A method for estimating the state of health of lithium-ion batteries

Abstract

The present invention discloses a method for estimating the state of health of a lithium-ion battery, which relates to the field of battery management technology. The method comprises: obtaining time series data during the cyclic charge and discharge process of a lithium-ion battery, constructing a second-order equivalent circuit model to extract dynamic features, fitting the open-circuit voltage curve using a multi-layer perceptron to obtain static features, and combining the dynamic features and static features into physical flow features; decoding the physical flow features using a physical decoder to obtain a first battery health state; decoding the data flow features using a data-driven decoder to obtain a second battery health state; determining the combined weight of the physical decoder and the data-driven decoder based on latent space features, and weightedly fusing the first battery health state and the second battery health state based on the combined weight to obtain the final battery health state of the lithium-ion battery. The present invention effectively solves the problem of the difficulty in coordinating weight distribution between physical constraints and data-driven learning in the PINN model.

Description

Technical Field

[0001] The present invention relates to the technical field of battery management, and in particular to a method for estimating the health status of a lithium-ion battery. Background Art

[0002] As one of the most widely used energy storage devices today, lithium-ion batteries are widely used in consumer electronics, electric vehicles, renewable energy storage, and other fields due to their high energy density, long cycle life, and lack of memory effect. With the rapid development of electric vehicles and energy storage systems, the demand for lithium-ion batteries has increased dramatically. Battery State of Health (SOH) estimation has become a key technology for ensuring safe battery operation, optimizing battery management strategies, and extending battery life. Battery SOH refers to the health of a battery relative to its initial or rated state, typically expressed as capacity retention or health percentage. Accurate SOH estimation is crucial to the performance of battery management systems. SOH estimation can help promptly detect battery aging or failure, thereby avoiding safety hazards, optimizing battery efficiency, and providing decision support for battery replacement or maintenance.

[0003] In recent years, with the rapid development of artificial intelligence and machine learning technologies, data-driven methods have gradually become an important research direction for SOH estimation. Deep learning, with its powerful nonlinear modeling capabilities, can extract complex features from battery operating data, providing a new solution for SOH estimation. For example, deep learning models such as Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Long Short-term Memory Networks (LSTM) have been successfully applied to battery SOH estimation. By processing multidimensional data (such as voltage, current, temperature, SOC, etc.), these methods can capture the long-term trends and nonlinear dynamics of battery degradation, thereby improving estimation accuracy. However, deep learning methods also face some challenges, such as dependence on high-quality labeled data, insufficient model generalization ability, and interpretability issues.

[0004] Physics-informed neural networks (PINN), an emerging deep learning framework, combine the advantages of data-driven approaches and physical models, providing new insights for SOH estimation in lithium-ion batteries. By embedding the physical laws of batteries, including electrochemical and thermodynamic equations, into the neural network's loss function, PINN maintains the model's physical consistency within a data-driven framework. This combination not only improves the accuracy and reliability of SOH estimation but also reduces reliance on labeled data, enhancing the model's generalization capabilities.

[0005] However, in practical applications, the simple weighting of physical constraints and data-driven optimization of PINN makes it difficult. As a result, a single decoder in a complex system cannot flexibly process different regions, lacks the ability to adapt to dynamic changes in physical parameters, is insufficiently interpretable, and is difficult to analyze the basis for model decision-making. Summary of the Invention

[0006] Based on this, it is necessary to provide a lithium-ion battery health status estimation method to address the above technical issues.

[0007] An embodiment of the present invention provides a method for estimating the health status of a lithium-ion battery, comprising:

[0008] Obtain timing data during the charge and discharge cycle of lithium-ion batteries;

[0009] A second-order equivalent circuit model of the lithium-ion battery is constructed based on time series data. Dynamic features reflecting the instantaneous physical state of the lithium-ion battery are extracted through the second-order equivalent circuit model. The open-circuit voltage curves of the lithium-ion battery at different states of charge during the charging and discharging process are fitted using a multi-layer perceptron model to obtain static features reflecting the aging state of the lithium-ion battery. The dynamic features are combined with the static features to obtain the physical flow features of the lithium-ion battery. The time dependency of the time series data is captured using a recurrent neural network to obtain the data flow features of the lithium-ion battery.

[0010] By implicitly projecting the physical flow features and data flow features into the same latent space, the latent space features of lithium-ion batteries are obtained.

