soc-soh joint estimation method based on double-cross physical guiding framework

Abstract

The application provides a SOC-SOH joint estimation method based on a double-cross physical guiding framework, data characteristics of battery charging and discharging are sent into a double-cross physical guiding model, a space-time characteristic encoder of the double-cross physical guiding model extracts a high-dimensional characteristic tensor in the data characteristics; the high-dimensional characteristic tensor is converted through two learnable projection heads, a double-flow mutual attention module calculates cross-attention information between the converted outputs, and dynamically regulates and fuses the cross-attention information through a dynamic interaction mechanism to obtain final state-of-charge dynamic interaction results and state-of-health dynamic interaction results; a physical constraint decoder maps the final state-of-charge dynamic interaction results and the state-of-health dynamic interaction results to state-of-charge prediction values and state-of-health prediction values; effectively solve the core problem that the existing method is difficult to balance short-time response capture and long-term degradation characterization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery state estimation, and in particular to a SOC-SOH joint estimation method based on a double-cross physical guiding framework. BACKGROUND

[0002] With the wide application of battery technology in new energy vehicles, energy storage systems and portable devices, higher requirements are put forward for the accuracy and real-time performance of battery state monitoring. The state of charge (SOC) and the state of health (SOH) of the battery are two key indicators for evaluating the performance of the battery, and their accurate estimation is of great significance for ensuring device safety, prolonging battery life and optimizing energy management. However, in practical applications, battery state estimation faces many challenges. First, the battery shows obvious dynamic response and transient characteristics during charging and discharging, such as voltage fluctuation, temperature gradient and current pulse, which requires the model to capture multi-scale and multi-dimensional time series information. At the same time, the battery aging process is often accompanied by nonlinear and complex degradation mechanisms, such as active material loss, SEI film thickening and lithium deposition, which cause the SOH decay curve to show a combination of sudden and gradual changes. In addition, due to the influence of sensor noise, uneven data sampling frequency and environmental factors in actual working conditions, traditional single data-driven or physical models have limitations in accurately capturing the internal state of the battery, and are prone to overfitting or insufficient extrapolation ability. In order to solve the above problems, it is necessary to introduce the prior knowledge of the internal physical mechanism of the battery on the basis of ensuring that the data is fully expressed, so as to improve the prediction accuracy and physical rationality of the model. Traditional methods mostly model SOC and SOH independently, ignoring the inherent coupling relationship between the two, and cannot simultaneously consider short-term charging and discharging dynamics and long-term aging evolution, and lack physical consistency constraints, resulting in that the mutual influence of battery charging and discharging behavior and degradation process cannot be fully reflected in actual applications. SUMMARY

[0003] In view of the needs in the prior art, the present application provides a SOC-SOH joint estimation method based on a double-cross physical guiding framework, aiming to solve the problem that existing methods are difficult to simultaneously capture the short-term dynamic response and long-term aging trend of the battery under complex working conditions.

[0004] The SOC-SOH joint estimation method based on the double-cross physical guiding framework comprises the following steps:

[0005] Step 1: Obtain the data features of battery charging and discharging;

[0006] Step 2: Send the data features into a double-cross physical guiding model, which includes a space-time feature encoder, a double-flow mutual attention module and a physical constraint decoder;

[0007] Step 3: The spatio-temporal feature encoder extracts a high-dimensional feature tensor in the data feature based on dilated convolution, multi-head self-attention and feedforward network;

[0008] Step 4: The high-dimensional feature tensor is converted by two parallel learnable projection heads in the feature-guided decoupling mechanism to generate two task-oriented feature streams, which are respectively denoted as the state-of-charge feature stream and the state-of-health feature stream;

[0009] Step 5: The dual-stream cross-attention module calculates the cross-attention information between the state-of-charge feature stream and the state-of-health feature stream, and dynamically regulates and fuses the cross-attention information through a dynamic interaction mechanism to obtain the final state-of-charge dynamic interaction result and the state-of-health dynamic interaction result;

[0010] Step 6: The physical constraint decoder maps the final state-of-charge dynamic interaction result and the state-of-health dynamic interaction result to the state-of-charge prediction value and the state-of-health prediction value based on the multi-layer perceptron and the physical correction term.

[0011] Further, in the spatio-temporal feature encoder, the data feature is processed by the multi-head self-attention layer and the feedforward network layer in turn after dilated convolution, and residual connections are provided in the multi-head self-attention layer and the feedforward network layer, and layer normalization is provided after the multi-head self-attention layer and the feedforward network layer.

