A battery state of charge and remaining useful life joint estimation method and system based on physical information guidance and dual timing attention

Abstract

The present application relates to a kind of physical information guiding and dual timing attention battery state of charge and residual service life joint estimation method and system, belong to electric vehicle battery management technical field.For the poor physical interpretability of existing method, state of charge and residual service life time scale asynchronous cause insufficient precision and cascade error accumulation problem, the present application constructs multidimensional input tensor by feature engineering, extracts global latent space feature using timing context encoder, designs dual-branch irregular timing attention mechanism to decouple fast and slow variables, introduces physical information state integral module to embed charge conservation constraint, combined with voltage reconstruction decoder to provide physical regularization, through multi-objective joint optimization to realize end-to-end training.The present application improves the physical consistency and mechanism reliability of model, considers long-period aging trend and short-time dynamic response, avoids cascade error, enhances generalization ability and robustness under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of electric vehicle battery management technology, and relates to a method and system for jointly estimating battery state of charge and remaining service life using physical information guidance and dual temporal attention. Background Technology

[0002] As the power source for electric vehicles, the safety, reliability, and economy of the power battery directly determine the overall performance of the vehicle. The battery management system is a core component for monitoring and managing the power battery. Accurately obtaining the battery's state of charge and remaining lifespan is a crucial prerequisite for achieving precise control of the battery's condition, extending its service life, and ensuring the safe operation of the vehicle.

[0003] State of charge (SOC) characterizes the percentage of a battery's rated capacity remaining under specific operating conditions and is a primary basis for drivers to determine driving range. Remaining useful life (SMB) characterizes the number of charge-discharge cycles or driving distance a battery will have remaining before reaching the end of its lifespan, and is an important basis for assessing battery aging and developing maintenance strategies. Due to the complexity of the electrochemical reactions within power batteries and the complex operating conditions they face in actual vehicles, such as varying currents, temperatures, and vibrations, the internal parameters of batteries exhibit strong nonlinearity, time-varying characteristics, and hysteresis, making high-precision estimation of SOC and SMB a significant challenge.

[0004] Existing battery state estimation methods can be broadly categorized into three types. The first type is the physical-driven method based on electrochemical and equivalent circuit models. This type of method describes the dynamic behavior of the battery by establishing a set of physicochemical equations or an electrical equivalent model within the battery. However, the internal reaction mechanisms of power batteries are extremely complex, and accurate physical models often contain a large number of parameters that are difficult to measure directly. Furthermore, the computational load is high, making it difficult to meet the real-time requirements of automotive embedded devices. The second type is the completely data-driven method. This type of method utilizes machine learning or deep learning algorithms to mine the mapping relationship between external battery characteristic data and state labels. Although this method avoids complex mechanism modeling to some extent, purely data-driven models are often considered black boxes, and their estimation results often lack physical interpretability. Moreover, when faced with entirely new operating conditions not covered by training data, their generalization ability significantly decreases, easily leading to erroneous estimates that violate physical common sense. The third type is the traditional segmented estimation method. Existing technologies typically treat state-of-charge estimation and remaining service life prediction as two independent tasks, constructing separate models for sequential processing. Since the state of charge is a fast time-varying dynamic variable, while the remaining lifetime is a slow time-varying degradation variable, the two are significantly asynchronous in time scale. Independent modeling not only ignores the inherent coupling relationship between the two, but also causes the error to be amplified step by step in the cascading propagation.

