Battery state of health estimation method and system based on soc domain reconstruction and feature decoupling

Abstract

The application discloses a battery state of health estimation method and system based on SOC domain reconstruction and feature decoupling, and belongs to the technical field of lithium ion battery state evaluation. In view of the problems that the existing data-driven model has low estimation accuracy and is prone to overfitting in a small sample scene under dynamic working conditions due to time sequence phase shift and feature coupling, the application proposes mapping original sensing data to an SOC domain to obtain sensing data in the SOC domain; intrinsic degradation trend features representing irreversible aging and transient operation disturbance features representing reversible polarization are extracted from the sensing data; the two types of features are fused to construct a joint state perception vector, which is input into a lightweight time sequence neural network model to obtain an SOH estimation value. The application eliminates working condition drift interference through SOC domain reconstruction, and distinguishes long-term degradation from short-term disturbance through feature decoupling, thereby significantly improving dynamic working condition adaptability and sparse data generalization capability, and the model is lightweight and suitable for online deployment of a battery management system with limited computing power.

Description

Technical Field

[0001] This invention relates to the field of lithium battery state assessment technology, and in particular to a battery health state estimation method and system based on SOC domain reconstruction and feature decoupling. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Battery state of health (SOH) is a core indicator for measuring the aging degree of lithium-ion batteries, and its accurate estimation is crucial for battery management system life prediction and safety early warning. Since actual health variables such as capacity and internal resistance cannot be directly measured online, they are typically inferred indirectly from external characteristic sensor signals such as voltage, current, and temperature. While existing data-driven methods do not require the construction of complex electrochemical models, their feature extraction paradigms mainly rely on the "time domain" or "cycle domain." However, in actual operation, fluctuations in charge / discharge rates, the randomness of the initial state, and the natural decay of available capacity all cause voltage and current trajectories to shift along the time axis. To unify the input length, existing technologies often employ mathematical processing techniques such as sequence truncation, zero-value padding, or equidistant resampling. However, this time-based interpolation operation severs the physical connection between external characteristic signals and the state of charge (SOC), failing to align data from different aging stages under the same electrochemical benchmark, thus introducing physical distortion at the source of feature generation.

[0004] On the other hand, the external response of a battery is the result of multi-scale process coupling: it includes both long-term irreversible capacity decay dominated by active material loss, lithium-ion loss, and SEI film thickening, and short-term reversible polarization fluctuations caused by dynamic load and environmental changes. Most existing deep learning models directly mix various health factors as equal variables as input, relying on network size to forcibly fit the overall aging trajectory, lacking a physical constraint mechanism to distinguish between long-term degradation and short-term disturbances.

[0005] Therefore, in cross-cell testing or in scenarios with only a very small number of discrete sample data, the model is prone to overfitting, resulting in serious prediction bias and loss of generalization ability. Summary of the Invention

[0006] To address the issues of data phase shift caused by time-domain feature extraction under dynamic operating conditions in existing technologies, and the difficulty of black-box models in distinguishing between long-term irreversible degradation and short-term reversible polarization disturbances, especially the problems of overfitting and prediction divergence in extremely small sample scenarios, this invention proposes a battery health state estimation method and system based on SOC domain reconstruction and feature decoupling. By mapping the non-stationary time-domain degradation trajectory to a fixed SOC space, the data misalignment caused by operating condition fluctuations is eliminated. The coupled electrical response is decomposed into intrinsic degradation trend features representing irreversible aging and reversible transient disturbance terms representing instantaneous polarization. This invention abandons the complex deep network architecture and adopts a lightweight GRU-MLP network model with low computational resource consumption to achieve the fusion of the two types of features, outputting a high-precision SOH estimate. This effectively suppresses the interference of operating condition disturbances on the estimation results and improves the generalization ability under sparse data conditions. It provides a solution for overcoming the computational barrier that makes it difficult to implement traditional complex deep learning models in vehicle battery management systems (BMS) and for efficient deployment on edge devices.

[0007] On the one hand, a battery health state estimation method based on SOC domain reconstruction and feature decoupling is provided, including: Acquire raw sensing data during the charge-discharge cycle of a lithium-ion battery, wherein the raw sensing data includes time series of voltage, current, and temperature. The original sensing data is mapped to the SOC domain to obtain sensing data in the SOC domain. The intrinsic degradation trend features characterizing irreversible aging of the battery are extracted from the sensing data in the SOC domain, and the transient operating disturbance features characterizing the reversible polarization intensity under the current operating conditions are also extracted. The intrinsic degradation trend features are fused with the transient operational disturbance features to construct a joint state perception vector; The joint state-aware vector is input into a time-series neural network model to obtain the battery health state (SOH) estimate.

[0008] Furthermore, the raw sensing data is mapped to the SOC domain, specifically including: using the ampere-hour integration method to convert the current time series into the cumulative charge capacity series, after monotonicity correction and zero-point shift, calculating the state of charge (SOC) based on the available capacity under the initial life cycle of the battery, and then mapping the voltage, current, and temperature data to a preset fixed SOC grid with equal spacing through linear interpolation.

[0009] Furthermore, intrinsic degradation trend characteristics characterizing irreversible aging of the battery are extracted from the sensing data in the SOC domain, including: Candidate features are constructed within the SOC domain, including: voltage statistical features, current statistical features, voltage versus time differential statistical features, local capacity increment features, and capacity differential IC curve features within the SOC interval. Pearson correlation coefficient and Spearman rank correlation coefficient were used to calculate the linear and nonlinear correlation between each candidate feature and the actual battery health state (SOH). Calculate the monotonicity score for each candidate feature sequence, which is used to quantify the purity of the evolution direction of the feature sequence within the global lifetime; The absolute values ​​of the Pearson correlation coefficient, the Spearman rank correlation coefficient, and the monotonicity score are normalized, and the three normalized indicators are weighted and summed to obtain the comprehensive score of each candidate feature. Several candidate features with the highest overall scores were selected as intrinsic degradation trend features.