[0012] Further, the cross-attention information includes attention from the state-of-charge to the state-of-health direction ;

[0013] wherein, is the mapping of the feature vector in the state-of-charge feature stream, and are the mappings of the feature vectors in the state-of-health feature stream; is a scaling factor; is the mapping of the feature vector in the state-of-health feature stream, and are the mappings of the feature vectors in the state-of-charge feature stream.

[0014] Further, the dynamic interaction mechanism is adaptively regulated by the gating function and the interaction score function, and the gating function value and the interaction score function value are respectively weighted and fused with , .

[0015] Further, Step 6 is specifically:

[0016] Step 6.1: The state-of-charge dynamic interaction results and the state-of-health dynamic interaction results are input into two multi-layer perceptrons, respectively, to obtain preliminary and ;

[0017] Step 6.2: In each time slice, the Coulomb integral constraint term and the Arrhenius degradation term are calculated by integrating or recursively propagating the cumulative charge and temperature and other variables in the current period.

[0018] Step 6.3: The preliminary and are fused with the Coulomb integral constraint term and the Arrhenius degradation term at the end of the decoder to form the final and .

[0019] Further, ;

[0020] wherein, is the rated capacity, is the Coulomb efficiency, when discretized, the integral can be replaced by step-by-step accumulation, and is clamped at each time step to prevent numerical drift; in the decoder, the Coulomb constraint often takes the form of a CoulombConstraint layer, which first estimates based on the current period's current, temperature, and other characteristics of the network, and then couples with the output of the MLP ; the simplified coupling method is:

[0021] ;

[0022] wherein, is a Sigmoid function to ensure that the output is in the [0, 1] interval; is an adjustable weight or gating factor to balance the data-driven prediction and the physical integral result.

[0023] Further, ;

[0024] wherein, is the cumulative discharge capacity or cycle number, represents the decay nonlinearity coefficient; the core of this equation is to dynamically calculate the decay amount based on the current temperature and cumulative capacity, and output a correction term multiplied or added to the MLP predicted ; the coupling method is:

[0025] ;

[0026] wherein, is a Sigmoid function.

[0027] further comprising: from the prediction error , physical consistency and cross-consistency three dimensions to constrain and optimize the Transformer model with double cross-physical guidance, the overall loss is:

[0028] ;

[0029] wherein and are hyperparameters for balancing the contribution of different sub-losses;

[0030] ;

[0031] ;

[0032] ;

[0033] wherein, and respectively represent the predicted values of the model for the state of charge and the state of health, and respectively represent the true values of the model for the state of charge and the state of health; is an adjustment parameter for balancing the deviation between the predicted capacity and the actual charge transfer calculated from the actually observed current integral value ; is an adjustment parameter for controlling and the consistency degree of the Arrhenius equation ; is the rated capacity of the new battery; represents the physical quantity inherent to the battery material, A is the exponential pre-factor in the Arrhenius model; T is the current battery temperature, and I is the current.

[0034] further comprising: step 1 specifically comprises the following steps:

[0035] Step 1.1: Establishing an experimental platform for battery cycle life test, testing the cycle life of lithium ion battery and obtaining the battery charge and discharge data;

[0036] Step 1.2: Multi-scale processing and derived feature extraction are performed on the original signal from the battery charge-discharge data, and the obtained features include voltage change rate, temperature gradient, cumulative throughput power, equivalent cycle number, sliding window statistics and relaxation period voltage recovery rate;

[0037] Step 1.3: All extracted features are processed by a dynamic normalization strategy, and data features of battery charge-discharge are obtained.

[0038] Further, the dynamic normalization strategy is mathematically expressed as follows:

[0039] ;

[0040] Wherein, is the output of the dynamic normalization strategy, is the original feature value at time step t; is a small constant to prevent division by 0; is the mean, is the standard deviation, which is expressed as:

[0041] ;

[0042] ;

[0043] is the dynamic sliding window length and is:

[0044] ;

[0045] Wherein, is the initial length of the window; is the growth rate coefficient; is the health state predicted by the current time step model, and then represents the degree of health decline.