[0005] Furthermore, existing methods often fail to effectively utilize the physical time information inherent in the sampling interval when processing non-uniformly sampled vehicle data, making it difficult for models to simultaneously capture short-term voltage fluctuations and long-term aging trends. Meanwhile, most deep learning models rely solely on end-label supervision during training, lacking consistency constraints on the battery's dynamic response process, further limiting the improvement of estimation accuracy. Therefore, how to integrate the advantages of physical mechanisms and data-driven approaches, solve the decoupling problem between state of charge and remaining lifetime on the time scale, and improve the physical consistency and generalization ability of models under complex operating conditions are urgent technical problems to be solved in the field of battery state estimation. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a method and system for jointly estimating battery state of charge and remaining useful life using physical information guidance and dual temporal attention.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for jointly estimating battery state of charge and remaining lifetime using physical information guidance and dual temporal attention includes the following steps: S1: Acquire the raw data generated by the power battery during operation. The raw data includes the total current of the battery pack, the total voltage of the battery pack, the cumulative mileage, the instantaneous vehicle speed, the maximum voltage of a single cell, the minimum voltage of a single cell, the sampling time interval between adjacent sampling points, the maximum temperature, the minimum temperature, and the cell voltage difference. S2: Perform feature engineering processing on the original data to construct a three-dimensional input tensor; S3: Input the three-dimensional input tensor into the temporal context encoder module, and output the global latent space state matrix and the local latent space state vector; S4: Input the global latent space state matrix and the local latent space state vector into the dual-branch irregular temporal attention mechanism module, and output the available capacity estimate and the window start point charge state estimate, respectively. S5: Input the available capacity estimate, the state of charge estimate at the window starting point, and the current data in the three-dimensional input tensor into the physical information state integration module, and recursively calculate the state of charge trajectory within the window based on the principle of charge conservation, and output the single-step capacity estimate and the single-step state of charge estimate. S6: Input the local latent space state vector, the charged state trajectory, and the current data into the voltage reconstruction decoder module, and output the reconstructed voltage value; S7: Construct a multi-objective joint loss function based on the single-step state of charge estimate, the single-step capacity estimate, the reconstructed voltage value, and the corresponding real label, and perform end-to-end training on the temporal context encoder module, the dual-branch irregular temporal attention mechanism module, the physical information state integration module, and the voltage reconstruction decoder module.

[0008] Furthermore, S2 specifically includes: Will t The principal input vector at time t is defined as:

[0009] in, Indicates the total current of the battery pack. Indicates the total voltage of the battery pack. Indicates cumulative mileage. Indicates instantaneous vehicle speed. Indicates the maximum voltage of a single unit. Indicates the minimum voltage of a single unit. Indicates the sampling time interval between adjacent sampling points; Will t The auxiliary feature vector at time step is defined as:

[0010] in, Indicates average temperature. Indicates the throughput per ampere-hour of a segment. Indicates the cell voltage difference. Indicates temperature difference; Will t The complete input column vector at time t is defined as:

[0011] in, T Indicates the transpose operation; Use a length of L The sliding window slices the complete input column vector to obtain the first... i Data slices :

[0012] Will B The data slices are stacked along the batch dimension to obtain a three-dimensional input tensor. The three-dimensional input tensor is then Z-score normalized.

[0013] Furthermore, the parameters in the auxiliary feature vector are defined as follows:

[0014] in, Indicates the maximum temperature. Indicates the minimum temperature;

[0015]

[0016]

[0017] in, Indicates the conversion factor for time units.

[0018] Furthermore, S3 specifically includes: The timing context encoder module consists of a two-layer gated recurrent unit (GRU); For the b Batch No. l Input tensor under each slice By updating the door Reset door Candidate hidden state and hidden state Perform recursive processing and update the formula as follows:

[0019]

[0020]

[0021]

[0022] in, This represents the Sigmoid activation function. This represents element-wise multiplication of vectors. This indicates updating the gate weight matrix. This represents the reset gate weight matrix. This represents the candidate hidden state weight matrix. This represents the update gate historical state weight matrix. This represents the weight matrix of the reset gate's historical states. This represents the candidate hidden state historical state weight matrix. This indicates updating the gate bias vector. This indicates the reset gate bias vector. This represents the candidate hidden state bias vector. Indicates the dimension of the hidden layer; Output the global latent space state matrix :

[0023] The hidden state at the last time step is taken as the local latent space state vector. .