[0011] The physical flow features, data flow features and latent space features are input into a dynamic gated decoder, wherein the dynamic gated decoder includes a physical decoder and a data-driven decoder arranged in parallel; the physical flow features are decoded by the physical decoder to obtain a first battery health state; the data flow features are decoded by the data-driven decoder to obtain a second battery health state; the combination weight of the physical decoder and the data-driven decoder is determined according to the latent space features, and the first battery health state and the second battery health state are weightedly fused according to the combination weight to obtain a final battery health state of the lithium-ion battery.

[0012] Optionally, the second-order equivalent circuit model of the lithium-ion battery includes: a voltage source, the positive electrode of which is sequentially connected to an ohmic internal resistance, a first resistor and a first capacitor in series to form a first-order RC network, and a second resistor R 2 and the second capacitor in series to form a second-order RC network;

[0013] The dynamic characteristics reflecting the instantaneous physical state of lithium-ion batteries are extracted through a second-order equivalent circuit model based on the following formula:

[0014] ;

[0015] ;

[0016] ;

[0017] ;

[0018] in, I is the loop current, R 1 is the first resistor, R 2 is the second resistor, U 1 is the voltage of the first resistor, U 2 is the voltage of the second resistor, or is the discharge efficiency of the battery, Q N is the rated capacity of the battery, U t is the terminal voltage, R 0 is the internal resistance in ohms.

[0019] Optionally, a multi-layer perceptron model is used to fit the open circuit voltage curves of the lithium-ion battery at different states of charge during the charging and discharging process to obtain static features reflecting the aging state of the lithium-ion battery, specifically including:

[0020] The multi-layer perceptron model consists of an input layer, multiple hidden layers and an output layer;

[0021] The target hidden layer is defined to contain multiple neurons, and the output value of the target neuron in the target hidden layer is determined based on the following formula:

[0022] ;

[0023] in, oh ij is the weight vector, Z (i-1) is the neuron vector of the previous hidden layer, b ij is the bias, ACT(·) is the activation function, i is the target hidden layer, jis the target neuron in the target hidden layer;

[0024] Determine the open circuit voltage based on the following formula:

[0025] ;

[0026] in, y is the open circuit voltage, ACT is the activation function, oh m is the weight vector of the target hidden layer; Z (m-1) is the output value of the neuron in the previous hidden layer, b m is bias;

[0027] Take the open circuit voltage as the static characteristic.

[0028] Optionally, a recurrent neural network is used to capture the time dependency of the time series data to obtain the data stream characteristics of the lithium-ion battery, specifically including:

[0029] Time series data is a sequence arranged in time order. X =( x 1, x 2,…, x k ), k =1, 2, ..., T ;

[0030] The recurrent neural network updates the hidden state of the current moment based on the input and the hidden state of the previous moment at each moment; the output and state of the recurrent neural network at each moment are calculated by the following formula:

[0031] ;

[0032] ;

[0033] in, x k is time series data, h k is the hidden layer state at the current moment, h k-1 is the hidden layer state at the previous moment, g (·) is the activation function, y k is the output of the recurrent neural network, W h and W x is the input weight matrix, b is the bias term, y k is the second battery health status; Wy is the output weight matrix;

[0034] The output of the recurrent neural network is used as the data stream feature of the lithium-ion battery.

[0035] Optionally, the physical flow features and data flow features are projected into the same latent space through implicit projection based on the following formula to obtain the latent space features of the lithium-ion battery:

[0036] ;

[0037] in, f ( t , x ) is the data flow feature, u ( t , x ) is the physical flow characteristic;

[0038] During the implicit projection process, the loss function is calculated as follows:

[0039] ;

[0040] in, 、 and u i for u ( t , x ), and for f ( t , x )'s matching points, N u is the total number of training data, N f is the total number of configuration points, MSE u is the error between the data stream features and the latent space, MSE f is the implicit constraint of the physical equations, l is the weight of the physical constraint.

[0041] Optionally, the combined weight of the physical decoder and the data-driven decoder is determined according to the latent space features based on the following formula:

[0042] ;

[0043] The first battery health state and the second battery health state are weightedly fused according to the combination weight based on the following formula to obtain the final battery health state of the lithium-ion battery:

[0044]

[0045] in, α is the combined weight, g (·) is the activation function, W is the weight matrix of the gating network, b is the bias of the gating network, z is the latent space feature, SOH h The first battery health state, SOH x It is the second battery health status.

