: Apply meta-learning, refine the initialization parameters of the parallel neural network through gradient descent, and perform pre-training using a small amount of data;
[0009] S5, TCN-Transformer Parallel Model Training: Use the data obtained in S2 to run the optimized TCN-Transformer parallel model and three deep learning algorithms TCN-Transformer, Transformer and CNN-LSTM, estimate SOH under the same conditions and compare the estimation results;
[0010] S6. Model Validation: The TCN-Transformer parallel model with fused meta-learning is validated under different temperature and cross-cell material conditions.
[0011] Preferably, in S2, the calculation process expression is:
[0012] Q = [Q1, Q2, ..., Q i ,…,Q n ];
[0013]
[0014] In the formula, Q represents the charging capacity, n represents the number of segments, S represents the voltage step size, Floor represents rounding down to the nearest integer, and V star Indicates the starting voltage, V stop The terminator represents the termination voltage, and ΔV represents the interval between voltage segments.
[0015] Preferably, in S2, the accurate state of health (SOH) estimate of the lithium-ion battery based on partial random charging data is used to estimate the given charging capacity sequence Q. i Divide the data into multiple segments and obtain the capacity increment order for each segment. The calculation formula is as follows:
[0016] Q i =[Q i1 Q i2 ,…,Q im ];
[0017] ΔQ i =[Q i1 Q i2 ,…,Q im ]-Q i1 ;
[0018] In the formula, m represents the number of segments into which the charging capacity is further divided, and ΔQ i This indicates the order of capacity increments for each segment.
[0019] Preferably, in S3, the specific steps for building a parallel model of the Temporal Convolutional Network (TCN) and the Transformer include:
[0020] S311. Extract the capacity increment sequence from all random partial charging segments in each cycle as a health feature for network training; extract the capacity increment sequence from a single random partial charging segment in each cycle.
[0021] S312. In a parallel hybrid network, an embedded attention mechanism is combined to capture the spatiotemporal characteristics of health data;
[0022] S313. Pre-train on a large number of training samples of the parallel hybrid network, apply meta-learning, refine the initialization parameters of the parallel network through gradient descent, fine-tune the network using the target dataset, and optimize the meta-model using the limited data available in the target domain.
[0023] S314. Use TCN for one-dimensional time series prediction and to preserve the time order of data, given a time series X and a corresponding output Y; where, when the kernel size of TCN is set to 2, the output can be represented as:
[0024] Y t =W1*X t +W2*X t-1 ;
[0025] In the formula, W1 and W2 represent the weights of the convolution kernel, and the weights are shared across all input sequences;
[0026] S315. Introduce gaps between kernel elements to expand the receptive field of the convolution process, connect and integrate residual blocks into the TCN, and directly add the input data after each convolution layer to the output to establish cross-layer connections.
[0027] S316. The Transformer architecture, based on sequence-to-sequence, is used to eliminate loops and convolution operations. Dot product attention and multi-head attention mechanisms are used to capture long-term dependencies and transform the connection results into the required dimensions.
[0028] Preferably, in S315, the residual block is calculated as follows:
[0029] Y = f(X + Ψ(X));
[0030] In the formula, Ψ(X) is the input of the residual block, f is the softmax activation function, and Y is the output of the residual block.
[0031] Preferably, in S316, the specific steps for converting the connection result into the required dimension are: position encoding the input data, using functions of different frequencies of sine and cosine to calculate position information, and labeling the data as relative or absolute data. The calculation formula is as follows:
[0032] PE(pos,2i)=sin(pos / 10000 2i / dmodel );
[0033] PE(pos,2i+1)=cos(pos / 10000 2i / dmodel );
[0034] In the formula, sin represents the sine function, cos represents the cosine function, pos represents the position of the current element, i represents different dimensions, dmodel represents the dimension of the model, and PE(pos,2i) represents the position code calculated at position pos and dimension 2i.
[0035] Preferably, in S316, when capturing long-term dependencies based on dot product attention, the dot product value is prevented from being in dimension d. k To prevent gradient vanishing and unstable gradients during attention calculation, scaling is performed when the value increases too much. The calculation formula is as follows:
[0036]
[0037] In the formula, Attention indicates that the self-attention mechanism is completed by multiple dot product operations, Q represents the query vector, K is the key vector, V represents the value vector, softmax represents the activation function, and d k This represents the dimension of the key vector.