[0024] Furthermore, S4 specifically includes: Capacity estimation branch: Process the global latent space state matrix using time-integral attention blocks to calculate capacity attention weights. :

[0025]

[0026] in, For stress score, To prevent tiny constants from being zero in logarithmic operations; The time-space integral degradation representation vector is obtained by weighted aggregation of the global latent space state matrix. It is then calibrated using a Channel Attention Block (CA Block) to output a degenerate representation vector for channel calibration. The normalized available capacity estimate is obtained by mapping through a multi-layer perceptron (MLP). ; Initial point charged state estimation branch: based on the local latent space state vector Using the global latent space state matrix as the key and value, the basic attention score is calculated as the query vector. And combined with an exponential decay factor based on the real time span Calculate the attention weights of the charged state :

[0027]

[0028]

[0029] in, Represents the query vector. Represents the key vector. Indicates the hidden layer dimension. This represents the learnable time decay constant; The global latent space state matrix is ​​weighted and aggregated to obtain the representation vector. After calibration by the channel attention block, the input to the multilayer perceptron is used to obtain the estimated state of charge at the window starting point. .

[0030] Furthermore, S5 specifically includes: The normalized available capacity estimate and the estimated state of charge at the starting point of the window Perform inverse normalization to obtain the inverse normalized capacity estimate. and inverse normalization window starting point charge state estimation :

[0031]

[0032] in, The standard deviation of the capacity labels in the training set. This represents the mean of the capacity labels in the training set. This represents the standard deviation of the charged state labels in the training set. This represents the mean of the charged state labels in the training set; The trajectory of the charged state within the window is recursively derived based on the ampere-hour integral equation:

[0033]

[0034] in, express t The cumulative charge consumption at any given time. Indicates the conversion factor for time units. This indicates that the gradient operation has been stopped. The last time step of the state-of-charge trajectory is taken as the single-step state-of-charge estimate, and the inverse normalized capacity estimate is taken as the single-step capacity estimate.

[0035] Furthermore, S6 specifically includes: exist t At time , the local latent space state vector The charged state trajectory and the current in the original data The data is stitched together along the feature channel dimension, and the terminal voltage at that moment is estimated using a multilayer perceptron.

[0036] in, This indicates a concatenation operation along the feature channel dimensions. This represents a voltage decoder multilayer sensor.

[0037] Furthermore, the multi-objective joint loss function in S7 is:

[0038] in, This indicates the estimated loss due to the state of charge at the end of the window. This indicates the estimated loss based on available capacity. Indicates voltage reconstruction loss, This represents the capacity loss weight hyperparameter. This represents the voltage reconstruction loss weighting hyperparameter; The , and The calculation formula is:

[0039]

[0040]

[0041] in, Indicates the first b The estimated state of charge of each sample slice at the end of the window. Indicates the first b The true label of the state of charge corresponding to each sample slice Indicates the first b Capacity estimates for each sample slice. Indicates the first b The true label of the volume corresponding to each sample slice Indicates the first b One sample in t Voltage reconstruction value at time t, Indicates the first b One sample in t The actual observed voltage value at that moment.

[0042] A joint estimation system for battery state of charge and remaining lifetime guided by physical information and dual temporal attention, comprising performing the method, including: The data acquisition and preprocessing module is used to execute S1 to S2 to obtain the three-dimensional input tensor; The temporal context encoding module is used to execute S3 and output the global latent space state matrix and the local latent space state vector. A dual-branch irregular temporal attention module is used to execute S4 and output the available capacity estimate and the window start point state of charge estimate; The physical information state integration module is used to execute S5 and output a single-step capacity estimate and a single-step state of charge estimate. The voltage reconstruction decoding module is used to execute S6 and output the reconstructed voltage value; The joint training module is used to execute S7 to perform end-to-end training on the system.