[0010] Furthermore, the monotonicity score is calculated according to the following formula: ; in, , These represent the candidate feature sequences at the th... Second and third The specific value extracted in the next loop. n The total number of cycles, It is a symbolic function.

[0011] Furthermore, extracting transient operation disturbance features specifically includes: defining a target feature interval within the SOC domain, calculating the mean values ​​of voltage, current, and temperature within the target feature interval for each charge-discharge cycle, and using the conditional expectations of the voltage, current, and temperature as transient operation disturbance features.

[0012] Furthermore, the temporal neural network model is a single-stream GRU-MLP network, wherein the GRU layer is used to receive the joint state-aware vector at the current time and recursively update the hidden degenerate state, and the MLP layer is connected after the GRU layer to map the hidden state to the SOH estimate; the GRU layer is a single-layer GRU, the MLP layer is a two-layer MLP, and the loss function of the model adopts Huber loss.

[0013] Furthermore, the intrinsic degradation trend feature and the transient operational disturbance feature are fused together, specifically by directly concatenating the intrinsic degradation trend feature vector and the transient operational disturbance feature vector along the feature dimension to form a joint state perception vector.

[0014] On the other hand, a battery health state estimation system based on SOC domain reconstruction and feature decoupling is provided, including: The data acquisition module is configured to acquire raw sensing data during the charge-discharge cycle of a lithium-ion battery, wherein the raw sensing data includes time series of voltage, current and temperature. The SOC domain reconstruction module is configured to: map the original sensing data to the SOC domain to obtain sensing data in the SOC domain. The feature decoupling module is configured to extract intrinsic degradation trend features characterizing irreversible aging of the battery from the sensing data in the SOC domain, and simultaneously extract transient operating disturbance features characterizing the reversible polarization intensity under the current operating conditions. The feature fusion module is configured to fuse the intrinsic degradation trend feature with the transient operational disturbance feature to construct a joint state perception vector; The estimation module is configured to input the joint state-aware vector into a temporal neural network model to obtain an estimated battery health state (SOH).

[0015] In another aspect, a computer device is also provided, including a computer-readable storage medium, a processor, and a computer program stored on the computer-readable storage medium and executable on the processor, wherein when the processor executes the program, it performs the method described in the first aspect.

[0016] In another aspect, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, performs the method described in the first aspect.

[0017] The above technical solution has the following advantages or beneficial effects: (1) By mapping the original sensing data from the time domain to the SOC domain, a unified electrochemical physical benchmark is established, eliminating the time phase shift caused by the fluctuation of charge and discharge rate, the randomness of the initial state and the capacity decay, and avoiding the physical distortion caused by mathematical processing such as truncation, filling or resampling in traditional time domain feature extraction.

[0018] (2) In the SOC domain, the electrical response of the battery is explicitly decoupled into intrinsic degradation trend characteristics (characterizing irreversible aging such as loss of active material and lithium ion loss) and transient operation disturbance characteristics (characterizing reversible polarization caused by fluctuations in current, voltage and temperature). This enables the time-series neural network model to clearly distinguish between long-term degradation and short-term disturbances, overcoming the shortcomings of existing black-box models that mix various health factors and rely on large-scale parameters to forcibly fit the degradation trajectory. Thus, overfitting and prediction divergence are significantly suppressed in cross-cell testing and extremely small sample scenarios.