[0046] The beneficial effects of the present application are: with the aid of the space-time feature encoder and the task decoupling mechanism, short-term dynamic behavior and long-term aging evolution features can be effectively distinguished, and heterogeneous modeling at the feature level can be realized; through the bidirectional cross-task attention mechanism and the gating interaction structure, the coupling perception ability between SOC and SOH is explicitly enhanced, and the multi-task collaborative estimation effect is improved; the core problem that existing methods are difficult to balance short-time response capture and long-term degradation description is effectively solved, and a solid support is provided for constructing a high-reliability intelligent battery management system. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 is the flowchart of the present application;

[0048] Figure 2 is the overall framework diagram of the double-cross physical guidance model in the present application;

[0049] Figure 3 SOC estimation results for multiple charge-discharge cycles of the B201 battery;

[0050] Figure 4 SOH estimation results for the B201 battery;

[0051] Figure 5 SOH estimation results for the B203 battery; DETAILED DESCRIPTION

[0052] The application will be described in detail below with reference to the drawings. The embodiments of the application are described in detail below, examples of which are shown in the drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are only used to explain the application, and cannot be interpreted as a limitation on the application. The left, middle, right, up, down and other orientation terms in the embodiments of the application are only relative concepts or are referenced to the normal use state of the product, and should not be considered as limiting.

[0053] The SOC-SOH joint estimation method based on the double-cross physical guidance framework includes the following steps:

[0054] Step 1: Obtain the data features of battery charge-discharge; specifically including the following steps:

[0055] Step 1.1: Establish an experimental platform for battery cycle life test, test the cycle life of lithium ion battery and obtain battery charge-discharge data;

[0056] The INR 18650 battery adopts a step charging method, first charges the battery to 4.2V voltage at 6C constant current; after a short rest, charges the battery to 4.2V voltage at 3C constant current, and then fully charges the battery at constant current and constant voltage (CC-CV) after a one-minute rest; the current in the CC stage is 0.5C, the cutoff current is 0.05C, and the cutoff voltage is 4.2V; After one hour of standing, discharge the battery repeatedly with six times CLTC-P (China Light Vehicle Test Cycle-Passenger Car) until the battery voltage drops to 2.7V; CLTC-P is a standard working condition based on traffic big data, including three intervals: low speed, medium speed, and high speed, with a total duration of 1800s; CLTC-P has high dynamicity and can reflect the aging of the battery in actual application, and capacity measurement is performed every 50 cycles;

[0057] Step 1.2: Multi-scale processing and derived feature extraction from raw signals in battery charge-discharge data, resulting features include voltage rate of change, temperature gradient, cumulative throughput, equivalent cycle number, sliding window statistics, and relaxation period voltage recovery rate;

[0058] Two differential features (dynamic response characterization):

[0059] (1) Voltage rate of change Reflects the speed of change of polarization voltage, directly related to lithium ion diffusion kinetics. SOC correlation: dV / dt mutation at the end of charging indicates that SOC is close to 100%; SOH correlation: the slope of dV / dt curve of aged battery decreases;

[0060] (2) Temperature gradient Indicates the temperature rise rate per unit power consumption; SOH correlation: the temperature gradient of aged battery increases (Joule heat increases due to increased internal resistance), where represents the temperature of the battery (cell temperature); represents the ambient temperature; represents the instantaneous input power of the battery, here represented by ;

[0061] Two cumulative features (long-term degradation metrics):

[0062] (3) Cumulative throughput Indicates a direct measure of total charge transfer; SOC correlation: provides baseline estimation through coulomb counting method; SOH correlation: There is a nonlinear mapping between and capacity decay;

[0063] (4) Equivalent cycle number Indicates the normalized aging cycle count, which is exponentially related to the capacity decay rate;

[0064] Cycle statistical features (local behavior patterns):

[0065] (5) Sliding window statistics Indicates the short-time statistical properties of voltage fluctuations; SOC correlation: is the mean voltage within the sliding window, which has a quasi-linear relationship with SOC (discharge platform region); SOH correlation: is the standard deviation of voltage within the sliding window, which increases to reflect the inhomogeneity of electrode material phase transition; where, is the size of the sliding window; is the voltage data within the window;

[0066] (6) Relaxation period voltage recovery rate ​represents the voltage relaxation rate after current interruption, the physical mechanism involves the competition between the dissipation of concentration polarization and the interface impedance, the relaxation rate of the aged battery decreases exponentially with the increase of SEI film impedance (R²>0.89), wherein, and respectively represent the values of the battery voltage recovery at times and ;