[0043] Furthermore, the system is deployed in a vehicle battery management system (BMS), which collects the raw data in real time and outputs the estimated state of charge and remaining life of the battery online through the trained system.

[0044] The beneficial effects of this invention are as follows: (1) This invention effectively overcomes the problem of poor physical interpretability in battery state estimation by deeply integrating physical mechanisms with deep learning. The physical information state integration module embeds the law of charge conservation as a priori constraint into the network architecture, which forces the estimation results of the model to follow the basic physical evolution law inside the battery. This significantly suppresses the abnormal output that may occur in pure black box models under unexpected operating conditions, which violates physical common sense, and greatly improves the consistency of the model's mechanism and engineering reliability.

[0045] (2) To address the inherent time-scale asynchrony challenge between the state of charge and remaining useful life, the dual-branch irregular temporal attention mechanism designed in this invention successfully decouples the fast and slow variables. The capacity estimation branch, by introducing a time interval as a physical prior, simulates the integral effect of the aging process, accurately capturing the slow decay trend over long periods. The initial point state of charge estimation branch, on the other hand, utilizes time-degradation attention to focus on recent observation data, effectively adapting to the drastic fluctuations in short-term dynamics. This divide-and-conquer strategy enables the model to simultaneously consider both long-term degradation trends and instantaneous dynamic responses, fundamentally solving the problem that a single model cannot accommodate features across multiple time scales.

[0046] (3) The joint estimation architecture proposed in this invention breaks the limitation of separating the state of charge estimation and the remaining useful life prediction in traditional methods. By sharing the underlying temporal feature extraction network and exchanging information in the high-level semantic space, the model can fully explore the inherent coupling relationship between the two tasks, avoid the error accumulation effect caused by cascaded estimation, and realize the synchronous output of high-precision state estimation results in a single forward inference, which significantly reduces computational redundancy.

[0047] (4) The introduction of the voltage reconstruction decoder module provides additional physical regularization constraints for the model. By forcing the model to learn the reverse reconstruction process from the latent space state to the terminal voltage, the model's ability to fit the external electrical characteristics of the battery is enhanced, so that the extracted latent space features not only contain label information, but also contain rich dynamic response information, which further improves the model's generalization performance and robustness under complex working conditions.

[0048] (5) By explicitly incorporating the sampling time interval into the attention weight calculation and ampere-hour integration process, the model can effectively utilize the time information in irregular sampling data, reduce the dependence on the stability of data acquisition frequency, and better adapt to the real vehicle operating environment. At the same time, by introducing a stopping gradient operation in the physical integration path, the interference of short-term current fluctuations on capacity estimation is cleverly isolated, ensuring the optimization stability of the slow time-varying capacity estimate, and further improving the convergence efficiency and estimation accuracy of the overall system.

[0049] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 For joint estimation workflow; Figure 2 For timing context encoder module; Figure 3 It is a two-branch irregular temporal attention module; Figure 4 For physical state integration module; Figure 5 This is a voltage reconstruction decoder module. Detailed Implementation

[0051] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0052] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0053] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0054] This invention provides a method for jointly estimating the state of charge and remaining life of a battery using physical information guidance and dual temporal attention. This method can be deployed in a vehicle battery management system (BMS) for real-time monitoring and state assessment of the power battery.

[0055] Example 1: Joint Estimation Method Based on Real Vehicle Data This embodiment describes in detail the specific workflow of the present invention in a real vehicle operating environment.

[0056] Step 1: Data Acquisition and Preprocessing (corresponding to...) Figure 1 ) The vehicle battery management system collects raw operating data of the power battery in real time. The system constructs the main input vector at time t, calculated using the following formula:

[0057] in, Indicates the total current of the battery pack. Indicates the total voltage of the battery pack. Indicates cumulative mileage. Indicates instantaneous vehicle speed. Indicates the maximum voltage of a single unit. Indicates the minimum voltage of a single unit. This indicates the sampling time interval between adjacent sampling points.