[0046] The above-mentioned lithium-ion battery health status estimation method provided by the embodiment of the present invention has the following beneficial effects compared with the prior art:

[0047] The present invention achieves deep collaboration between physical constraints and data-driven learning through a dynamic gated decoder, and adaptively adjusts the weights of the dual decoders using latent space features as a medium, replacing the traditional simple weighting method, effectively solving the problem of difficult coordination of weight distribution between physical constraints and data-driven learning in the PINN model; physical flow features explicitly encode the battery aging mechanism, and data flow features implicitly mine the temporal correlation rules. The aligned mapping of the two in the latent space not only significantly reduces optimization conflicts and improves the adaptability to complex working conditions, but also enables health status assessment to have both physical traceability and data-driven accuracy, thereby enhancing the model's responsiveness to changes in physical parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A model framework diagram of a lithium-ion battery health status estimation method provided in one embodiment;

[0049] Figure 2 A network structure diagram of a lithium-ion battery health status estimation method provided in one embodiment;

[0050] Figure 3 A second-order RC equivalent circuit diagram of a lithium-ion battery health status estimation method provided in one embodiment. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0052] In one embodiment, a method for estimating the health status of a lithium-ion battery is provided, such as Figure 1 As shown, the method includes:

[0053] Obtain time series data during the cyclic charge and discharge process of lithium-ion batteries.

[0054] A second-order equivalent circuit model of a lithium-ion battery is constructed based on time series data. Dynamic features reflecting the instantaneous physical state of the lithium-ion battery are extracted from this model. A multi-layer perceptron model is used to fit the open-circuit voltage curves of the lithium-ion battery at different states of charge during the charge and discharge process, yielding static features reflecting the battery's aging state. Dynamic and static features are combined to derive the physical flow characteristics of the lithium-ion battery. A recurrent neural network is used to capture the temporal dependencies of the time series data, yielding the data flow characteristics of the lithium-ion battery.

[0055] The physical flow features and data flow features are projected into the same latent space through implicit projection to obtain the latent space features of lithium-ion batteries.

[0056] The physical flow features, data flow features, and latent space features are input into a dynamic gated decoder, which includes a physical decoder and a data-driven decoder arranged in parallel. The physical decoder decodes the physical flow features to obtain a first battery state of health (SOH). The data-driven decoder decodes the data flow features to obtain a second battery state of health (SOH). A combined weight of the physical decoder and the data-driven decoder is determined based on the latent space features. The first and second SOHs are weighted and fused based on the combined weights to obtain the final SOH of the lithium-ion battery.

[0057] like Figure 2 As shown, the implementation process includes:

[0058] (1) Equivalent circuit model.

[0059] The equivalent-circuit model (ECM) consists of three parts: a voltage source, an ohmic internal resistance, and an RC network. The ideal voltage source represents the open-circuit voltage of the battery. U oc , which has a nonlinear relationship with the battery SOC, n The RC network of the first order characterizes the dynamic performance of the battery. n With the increase of , the accuracy of battery behavior characterization is significantly improved, but the difficulty of parameter identification increases exponentially. The second-order ECM achieves a good balance between complexity and accuracy. Therefore, the present invention selects Figure 3 The second-order ECM shown is used to build the model. The second-order equivalent circuit model of the lithium-ion battery includes: voltage source U oc , voltage source U oc The positive pole is connected to the ohmic internal resistance R 0. First resistor R 1 and the first capacitor C 1 The first-order RC network obtained by connecting in series with the second resistor R2 and the second capacitor C 2A second-order RC network obtained by connecting 2 in series.

[0060] According to Kirchhoff's law and current integration method, the mathematical expression that reflects the dynamic characteristics of the instantaneous physical state of the lithium-ion battery is extracted through the second-order equivalent circuit model as follows:

[0061] (1)

[0062] (2)

[0063] (3)

[0064] (4)

[0065] in, I is the loop current, R 1 is the first resistor, R 2 is the second resistor, U 1 is the voltage of the first resistor, U 2 is the voltage of the second resistor, or is the discharge efficiency of the battery, Q N is the rated capacity of the battery, U t is the terminal voltage, R 0 is the ohm internal resistance. Under the condition that the battery parameters and operating current remain unchanged during each sampling period, the discretized battery model can be expressed as:

[0066] (5)

[0067] (6)

[0068] in, k represents time discretization; Δ t is the sampling interval; t 1= R 1 C 1, τ2 = R 2 C 2.

