High-fidelity estimation method for lithium ion battery health state under complex working condition

Abstract

The invention discloses a high-fidelity estimation method for a lithium ion battery health state under a complex working condition, and belongs to the technical field of lithium battery health state monitoring, and the method comprises the steps: extracting short-time voltage and current segments from a constant current-constant voltage charging stage of a battery as model input signals; mapping the short-time signal data into an entropy domain feature sequence to enhance degradation characterization; constructing a data fitting item containing double weighting, an experience degradation constraint item and a solid-phase interfacial film growth mechanism constraint item as degradation loss constraints; establishing a physical information neural network model integrating an attention mechanism and time sequence modeling to output battery health state baseline estimation; and constructing an explicit sparse dynamical model to realize residual correction of the complex data set. By adopting the method, collaborative improvement of battery health state estimation precision and generalization robustness can be realized, and degradation mechanism consistency and training stability are considered.

Description

Technical Field

[0001] This invention relates to the field of lithium battery health status monitoring technology, and in particular to a high-fidelity estimation method for the health status of lithium-ion batteries under complex operating conditions. Background Technology

[0002] As the number of charge-discharge cycles increases, the internal impedance of lithium-ion batteries increases, leading to a decline in capacity and power performance. Therefore, accurately assessing battery health to understand the degradation process is crucial for ensuring the safe operation of battery systems and maximizing their lifespan.

[0003] Existing methods for estimating battery health status can generally be divided into two main categories: direct measurement methods and indirect analysis methods. Direct measurement methods obtain data on the battery under specific operating or laboratory conditions through experimental means and infer the battery state based on this data. For example, the ampere-hour integration (coulomb counting) method estimates the state by integrating the current to obtain the capacity, but it is susceptible to accumulated errors under complex operating conditions. Other methods, such as electrochemical impedance spectroscopy, internal resistance measurement, cycle counting, and destructive testing, also generally face trade-offs between accuracy, cost, and deployability, limiting their engineering application in battery management systems.

[0004] Indirect analysis methods mainly include model-based methods and data-driven methods. Model-based methods utilize electrochemical mechanism models or equivalent circuit models combined with filtering and parameter identification to achieve state estimation. These methods have a certain degree of interpretability, but they usually rely on precise calibration and are difficult to adapt to the strong nonlinearity and multi-factor coupling characteristics of battery degradation, resulting in high application costs.

[0005] Data-driven methods learn performance degradation patterns from historical test data to achieve state estimation. Deep learning methods can automatically learn nonlinear correlations in time-series data using hierarchical network structures, thereby improving estimation accuracy. However, they generally face problems such as strong data dependence, insufficient generalization across operating conditions, and uninterpretable mechanisms, making it difficult to simultaneously meet the requirements of high accuracy and high reliability. Under different battery types, different usage conditions, and different aging modes, the degradation mechanisms vary significantly, and a single data model often cannot simultaneously meet the engineering requirements of high accuracy and high reliability.

[0006] In recent years, physical information neural networks (PINs) that integrate physical models with deep learning techniques have emerged as a new research approach. They embed physical priors or mechanistic constraints into the network structure or loss function to enhance mechanistic consistency and stability. Nevertheless, existing PIN methods still have shortcomings: insufficient fusion of physical features and deep features, limited ability to characterize complex decay processes, and a lack of systematic improvement methods for generalizing the model structure and constraint forms under different chemical systems and operating conditions. Summary of the Invention

[0007] The purpose of this invention is to provide a high-fidelity estimation method for the health status of lithium-ion batteries under complex operating conditions. Through an integrated framework of "entropy domain degradation characterization - mechanism constraint - fusion modeling", the method achieves a synergistic improvement in the accuracy, robustness and interpretability of battery health status estimation, and effectively reduces operating costs.

[0008] To achieve the above objectives, this invention provides a high-fidelity estimation method for the health status of lithium-ion batteries under complex operating conditions, comprising the following steps: S1. General input signals: Short-time voltage and current segments are extracted from the constant current-constant voltage charging stage of the lithium-ion battery and used as input signals for the model. S2. Entropy domain feature extraction: Mapping short-time signal data into an entropy domain feature sequence to enhance degradation characterization; S3. Physical loss constraint: Construct a physical degradation loss function that includes a double-weighted data fitting term, an empirical degradation constraint term, and a solid-phase interface film growth mechanism constraint term; S4. Model Architecture: Establish a physical information neural network model that integrates attention mechanism and time series modeling to output the baseline estimate of battery health state (SOH); S5. Residual Adaptive Correction: Construct an explicit sparse dynamic model to achieve residual correction for complex datasets.