[0019] (3) A lightweight GRU-MLP single-stream network structure is adopted, which integrates two types of features in the hidden state space. While ensuring high accuracy, the number of model parameters and computational complexity are greatly reduced, making it suitable for deployment in resource-constrained battery management systems. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0021] Figure 1 This is a flowchart of the method according to Embodiment 1 of the present invention; Figure 2 This is a diagram illustrating the comprehensive degradation evolution of the NASA battery dataset from Embodiment 1 of the present invention, wherein... Figure 2 (a) in the figure is a graph showing the evolution of charging voltage over time at different number of cycles; Figure 2 (b) in the figure is a graph showing the evolution of charging current over time for different number of cycles; Figure 2 (c) in the figure is a graph showing the evolution of charging temperature over time at different number of cycles; Figure 2 (d) in the figure shows the trend distribution of battery discharge capacity decay with the number of cycles; Figure 3 This is a normalized evolution trajectory diagram of different ranking features in the battery according to Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the state-aware GRU-MLP prediction model according to Embodiment 1 of the present invention. Figure 5 This is a comparison diagram of the evolution trajectory of intrinsic degradation descriptors for the B0007 dataset in Embodiment 1 of the present invention, wherein... Figure 5 (a) represents the voltage minimum under traditional cycle-based reconstruction. Evolutionary trajectory diagram; Figure 5 (b) represents the 10th percentile of the voltage under traditional cycle-based reconstruction. Evolutionary trajectory diagram; Figure 5 (c) represents the 25th percentile of the voltage under traditional cycle-based reconstruction. Evolutionary trajectory diagram; Figure 5 In the figure, (d) represents the voltage minimum under SOC domain reconstruction. Evolutionary trajectory diagram; Figure 5 (e) represents the 10th percentile of the voltage under SOC domain reconstruction. Evolutionary trajectory diagram; Figure 5 (f) represents the 25th percentile of the voltage under SOC domain reconstruction. Evolutionary trajectory diagram; Figure 6 This is a diagram illustrating the projection transformation process of the charging voltage curves from the original sampling domain to the SOC reference domain under different aging stages in Embodiment 1 of the present invention. Figure 6 (a) in the figure shows the characteristic phase shift phenomenon under the traditional original sampling domain; Figure 6 (b) in the diagram is a schematic diagram of constructing a unified physical reference frame to achieve multivariable alignment; Figure 7 This is a comprehensive comparison chart of the battery SOH estimation accuracy and error distribution under the traditional cyclic domain and the proposed SOC reconstruction in Embodiment 1 of the present invention. Figure 7 (a) in the diagram shows the global deviation between the predicted and actual SOH values ​​of the B0005 battery under the traditional cyclic domain. Figure 7 (b) in the diagram shows the global deviation between the predicted and actual SOH values ​​of the B0006 battery under the traditional cyclic domain. Figure 7 (c) in the diagram shows the global deviation between the predicted and actual SOH values ​​of the B0007 battery under the traditional cyclic domain. Figure 7 (d) in the diagram represents the global deviation between the predicted and actual SOH values ​​of the B0018 battery under the traditional cyclic domain. Figure 7 (e) in the figure is a global deviation diagram of the predicted and actual SOH values ​​of the B0005 battery under the proposed SOC reconstruction. Figure 7 (f) in the figure is a global deviation diagram of the predicted and actual SOH values ​​of the B0006 battery under the proposed SOC reconstruction. Figure 7 (g) in the figure is a global deviation diagram of the predicted and actual SOH values ​​of the B0007 battery under the proposed SOC reconstruction. Figure 7 (h) in the figure represents the global deviation pattern between the predicted and actual SOH values ​​of the B0018 battery under the proposed SOC reconstruction. Figure 7 In the diagram, (i) represents the distribution characteristics of the absolute error (AE) and absolute percentage error (APE) of battery B0005. Figure 7 In the figure, (j) is the distribution characteristic diagram of absolute error (AE) and absolute percentage error (APE) of B0006 battery; Figure 7 In the diagram, (k) represents the distribution characteristics of the absolute error (AE) and absolute percentage error (APE) of the B0007 battery. Figure 7 (l) in the figure is a distribution characteristic diagram of the absolute error (AE) and absolute percentage error (APE) of the B0018 battery; Figure 8 This is an ablation analysis and multidimensional performance evaluation diagram of the decoupling of the physical mechanism of the framework proposed in Embodiment 1 of the present invention, wherein, Figure 8 (a) in the figure is the scatter regression result of the SOC Trend-Only model on the B0005 cell; Figure 8(b) in the figure is the scatter regression result of the SOC Trend-Only model on the B0006 cell; Figure 8 (c) in the figure is the scatter regression result of the SOC Trend-Only model on the B0007 cell; Figure 8 (d) in the figure is the scatter regression result of the SOC Trend-Only model on the B0018 cell; Figure 8 (e) in the figure is the Taylor plot of the B0005 battery cell under three different models; Figure 8 (f) in the figure is the Taylor plot of the B0006 battery cell under three different models; Figure 8 (g) in the figure represents the Taylor plot of the B0007 battery cell under three different models; Figure 8 (h) in the figure represents the Taylor plot of the B0018 battery cell under three different models; Figure 9 The diagram shows the SOH estimation and corresponding absolute error distribution of the framework proposed in Embodiment 1 of the present invention under different operating conditions. Figure 9 (a) shows the SOH estimation results and absolute error distribution of Cell 1 at a 0.5C rate; Figure 9 (b) shows the SOH estimation results and absolute error distribution of Cell 1 at a 0.25C rate; Figure 9 (c) in the figure shows the SOH estimation results of Cell 2 at a 0.5C rate and its absolute error distribution. Figure 9 (d) in the figure shows the SOH estimation results of Cell 2 at a 0.25C rate and its absolute error distribution. Figure 9 (e) in the figure shows the SOH estimation results of Cell 3 at a 0.5C rate and its absolute error distribution. Figure 9 (f) in the figure shows the SOH estimation results of Cell 3 at a 0.25C rate and its absolute error distribution. Figure 10 This is a schematic diagram of the overall method framework of Embodiment 1 of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. Those skilled in the art should understand that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0023] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] Accurate state of health (SOH) estimation is crucial for the safe operation of electric vehicles and energy storage systems. However, existing data-driven models often extract features using time or cycle count as indices, which can easily lead to time-series phase shifts due to differences in charge-discharge and capacity decay under dynamic operating conditions. Furthermore, conventional feature mixing input methods struggle to distinguish between long-term irreversible degradation and short-term polarization fluctuations, causing models to be prone to overfitting and prediction divergence in scenarios with discontinuous sparse data (extremely small samples).

[0025] To address the aforementioned issues, this invention proposes a battery health state estimation method and system based on SOC domain reconstruction and feature decoupling. By mapping the non-stationary time-domain degradation trajectory to a fixed SOC space, data misalignment caused by operating condition fluctuations is eliminated. The electrical response is decomposed into an "intrinsic degradation trend" representing irreversible aging and a "transient operating disturbance" representing instantaneous polarization. Subsequently, a lightweight state transition model using a single-stream gated cyclic unit-multilayer perceptron (GRU-MLP) is constructed to achieve the fusion of these two types of multi-scale features in the hidden state space. This invention provides a reliable lightweight estimation scheme for battery management systems (BMS) where computing power is limited by data sparsity.

[0026] Example 1 This embodiment provides a battery health state estimation method based on SOC domain reconstruction and feature decoupling. Figure 1 The following is an overall flowchart of the method according to Embodiment 1 of the present invention, which includes the following steps: S101: Acquire raw sensing data during the charge and discharge cycle of a lithium-ion battery. The raw sensing data includes time series of voltage, current, and temperature. S102: Map the raw sensing data to the SOC domain to obtain the sensing data in the SOC domain; S103: Extract intrinsic degradation trend features characterizing irreversible aging of the battery from the sensing data in the SOC domain, and extract transient operating disturbance features characterizing the reversible polarization intensity under the current operating conditions. S104: Intrinsic degradation trend features are fused with transient operational disturbance features to construct a joint state perception vector; S105: Input the joint state-aware vector into the time-series neural network model to obtain the battery health state (SOH) estimate.