[0067] Step 1.3: Perform dynamic normalization strategy processing on all extracted features, and obtain the data features of battery charging and discharging; wherein the multi-source features (including voltage, current, temperature and their differential / cumulative features) after extraction may exhibit significant scale difference and non-stationarity under different aging stages and different working conditions; in order to improve the generalization ability of the model in the long-term use process, this paper introduces a dynamic normalization strategy (Dynamic Normalization) to perform time-varying standardization processing on all input features, and combines a SOH-guided adaptive window strategy to realize dynamic leveling and change smoothing of feature scale; traditional normalization methods (such as Z-score, Min-Max) often use fixed statistics to perform global normalization on the whole data, however, for the aged battery, the feature distribution itself changes over time, which makes it difficult for static normalization methods to adapt to the non-stationarity of feature distribution, which may lead to systematic deviation of estimation error; therefore, this paper adopts a dynamic normalization strategy updated within a sliding window, which is mathematically expressed as follows:

[0068] ;

[0069] wherein, is the output of the dynamic normalization strategy, is the original feature value at time step t; is a small constant to prevent division by 0; is the mean, is the standard deviation, which is expressed as:

[0070] ;

[0071] ;

[0072] For dynamic sliding window length, due to the fact that the state of health (SOH) of the battery presents a phased change in the degradation process, using a fixed window length may lead to the following problems: in the early stage of battery aging, the SOH is stable, and a short window can quickly respond to instantaneous changes; but in the later stage of aging, the SOH rapidly declines, and a short window cannot effectively smooth the mutation, which will lead to dramatic fluctuations in normalization; therefore, this paper proposes a strategy based on the dynamic adjustment of the window length according to the SOH value, and

[0073] ;

[0074] wherein, is the initial length of the window (a tunable hyperparameter); is the growth rate coefficient; is the state of health predicted by the model at the current time step, and then represents the degree of health decline;

[0075] Step 2: send the data features into a double-cross physical guidance model, which includes a spatio-temporal feature encoder, a double-flow mutual attention module, and a physical constraint decoder;

[0076] Step 3: the spatio-temporal feature encoder extracts high-dimensional feature tensors from the data features based on dilated convolution, multi-head self-attention, and feedforward network; in the spatio-temporal feature encoder, the data features are processed by the multi-head self-attention layer and the feedforward network layer in turn after dilated convolution, and residual connections are set in both the multi-head self-attention layer and the feedforward network layer, and layer normalization is set after both the multi-head self-attention layer and the feedforward network layer;

[0077] wherein, the dilated convolution (Dilated Convolution) performs local perception and feature extraction on the original input; specifically, let the input sequence be wherein, is the length of the time series, is the feature dimension, represents a set of real numbers, then the output of the dilated convolution can be represented as:

[0078] ;

[0079] wherein, represents the size of the convolution kernel, is the dilation rate (dilation rate), is the learnable convolution kernel parameter, is usually a nonlinear activation function; on this basis, by stacking or parallel multi-layer dilated convolution, multi-scale sampling of the sequence can be performed at different dilation rates, thereby capturing the feature changes of the battery data in short, medium, and long time ranges;

[0080] Let the output of the dilated convolution be denoted as , is the length of the time series, denotes the dimension of the new feature map, then the multi-head self-attention can be expressed as:

[0081] ;

[0082] where, are obtained by linear transformation, Through the multi-head self-attention mechanism, multiple attention patterns can be learned in different subspaces in parallel, so that the model can not only pay attention to global trends, but also capture local dependencies or special events.

[0083] After the multi-head self-attention output, two layers of fully connected networks are used for non-linear mapping, which can be generally written as:

[0084]

[0085] where, is the output of the multi-head self-attention, , , , are learnable parameters, and GELU is a commonly used activation function.

[0086] Step 4: The high-dimensional feature tensor is converted by two parallel learnable projection heads in the feature-guided decoupling mechanism, generating two task-oriented feature flows, namely the state of charge feature flow and the state of health feature flow. The structure of the two learnable projection heads (Projection Heads) is a fully connected network, and they have independent parameter update paths for learning the semantic space representation most suitable for the corresponding task. The core idea of this design is to capture unified time series features based on shared bottom-layer encoders, and then complete the exclusive customization of upper-layer representations through the task shunting module, thereby enhancing the collaborative modeling capability between multiple tasks.