[0038] Preferably, the multi-head attention mechanism is based on the self-attention mechanism, and the calculation formula is as follows:
[0039] Multihead(Q,K,V)=Concat(head1,...,head n W o ;
[0040] In the formula, head1,...,head n W represents the output result of each header. o This represents a weight matrix used to transform the concatenated result to the dimensions required by other parts of the model. Concat means concatenating the output results along the feature dimension. Multihead means the steps of concatenating and transforming the outputs of multiple independent attention heads.
[0041] Preferably, the TCN consists of causal convolution, dilated convolution, and residual connections, while the Transformer consists of an encoder, a decoder, a self-attention mechanism, a positional encoder, and a fully connected layer. The TCN branch uses dilated convolution to model local temporal dependencies, and the Transformer uses a self-attention mechanism to capture long-range dependencies.
[0042] Preferably, in S3, the specific steps for fusing local and global features are as follows:
[0043] S321. Concatenate the feature vectors of TCN and Transformer to form a unified representation. The calculation formula is as follows:
[0044]
[0045] In the formula, d T d′ is the output size of the TCN branch. T F represents the output size of the transformer branch. TCN F represents a TCN neural network. Transformer Describing the Transformer model, F conact This means concatenating the two feature tensors output by the two models according to the feature number dimension;
[0046] S322. Integrate features by applying a linear transformation to project the connected features into a low-dimensional space. The calculation formula is as follows:
[0047] F fused =F concat w fusion +b fusion ;
[0048] In the formula, w fusion and b fusion These are the weights and biases of the linear layer;
[0049] S323. Estimate SOH by passing the fused features through a fully connected layer. The calculation formula is as follows:
[0050] SOH=FC(F fused ) = F fused w out +b out ;
[0051] In the formula, w out and b out These are the weights and biases of the output layer.
[0052] Therefore, this invention proposes a method for estimating the state of health of lithium-ion batteries that considers temperature and charging current, with the following advantages:
[0053] (1) The TCN-Transformer model constructed in this invention can quickly adapt to new tasks and achieve high performance with limited data and iterations. This model promotes deep learning of domain knowledge in the source dataset, facilitates efficient knowledge transfer, and enhances the flexibility of health feature extraction in real-world applications.
[0054] (2) The parallel structure of the present invention effectively captures local and global dependencies, overcomes the information bottleneck of serial architecture, reduces latency, improves computational efficiency, and enhances the robustness and adaptability of SOH estimation under different operating conditions.
[0055] (3) The feature extraction and SOH estimation methods proposed in this invention are superior to other deep learning algorithms, and have higher accuracy and robustness in battery aging behavior modeling.
[0056] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0057] Figure 1 This is a schematic diagram of the health features extracted in the first cycle of a lithium-ion battery health state estimation method that takes into account temperature and charging current according to the present invention.
[0058] Figure 2 This is a schematic diagram of the state of health estimation method for lithium-ion batteries that takes into account temperature and charging current, and the results of different SOH estimation algorithms according to the present invention.
[0059] Figure 3 This is a schematic diagram comparing the error of a lithium-ion battery state of health estimation method that considers temperature and charging current with different SOH estimation algorithms according to the present invention.
[0060] Figure 4 This is a flowchart illustrating a method for estimating the health status of a lithium-ion battery that takes into account temperature and charging current, according to the present invention. Detailed Implementation
[0061] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of this application.
[0062] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.
[0063] like Figures 1-4 As shown, a method for estimating the state of health of a lithium-ion battery considering temperature and charging current according to an embodiment of the present invention specifically includes the following steps:
[0064] S1. Obtain data on different charging segments of lithium-ion batteries: Conduct cyclic charge and discharge tests on the battery through experiments until the terminal voltage reaches 4.2V and the current drops to the cutoff current of 0.02C, and record the battery charging voltage and capacity data in real time.
[0065] S2. Data Processing: The voltage obtained in S1 is sliced to obtain charging capacity sequences corresponding to different voltage sequences, and the capacity increment is used as a health feature; the calculation process expression is:
[0066] Q = [Q1, Q2, ..., Q i ,…,Q n ];
[0067]
[0068] In the formula, Q represents the charging capacity, n represents the number of segments, S represents the voltage step size, Floor represents rounding down to the nearest integer, and V star Indicates the starting voltage, V stop The terminator represents the termination voltage, and ΔV represents the interval between voltage segments.