[0027] In step S101, in order to comprehensively evaluate the generalization ability and robustness of the proposed method under different battery chemistry systems, degradation mechanisms and complex operating conditions, this embodiment constructs a cross-dimensional multi-source battery aging verification system. This system includes two independent data sources: the NASA public aging dataset used for benchmark verification and ablation experiments, and a self-built lithium iron phosphate (LFP) accelerated aging dataset for extremely small sample scenarios.

[0028] (I) NASA Dataset: This embodiment uses a commercially available 18650 lithium cobalt oxide (LiCoO2) battery with a rated nominal capacity of 2Ah. To verify the model's adaptability under different depths of discharge and differentiated aging trajectories, four cells from the same batch, B0005, B0006, B0007, and B0018, were selected as the research objects. The test followed the standard constant current-constant voltage (CC-CV) charging protocol: first, a constant current of 1.5A was used to charge to the upper cutoff voltage of 4.2V, followed by a constant voltage charging stage until the charging current decayed to 20mA. The discharge stage was carried out with a constant current of 2A. To introduce different electrochemical stress boundaries, the lower cutoff voltages of the four cells were set as follows: 2.7V for B0005, 2.5V for B0006, 2.2V for B0007, and 2.5V for B0018. In order to fit the physical state perception framework of the present invention, this embodiment only extracts the voltage, current and temperature multi-dimensional sensing sequences during the CC-CV charging process as the effective input of the model.

[0029] Figure 2 Figures (a) to (d) illustrate the comprehensive degradation evolution of the NASA battery dataset. With increasing cycle count, the battery exhibits significant degradation characteristics: externally, this manifests as a non-linear decrease in discharge capacity; internally, due to the increasing polarization resistance with aging, the voltage slope of the constant current charging stage increases, the constant voltage stage is initiated earlier, and the inflection point of charging current decay shifts significantly earlier. The temperature rise curve shows that the internal heat generation rate of the battery evolves with the degree of degradation, and the temperature peak shifts along the time axis. Notably, due to ambient temperature fluctuations and dynamic discharge stress interference, the capacity trajectory exhibits a typical capacity regeneration phenomenon. This deep coupling between the long-term degradation trend and short-term operating condition fluctuations in the original signal provides an objective data basis for the physical feature decoupling of this invention.

[0030] (ii) Self-built LFP minimal sample aging dataset: In addition to the NASA public dataset, this embodiment also conducted a cycle aging experiment on a commercial LFP battery with a nominal capacity of 15Ah. Because LFP batteries have an extremely flat charge and discharge voltage plateau, the mapping distinction of their internal degradation state on the external sensing curve is low, and the feature extraction difficulty is significantly higher than that of the lithium cobalt oxide system, which is more effective in testing the generalization performance of the prediction architecture.

[0031] The experiment adopted a cyclic protocol of alternating "Reference Performance Test (RPT) + Accelerated Aging" and ran for a total of 1550 cycles. The steps of a single cycle are as follows: (1) Reference characteristic test: In the voltage range of 2.5V-3.65V, constant current charge and discharge at 0.5C and 0.25C rates were performed in sequence to calibrate the current capacity and obtain a high-precision response curve; (2) Accelerated aging: constant current discharge at 2C rate to 2.5V, then constant current charge at 1C rate to 3.6V, and after running 50 aging cycles, the experiment returned to the RPT step.

[0032] To address the limitation of continuously acquiring high-fidelity data throughout the entire lifecycle in practical engineering, this embodiment constructs a very small sample test set. During the aging process, only the charging response trajectory of the RPT stage is extracted as input at fixed cycle intervals. Due to the differences in degradation rates between individual cells, the actual number of cycles corresponding to the end of the lifespan of each battery is different, resulting in the final effective sample size being distributed between 18 and 31 cycles. This large-interval sampling strategy eliminates the strong temporal correlation between adjacent cycles, simulating the data sparsity in real-world scenarios and objectively verifying the predictive stability of the model under discontinuous and discrete operating conditions.

[0033] In step S102, to address the timing misalignment problem caused by inconsistent charging and discharging durations and start / end points under dynamic operating conditions, this invention first performs anomaly cleaning and capacity monotonicity correction on the original sensor data (voltage, current, temperature). Subsequently, using ampere-hour integration and linear interpolation algorithms, the non-stationary time-domain trajectory is mapped to a continuous and equally spaced SOC domain. This operation physically eliminates the reconstruction error caused by dynamic operating conditions, ensuring the consistency of the benchmark for subsequent feature extraction.

[0034] Specifically, to eliminate the interference of timescale bias on degradation characterization, the original time-series data are first reconstructed to the SOC domain to establish a unified electrochemical benchmark. Traditional degradation factors are typically defined at the original sampling time. t Intra-domain extraction, mathematically represented as To unify the physical reference frame, the time series is mapped to a cumulative fill capacity series using the ampere-hour integration method. For any ampere-hour... In each cycle, the accumulated charge capacity is... The discretization calculation formula is as follows: (1); in, This is the charging current. The sampling time interval is used to eliminate non-physical fluctuations caused by measurement noise. Perform monotonicity correction and zero-point translation alignment: (2); in, This represents the effective filling capacity after zero-point shift and monotonicity correction. This represents the initial integral offset at t=.

[0035] Subsequently, based on the available capacity of the battery during its initial lifespan. As a globally unified metric, the transient nominal state of charge during the charging process is calculated. : (3); Specifically, the original time-domain observation data set The set of electrochemical states converted to the SOC domain Furthermore, to ensure the physical consistency of data sampling at different aging stages of the battery, a value range of [value range missing] is constructed. Equal-spacing fixed reference grid A linear interpolation algorithm is employed to uniformly map the electrochemical state set onto the fixed reference grid, thereby overcoming the time-scale scaling interference caused by battery capacity decay. After completing the SOC domain reconstruction and alignment with the reference grid, the battery's external characteristic curves are mapped to a unified physical reference system.

[0036] In S103, the present invention abandons the traditional global feature aliasing input method and decomposes the input information into two independent subspaces: long-term intrinsic degradation trend and transient operation disturbance through the physical perception decoupling module.