[0069] (2) Open circuit voltage modeling based on multi-layer perceptron model.

[0070] Since neural networks have the ability to approximate any continuous function and estimate the complex nonlinear characteristics of the system, compared with traditional predefined lookup tables and polynomial fitting methods, this method shows greater flexibility in capturing battery dynamic characteristics. U oc-The SOC curve will shift with aging and environmental factors. The present invention comprehensively considers the complexity of the problem and computational efficiency and adopts a multilayer perceptron (MLP) model to establish an open circuit voltage model.

[0071] (7)

[0072] in, U oc is the open circuit voltage; SOC is the input value of MLP; f MLP is the trained MLP network model.

[0073] MLP consists of the input layer, m hidden layers and an output layer. i The hidden layers contain d neurons, then the layer j The output value of a neuron is calculated as:

[0074] (8)

[0075] in, oh ij is the weight vector, Z (i-1) is the neuron vector of the previous hidden layer, b ij is the bias, ACT(·) is the activation function, m is the total number of hidden layers, d is the total number of neurons in the target hidden layer, i is the target hidden layer, j is the target neuron in the target hidden layer.

[0076] The open circuit voltage is:

[0077] (9)

[0078] in, y is the output value of the open circuit voltage model, ACT is the activation function, oh m is the weight vector of the target hidden layer; Z (m-1) is the output value of the neuron in the previous hidden layer, b m is the bias; the open circuit voltage is taken as the static characteristic.

[0079] As the network depth m As the number of parameters increases, the nonlinear representation capability of the model is enhanced, but the geometric growth of the parameter scale will lead to a significant increase in computational complexity. Therefore, it is necessary to determine the optimal network depth through rigorous complexity analysis and experimental verification.

[0080] Combining dynamic features with static features to obtain the physical flow characteristics of lithium-ion batteries u ( t , x ).

[0081] (3) Temporal feature extraction based on RNN.

[0082] Recurrent Neural Networks (RNNs), a classic deep learning model, can capture the temporal dependencies of time series data. Due to their simple structure and low computational complexity, this paper uses RNNs to extract time series features.

[0083] RNN accepts experimental data as input, assuming that the experimental data is a sequence X =( x 1, x 2,…, x k ), k =1, 2, ..., T The RNN updates its hidden state at each moment based on the input and the previous hidden state. The layers of the RNN are connected in a cyclic manner, and the output and state of each time step are generated based on the input and state of the previous step, as shown below:

[0084] (10)

[0085] (11)

[0086] in, x k is time series data, h k is the hidden layer state at the current moment, h k-1 is the hidden layer state at the previous moment, g (·) is the activation function, y k is the output of the recurrent neural network, W h and W x is the input weight matrix, b is the bias term, y k is the second battery health status; W y is the output weight matrix; the output of the recurrent neural network is used as the data flow feature of the lithium-ion battery f ( t , x ).

[0087] (4) Fusion of physical flow characteristics and data flow characteristics.

[0088] The data stream features are transformed into f ( t , x ) and physical flow characteristics u ( t , x ) is projected into a shared latent space:

[0089] (12)

[0090] By implicitly projecting, data flow features and physical flow features are combined in a latent space to ensure that they can be jointly optimized and satisfy physical constraints. In order to enhance the combination of data flow features and physical flow features, the loss function is defined as:

[0091] (13)

[0092] in, 、 and u i for u ( t , x ), and for f ( t , x )'s matching points, N u is the total number of training data, N f is the total number of configuration points, MSE u is the error between the data stream features and the latent space; and MSE f is the implicit constraint of the physical equation; l is the weight of the physical constraint, which is used to balance the relationship between data flow characteristics and physical constraints.

[0093] (5) Dynamic gating mechanism.