[0069] In S2, to achieve accurate State of Health (SOH) estimation of lithium-ion batteries based on partially random charging data, the given charging capacity sequence Q is... i Divide the data into multiple segments and obtain the capacity increment order for each segment. The calculation formula is as follows:
[0070] Q i =[Q i1 Q i2 ,…,Q im ];
[0071] ΔQ i =[Q i1 Q i2 ,…,Q im ]-Q i1 ;
[0072] In the formula, m represents the number of segments into which the charging capacity is further divided, and ΔQ i This indicates the order of capacity increments for each segment.
[0073] S3. Construct a parallel model of Temporal Convolutional Network (TCN) and Transformer to integrate local and global features: Integrate the parallel temporal convolutional network (TCN-Transformer) architecture with meta-learning, while taking into account the changes in temperature and charging current.
[0074] TCN consists of causal convolutions, dilated convolutions, and residual connections, while Transformer consists of an encoder, a decoder, a self-attention mechanism, a positional encoder, and fully connected layers. The TCN branch uses dilated convolutions to model local temporal dependencies, while Transformer uses a self-attention mechanism to capture long-range dependencies.
[0075] In S3, the specific steps for building a parallel model of Temporal Convolutional Network (TCN) and Transformer include:
[0076] S311. Extract the capacity increment sequence from all random partial charging segments within each cycle as a health feature for network training; extract the capacity increment sequence from a single random partial charging segment within each cycle, enhancing the flexibility of health feature extraction in real-world applications.
[0077] S312. In a parallel hybrid network, an embedded attention mechanism is combined to capture the spatiotemporal characteristics of health data;
[0078] S313. Pre-training on a large number of training samples to develop a robust model with optimized parameters, applying meta-learning, and refining the initialization parameters of the parallel network through gradient descent, enabling the model to quickly adapt to new tasks and achieve high performance with limited data and iterations. This process promotes deep learning of the domain knowledge of the source dataset and facilitates efficient knowledge transfer.
[0079] Fine-tuning the network using the target dataset and optimizing the meta-model with the limited data available in the target domain enhances the network's adaptability while maintaining accuracy.
[0080] S314. Use TCN for one-dimensional time series prediction and to preserve the time order of data, given a time series X and a corresponding output Y; where, when the kernel size is set to 2, the output can be represented as:
[0081] Y t =W1*X t +W2*X t-1 ;
[0082] In the formula, W1 and W2 represent the weights of the convolution kernel, and the weights are shared across all input sequences;
[0083] S315. By introducing gaps between kernel elements to expand the receptive field of the convolution process, and integrating residual connections into the TCN, the input data after each convolution layer is directly added to the output to establish cross-layer connections. This method enables the network to capture dependencies over a longer time range without increasing the number of parameters, alleviates the gradient vanishing problem, makes the network easier to train, and also enhances the feature learning ability. When processing long-term sequence data, it improves performance and convergence speed.
[0084] S316. The Transformer architecture, based on sequence-to-sequence, is used to eliminate loops and convolution operations. Dot product attention and multi-head attention mechanisms are used to capture long-term dependencies and transform the connection results into the required dimensions.
[0085] In S315, the residual block is calculated as follows:
[0086] Y = f(X + Ψ(X));
[0087] In the formula, Ψ(X) is the output of the residual block, f is the softmax activation function, and Y is the output of the residual block.
[0088] In S316, the specific steps for converting the connection result into the required dimension are as follows: Position encoding of the input data is performed using functions of different frequencies of sine and cosine to calculate position information, and the data is labeled as relative or absolute data. The calculation formula is as follows:
[0089] PE(pos,2i)=sin(pos / 10000 2i / dmodel );
[0090] PE(pos,2i+1)=cos(pos / 10000 2i / dmodel );
[0091] In the formula, sin represents the sine function, cos represents the cosine function, pos represents the position of the current element, i represents different dimensions, dmodel represents the dimension of the model, and PE(pos,2i) represents the position code calculated at position pos and dimension 2i.
[0092] In S316, when capturing long-term dependencies based on dot product attention, it is necessary to prevent the dot product value from being in dimension d. k To prevent gradient vanishing and unstable gradients during attention calculation, scaling is performed when the value increases too much. The calculation formula is as follows:
[0093]
[0094] In the formula, Attention represents the autonomous attention mechanism accomplished by multiple dot product operations, Q represents the query vector, K is the key vector, V represents the value vector, softmax represents the activation function, and d k This represents the dimension of the key vector.