[0037] (a) Extraction of intrinsic degradation trend features: To quantify the kinetic evolution during battery aging, 20 candidate intrinsic degradation features were extracted from the sensing data within the SOC domain obtained in step S102. These features capture the nonlinear degradation behavior caused by mechanisms such as ohmic resistance increase, intensified electrochemical polarization, and loss of active material from multiple dimensions, specifically covering the following five categories: (1) Voltage statistical degradation characteristics: Extract the average voltage within a specified SOC range. Voltage standard deviation Voltage minimum value and 10%, 25%, 50% quantiles , , .

[0038] (2) Current statistical degradation characteristics: Extract the mean value of the current within the corresponding interval. Standard deviation and 10%, 25%, 50% quantiles , , .

[0039] (3) Polarization derivative characteristics: Extracting the voltage-time differential mean Standard deviation and 10%, 25%, 50% quantiles , , It is used to quantify the decay of charging acceptance capability.

[0040] (4) Local capacity increment degradation characteristics: Calculate the charge capacity within a specific electrochemical phase transition range. For example, the capacity increment in the SOC range of 30% to 40%. and a capacity increase of 40% to 50%. wait.

[0041] (5) Capacitance Differential IC Degradation Characteristics: Extraction The height of the main peak of the curve and the corresponding state of charge position .

[0042] The aforementioned 20 candidate features exhibit high information redundancy and multicollinearity. Inputting all of them into the prediction network would increase the model's parameter optimization burden and easily lead to overfitting in scenarios with extremely small sample sizes. Therefore, this embodiment constructs a multi-dimensional quantitative evaluation system based on correlation and evolutionary monotonicity to comprehensively score and filter the feature pool. Pearson correlation coefficients are used respectively. Spearman rank correlation coefficient Quantify candidate feature sequences from both linear and nonlinear dimensions. The degree of correlation between the battery and the actual state of health (SOH).

[0043] Considering that irreversible battery capacity degradation should exhibit a monotonic evolution characteristic over the entire life cycle, a monotonicity score is introduced. The purity of the evolution direction of a quantitative feature sequence is calculated using the following formula: (4); in, n The total number of cycles, For symbolic functions, , These represent the candidate feature sequences at the th... Second and third The specific extracted value refers to the physical quantity measure used to characterize aging extracted from the battery's single charge-discharge cycle data, for example: the first The constant current charging time of each cycle, the voltage variance or capacity increment within a specific SOC range (such as the peak height of the IC curve).

[0044] To eliminate the dimensional differences among the evaluation indicators, the absolute values ​​of the Pearson and Spearman correlation coefficients, as well as the monotonicity scores, were normalized and mapped to the interval [0, 1]. Equal weights were then assigned to calculate the comprehensive score. : (5); in, This represents the final comprehensive evaluation score for the candidate feature; , and These represent the normalized values ​​of the candidate feature under three different evaluation dimensions.

[0045] In this embodiment, to eliminate the differences in units and value ranges among different evaluation indicators, before calculating the comprehensive evaluation score, each original evaluation indicator needs to be normalized to a minimum and maximum value, mapping them uniformly to a numerical range of 0 to 1. Taking the calculation of as an example, the specific calculation formula is as follows: ; In the formula, This represents the absolute value of the original evaluation index of the current candidate feature; and These are the maximum and minimum absolute values ​​of the indicator among all candidate features, respectively. Similarly, and The specific calculation method and The same applies. Finally, by calculating the arithmetic mean of the above three normalized indicators, the comprehensive evaluation score for each candidate feature can be obtained. .based on The size of all candidate features is used to sort them in descending order, and the top-ranked features are selected as the intrinsic degradation trend features that are finally input into the temporal neural network model.

[0046] Based on this comprehensive score, all candidate features are sorted in descending order, such as... Figure 3 As shown, Figure 3 The normalized evolution trajectories of batteries with different ranking characteristics were compared. The top three intrinsic degradation factors were identified. , , The degradation trend is smooth and monotonous with increasing cycle count, indicating that it can effectively track the aging process of the battery; while the three characteristics that ranked last... , , The large fluctuations and weak consistent evolution patterns throughout its life cycle indicate that it is susceptible to local measurement noise and cannot characterize long-term degradation mechanisms.

[0047] Furthermore, this invention also analyzed the quantitative evaluation results of 20 candidate features across four dimensions: absolute value of Pearson correlation coefficient, absolute value of Spearman correlation coefficient, monotonicity, and final comprehensive score. The analysis shows that voltage-derived features... and It performed best, with Pearson and Spearman correlation coefficients both exceeding 0.98, and a final score greater than 0.95; some incremental capacity features and differential voltage The relevance scores of the features were generally low. In addition, among all the evaluation indicators, the monotonicity score had a low overall distribution (the highest value was only 0.52), indicating that this indicator has a more stringent screening effect and can effectively eliminate inferior features with high global relevance but large local fluctuations.

[0048] Based on the above multidimensional evaluation criteria, the top-ranked features were selected as input variables for the subsequent SOH estimation model.

[0049] (II) Extraction and Characterization of Reversible Transient Perturbations: In actual operating conditions, the dynamic response of a battery is the result of the coupling effect of long-term irreversible aging (such as loss of active material) and short-term reversible perturbations (such as fluctuations in ambient temperature and charge / discharge rate). These transient perturbations can induce varying degrees of polarization overpotential within a single cycle by altering electrolyte conductivity and solid-phase diffusion kinetics, thereby causing an aging-independent shift in the extracted intrinsic degradation factor. To achieve accurate health state estimation, such reversible perturbations must be decoupled from the overall degradation trend.