[0087] where, the high-dimensional feature tensor can be denoted as:

[0088] where, T here represents the number of time steps, and d represents the dimension of the encoded features. To further serve the dual-task modeling goal (SOC and SOH), the state of charge feature flow is: , and the state of health feature flow is: ;

[0089] Step 5: The dual-stream mutual attention module calculates the cross-attention information between the charging state feature stream and the health state feature stream, and dynamically regulates and fuses the cross-attention information through a dynamic interaction mechanism to obtain the final dynamic interaction results of the charging state and the health state; the cross-attention information includes the attention from the charging state to the health state. And attention in the direction from healthy state to charging state ;in, For charging state characteristic flow ( Feature vector mapping in ) and All are health status feature flows ( Feature vector mapping in ) This is the scaling factor; For health status feature flow ( Feature vector mapping in ) and All are charging state characteristic flows ( The feature vector mapping in ); the dynamic interaction mechanism adaptively regulates through gating functions and interaction scoring functions, and respectively maps the gating function values ​​and interaction scoring function values ​​with , Weighted fusion is performed so that the model can dynamically determine "how much" and "how" to fuse information from each other's feature streams under different aging stages and operating conditions.

[0090] Gating Function

[0091] First, define an average gating function. To characterize the interaction intention or intensity of the two feature streams at the current stage; the sigmoid function is used to map the learnable parameter W to the interval of 0~1:

[0092] ;

[0093] Let W receive the predicted values ​​of SOC and SOH as inputs simultaneously, and the gating will better reflect the actual battery state; if the value of G is large, it means that the model relies more on cross-attention fusion; if the value of G is small, it means that the independent features of the current stream are more important.

[0094] Interactive scoring function (score)

[0095] To enable gating to have greater decision-making flexibility in different scenarios, an interactive scoring function is further defined. ,in and respectively represent the current predicted SOC and SOH of the model; measure the consistency or complementarity of SOC and SOH:

[0096]

[0097] wherein, represents a nonlinear mapping of the SOC, SOH difference or correlation, is an adjustable hyperparameter; if the difference is large or the complementarity is strong, the score value may be higher; if the correlation is weak or the difference is too small, the score value is low;

[0098] Dynamic fusion strategy

[0099] After obtaining G and score, the cross-attention output can be weighted and fused to form the final dynamic interaction result. For the SOC stream, the following method can be used for coupling:

[0100]

[0101] wherein, represents element-wise multiplication, which is used to further adjust the cross-attention output according to the score, and similarly, the SOH stream can also use a similar gating weighting strategy; through such a dynamic mechanism, the model can reduce the influence of mutual attention on the final output in the early aging stage (SOC and SOH relationship is relatively loose); and in the later aging stage (capacity decay intensifies, the correlation between the two rises), the mutual attention weight is increased, so as to capture more significant coupling effect;

[0102] Step 6: The physical constraint decoder maps the final state of charge dynamic interaction result and the state of health dynamic interaction result to the state of charge prediction value and the state of health prediction value based on the multi-layer perceptron and the physical correction term; Specifically:

[0103] Step 6.1: input the state of charge dynamic interaction result and the state of health dynamic interaction result into two multi-layer perceptrons respectively to obtain the preliminary and ;

[0104] Step 6.2: in each time slice, by integrating or recursively calculating the cumulative charge and temperature variables in the current period, calculate the coulomb integral constraint term and the Arrhenius degradation term ;

[0105] Step 6.3: input the preliminary and respectively with the coulomb integral constraint term and the Arrhenius degradation term Fusion is performed at the decoder end to form the final With ;

[0106] where,

[0107] Based on the Coulomb's law constraint, its continuous form is:

[0108] ;

[0109] where, is the rated capacity, For the discrete implementation of Coulomb efficiency, the integral can be replaced by step-by-step accumulation, and at each time step, is clamped to prevent numerical drift; in the decoder, the Coulomb constraint often takes the form of a CoulombConstraint layer, which first estimates based on the current period's current, temperature, and other characteristics of the network, and then couples with the output of the MLP; the simplified coupling method is:

[0110] ;

[0111] where, is the Sigmoid function to ensure that the output is in the [0, 1] interval; is an adjustable weight or gating factor to balance data-driven prediction and physical integral results; in this way, the network can automatically correct the SOC deviation under most working conditions and maintain physical consistency with charge conservation;

[0112] Based on the Arrhenius model, the capacity decay and life decay rate are related to temperature and cycle depth:

[0113] ;

[0114] where, is the cumulative discharge amount or cycle number, represents the decay nonlinearity coefficient; the core of this equation is to dynamically calculate the decay amount based on the current temperature and cumulative capacity (or cycle depth), and output a correction term that can be multiplied or added to the predicted by the MLP; the coupling method is:

[0115] ;

[0116] where, For the Sigmoid function, through such a joint output layer, the model can strictly follow the charge conservation, temperature-reaction rate, etc. electrochemical priori, while maintaining high prediction accuracy, providing more physically credible results for subsequent applications.