[0009] Preferably, the short-time voltage segment in S1 is the interval data of the constant current charging stage of the battery, and the short-time current segment is the interval data of the constant voltage stage, and they are time-aligned and interpolated and resampled to form a unified input charging cycle signal.

[0010] Preferably, the entropy domain feature extraction in S2 employs a refined composite multi-scale Hilbert cumulative residual entropy algorithm. The short-time charging cycle signal is sequentially subjected to scale decomposition, Fourier transform, Hilbert filter, and inverse Fourier transform to obtain an analytical signal, which is then normalized in the frequency domain. By calculating the cumulative residual entropy and performing refined composite and averaging along the scale dimension, it is mapped to an entropy domain feature sequence; the formula is shown below: ; in, The cumulative distribution function; The input charging cycle signal; The scaling factor; Indicated in scale Next After Hilbert transform and frequency normalization, the subsequences are indexed in the discrete index. The probability distribution value at that point, i.e., the normalized value of the power spectrum.

[0011] Preferably, S3 constructs the physical degradation loss function, specifically including the following: S31. Construct a double-weighted data fitting loss, which is specifically expressed as follows: ; in, and These represent the actual and estimated SOH values, respectively. The time weight is set to increase linearly from 0.3 to 1; , These represent batch size and sequence length, respectively; trend-aware weights. Defined as: , and These represent the changes in the true value and the estimated value at adjacent time points, respectively. S32. Construct a monotonic regularization constraint loss to penalize the rising increment of the SOH predicted trajectory, which is specifically expressed as follows: ; in, , These represent the estimated values ​​at adjacent time points; This is the monotonicity tolerance threshold, used to allow extremely small numerical fluctuations without penalty; This is an enhancement factor used to adjust the intensity of the penalty; S33. Construct a global slope constraint loss to suppress long-term trend deviations from the overall downward trend, specifically expressed as follows: ; in, , They represent the first The sample at time 1 and the sample at time 2 The estimated value at each moment; This indicates taking the maximum value; S34. Construct a mechanism-constrained loss to characterize the effects of solid-phase interfacial (SEI) film growth on cyclic lithium loss and capacity decay, specifically expressed as follows: ; in, These represent the nominal capacity of the battery and Remaining capacity at any given time; This represents the estimated degradation of the battery's health status due to the growth of the solid-phase interface film. Let be the molar gas constant, with a value of . ; Let be the Faraday constant, and let its value be . ; The SEI growth coupling coefficient has a value of [value missing]. ; The reaction density for SEI formation is set to a value of ; Let be the reaction order, and take the value . ; This is the reaction rate proportionality coefficient, with a value of [value missing]. ; These represent absolute temperature, characteristic current density, and SEI film overpotential parameters, respectively, and are considered as adjustable variables inverted through iterative training in the network. S35. Construct the total loss function, and to address the issue of inconsistencies in the dimensions and numerical scales between the data fitting term and each physical constraint term, introduce adjustable weight coefficients and dynamic scaling factors to achieve adaptive balance. Specifically, this is expressed as follows: ; Among them, the scaling factors of each part Constrained Within the range, As an auxiliary loss term, To smooth out the terms and prevent the denominator from being too small; These are adjustable weighting coefficients for each loss term.

[0012] Preferably, the physical information neural network model in S4 includes a multi-head attention feature extraction module and a temporal modeling module; specifically, it includes the following: S41. The multi-head attention feature extraction module is a Transformer encoding structure that includes an input layer, an encoding layer, a multi-head attention module, a feedforward network module, and an output layer. It captures the global dependency characteristics of the time series and transforms the current and voltage entropy domain feature sequences into encoded feature representations. S42, the timing modeling module is a BiGRU structure consisting of an input layer, a two-layer propagation layer, an activation layer, and an output layer. It models local nonlinear timing features through forward and backward gated recursive parallel modeling, and concatenates the bidirectional hidden states to obtain a timing representation containing contextual information; the bidirectional hidden states... The expression is: ; in, , These represent the forward GRU hidden state and the reverse GRU hidden state, respectively. S43. The outputs of the multi-head attention feature extraction module and the temporal modeling module are indexed and aligned before being output by the fully connected layer as the SOH baseline estimate.

[0013] Preferably, in S41, the specific steps are as follows: S411. Encode each position by combining sine and cosine functions to ensure that the model understands the relative and absolute positional relationships between the data. S412. Calculate the weights of the self-attention mechanism to capture the dependencies between data points in the input sequence. The formula for the self-attention mechanism is: ; in, It is a query matrix; It is a key matrix; It is a value matrix; It is the dimension of the key; The dot product between the query and the key is used to obtain the attention weights through the softmax function, and finally, the result is multiplied by the value matrix. Multiply them to get the weighted sum; S413, Utilizing multi-head attention mechanisms to... The space is divided into multiple subspaces, and the attention outputs of each subspace are computed in parallel. The attention calculation formula is as follows: ; in, The number of heads is represented by a linear transformation weight matrix. The outputs of all heads are concatenated and mapped to the final dimension; S414: A feedforward neural network is used to perform a nonlinear transformation on the representation at each position. The feedforward neural network consists of two linear transformations and an activation function. The expression for the nonlinear transformation is: ; in, The feedforward network receives input feature vectors; and It is a weight matrix; and It is a bias term; ReLU is a non-linear activation function to increase the non-linear expressive power of the model.