[0050] This invention considers temperature T, current I, and voltage V as the main components constituting the operating conditions. To quantify the intensity of transient disturbances within a specific electrochemical phase transition range, a target characteristic range is first defined in the SOC domain. Assuming in the first During the next charge-discharge cycle, a certain macroscopic state variable The continuous trajectory function that evolves with SOC is: Then the operating state of this variable can be represented by integral normalization: (6); in, This is the current charge / discharge cycle sequence number. To represent the continuous integral variable of the state of charge, and These are the effective lower and upper bounds of the target SOC interval, respectively. Indicates the first State variables in the next loop A continuous function that evolves with SOC.

[0051] Considering that the data acquired by actual sensors is a discrete time series, to ensure the robustness and computability of the algorithm, the above continuous integral is mapped to a discrete sampling space. Define the... The SOC sampling sequence for the next cycle is Introducing indicator functions to construct an effective index set that satisfies the target interval constraints. The mean of the sensed variables is calculated over this finite set, and this mean is used as the transient perturbation component characterizing the current cycle: (7); (8); in, For the set of natural numbers, For discrete sampling point index, , For the first The total number of sampling points in the second cycle, for the first... The loop iterates through. One sampling point, This represents the discrete SOC values ​​recorded by the sensor. This represents the measured value of the corresponding state variable (such as voltage, current, or temperature); This indicates the index set The potential is the total number of valid sampling points falling within the target interval.

[0052] Specifically, the above formula (8) is applied to the three state variables of voltage, current and temperature respectively, that is, let the formula... Take the first one in sequence In the second loop Voltage measurements at each sampling point Current measurement value and temperature measurement value Calculate their conditional mean values ​​within the target SOC range to obtain three transient operating disturbance components: range average voltage. Interval average current and the average temperature of the interval .

[0053] The three components extracted through this mathematical framework each have a clear physical meaning: among them, It characterizes the average thermodynamic potential and the degree of global polarization accumulation within this range; The dynamic stress intensity that induces polarization was quantified; These variables characterize the thermodynamic conditions affecting the reaction rate. Together, they constitute a feature set containing information about actual operating conditions. Introducing them into the model as supplementary inputs can adaptively offset the feature shift caused by reversible polarization.

[0054] In S104, traditional battery SOH data-driven prediction frameworks typically rely solely on health factor sequences characterizing long-term degradation as model input. However, as mentioned earlier, the external observation response during actual battery operation is a result of the deep coupling between "intrinsic irreversible aging" and "short-term reversible transient fluctuations." If the deep learning model lacks the ability to perceive the current operating conditions, it is prone to misjudging short-term polarization fluctuations as permanent degradation of the battery itself, leading to oscillations in the predicted trajectory and reduced accuracy. To address this limitation, this invention proposes a state-aware joint feature fusion mechanism that reconstructs the decoupled "intrinsic degradation trend" and "transient operating disturbances" at the model input layer.

[0055] Specifically, regarding the first In the next charging cycle, the first three intrinsic degradation characteristics selected by the multi-dimensional evaluation system in step S102 constitute the benchmark vector for battery degradation. , R 3 Representing a three-dimensional real vector space, this reference vector consists of three intrinsically degenerate eigenvalues; the extracted transient disturbance mean values ​​(mean voltage, mean current, and mean temperature) constitute the transient disturbance vector characterizing the current reversible polarization intensity. Before feeding features into the deep neural network, all variables are first normalized to prevent gradient dominance during network training. Then, the two are fused along the feature dimension to construct a joint state-aware vector for the current loop. : (9); in, Indicates the first The joint state-aware vector constructed in the next iteration; and They represent the first time. The baseline vector characterizing the battery's bulk degradation and the perturbation vector characterizing the transient polarization intensity under the next cycle; This represents the vector concatenation operator; This represents the transpose operation of a vector; , and respectively represent the aforementioned through multiple The first, second, and third intrinsic degradation features selected by the dimensional evaluation system are in the 1st... The specific calculation values ​​are obtained in the next loop; , and They represent the first time. The interval average voltage, interval average current and interval average temperature calculated under each cycle are the three transient operating disturbance components mentioned above. This represents the six-dimensional real space, i.e., the fused joint state-aware vector. It is a column vector containing 6 elements.

[0056] In this reconstructed joint feature space, To transmit long-term, irreversible aging evolution patterns to neural networks, and This then acts as an adaptive "calibration signal." When actual operating conditions change abruptly, It can serve as a priori boundary conditions, guiding the model to identify and counteract physical disturbances caused by external environmental variables, thus avoiding overestimation or underestimation of the true state of energy (SOH). Finally, along the time evolution axis of battery life, continuous eigenvectors are assembled into a high-dimensional time series matrix. This serves as the input for subsequent time-series deep learning models.

[0057] In S105, the decoupled joint state-aware vector is fed into a lightweight GRU-MLP combined network. The single-layer GRU unit utilizes its gating mechanism to filter transient random noise and extract deep temporal degradation dynamics; the MLP head acts as a nonlinear mapping terminal, mapping the high-dimensional hidden state to the SOH estimate. This lightweight architecture significantly reduces the parameter dimensionality, enabling computational efficiency to be adapted for online applications in automotive BMS, and also providing theoretical support for robust prediction in extremely small sample scenarios.

[0058] Specifically, in constructing the joint state-aware vector Subsequently, an efficient time series model is needed to uncover the dynamic correlation between long-term degradation dependencies and short-term perturbations in the data. Although LSTM performs excellently in time series tasks, its complex gating mechanism leads to high computational and memory overhead, making it difficult to meet the lightweight deployment requirements of BMS. Therefore, this embodiment adopts a lightweight architecture of GRU cascaded MLP to implement SOH prediction. In this architecture, the GRU plays the core role of the degradation state extractor, receiving the joint sensing input at the current time through internal update and reset gates. In conjunction with historical memories from the previous moment, the hidden degradation state of the battery is recursively updated. This process effectively captures the intrinsic degradation dynamics inherent in the time series while smoothing out noise introduced by transient perturbations. Its core formula is as follows: Update Gate: (10); Reset Door: (11); Candidate hidden states and final state output: (12); (13); in, for The joint sensing input feature vector of the time-mapping model and They are respectively Time and The hidden state vector at time step is used to characterize the battery degradation memory that evolves over time; and These represent the state vectors of the update gate and the reset gate, respectively. for The candidate hidden state vector at time step 1. , , To learn the weight matrix, , , For the bias term vector, and These represent the Sigmoid and hyperbolic tangent nonlinear activation functions, respectively. This represents element-wise multiplication of a matrix.