[0094] To determine the decoder type to use at each time point, the present invention designs a gating mechanism. This mechanism dynamically adjusts the combined weights of the data-driven decoder and the physical decoder using latent space features. Specifically, the combined weights are calculated using a neural network and used to adjust the weighting coefficients of the data-driven decoder and the physical decoder. The formula is as follows:

[0095] (14)

[0096] in,α is the combined weight; g (·) is the activation function, which limits the output to the range of [0, 1]; W is the weight matrix of the gating network, b is the bias of the gating network.

[0097] Neural networks include, but are not limited to, RNN, Long Short-Term Memory (LSTM), or Gated Recurrent Unit (GRU).

[0098] (6) SOH estimation.

[0099] After the data-driven decoder and the physical decoder output SOH respectively, a weighted mechanism is used to fuse the two to obtain the final SOH. The specific implementation is as follows:

[0100] (15)

[0101] Among them, SOH h The first battery health state, SOH x It is the second battery health status.

[0102] Provide the specific implementation process of the present invention:

[0103] Step 1: Data collection and preprocessing

[0104] First, collect experimental data of lithium-ion battery cycle charge and discharge, including voltage, current, temperature, time and capacity.

[0105] The experimental data are normalized by using the minimum-maximum normalization method, which is achieved by mapping the data proportionally to the specified interval. x For example, the normalized value x norm The calculation is as follows:

[0106] (15)

[0107] in, x min and x max are the minimum and maximum values ​​of the data respectively. In this way, all features will be mapped to the range of [0, 1].

[0108] Step 2: Equivalent circuit model

[0109] The ECM consists of three parts: a voltage source, an ohmic internal resistor, and an RC network. The ideal voltage source represents the open circuit voltage of the battery. Uoc , which has a nonlinear relationship with the battery SOC, n The RC network of the first order characterizes the dynamic performance of the battery. n With the increase of , the accuracy of battery behavior characterization is significantly improved, but the difficulty of parameter identification increases exponentially. The second-order ECM achieves a good balance between complexity and accuracy. Therefore, the present invention selects Figure 3 The second-order ECM shown is used to construct the model.

[0110] According to Kirchhoff's law and current integration method, the mathematical expression of the model is expressed as:

[0111] (16)

[0112] (17)

[0113] (18)

[0114] (19)

[0115] in, I is the loop current, and the specified discharge time is I is positive; U 1. U 2 are R 1. R 2 voltage; or is the discharge efficiency of the battery; Q N is the rated capacity of the battery; U t is the terminal voltage, R 0 is the ohm resistance. Under the condition that the battery parameters and operating current remain unchanged during each sampling period, the discretized battery model can be expressed as:

[0116] (20)

[0117] (twenty one)

[0118] in, k represents time discretization; Δ t is the sampling interval; t 1= R 1 C 1, τ2 = R 2 C 2.

[0119] Step 3: Open circuit voltage modeling based on multi-layer perceptron neural network

[0120] Since neural networks have the ability to approximate any continuous function and estimate the complex nonlinear characteristics of the system, compared with traditional predefined lookup tables and polynomial fitting methods, this method shows greater flexibility in capturing battery dynamic characteristics. Uoc The SOC curve will shift with aging and environmental factors. The present invention comprehensively considers the complexity of the problem and computational efficiency and uses a multi-layer perceptron to establish an open-circuit voltage model.

[0121] The MLP network consists of an input layer, m hidden layers and an output layer. i The hidden layers contain d neurons, then the layer j The output value of a neuron is calculated as:

[0122] (twenty two)

[0123] in, oh ij is the weight vector; Z (i-1) It is i -1 neuron vector; b ij is the bias; ACT is the activation function. The output value is:

[0124] (twenty three)

[0125] As the network depth m As the number of parameters increases, the nonlinear representation capability of the model is enhanced, but the geometric growth of the parameter scale will lead to a significant increase in computational complexity. Therefore, it is necessary to determine the optimal network depth through rigorous complexity analysis and experimental verification.

[0126] Step 4: RNN-based temporal feature extraction

[0127] As a classic deep learning model, RNN can capture the temporal dependencies in sequence data. Its structure is simple and computational complexity is low. Therefore, the present invention uses the RNN structure to extract temporal features.