[0058] The system simultaneously constructs auxiliary feature vectors, calculated using the following formula:

[0059] in, Indicates average temperature. Indicates the throughput per ampere-hour of a segment. Indicates the cell voltage difference. This represents the temperature difference. The specific calculation method for the auxiliary features is as follows:

[0060]

[0061]

[0062]

[0063] In the formula, Indicates the maximum temperature. Indicates the minimum temperature. Indicates the conversion factor for time units.

[0064] At time t, the system concatenates the main input vector with the auxiliary feature vector to construct the complete input column vector:

[0065] in, T This indicates the transpose operation.

[0066] Subsequently, the system adopted a length of L A sliding window slices a continuous time series. i The formula for calculating each data slice is:

[0067] Will B The slices are stacked to form a three-dimensional input tensor:

[0068] To eliminate dimensional differences, the system performs Z-score normalization on the 3D input tensor, calculated using the following formula:

[0069] in, Indicates the first c The mean of each feature on the training set Indicates the first c The standard deviation of each feature on the training set. The processed tensor is then fed into the temporal context encoder module.

[0070] Step 2: Temporal Feature Extraction (corresponding to...) Figure 2 ) The temporal context encoder module consists of a two-level gated recurrent unit (GRU). For each time step of data in the input tensor... The module calculates recursively using the following formula: Update the retention ratio of historical information for updated door control:

[0071] Reset the degree of forgetting historical information of the door control:

[0072] Candidate hidden states are generated by combining the current input with the historical states after reset:

[0073] The final hidden state is obtained by weighted fusion of the previous hidden state and the candidate hidden state using the update gate:

[0074] in, This represents the Sigmoid activation function. This represents element-wise multiplication of vectors. This indicates updating the gate weight matrix. This represents the reset gate weight matrix. This represents the candidate hidden state weight matrix. This represents the update gate historical state weight matrix. This represents the weight matrix of the reset gate's historical states. This represents the candidate hidden state historical state weight matrix. This indicates updating the gate bias vector. This indicates the reset gate bias vector. This represents the candidate hidden state bias vector. Indicates the dimension of the hidden layer.

[0075] After processing by a two-layer gated loop unit, the module outputs the global latent space state matrix:

[0076] The hidden state at the last time step is extracted as the local latent space state vector. .

[0077] Step 3: Processing of bi-branch irregular temporal attention (corresponding to) Figure 3 ) The global latent space state matrix is ​​input into the bi-branch irregular temporal attention mechanism module.

[0078] The capacity estimation branch focuses on capturing slow, time-varying aging trends. This branch utilizes time-integral attention blocks to calculate stress scores. The capacity attention weights are calculated using the following formula:

[0079]

[0080] in, It represents a tiny constant to prevent logarithmic operations from being zero.

[0081] By performing weighted aggregation on the global latent space state matrix, the time-space integral degradation representation vector is obtained:

[0082] Subsequently, the Channel Attention Block (CA Block) calibrates the vector, and the channel weights are calculated using the following formula:

[0083] in, Represents the ReLU function. This represents the Sigmoid function. and The weight matrix is ​​a learnable matrix. and This is the bias vector.

[0084] The calibrated degradation characterization vector is:

[0085] Finally, a normalized estimate of the available capacity is obtained through multi-layer perceptron (MLP) mapping:

[0086] The initial-point charged state estimation branch focuses on capturing fast time-varying dynamic responses. This branch uses the local latent space state vector... Using the global latent space state matrix as the key and value, the base attention score is calculated:

[0087]

[0088]

[0089] in, Represents the query vector. Represents the key vector. and It is a learnable linear mapping matrix.

[0090] Construct an exponential decay factor based on the actual time span:

[0091] in, is a learnable time decay constant.