[0059] The hidden states of the temporally degenerate dynamics are recursively computed by GRU. Subsequently, the MLP, acting as a nonlinear mapping terminator, transforms the high-dimensional abstract degenerate state. Mapped to the physical space, the output is the health status predictor variable SOH. The network mapping of an MLP can be represented as: (14); in, Indicates the first Estimated SOH value of the lithium battery under one cycle; and These represent the weight matrix and bias vector of the MLP hidden layer, respectively. and These represent the weight matrix and bias vector of the MLP output layer, respectively. This represents the linear rectified activation function, used to perform nonlinear mapping on hidden layer features to capture the nonlinear dynamic characteristics during battery aging. This indicates matrix multiplication or dot product operations.

[0060] like Figure 4 As shown, Figure 4This is a schematic diagram of the state-aware GRU-MLP prediction model proposed in this embodiment. Thanks to the physical mechanism-guided feature decoupling and state-aware input reconstruction proposed in this invention, the burden on deep learning models to mine degradation patterns from raw data is significantly reduced. The proposed prediction framework adopts a very simple topology, requiring only a single-layer GRU and a two-layer MLP to achieve high-precision mapping of battery SOH. This lightweight architecture reduces the number of model parameters, alleviates the overfitting risk caused by limited battery capacity data, and simultaneously reduces BMS computing power and memory overhead, providing strong support for the online health estimation needs of electric vehicles.

[0061] To further demonstrate the effectiveness of the proposed battery health state estimation method based on SOC domain reconstruction and feature decoupling, this embodiment experimentally verifies its effectiveness in hybrid modeling of lithium-ion batteries within the SOC reconstruction coordinate system. Unlike traditional empirical modeling paradigms based on cycle counting, this invention constructs a state transition evolution operator capable of accurately characterizing battery degradation trajectories across time scales through intrinsic feature reconstruction of the SOC domain. The core logic of the experiment lies in demonstrating that this physically aligned dynamic modeling approach exhibits significantly greater robustness compared to traditional time-domain models when capturing key features such as nonlinear capacity decay, capacity regeneration, and cross-sample generalization.

[0062] Therefore, this embodiment selects four typical aging sequences (B0005, B0006, B0007, and B0018) from the NASA public dataset as validation benchmarks. These four sets of samples cover multiple trajectories from linear steady-state decay to later accelerated degradation, providing sufficient sample diversity for evaluating the robustness of the state transition model under different battery evolution paths. By comparing with the traditional cyclic domain alignment paradigm, this embodiment deeply analyzes the underlying physical logic of SOC reconstruction in eliminating temporal scale fluctuations and unifying electrochemical degradation benchmarks. The evaluation system includes mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2 This study examines statistical indicators such as [missing information], and focuses on the synergistic relationship between the long-term temporal dynamics extracted by the lightweight GRU module and the observation mapping mechanism of the MLP layer, clarifying how the model effectively decouples "intrinsic evolutionary trends" and "transient operational perturbations" at the physical level. This architecture improves the physical interpretability and long-term evolutionary stability of the prediction results while maintaining extremely low parameter complexity.

[0063] (I) Evaluation Indicators and Experimental Setup: To quantify the estimation accuracy of the state transition model, this embodiment uses MAE, RMSE, MAPE, and R... 2 As an evaluation metric, MAE reflects the average level of error between predicted and actual values ​​and exhibits good robustness.

[0064] (15); in, and These represent the estimated SOH and the actual SOH of the battery, respectively. This represents the average SOH of all actual batteries. This represents the total number of samples.

[0065] RMSE amplifies the weights of larger deviations through squared terms, enabling it to sensitively capture prediction fluctuations at the end of battery aging or at capacity regeneration points. (16); MAPE is used to evaluate the ratio of error to actual capacity, making it easy to measure the relative prediction accuracy between cells of different specifications.

[0066] (17); The explanatory power of the characterization model for the evolution of degenerate state trajectories. The closer the value is to 1, the more accurately the state transition operator can reproduce the physical trend of electrochemical degradation. (18); To verify the model's performance, this embodiment selected four physically representative battery individuals for experiments. Among them, B0006 exhibited significant internal resistance growth and accelerated capacity degradation, while B0018's charge / discharge protocol showed a clear distributional shift in time scale compared to the other individuals. This data-level heterogeneity provides ideal samples for testing the model's generalization ability under complex dynamic conditions. This invention integrates degradation sequences with different evolution rates, aiming to enable the GRU layer to learn a universally applicable intrinsic state transition operator from multi-dimensional degradation paths. This allows for the accurate capture of capacity evolution and degradation rates, validating the proposed SOC domain reconstruction and feature decoupling battery health state estimation method (hereinafter referred to as SOC domain reconstruction and feature decoupling method). 2 The effectiveness of -HD).

[0067] To demonstrate that the improved estimation accuracy does not stem from the overfitting ability of deep neural networks, but rather from the SOC space reconstruction and state transition logic proposed in this invention, this embodiment specifically employs a lightweight computational architecture. Specifically, this invention selects GRU-MLP as the algorithm carrier, constructing physically consistent input features and focusing on examining the accuracy of the state transition operator in describing intrinsic degradation trajectories. This method abstracts the complex electrochemical evolution process into a series of linearly consistent state transition sequences. Experimental results show that, with a very small parameter scale and mapping space, the model remains highly sensitive to nonlinear phenomena such as accelerated degradation and stochastic capacity regeneration. This design not only enhances the interpretability of the degradation kinetics model but also significantly reduces computational complexity and memory overhead at the engineering application level, enabling real-time deployment of embedded battery management systems. The core parameter configuration is shown in Table 1.