[0014] Preferably, the specific steps in S43 are as follows: S431. Align the multi-head attention encoding output sequence with the BiGRU output sequence at the same time step and concatenate them to obtain the fused feature sequence expression: ; in, Indicates the Transformer structure at the 1st... The attention-encoded feature vectors output at each time step; BiGRU in the first A bidirectional gated temporal feature vector output at each time step; The vector concatenation operator represents the fusion feature that simultaneously contains global dependency representations and local temporal representations. ; S432. The fully connected layer integrates the input features extracted by the previous layer into a high-dimensional representation of the current task, and outputs the SOH baseline estimate. The output calculation formula is as follows: ; in, The input is the feature fusion vector; This is a weight matrix used to represent the connection weights between input and output nodes; This is a bias vector used to adjust the activation threshold; It is an activation function used to introduce nonlinearity, enabling the network to express more complex feature relationships.

[0015] Preferably, the explicit sparse dynamics model described in S5 sparsely represents the residual evolution relationship by constructing a candidate basis function library, and uses regression optimization with sparse constraints to solve the model parameters to obtain the predicted residual values ​​for SOH correction; specifically, it includes the following: S51. Construct the residual sequence based on the SOH baseline estimate. Its expression is: ; in, This is the cycle time; For residuals; For true SOH; Output SOH for the baseline model; S52. Construct a candidate basis function library for characterizing residual evolution relationships. The residuals and their combination with auxiliary variables are used to form a candidate library vector. Its expression is: ; in, As an auxiliary variable; For candidate basis function library mapping, belonging to 3D real vector; Includes polynomials, cross terms, and , Nonlinear functions, including those used to cover the residual evolution form; S53, using the following: - The mixed regularization form of the regularization term constructs the sparse-constrained regression objective function, whose expression is: ; in, for k The residual at time +1; Let be the sparse coefficient vector to be determined; , To control the trade-off between sparsity and smoothness; S54. The discrete evolution relationship of the residuals can be expressed using a sparse linear combination, and its expression is as follows: ; in, This is the sparse coefficient vector after integration. To correct residuals; S55. The predicted residuals are superimposed on the baseline estimate to obtain the corrected SOH, the expression of which is: ; in, This is the estimated value after residual correction. This is the baseline estimate.

[0016] Therefore, the present invention employs the above-mentioned high-fidelity estimation method for the health status of lithium-ion batteries under complex operating conditions, which has the following beneficial effects: 1) Based on the refined composite multi-scale Hilbert cumulative residual entropy algorithm, entropy domain features are extracted from short-time constant current and constant voltage charging segments, which can robustly characterize degradation information and improve robustness and cross-condition generalization ability under noise disturbance and operating condition changes. 2) By jointly constructing a physical degradation joint loss function through dual-weighted fitting loss, empirical degradation constraint and SEI mechanism constraint, the consistency and interpretability of the mechanism can be enhanced, and the training stability can be improved. 3) Construct a hybrid time series model of Transformer-BIGRU, use multi-head attention mechanism to capture the global dependency characteristics of time series, and use bidirectional gated recurrent units to deeply mine local time series features, significantly enhancing the model's tracking accuracy for complex degenerate trajectories; 4) Explicit sparse residual dynamic correction can adaptively compensate for unmodeled errors, further improving the high-fidelity performance and reliability of health state estimation.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a diagram of the multi-head attention mechanism introduced in this embodiment of the invention; Figure 3 This is a diagram of the gate control unit introduced in the embodiments of the present invention; Figure 4 This is a diagram of the BiGRU architecture introduced in the embodiments of the present invention; Figure 5 This is a sparse residual correction diagram according to an embodiment of the present invention; Figure 6 This is a comparison chart of the true and estimated values ​​on different methods in the XJTU dataset of this invention embodiment; Figure 7 This is a comparison chart of the true and estimated values ​​on different methods in the HUST dataset of this invention embodiment; Figure 8 This is a comparison chart of the actual values ​​and estimated values ​​using different methods in the MIT dataset of this invention. Detailed Implementation

[0019] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] Example 1 like Figures 1-5 As shown, this invention provides a high-fidelity estimation method for the health status of lithium-ion batteries under complex operating conditions, comprising the following steps: S1. Obtain three publicly available battery datasets (XJTU dataset, HUST dataset, and MIT dataset) to capture short-time voltage segments, current segments, and health state characterization sequences of lithium-ion batteries during the constant current-constant voltage (CC-CV) charging phase. Integrate these sequences by charging cycle as input signals for the model. The datasets are based on their respective sensor sampling intervals. Specifically, they include: S11, Constant Current Stage Voltage Segment: Described as the cutoff charging voltage. Select The voltage signal corresponding to the interval.