[0092] Calculate the attention weights for the charged state:

[0093] Weighted aggregation yields the representation vector:

[0094] After channel attention block calibration, the estimated state of charge at the window starting point is obtained through multilayer perceptron mapping:

[0095] Step 4: Physical Information State Integration (corresponding to) Figure 4 ) The physical information state integration module receives the normalized capacity estimate. and the estimated state of charge at the starting point The module first performs inverse normalization on these two estimates:

[0096]

[0097] in, and Let these represent the standard deviation and mean of the capacity labels in the training set, respectively. and represents the standard deviation and mean of the charged state labels in the training set, respectively.

[0098] Subsequently, based on the principle of charge conservation, the module recursively calculates the charged state trajectory within the window using the ampere-hour integral equation:

[0099]

[0100] in, This represents the cumulative charge consumption at time t. Indicates the conversion factor for time units. This indicates that the gradient operation is stopped.

[0101] Finally, the module outputs a single-step capacity estimate. and the single-step state-of-charge estimate at the end of the window. .

[0102] Step 5: Voltage Reconstruction and Joint Optimization (corresponding to...) Figure 5 ) During the training phase, the voltage reconstruction decoder module is activated. The module converts the local latent space state vector... The recursive obtained charge state trajectory and the original input current The data is stitched together along the feature channel dimension, and the terminal voltage at that moment is estimated using a multilayer perceptron.

[0103] in, This indicates a concatenation operation along the feature channel dimensions. This represents a voltage decoder multilayer sensor.

[0104] The system constructs a multi-objective joint loss function, which consists of the following three parts: Loss estimation at the end of the window's state of charge:

[0105] Estimated loss of available capacity:

[0106] Voltage reconstruction loss:

[0107] in, This represents the estimated state of charge of the b-th sample slice at the end of the window. Indicates the first b The true label of the state of charge corresponding to each sample slice Indicates the first b Capacity estimates for each sample slice. Indicates the first b The true label of the volume corresponding to each sample slice Indicates the first b One sample in t Voltage reconstruction value at time t, Indicates the first b One sample in t The actual observed voltage value at that moment.

[0108] The final total loss is the weighted sum of the three:

[0109] in, This represents the capacity loss weight hyperparameter. This represents the voltage reconstruction loss weight hyperparameter.

[0110] By minimizing the joint loss function, the system's temporal context encoder module, dual-branch irregular temporal attention mechanism module, physical information state integration module, and voltage reconstruction decoder module undergo end-to-end closed-loop training until the model converges.

[0111] Example 2: Cloud-based offline training and vehicle-mounted online deployment This embodiment describes the implementation process of the present invention in the collaborative working mode of cloud server and vehicle terminal.

[0112] On the cloud server side, the system executes the training process described in Example 1 using large-scale historical real-vehicle data. The cloud server has powerful computing capabilities, enabling it to process massive amounts of data and train a high-precision joint estimation model through a multi-objective joint optimization strategy. After training, the model parameters are fixed and deployed to the vehicle battery management system (BMS).