[0117] In addition, from the prediction error , physical consistency and cross consistency Three dimensions constrain and optimize the double-cross physical guided Transformer model, and the overall loss is:

[0118] ;

[0119] Among them and are hyperparameters, used to balance the contribution of different sub-losses;

[0120] ;

[0121] ;

[0122] ;

[0123] Among them, and respectively represent the predicted values of the model for the state of charge and the state of health, and respectively represent the true values of the model for the state of charge and the state of health; is an adjustment parameter, used to balance the deviation between the predicted capacity and the actual charge transfer calculated by the integral value of the actual observed current This term ensures that the model output conforms to the principle of charge conservation and avoids serious deviation of SOC estimation due to pure data-driven; is an adjustment parameter, used to control and the consistency degree of Arrhenius equation ; is the rated capacity of the new battery; represents the inherent physical quantity of the battery material, A is the exponential pre-factor in the Arrhenius model, which is set as a constant here; T here is the current battery temperature, and I is the current.

[0124] The joint verification experiment of the invention method is carried out, and the results are shown in Table 1, which lists the SOC-SOH joint estimation error indicators of multiple experimental samples such as battery B201, B202, B203, etc., including MAE, RMSE, MAXE.

[0125] Figure 3 The SOC prediction results of battery B201 during multiple charge-discharge cycles are shown, with the red curve representing the true SOC and the blue curve representing the model prediction value. It can be observed that the model can accurately track the true curve within a wide charge-discharge range, especially during the discharge plateau and rapid charging phase, with the error between the two remaining at a low level, only showing small deviations at individual peaks and valleys. This deviation is usually related to instantaneous rate changes or polarization resistance fluctuations, indicating that the model can maintain good robustness when facing more complex dynamic conditions. Overall, the SOC prediction curve shows a highly consistent trend with the true value over time, proving the effectiveness of the multi-task fusion and physical constraint strategy in capturing charge conservation and charge-discharge behavior.

[0126] In combination with Figure 4 and Figure 5 It can be seen that the overall error level of each battery in SOC prediction is low, with the maximum error generally controlled within an acceptable range, indicating that the model can maintain good fitting degree under different cycle conditions and health levels. For SOH, although the decay mechanism is more complex, the error can still be maintained within a small range, especially during the early and middle stages of decay, with the prediction curve and true value being basically consistent. In summary, the joint estimation method, after integrating multi-task learning and physical constraints, not only effectively captures the capacity decay law of the battery at different aging stages, but also accurately depicts the current state of charge, achieving stable and reliable joint prediction in most working conditions.

[0127] In summary, the joint estimation method, after integrating multi-task learning and physical constraints, not only effectively captures the capacity decay law of the battery at different aging stages, but also accurately depicts the current state of charge, achieving stable and reliable joint prediction in most working conditions.

[0128] Table 1: Evaluation index results of SOC-SOH joint estimation

[0129]

[0130] In the SOH model comparison estimation experiment, we used the "double cross physical guidance model" (DCPGM) and compared its performance with three classic comparison models: support vector machine (SVM), extreme learning machine (ELM), and bidirectional long short-term memory network (BiLSTM). Through the performance comparison of these models, we verified the advantages of DCPGM in SOH estimation.