[0022] S12, Constant Voltage Stage Current Segment: Based on the cutoff charging current... Select The current signal corresponding to the interval.

[0023] S13. Health status characterization sequence: The health status is characterized by the battery capacity decay. The ratio of the maximum discharge capacity of the cycle to its nominal capacity is selected as the corresponding status value.

[0024] S2. A refined composite multi-scale Hilbert cumulative residual entropy algorithm is used to map short-time voltage / current signal data into an entropy domain feature sequence to enhance degradation characterization. The algorithm sequentially performs scale decomposition, Fourier transform, Hilbert filter, and inverse Fourier transform to obtain the analytic signal, which is then normalized in the frequency domain. The cumulative residual entropy is calculated and refined composited and averaged across the scale dimension to map it into an entropy domain feature sequence. The dataset is divided into an 80% training set and a 20% test set based on battery type. The formula is shown below: ; in, The cumulative distribution function; The input charging cycle signal; The scaling factor; Indicated in scale Next After Hilbert transform and frequency normalization, the subsequences are indexed in the discrete index. The probability distribution value at that point, i.e., the normalized value of the power spectrum.

[0025] S3 constructs a physical degradation loss function that includes a double-weighted data fitting term, an empirical degradation constraint term, and a solid-phase interface film growth mechanism constraint term to enhance cross-condition stability and physical consistency. Specifically, it includes: The data fitting loss with dual weighting is constructed and expressed as follows: ; in, and These represent the actual and estimated SOH values, respectively. The time weight is set to increase linearly from 0.3 to 1; , These represent batch size and sequence length, respectively; trend-aware weights. Defined as: , and These represent the changes in the true value and the estimated value at adjacent time points, respectively.

[0026] S32. Construct a monotonic regularization constraint loss to penalize the rising increment of the SOH predicted trajectory, which is specifically expressed as follows: ; in, , These represent the estimated values ​​at adjacent time points; This is the monotonicity tolerance threshold, used to allow extremely small numerical fluctuations without penalty; This is an enhancement factor used to adjust the intensity of the penalty.

[0027] S33. Construct a global slope constraint loss to suppress long-term trend deviations from the overall downward trend, specifically expressed as follows: ; in, , They represent the first The sample at time 1 and the sample at time 2 The estimated value at each moment; This indicates taking the maximum value.

[0028] S34. Construct a mechanism-constrained loss to characterize the effects of solid-phase interfacial (SEI) film growth on cyclic lithium loss and capacity decay, specifically expressed as follows: ; in, These represent the nominal capacity of the battery and Remaining capacity at any given time; This represents the estimated degradation of the battery's health status due to the growth of the solid-phase interface film. Let be the molar gas constant, with a value of . ; Let be the Faraday constant, and let its value be . ; The SEI growth coupling coefficient has a value of [value missing]. ; The reaction density for SEI formation is set to a value of ; Let be the reaction order, and take the value . ; This is the reaction rate proportionality coefficient, with a value of [value missing]. ; These represent absolute temperature, characteristic current density, and SEI film overpotential parameters, respectively, and are considered as adjustable variables retrieved through iterative training in the network. In this embodiment, their value ranges are set to [values ​​to be inserted here]. .

[0029] S35. Construct the total loss function, and to address the issue of inconsistencies in the dimensions and numerical scales between the data fitting term and each physical constraint term, introduce adjustable weight coefficients and dynamic scaling factors to achieve adaptive balance. Specifically, this is expressed as follows: ; Among them, the scaling factors of each part Constrained Within the range, As an auxiliary loss term, To smooth out the terms and prevent the denominator from being too small; These are adjustable weighting coefficients for each loss term.