[0128] RNN accepts time series data as input, assuming the data is a sequence X =( x 1, x 2,…, x k ), k =1, 2, ..., T The RNN updates its hidden state at each moment based on the input and the previous hidden state. The layers of the RNN are connected in a cyclic manner, and the output and state of each time step are generated based on the input and state of the previous step, as shown below:

[0129] (twenty four)

[0130] (25)

[0131] in, x k Represents input; h k Represents the hidden layer state; g (·) is the activation function; W h and W x is the input weight matrix; b is the bias term; y k is the output; W y is the output weight matrix.

[0132] Step 5: Fusion of physical flow features and data flow features

[0133] The data stream features are transformed into f ( t , x ) and physical flow characteristics u ( t , x ) is projected into a shared latent space:

[0134] (26)

[0135] By implicitly projecting, data flow features and physical flow features are combined in a latent space to ensure that they can be jointly optimized and satisfy physical constraints. In order to enhance the combination of data flow features and physical flow features, the loss function is defined as:

[0136] (27)

[0137] in, 、 and u i yes u ( t , x )’s initial and boundary training data; and yes f ( t , x )’s matching points; N u and N f is the total number of training data and configuration points; MSEu is the error between the data stream features and the latent space; and MSE f is the implicit constraint of the physical equation; l is the weight of the physical constraint, which is used to balance the relationship between data flow characteristics and physical constraints.

[0138] Step 6: Dynamic Gating Mechanism

[0139] In order to determine the type of decoder to be used at each time point, the present invention designs a gating mechanism. This mechanism dynamically adjusts the combined weights of the data-driven decoder and the physical decoder through latent space features. Specifically, the combined weights are calculated by a neural network. α , the combined weight is used to adjust the weight coefficients of the data-driven decoder and the physical decoder, and the formula is as follows:

[0140] (28)

[0141] in, α represents the combined weight; g (·) is the activation function, which limits the output to the range of [0, 1]; W is the weight matrix of the gating network; b is the bias of the gating network.

[0142] Step 7: SOH estimation

[0143] After the data-driven decoder and the physical decoder output SOH respectively, a weighted mechanism is used to fuse the two to obtain the final SOH. The specific implementation is as follows:

[0144] (29)

[0145] Step 8: SOH Assessment

[0146] In order to evaluate the performance of transfer learning in transferring previously learned knowledge to the target task, this paper uses four different regression evaluation metrics to evaluate the performance of the model in different scenarios. R 2 , Mean Absolute Error (MAE), Mean Bias Error (MBE) and Root Mean Square Error (RMSE).

[0147] (1) R -Square

[0148] (30)

[0149] R 2A measure of the quality of the model fit, indicating the proportion of data variation explained by the model. Its value ranges from 0 to 1, with values ​​closer to 1 indicating a better model fit. R 2 The larger it is, the stronger the predictive ability of the model.

[0150] (2) Mean absolute error

[0151] (31)

[0152] MAE is the average of the differences between the predicted values ​​and the true values. It is used to measure the mean absolute error between the predicted values ​​and the true values. The smaller the MAE value, the more accurate the model's predictions.

[0153] (3) Mean deviation error

[0154] (32)

[0155] MBE is the average difference between the predicted value and the true value. If the MBE is positive, it means that the model tends to overestimate; if it is negative, it means that the model tends to underestimate.

[0156] (4) Root mean square error

[0157] (33)

[0158] RMSE is the average of the square roots of the errors and can penalize large errors. The smaller the RMSE, the more accurate the model's predictions.

[0159] The above-described embodiments merely illustrate several implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.

Claims

1. A method for estimating the health status of a lithium-ion battery, characterized in that: include: Obtain timing data during the charge and discharge cycle of lithium-ion batteries; A second-order equivalent circuit model of the lithium-ion battery is constructed based on time series data. The equivalent circuit model consists of three parts: a voltage source, an ohmic internal resistance, and an RC network. The ideal voltage source represents the battery's open-circuit voltage, which has a nonlinear relationship with the battery's SOC. The second-order RC network characterizes the battery's dynamic performance. The second-order equivalent circuit model is used to extract dynamic characteristics reflecting the lithium-ion battery's instantaneous physical state. The open circuit voltage curves of lithium-ion batteries at different states of charge during the charge and discharge process are fitted using a multi-layer perceptron model to obtain static features reflecting the aging state of lithium-ion batteries, with the open circuit voltage being used as the static feature. Combining dynamic features with static features, the physical flow characteristics of lithium-ion batteries are obtained; The time dependency of time series data is captured by recurrent neural networks to obtain the data stream characteristics of lithium-ion batteries; By implicitly projecting the physical flow features and data flow features into the same latent space, the latent space features of lithium-ion batteries are obtained. Inputting the physical flow features, the data flow features, and the latent space features into a dynamic gated decoder, wherein the dynamic gated decoder includes a physical decoder and a data-driven decoder arranged in parallel; decoding the physical flow features by the physical decoder to obtain a first battery health state; Decoding the data stream feature by a data-driven decoder to obtain a second battery health state; The combination weight of the physical decoder and the data-driven decoder is determined according to the latent space characteristics, and the first battery health state and the second battery health state are weightedly fused according to the combination weight to obtain the final battery health state of the lithium-ion battery.