[0131] Table 2: SOH comparison experiment results

[0132]

[0133] As shown in Table 2, the experimental results show that DCPGM performs well in SOH estimation; specifically, the MAE of DCPGM is 0.003, the RMSE is 0.004, and the MAXE is 0.024, which are significantly lower than those of the other three comparison models, especially in the MAXE (maximum error) indicator, DCPGM successfully controls the maximum error at a low level; the model can accurately capture the state of health of the battery at different stages of degradation, especially in the early and middle stages of degradation, showing strong stability and accuracy; this indicates that DCPGM can effectively deal with the complex changes in battery health, especially in the early and middle stages of battery degradation, and can stably provide accurate SOH predictions;

[0134] In contrast, the SVM and ELM models perform poorly in SOH estimation, especially when the battery degradation is more significant, with the maximum error (MAXE) of SVM being 0.052, which is much higher than that of DCPGM. This indicates that SVM cannot accurately track the health status of the battery when dealing with rapid changes in battery health, and there is a large prediction bias; the MAXE of ELM is 0.061, although ELM has high training efficiency, but the accuracy in SOH estimation is still insufficient, especially in the early stage of battery aging, its performance cannot be compared with DCPGM;

[0135] In summary, DCPGM has obvious advantages in SOH estimation, through the double-cross physical guidance framework, DCPGM can better capture the nonlinear characteristics in the battery degradation process, and maintain high prediction accuracy in different stages of battery degradation; compared with traditional SVM and ELM models, DCPGM can more accurately reflect the health status of the battery, especially during the battery degradation process; although BiLSTM can also process time series data, it does not fully consider the coupling relationship between SOC and SOH, so its accuracy is slightly inferior to DCPGM. Therefore, DCPGM exhibits excellent performance in battery health state estimation, especially in the complex battery degradation process, showing higher accuracy and stability;

[0136] In the SOC model comparison estimation experiment, we used the double-cross physical guidance model (DCPGM) and compared its performance with three other comparison models: convolutional neural network (CNN), long short-term memory network (LSTM), and particle swarm optimization combined with long short-term memory network (PSO-LSTM); through the performance comparison of these models, we verified the advantages of DCPGM in SOC estimation;

[0137] Table 3: SOC comparison experiment results

[0138]

[0139] As shown in Table 3, the experimental results show that DCPGM performs best in SOC estimation, with MAE of 0.003, RMSE of 0.004, and MAXE of 0.019, all of which are significantly lower than other comparative models. DCPGM can accurately capture the short-term dynamic response of the battery through the double-cross physical guidance framework, especially in the fast charging and discharging process of the battery, showing high accuracy and stability; this shows that DCPGM not only can better handle the time series data of the battery, but also can maintain high precision under complex battery conditions;

[0140] In contrast, CNN also performs well in SOC estimation, but still lags behind DCPGM. The MAE of CNN is 0.004, the RMSE is 0.005, and the MAXE is 0.050, although CNN can effectively extract features in the battery charging and discharging process, its prediction accuracy is low under complex battery conditions, especially when the battery is aging or fast charging, the error is relatively large; the LSTM model can better handle the time series data of the battery, with MAE of 0.003, RMSE of 0.004, and MAXE of 0.048; the performance of the LSTM model is close to that of DCPGM, but in some complex dynamic conditions, the prediction accuracy of LSTM is slightly insufficient, especially in the process of rapid change of the battery, the error is large; PSO-LSTM optimizes the hyperparameters of LSTM through particle swarm optimization, its MAE and RMSE are close to DCPGM, but its MAXE performance is slightly inferior, with a maximum error of 0.045; although PSO-LSTM can improve the accuracy of LSTM, it still has some error when dealing with complex dynamic characteristics of the battery;

[0141] In summary, DCPGM has obvious advantages in SOC estimation, by combining the learning ability of physical guidance and data-driven, DCPGM can accurately capture the short-term dynamic response and long-term degradation characteristics of the battery, especially in the fast charging and discharging process, maintaining high accuracy and stability; in contrast, other models have larger errors when dealing with complex battery conditions, especially in the case of battery aging or rapid change, DCPGM performs better.