[0030] S4. Establish a physical information neural network model that integrates attention mechanism and temporal modeling to output the SOH baseline estimate, specifically including: S41. Input the battery entropy domain feature data into the physical information neural network model. The multi-head attention feature extraction module is a Transformer encoding structure including an input layer, an encoding layer, a multi-head attention module, a feedforward network module, and an output layer, as follows: Figure 1 , 2 As shown, the global dependency characteristics of the time series are captured, and the current-voltage entropy domain feature sequence is transformed into an encoded feature representation. The specific steps are as follows: S411. Encode each position by combining sine and cosine functions to ensure that the model understands the relative and absolute positional relationships between the data. S412. Calculate the weights of the self-attention mechanism to capture the dependencies between data points in the input sequence. The formula for the self-attention mechanism is: ; in, It is a query matrix; It is a key matrix; It is a value matrix; It is the dimension of the key; The dot product between the query and the key is used to obtain the attention weights through the softmax function, and finally, the result is multiplied by the value matrix. Multiply them to get the weighted sum; S413, Utilizing multi-head attention mechanisms to... The space is divided into multiple subspaces, and the attention outputs of each subspace are computed in parallel. The attention calculation formula is as follows: ; in, The number of heads is represented by a linear transformation weight matrix. The outputs of all heads are concatenated and mapped to the final dimension; S414: A feedforward neural network is used to perform a nonlinear transformation on the representation at each position. The feedforward neural network consists of two linear transformations and an activation function. The expression for the nonlinear transformation is: ; in, The feedforward network receives input feature vectors; and It is a weight matrix; and It is a bias term; ReLU is a non-linear activation function to increase the non-linear expressive power of the model.

[0031] S42, the timing modeling module is a bidirectional gated recurrent unit (BiGRU) structure including an input layer, a two-layer propagation layer, an activation layer, and an output layer. For example... Figure 1 , 3 As shown in Figure 4, local nonlinear temporal features are modeled in parallel using forward and backward gated recursion, and the bidirectional hidden states are concatenated to obtain a temporal representation containing contextual information. Bidirectional hidden states The expression is: ; in, , These represent the forward GRU hidden state and the reverse GRU hidden state, respectively. S43. The outputs of the multi-head attention feature extraction module and the temporal modeling module, after index alignment, are output by the fully connected layer as the SOH baseline estimate. The specific steps are as follows: S431. Align the multi-head attention encoding output sequence with the BiGRU output sequence at the same time step and concatenate them to obtain the fused feature sequence expression: ; in, Indicates the Transformer structure at the 1st... The attention-encoded feature vectors output at each time step; BiGRU in the first A bidirectional gated temporal feature vector output at each time step; The vector concatenation operator represents the fusion feature that simultaneously contains global dependency representations and local temporal representations. .

[0032] S432. The fully connected layer integrates the input features extracted by the previous layer into a high-dimensional representation of the current task, and outputs the SOH baseline estimate. The output calculation formula is as follows: ; in, The input is the feature fusion vector; This is a weight matrix used to represent the connection weights between input and output nodes; This is a bias vector used to adjust the activation threshold; It is an activation function used to introduce nonlinearity, enabling the network to express more complex feature relationships.

[0033] S5. Construct an explicit sparse dynamic model to achieve residual correction for complex datasets. For example... Figure 5As shown. Its core is to represent the residual evolution relationship as a sparse combination on a candidate basis function library, and to introduce a block resampling ensemble strategy to weaken the temporal correlation of the residual sequences, specifically including: S51. Construct the residual sequence based on the SOH baseline estimate. Its expression is: ; in, This is the cycle time; For residuals; For true SOH; Output SOH for the baseline model.

[0034] S52. Construct a candidate basis function library for characterizing residual evolution relationships. The residuals and their combination with auxiliary variables are used to form a candidate library vector. Its expression is: ; in, As an auxiliary variable; For candidate basis function library mapping, belonging to 3D real vector; Includes polynomials, cross terms, and optional terms. , Nonlinear functions are used to cover the residual evolution form.

[0035] S53, using the following: - The mixed regularization form of the regularization term constructs the sparse-constrained regression objective function, whose expression is: ; in, for k The residual at time +1; Let be the sparse coefficient vector to be determined; , To control the trade-off between sparsity and smoothness.

[0036] S54. The discrete evolution relationship of the residuals can be expressed using a sparse linear combination, and its expression is as follows: ; in, This is the sparse coefficient vector after integration. To correct the residuals.

[0037] S55. The predicted residuals are superimposed on the baseline estimate to obtain the corrected SOH, the expression of which is: ; in, This is the estimated value after residual correction. This is the baseline estimate.

[0038] It should be noted that the model performance evaluation in this embodiment is geared towards the continuous value regression task of battery state of health (SOH), and uses mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2 The estimation accuracy and goodness of fit are quantitatively evaluated, and the calculation formula is as follows: ; in, For the first The true SOH of each sample; For the corresponding predicted SOH; This represents the average value of the true SOH. The number of samples to be predicted is used. Among the above indicators, the smaller the MAE, RMSE, and MAPE, the lower the error. The closer to 1, the higher the degree of fit.