[0113] At the vehicle end, the onboard battery management system collects real-time raw operating data and constructs an input tensor according to steps one and two in Example 1. The system loads the model parameters trained in the cloud and, through forward inference, directly outputs the estimated state of charge and remaining lifespan at the current moment. Because the model embeds a physical information state integration module, even under conditions of unstable sampling frequency or drastic changes in operating conditions, the system can still output stable estimation results that conform to physical laws, providing the driver with accurate range warnings and providing decision-making basis for battery thermal management and charging strategies.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. 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 be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for jointly estimating the state of charge and remaining useful life of a battery using physical information guidance and dual temporal attention, characterized in that: Includes the following steps: S1: Acquire the raw data generated by the power battery during operation. The raw data includes the total current of the battery pack, the total voltage of the battery pack, the cumulative mileage, the instantaneous vehicle speed, the maximum voltage of a single cell, the minimum voltage of a single cell, the sampling time interval between adjacent sampling points, the maximum temperature, the minimum temperature, and the cell voltage difference. S2: Perform feature engineering processing on the original data to construct a three-dimensional input tensor; S3: Input the three-dimensional input tensor into the temporal context encoder module, and output the global latent space state matrix and the local latent space state vector; S4: Input the global latent space state matrix and the local latent space state vector into the dual-branch irregular temporal attention mechanism module, and output the available capacity estimate and the window start point charge state estimate, respectively. S5: Input the available capacity estimate, the state of charge estimate at the window starting point, and the current data in the three-dimensional input tensor into the physical information state integration module, and recursively calculate the state of charge trajectory within the window based on the principle of charge conservation, and output the single-step capacity estimate and the single-step state of charge estimate. S6: Input the local latent space state vector, the charged state trajectory, and the current data into the voltage reconstruction decoder module, and output the reconstructed voltage value; S7: Construct a multi-objective joint loss function based on the single-step state of charge estimate, the single-step capacity estimate, the reconstructed voltage value, and the corresponding real label, and perform end-to-end training on the temporal context encoder module, the dual-branch irregular temporal attention mechanism module, the physical information state integration module, and the voltage reconstruction decoder module.

2. The method for jointly estimating battery state of charge and remaining useful life using physical information guidance and dual temporal attention as described in claim 1, characterized in that: S2 specifically includes: Will t The principal input vector at time t is defined as: in, Indicates the total current of the battery pack. Indicates the total voltage of the battery pack. Indicates cumulative mileage. Indicates instantaneous vehicle speed. Indicates the maximum voltage of a single unit. Indicates the minimum voltage of a single unit. Indicates the sampling time interval between adjacent sampling points; Will t The auxiliary feature vector at time step is defined as: in, Indicates average temperature. Indicates the throughput per ampere-hour of a segment. Indicates the cell voltage difference. Indicates temperature difference; Will t The complete input column vector at time t is defined as: in, T Indicates the transpose operation; Use a length of L The sliding window slices the complete input column vector to obtain the first... i Data slices : Will B The data slices are stacked along the batch dimension to obtain a three-dimensional input tensor. The three-dimensional input tensor is then Z-score normalized.

3. The method for jointly estimating battery state of charge and remaining useful life using physical information guidance and dual temporal attention as described in claim 2, characterized in that: The parameters in the auxiliary feature vector are defined as follows: in, Indicates the maximum temperature. Indicates the minimum temperature; in, Indicates the conversion factor for time units.

4. The method for jointly estimating battery state of charge and remaining useful life using physical information guidance and dual temporal attention as described in claim 1, characterized in that: S3 specifically includes: The timing context encoder module is composed of a two-layer gated loop unit (GRU). For the b Batch No. l Input tensor under each slice By updating the door Reset door Candidate hidden state and hidden state Perform recursive processing and update the formula as follows: in, This represents the Sigmoid activation function. This represents element-wise multiplication of vectors. This indicates updating the gate weight matrix. This represents the reset gate weight matrix. This represents the candidate hidden state weight matrix. This represents the update gate historical state weight matrix. This represents the weight matrix of the historical states of the reset gate. This represents the candidate hidden state historical state weight matrix. This indicates updating the gate bias vector. This indicates the reset gate bias vector. This represents the candidate hidden state bias vector. Indicates the dimension of the hidden layer; Output the global latent space state matrix : The hidden state at the last time step is taken as the local latent space state vector. .