2. A lithium-ion battery health status estimation method according to claim 1, characterized in that: The second-order equivalent circuit model of the lithium-ion battery includes: a voltage source, a positive electrode of the voltage source is sequentially connected to an ohmic internal resistance, a first resistor and a first capacitor in series to form a first-order RC network, and a second resistor R 2 and the second capacitor in series to form a second-order RC network; The dynamic characteristics reflecting the instantaneous physical state of lithium-ion batteries are extracted through a second-order equivalent circuit model based on the following formula: ； ； ； ； in, I is the loop current, R 1 is the first resistor, R 2 is the second resistor, U 1 is the voltage of the first resistor, U 2 is the voltage of the second resistor, η is the discharge efficiency of the battery, Q N is the rated capacity of the battery, U t is the terminal voltage, R 0 is the internal resistance in ohms.

3. A lithium-ion battery health status estimation method according to claim 2, characterized in that: The multi-layer perceptron model is used to fit the open circuit voltage curve of the lithium-ion battery at different states of charge during the charging and discharging process to obtain static characteristics reflecting the aging state of the lithium-ion battery, specifically including: The multi-layer perceptron model consists of an input layer, multiple hidden layers and an output layer; The target hidden layer is defined to contain multiple neurons, and the output value of the target neuron in the target hidden layer is determined based on the following formula: ； in, ω ij is the weight vector, Z (i-1) is the neuron vector of the previous hidden layer, b ij is the bias, ACT(·) is the activation function, i is the target hidden layer, j is the target neuron in the target hidden layer; Determine the open circuit voltage based on the following formula: ； in, y is the open circuit voltage, ACT is the activation function, ω m is the weight vector of the target hidden layer; Z (m-1) is the output value of the neuron in the previous hidden layer, b m For bias.

4. The method for estimating the health status of a lithium-ion battery according to claim 1, wherein: The method of capturing the time dependency of time series data through a recurrent neural network to obtain the data stream characteristics of the lithium-ion battery specifically includes: Time series data is a sequence arranged in time order. X =( x 1, x 2,…, x k ), k =1, 2, ..., T ; The recurrent neural network updates the hidden state of the current moment based on the input and the hidden state of the previous moment at each moment; the output and state of the recurrent neural network at each moment are calculated by the following formula: ； ； in, x k is time series data, h k is the hidden layer state at the current moment, h k-1 is the hidden layer state at the previous moment, g (·) is the activation function, y k is the output of the recurrent neural network, W h and W x is the input weight matrix, b is the bias term, y k is the second battery health status; W y is the output weight matrix; The output of the recurrent neural network is used as the data stream feature of the lithium-ion battery.

5. The method for estimating the health status of a lithium-ion battery according to claim 1, wherein: Based on the following formula, the physical flow features and data flow features are projected into the same latent space through implicit projection to obtain the latent space features of lithium-ion batteries: ； in, f ( t , x ) is the data flow feature, u ( t , x ) is the physical flow characteristic; According to the loss function, the data flow features and physical flow features are combined to obtain the latent space features.

6. A lithium-ion battery health status estimation method according to claim 1, characterized in that: The combined weight of the physical decoder and the data-driven decoder is determined based on the latent space features according to the following formula: ； The first battery health state and the second battery health state are weightedly fused according to the combination weight based on the following formula to obtain the final battery health state of the lithium-ion battery: in, α is the combined weight, g (·) is the activation function, W is the weight matrix of the gating network, b is the bias of the gating network, z is the latent space feature, SOH h The first battery health state, SOH x It is the second battery health status.

Images