[0039] This analysis is based on three representative publicly available datasets: XJTU (including 2C / 3C / R2.5 operating conditions), HUST (multi-stage discharge protocol), and MIT (various fast charging strategy subsets). Table 1 details the chemical composition, nominal capacity, cutoff voltage, temperature, and number of batteries for each dataset.

[0040] Table 1 Dataset Information

[0041] To ensure the fairness of the examples and the validity of the results, the experimental facility was equipped with an RTX 5050 GPU with 32GB of memory, running Windows 11, and using PyTorch 2.8.9 as the deep learning framework. The model parameters were thoroughly debugged to ensure efficient and stable model training and inference processes. Table 2 details the specific parameter settings.

[0042] Table 2 Model Parameter Settings

[0043] To verify the effectiveness and superiority of the present invention in estimating the health status of batteries in cyclic charge-discharge sequence data, comparative experiments, ablation experiments, and trajectory verification were conducted to systematically evaluate the performance of the method of the present invention from the aspects of quantitative indicators, module contribution, and full-lifetime degradation trajectory tracking capability.

[0044] 1. Comparative Experiment Based on three battery charge-discharge cycle datasets, comparative experiments were conducted with classic and commonly used methods, where "Proposed" represents the method of this invention. The experimental results are shown in Table 3, demonstrating that the estimation performance of this invention significantly outperforms other methods.

[0045] Table 3. Evaluation metrics of the comparison method on three battery datasets.

[0046] Results show that this method captures global dependencies through the Transformer's multi-head attention mechanism, models local features using BiGRU, and adaptively corrects systematic prediction biases under complex conditions by introducing physical degradation constraints and employing an explicit sparse residual correction strategy during training. Compared with other models, this method significantly reduces MAE, RMSE, and MAPE on three types of battery datasets, and achieves higher R-values. 2 The result is closer to 1, demonstrating higher estimation accuracy and verifying the effectiveness and superiority of this method in the high-fidelity estimation task of battery SOH.

[0047] 2. Ablation test The ablation experiment aims to analyze the specific contribution of each module of the present invention to the model performance by progressively removing them. Table 4 shows the experimental results of the performance metrics of the present invention after progressively removing modules on three battery charge-discharge cycle datasets. Here, Proposed(ALL) represents the method of the present invention, Only-Transformer represents the estimation method without BiGRU local modeling, Only-BiGRU represents the estimation method without Transformer global dependencies, without_SRC represents the estimation method without sparse residual correction strategies, and Without_SRC_physics represents the estimation method without physical degradation constraints and sparse residual correction strategies.

[0048] Table 4 Ablation Experiment Results

[0049] The results of the implementation examples demonstrate that the global dependency modeling module of Transformer, the local temporal modeling module of BiGRU, sparse residual correction (SRC), and physical constraints all play crucial roles in the performance of battery state of health estimation. Compared with control configurations that retain only a single module or remove SRC / physical constraints, the evaluation metrics of this invention on the XJTU, HUST, and MIT datasets are significantly better than other configurations, verifying the effectiveness and superiority of this method in high-fidelity battery SOH estimation tasks under complex operating conditions.

[0050] 2. Trajectory Verification To further verify the ability of this method to dynamically characterize the evolution of battery health state, several representative battery samples were selected from three types of datasets for different operating conditions. The estimated SOH trajectories were then fitted with the actual trajectories. The results are as follows: Figure 6-8 As shown. Among them, Figure 6 The XJTU dataset samples shown exhibit long-term degradation trends, periodic fluctuations, and local "step-like" perturbation characteristics, indicating that this method can accurately characterize the overall change process and better reflect local peak and trough changes and trend reversals within the magnified local area, thereby reducing estimation errors caused by local oversmoothing or excessive oscillation. Figure 7 The HUST dataset samples shown exhibit a continuous downward trend with local perturbations, indicating that the estimated trajectory obtained by this method closely matches the actual trajectory, effectively reflecting the phased decay pattern during battery aging and demonstrating strong robustness. Figure 8 The MIT dataset samples shown exhibit longer cycle lifetimes and more complex local variation characteristics. Some samples show subtle local anomalies in the mid-to-late stages and rapid decay at the end. Under these circumstances, our method can still maintain high trajectory consistency when handling local mutations, short-term abnormal fluctuations, and inflection point changes. Figure 6-8 This method can not only characterize the overall degradation trend of battery state of health (SOH) well, but also effectively capture local fluctuations and key turning points, thus verifying its effectiveness and superiority in estimating battery health status under complex operating conditions.