[0095] The multi-head attention mechanism is based on the acquisition of autonomous willpower, and the calculation formula is as follows:
[0096] Multihead(Q,K,V)=Concat(head1,...,head n Wo ;
[0097] In the formula, head1,...,head n W represents the output result of each header. o This represents a weight matrix used to transform the concatenated result to the dimensions required by other parts of the model. Concat means concatenating the output results along the feature dimension. Multihead means the steps of concatenating and transforming the outputs of multiple independent attention heads.
[0098] In S3, the specific steps for fusing local and global features are as follows:
[0099] S321. Concatenate the feature vectors of TCN and Transformer to form a unified representation. The calculation formula is as follows:
[0100]
[0101] In the formula, d T d′ is the output size of the TCN branch. T F represents the output size of the transformer branch. TCN F represents a TCN neural network. Transformer Describing the Transformer model, F conact This means concatenating the two feature tensors output by the two models according to the feature number dimension;
[0102] S322. Integrate features by applying a linear transformation to project the connected features into a low-dimensional space. The calculation formula is as follows:
[0103] F fused =F concat w fusion +b fusion ;
[0104] In the formula, w fusion and b fusion These are the weights and biases of the linear layer;
[0105] S323. Estimate SOH by passing the fused features through a fully connected layer. The calculation formula is as follows:
[0106] SOH=FC(F fused ) = F fused w out +b out ;
[0107] In the formula, w out and b out These are the weights and biases of the output layer.
[0108] S4. Initialize parameters and pre-train the model: Apply meta-learning, refine the initialization parameters of the parallel network through gradient descent, and pre-train using a small amount of data;
[0109] S5. TCN-Transformer Parallel Model Training: Use the data obtained in step S2 to run the optimized TCN-Transformer parallel model and three deep learning algorithms, TCN-Transformer, Transformer, and CNN-LSTM, to estimate SOH under the same conditions and compare the estimation results.
[0110] S6. Model Validation: The model is validated under different temperatures and conditions across battery materials.
[0111] Therefore, this invention provides a method for estimating the state of health (SOH) of lithium-ion batteries that considers temperature and charging current. This method improves the adaptability of the SOH of lithium-ion batteries under different operating temperatures, charging currents, and battery chemistry properties, and effectively reduces the maximum estimation error under cross-temperature and charging current conditions. When applied to different battery types, the need for aging data is reduced by 50%, demonstrating robust generalization, high accuracy, and strong potential for practical applications.
[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A lithium-ion battery state-of-health estimation method considering temperature and charging current, characterized by, The method includes the following steps: S1. Obtain data on different charging segments of lithium-ion batteries: Conduct cyclic charge and discharge tests on the battery through experiments until the terminal voltage reaches 4.2V and the current drops to the cutoff current of 0.02C. Record the battery charging voltage and capacity data in real time, and collect battery aging data under different temperature and charging current conditions. S2, Data Processing: The voltage obtained in S1 is sliced to obtain the charging capacity sequence corresponding to different voltage sequences, and the capacity increment is used as a health feature. S3. Construct a parallel model of Temporal Convolutional Network (TCN) and Transformer, and integrate local and global features: Integrate the parallel neural network TCN-Transformer with meta-learning, and consider the changes in temperature and charging current. The consideration of temperature and charging current is achieved by using battery aging data under different temperature and charging current conditions, and using the meta-learning mechanism to adapt the model to new temperature or current conditions. In S3, the specific steps for building a parallel model of Temporal Convolutional Network (TCN) and Transformer include: S311. Extract the capacity increment sequence from all random partial charging segments in each cycle as a health feature for network training; extract the capacity increment sequence from a single random partial charging segment in each cycle. S312. In a parallel hybrid network, an embedded attention mechanism is combined to capture the spatiotemporal characteristics of health data; S313. Pre-train on a large number of training samples of the parallel hybrid network, apply meta-learning, refine the initialization parameters of the parallel network through gradient descent, fine-tune the network using the target dataset, and optimize the meta-model using the limited data available in the target domain. S314, using the TCN to make one-dimensional time series prediction and reserve data time sequence, and given a time series and the corresponding output ; wherein the kernel size of the TCN is set to 2, the output can be represented as: ； wherein and denote weights of the convolution kernel, which are shared across all input sequences. S315. Introduce gaps between kernel elements to expand the receptive field of the convolution process, connect and integrate residual blocks into the TCN, and directly add the input data after each convolution layer to the output to establish cross-layer connections. S316. The Transformer architecture based on