[0142] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above examples, and the above examples and descriptions in the specification are only to illustrate the principles of the present application, without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A SOC-SOH combined estimation method based on a double-cross physical guidance framework, characterized in that: The method comprises the following steps: Step 1: obtaining data features of battery charging and discharging; specifically comprising the following steps: Step 1.1: establishing an experimental platform for battery cycle life test, testing the cycle life of lithium ion battery and obtaining battery charging and discharging data; Step 1.2: performing multi-scale processing and derivative feature extraction on the original signal from the battery charging and discharging data, and obtaining features including voltage change rate, temperature gradient, cumulative throughput, equivalent cycle number, sliding window statistics and relaxation period voltage recovery rate; Step 1.3: processing all the extracted features by dynamic normalization strategy, and obtaining data features of battery charging and discharging; The dynamic normalization strategy is mathematically expressed as follows: ； wherein, is the output of the dynamic normalization strategy, is the original feature value at time step t; is a small constant to prevent division by zero; is the mean, is the standard deviation, respectively expressed as: ； ； is the dynamic sliding window length and is ； wherein, is the window initial length; is the growth rate coefficient; is the current time step model prediction of the health state, then represents the degree of health degradation; Step 2: sending the data features into a double-cross physical guidance model, which comprises a space-time feature encoder, a double-flow mutual attention module and a physical constraint decoder; Step 3: the space-time feature encoder extracts high-dimensional feature tensors in the data features based on dilated convolution, multi-head self-attention and feedforward network; Step 4: the high-dimensional feature tensors are converted by two parallel learnable projection heads in the feature guidance decoupling mechanism to generate two task-oriented feature flows, which are respectively denoted as charging state feature flow and health state feature flow; Step 5: the double-flow mutual attention module calculates the cross-attention information between the charging state feature flow and the health state feature flow, and dynamically regulates and fuses the cross-attention information through dynamic interaction mechanism to obtain the final charging state dynamic interaction result and health state dynamic interaction result; Wherein, the cross-attention information includes attention from the charging state to the health state And attention from the health state to the charging state is a feature vector mapping in the charging state feature stream, is a feature vector mapping in the health state feature stream; is a scaling factor; is a feature vector mapping in the charging state feature stream, The dynamic interaction mechanism is adaptively regulated through a gating function and an interaction score function, and the gating function value and the interaction score function value are respectively weighted and fused with ​​​​​ Step 6: the physical constraint decoder maps the final charging state dynamic interaction result and health state dynamic interaction result to the charging state prediction value and health state prediction value based on multilayer perceptron and physical correction term; specifically: Step 6.1: The state of charge dynamic interaction results and the state of health dynamic interaction results are input into two multi-layer perceptrons, respectively, to obtain preliminary and ; Step 6.2: Calculate the Coulomb integral constraint term in each time slice by integrating or recursively the cumulative charge of the current period and the temperature variable, respectively and the Arrhenius degradation term ; Step 6.3: The preliminary and are fused at the end of the decoder to form the final and Arrhenius degradation terms with ;​ ； where, is the rated capacity, For Coulomb efficiency discretization, the integral can be replaced by a step-wise accumulation, and at each time step is clipped to prevent numerical drift; in the decoder, the Coulomb constraint often comes in the form of a CoulombConstraint layer, which first estimates from the current and temperature of the network at the current time step, and then couples with the output of the MLP A simplified coupling way is: ； where, is a Sigmoid function to ensure the output lies in the interval [0, 1]; is an adjustable weight or gating factor to balance the data-driven prediction with the physical integration result; ； wherein, is the cumulative discharge capacity or cycle number, represents the attenuation nonlinear coefficient; the core of the equation is to dynamically calculate the attenuation amount according to the current temperature and the cumulative capacity, and output a correction term multiplied or added with the MLP predicted ; the coupling mode is: ； wherein, is a sigmoid function.

2. The SOC-SOH co-estimation method based on the dual-cross physical guidance framework according to claim 1, characterized in that: In the space-time feature encoder, the data features are processed by dilated convolution, multi-head self-attention layer and feedforward network layer in turn, and residual connection is set in the multi-head self-attention layer and the feedforward network layer, and layer normalization is set after the multi-head self-attention layer and the feedforward network layer.

3. The SOC-SOH co-estimation method based on the dual-cross physical guidance framework according to claim 1, characterized in that: from the prediction error , physical consistency and cross-consistency The three dimensions constrain and optimize the Transformer model with double-cross physical guidance, and the overall loss is: ； wherein with is a hyper-parameter for balancing the contribution of different sub-losses; ； ； ； in, and These represent the model's predicted values ​​for the charging state and the health state, respectively. and These represent the model's actual values ​​for the charging state and the health state, respectively. To adjust parameters for balancing prediction capacity Compared with the integral value of the actually observed current The calculated deviation between the actual charge transfers; To adjust parameters for control With the Arrhenius equation The degree of consistency; The rated capacity of the new battery; The values ​​represent the inherent physical quantities of the battery materials, where A is the exponential pre-factor in the Arrhenius model; T is the current battery temperature; and I is the current.

Images