[0051] Therefore, this invention adopts the above-mentioned high-fidelity estimation method for the health status of lithium-ion batteries under complex operating conditions, which combines general charging segment input, entropy domain feature extraction, physical degradation constraints and explicit sparse residual correction. This helps to improve the accuracy and stability of estimation under complex scenarios such as changes in charging strategy, noise disturbances and increased uncertainty in the later stages of degradation, while taking into account the consistency of mechanism and interpretability, thereby providing reliable support for online status monitoring and health management of battery management systems.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A high-fidelity estimation method for the state of health of lithium-ion batteries under complex operating conditions, characterized in that, Includes the following steps: S1. General input signals: Short-time voltage and current segments are extracted from the constant current-constant voltage charging stage of the lithium-ion battery and used as input signals for the model. S2. Entropy domain feature extraction: Mapping short-time signal data into an entropy domain feature sequence to enhance degradation characterization; S3. Physical loss constraint: Construct a physical degradation loss function that includes a double-weighted data fitting term, an empirical degradation constraint term, and a solid-phase interface film growth mechanism constraint term; S4. Model Architecture: Establish a physical information neural network model that integrates attention mechanism and time series modeling to output the baseline estimate of battery health state (SOH); S5. Residual Adaptive Correction: Construct an explicit sparse dynamic model to achieve residual correction for complex datasets.

2. The high-fidelity estimation method for the health status of lithium-ion batteries under complex operating conditions according to claim 1, characterized in that, The short-time voltage segment mentioned in S1 is the interval data of the constant current charging stage of the battery, and the short-time current segment is the interval data of the constant voltage stage. After time alignment and interpolation resampling, a unified input charging cycle signal is formed.

3. The high-fidelity estimation method for the health status of lithium-ion batteries under complex operating conditions according to claim 2, characterized in that, The entropy domain feature extraction described in S2 employs a refined composite multi-scale Hilbert cumulative residual entropy algorithm. This algorithm sequentially performs scale decomposition, Fourier transform, Hilbert filter, and inverse Fourier transform on the short-time charging cycle signal to obtain an analytical signal, which is then normalized in the frequency domain. By calculating the cumulative residual entropy and performing refined composite and averaging along the scale dimension, it is mapped to an entropy domain feature sequence; the formula is shown below: ； in, The cumulative distribution function; The input charging cycle signal; The scaling factor; Indicated in scale Next After Hilbert transform and frequency normalization, the subsequences are indexed in the discrete index. The probability distribution value at that point, i.e., the normalized value of the power spectrum.

4. The high-fidelity estimation method for the health status of lithium-ion batteries under complex operating conditions according to claim 3, characterized in that, S3 constructs the physical degradation loss function, specifically including the following: S31. Construct a double-weighted data fitting loss, which is specifically expressed as follows: ； in, and These represent the actual and estimated SOH values, respectively. The time weight is set to increase linearly from 0.3 to 1; , These represent batch size and sequence length, respectively; trend-aware weights. Defined as: , and These represent the changes in the true value and the estimated value at adjacent time points, respectively. S32. Construct a monotonic regularization constraint loss to penalize the rising increment of the SOH predicted trajectory, which is specifically expressed as follows: ； in, , These represent the estimated values ​​at adjacent time points; This is the monotonicity tolerance threshold, used to allow extremely small numerical fluctuations without penalty; This is an enhancement factor used to adjust the intensity of the penalty; S33. Construct a global slope constraint loss to suppress long-term trend deviations from the overall downward trend, specifically expressed as follows: ； in, , They represent the first The sample at time 1 and the sample at time 2 The estimated value at each moment; This indicates taking the maximum value; S34. Construct a mechanism-constrained loss to characterize the effects of SEI growth on cyclic lithium loss and capacity decay, specifically expressed as follows: ； in, These represent the nominal capacity of the battery and Remaining capacity at any given time; This represents the estimated degradation of the battery's health status due to the growth of the solid-phase interface film. Let be the molar gas constant, with a value of . ; Let be the Faraday constant, and let its value be . ; The SEI growth coupling coefficient has a value of [value missing]. ; The reaction density for SEI formation is set to a value of ; Let be the reaction order, and take the value . ; This is the reaction rate proportionality coefficient, with a value of [value missing]. ; These represent absolute temperature, characteristic current density, and SEI film overpotential parameters, respectively, and are considered as adjustable variables inverted through iterative training in the network. S35. Construct the total loss function, and to address the issue of inconsistencies in the dimensions and numerical scales between the data fitting term and each physical constraint term, introduce adjustable weight coefficients and dynamic scaling factors to achieve adaptive balance. Specifically, this is expressed as follows: ； Among them, the scaling factors of each part Constrained Within the range, As an auxiliary loss term, To smooth out the terms and prevent the denominator from being too small; These are adjustable weighting coefficients for each loss term.