sequence-to-sequence is adopted to eliminate loops and convolution operations. It captures long-term dependencies based on dot product attention and multi-head attention mechanisms and transforms the connection results into the required dimensions. S4. Initialize parameters and perform model pre-training: Apply meta-learning, refine the initialization parameters of the parallel neural network through gradient descent, and perform pre-training using a small amount of data; S5, TCN-Transformer Parallel Model Training: Use the data obtained in S2 to run the optimized TCN-Transformer parallel model and three deep learning algorithms TCN-Transformer, Transformer and CNN-LSTM, estimate SOH under the same conditions and compare the estimation results; S6. Model Validation: The TCN-Transformer parallel model with fused meta-learning is validated under different temperature and cross-cell material conditions. 2.The method of claim 1, wherein, In S2, the calculation process is expressed as follows: ； ； wherein, represents the charging capacity, represents the number of segments, represents the voltage step, represents rounding down to the nearest integer, represents the starting voltage, represents the ending voltage, represents the interval between segment voltages. 3.The method of claim 2, wherein, In S2, based on the accurate lithium-ion battery state of health SOH estimation of partial random charging data, the given charging capacity sequence Q i is divided into multiple segments, the capacity increment sequence of each segment is obtained, and the calculation formula is: ； ； In the formula, represents the number of segments into which the charging capacity is further divided, represents the order of capacity increments for each segment. 4.The method of claim 1, wherein, In S315, the residual block is calculated as follows: ； In the formula, For the input of the residual block, for Activation function This is the output of the residual block.
5. The method for estimating the state of health of a lithium-ion battery considering temperature and charging current according to claim 1, characterized in that, In S316, the specific steps for converting the connection result into the required dimension are as follows: Position encoding of the input data is performed using functions of different frequencies of sine and cosine to calculate position information, and the data is labeled as relative or absolute data. The calculation formula is as follows: ； ； In the formula, Represents the sine function. Represents the cosine function. Indicates the position of the current element. Representing different dimensions, Indicates the dimension of the model. Indicates the location The positional code is calculated on dimension 2i.
6. The method for estimating the state of health of a lithium-ion battery considering temperature and charging current according to claim 1, characterized in that, In S316, when capturing long-term dependencies using dot product attention, it is important to prevent the dot product value from changing in dimension. To prevent gradient vanishing and unstable gradients during attention calculation, scaling is performed when the value increases too much. The calculation formula is as follows: ； In the formula, This indicates that the self-attention mechanism is accomplished by multiple dot product operations. Represents the query vector. It is a key vector. Represents a value vector. Represents the activation function. This represents the dimension of the key vector.
7. The method for estimating the state of health of a lithium-ion battery considering temperature and charging current according to claim 6, characterized in that, The multi-head attention mechanism is based on the self-attention mechanism, and the calculation formula is: ； In the formula, This indicates the output result of each header. This represents a weight matrix used to transform the concatenated result to the dimensions required by other parts of the model. This indicates that the output results will be concatenated along the feature dimension. This describes the steps of splicing and transforming the outputs of multiple independent attention heads.
8. The method for estimating the state of health of a lithium-ion battery considering temperature and charging current according to claim 1, characterized in that, TCN consists of causal convolutions, dilated convolutions, and residual connections, while Transformer consists of an encoder, a decoder, a self-attention mechanism, a positional encoder, and fully connected layers. The TCN branch uses dilated convolutions to model local temporal dependencies, while Transformer uses a self-attention mechanism to capture long-range dependencies.
9. The method for estimating the state of health of a lithium-ion battery considering temperature and charging current according to claim 1, characterized in that, In S3, the specific steps for fusing local and global features are as follows: S321. Concatenate the feature vectors of TCN and Transformer to form a unified representation. The calculation formula is as follows: ； In the formula, For TCN branch output size, For the output dimensions of the transformer branch, This represents a TCN neural network. Represents the Transformer model. This means concatenating the two feature tensors output by the two models according to the feature number dimension; S322. Integrate features by applying a linear transformation to project the connected features into a low-dimensional space. The calculation formula is as follows: ； In the formula, and These are the weights and biases of the linear layer; S323. Estimate SOH by passing the fused features through a fully connected layer. The calculation formula is as follows: ； In the formula, and These are the weights and biases of the output layer.