5. The method for jointly estimating battery state of charge and remaining useful life using physical information guidance and dual temporal attention as described in claim 1, characterized in that: S4 specifically includes: Capacity estimation branch: Process the global latent space state matrix using time-integral attention blocks to calculate capacity attention weights. : in, For stress score, To prevent tiny constants from being zero in logarithmic operations; The time-space integral degradation representation vector is obtained by weighted aggregation of the global latent space state matrix. It is calibrated using a channel attention block (CA Block) and outputs a degradation representation vector for channel calibration. The normalized available capacity estimate is obtained by multilayer perceptron (MLP) mapping. ; Initial point charged state estimation branch: based on the local latent space state vector Using the global latent space state matrix as the key and value, the basic attention score is calculated as the query vector. And combined with an exponential decay factor based on the real time span Calculate the attention weights of the charged state : in, Represents the query vector. Represents the key vector. Indicates the hidden layer dimension. This represents the learnable time decay constant; The global latent space state matrix is ​​weighted and aggregated to obtain the representation vector. After calibration by the channel attention block, the input to the multilayer perceptron is used to obtain the estimated state of charge at the window starting point. .

6. The method for jointly estimating battery state of charge and remaining useful life using physical information guidance and dual temporal attention as described in claim 1, characterized in that: S5 specifically includes: The normalized available capacity estimate and the estimated state of charge at the starting point of the window Perform inverse normalization to obtain the inverse normalized capacity estimate. and inverse normalization window starting point charge state estimation : in, The standard deviation of the capacity labels in the training set. This represents the mean of the capacity labels in the training set. This represents the standard deviation of the charged state labels in the training set. This represents the mean of the charged state labels in the training set; The trajectory of the charged state within the window is recursively derived based on the ampere-hour integral equation: in, express t The cumulative charge consumption at any given time. Indicates the conversion factor for time units. This indicates that the gradient operation has been stopped. The last time step of the state-of-charge trajectory is taken as the single-step state-of-charge estimate, and the inverse normalized capacity estimate is taken as the single-step capacity estimate.

7. The method for jointly estimating battery state of charge and remaining useful life using physical information guidance and dual temporal attention as described in claim 1, characterized in that: S6 specifically includes: exist t At time , the local latent space state vector The state of charge trajectory and the current in the original data The data is stitched together along the feature channel dimension, and the terminal voltage at that moment is estimated using a multilayer perceptron. in, This indicates a concatenation operation along the feature channel dimensions. This represents a voltage decoder multilayer sensor.

8. The method for jointly estimating battery state of charge and remaining useful life using physical information guidance and dual temporal attention as described in claim 1, characterized in that: The multi-objective joint loss function in S7 is: in, This indicates the estimated loss due to the state of charge at the end of the window. This indicates the estimated loss based on available capacity. Indicates voltage reconstruction loss, This represents the capacity loss weight hyperparameter. This represents the voltage reconstruction loss weighting hyperparameter; The , and The calculation formula is: in, Indicates the first b The estimated state of charge of each sample slice at the end of the window. Indicates the first b The true label of the state of charge corresponding to each sample slice Indicates the first b Capacity estimates for each sample slice. Indicates the first b The true label of the volume corresponding to each sample slice Indicates the first b One sample in t Voltage reconstruction value at time t, Indicates the first b One sample in t The actual observed voltage value at a given time.

9. A joint estimation system for battery state of charge and remaining useful life guided by physical information and dual temporal attention, executing the method of any one of claims 1 to 8, characterized in that: include: The data acquisition and preprocessing module is used to execute S1 to S2 to obtain the three-dimensional input tensor; The temporal context encoding module is used to execute S3 and output the global latent space state matrix and the local latent space state vector. A dual-branch irregular temporal attention module is used to execute S4 and output the available capacity estimate and the window start point state of charge estimate; The physical information state integration module is used to execute S5 and output a single-step capacity estimate and a single-step state of charge estimate. The voltage reconstruction decoding module is used to execute S6 and output the reconstructed voltage value; The joint training module is used to execute S7 to perform end-to-end training on the system.

10. The joint estimation system for battery state of charge and remaining useful life guided by physical information and dual temporal attention as described in claim 9, characterized in that: The system is deployed in a vehicle battery management system (BMS), which collects the raw data in real time and outputs the estimated state of charge and remaining life of the battery online through the trained system.

Images