5. The high-fidelity estimation method for the health status of lithium-ion batteries under complex operating conditions according to claim 4, characterized in that, The physical information neural network model described in S4 includes a multi-head attention feature extraction module and a temporal modeling module; specifically, it includes the following: S41. The multi-head attention feature extraction module is a Transformer encoding structure that includes an input layer, an encoding layer, a multi-head attention module, a feedforward network module, and an output layer. It captures the global dependency characteristics of the time series and transforms the current and voltage entropy domain feature sequences into encoded feature representations. S42, the temporal modeling module is a BiGRU structure consisting of an input layer, a two-layer propagation layer, an activation layer, and an output layer. It models local nonlinear temporal features through forward and backward gated recursive parallel modeling and concatenates the bidirectional hidden states to obtain a temporal representation containing contextual information; the bidirectional hidden states... The expression is: ； in, , These represent the forward GRU hidden state and the reverse GRU hidden state, respectively. S43. The outputs of the multi-head attention feature extraction module and the temporal modeling module are indexed and aligned before being output by the fully connected layer as the SOH baseline estimate.

6. The high-fidelity estimation method for the health status of lithium-ion batteries under complex operating conditions according to claim 5, characterized in that, In S41, the specific steps are as follows: S411. Encode each position by combining sine and cosine functions to ensure that the model understands the relative and absolute positional relationships between the data. S412. Calculate the weights of the self-attention mechanism to capture the dependencies between data points in the input sequence. The formula for the self-attention mechanism is: ； in, It is a query matrix; It is a key matrix; It is a value matrix; It is the dimension of the key; The dot product between the query and the key is used to obtain the attention weights through the softmax function, and finally, the result is multiplied by the value matrix. Multiply them to get the weighted sum; S413, Utilizing multi-head attention mechanisms to... The space is divided into multiple subspaces, and the attention outputs of each subspace are computed in parallel. The attention calculation formula is as follows: ； in, The number of heads is represented by a linear transformation weight matrix. The outputs of all heads are concatenated and mapped to the final dimension; S414: A feedforward neural network is used to perform a nonlinear transformation on the representation at each position. The feedforward neural network consists of two linear transformations and an activation function. The expression for the nonlinear transformation is: ； in, The feedforward network receives input feature vectors; and It is a weight matrix; and It is a bias term; ReLU is a non-linear activation function to increase the non-linear expressive power of the model.

7. The high-fidelity estimation method for the health status of lithium-ion batteries under complex operating conditions according to claim 6, characterized in that, The specific steps in S43 are as follows: S431. Align the multi-head attention encoding output sequence with the BiGRU output sequence at the same time step and concatenate them to obtain the fused feature sequence expression: ； in, Indicates the Transformer structure at the 1st... The attention-encoded feature vectors output at each time step; BiGRU in the first A bidirectional gated temporal feature vector output at each time step; The vector concatenation operator represents the fusion feature that simultaneously contains global dependency representations and local temporal representations. ; S432. The fully connected layer integrates the input features extracted by the previous layer into a high-dimensional representation of the current task, and outputs the SOH baseline estimate. The output calculation formula is as follows: ； in, The input is the feature fusion vector; This is a weight matrix used to represent the connection weights between input and output nodes; This is a bias vector used to adjust the activation threshold; It is an activation function used to introduce nonlinearity, enabling the network to express more complex feature relationships.

8. The high-fidelity estimation method for the health status of lithium-ion batteries under complex operating conditions according to claim 7, characterized in that, The explicit sparse dynamics model described in S5 sparsely represents the residual evolution relationship by constructing a candidate basis function library, and uses regression optimization with sparse constraints to solve the model parameters to obtain residual predictions for SOH correction; specifically, it includes the following: S51. Construct the residual sequence based on the SOH baseline estimate. Its expression is: ； in, This is the cycle time; For residuals; For true SOH; Output SOH for the baseline model; S52. Construct a candidate basis function library for characterizing residual evolution relationships. The residuals and their combination with auxiliary variables are used to form a candidate library vector. Its expression is: ； in, As an auxiliary variable; For candidate basis function library mapping, belonging to 3D real vector; Includes polynomials, cross terms, and , Nonlinear functions, including those used to cover the residual evolution form; S53, using the following: - The mixed regularization form of the regularization term constructs the sparse-constrained regression objective function, whose expression is: ； in, for k The residual at time +1; Let be the sparse coefficient vector to be determined; , To control the trade-off between sparsity and smoothness; S54. The discrete evolution relationship of the residuals can be expressed using a sparse linear combination, and its expression is as follows: ； in, This is the sparse coefficient vector after integration. To correct residuals; S55. The predicted residuals are superimposed on the baseline estimate to obtain the corrected SOH, the expression of which is: ； in, This is the estimated value after residual correction. This is the baseline estimate.

Images