[0068] Table 1. Core parameter configuration.

[0069]

[0070] The formula for the Huber Loss function is as follows: (19); in, For loss function, This embodiment sets =1, so that it maintains the sensitivity of mean square error when the error is small, and maintains the robustness of mean absolute error when the error is large.

[0071] (II) Analysis of the evolution of the degenerate state: Unlike traditional static feature mapping, this invention constructs SOH estimation as a state transition dynamic model. Based on the degraded conditions reconstructed from SOC, the GRU captures the current state. State at the next moment By mapping nonlinear dynamics between the two, the model learns the "transition operator" of battery degradation, enabling it to automatically update battery health information and predict the next state of the degradation trajectory. This ability to predict in advance allows the model to keenly capture changes in the degradation rate.

[0072] In the early feature evaluation stage, based on a comprehensive quantitative score of correlation and monotonicity, the voltage minimum value under a specific SOC benchmark in dataset B0007 was determined. ), the first 10% quantile voltage ( ) and lower quartile voltage ( These parameters represent the optimal degradation state. As trends in quantifying irreversible capacity decay of the battery, they form the basis for the evolution of the state-space model.

[0073] like Figure 5 As shown, in order to intuitively verify the supporting role of SOC reconstruction in degradation dynamics modeling, Figure 5 A comparative plot showing the evolution trajectory of intrinsic degradation descriptors for the B0007 dataset is presented. Figure 5 In the figure, (a), (b), and (c) represent the voltage minimum values ​​under traditional cycle-based reconstruction, respectively. 10th percentile of voltage and voltage 25th percentile The evolution trajectory of the signal shows a "non-physical drift" characteristic in the blue trend line during the first 50 cycles. Particularly between cycles 25 and 50, the eigenvalues ​​exhibit non-monotonic jumps. From an electrochemical mechanism perspective, this phenomenon does not originate from a sudden change in the battery's internal state, but rather from the continuous decay of usable capacity, leading to a severe phase shift in the voltage signal extracted within a fixed time window.

[0074] In comparison, Figure 5 (d), (e), and (f) in the figure illustrate the voltage minimum under SOC domain reconstruction proposed in this invention. 10th percentile of voltage and voltage 25th percentile The evolution trajectory. After physical space reconstruction, the red trend line shows a monotonically increasing trend throughout the entire life cycle, which is consistent with the physical law of battery internal resistance climbing. Even in the face of relatively large capacity decay in the early stage of B0007, by setting the characteristics at a unified SOC benchmark, the time scale scaling interference caused by the reduction of available capacity can be mitigated.

[0075] Furthermore, the gray scatter plots in the figure record the transient operating condition components characterized by the average voltage, average current, and average temperature within the target range. Comparative results show that, under a unified electrochemical reference frame, this scheme can effectively decouple transient fluctuations caused by complex operating conditions from the irreversible intrinsic degradation trend. This improved physical consistency not only enhances the model's interpretability but also provides data support for maintaining good generalization ability when facing individual cells with different aging rates.

[0076] To further reveal the proposed S 2 -The inherent mechanism of HD frames in eliminating physical inconsistencies Figure 6The process of projecting the charging voltage curve from the original sampling domain to the SOC reference domain under different aging stages is shown. Here, cycle1, cycle56, etc. represent the number of battery cycles. One cycle is represented from the battery being fully charged to being depleted. The battery's first cycle Cycle1 (early life), the battery's 56th cycle Cycle 56, the battery's 112th cycle Cycle 112 (mid-aging stage), and the battery's 167th cycle Cycle 167 (end of life) are used as typical examples. Figure 6 (a) illustrates the characteristic phase shift phenomenon in the traditional original sampling domain. In this domain, the influence of degradation on signal features exhibits obvious coupling characteristics. With the increase of cycle number, the battery's charge acceptance decreases due to the combined effects of active lithium loss (LLI) and increased internal resistance, resulting in a significant shortening of the time to reach the cutoff voltage. The voltage curve shows a significant "length contraction" on the horizontal axis. If a fixed sampling point index is used as the basis for feature extraction, the voltage values ​​at different aging levels will exhibit significant phase shifts.

[0077] like Figure 6 As shown in (b), this embodiment achieves multivariate alignment by constructing a unified physical reference frame. Specifically, the original voltage, current, and temperature sequences are reconstructed into the SOC domain after ampere-hour integration and capacity correction. Within this space, the observed variables are freed from the time-scale scaling interference caused by capacity reduction. Taking the voltage curve as an example, its vertical rise directly corresponds to the increase in battery internal resistance. This reconstruction mechanism ensures that all sensor data are under the same electrochemical reference, providing a prerequisite for accurately extracting the characteristics of each degradation stage.

[0078] Beyond the visually intuitive improvement in physical consistency, the quantitative evaluation results further confirm the superiority of SOC reconstruction. Compared to traditional cyclic domain features, the intrinsic degradation features extracted from the SOC domain all showed improved scores in the comprehensive quantitative evaluation based on Pearson, Spearman, and Monotonicity. These intrinsic trend components, possessing high signal-to-noise ratios and clear physical meanings, reduce the mapping difficulty of subsequent algorithms.

[0079] (III) Evaluation and Discussion of Prediction Accuracy: To objectively evaluate the performance advantages of the proposed S²-HD framework in SOH estimation, this embodiment uses a traditional cycle-based model as a benchmark control. The cycle-based model directly inputs traditional time-domain or cyclic-domain features into the GRU-MLP network without SOC domain reconstruction and feature decoupling. Under strict variable control, using the exact same lightweight GRU-MLP network topology and hyperparameter configuration shown in Table 1, the experiments analyzed four typical aging datasets: B0005, B0006, B0007, and B0018. The model's ability to track global decay trajectories and its robustness against local nonlinear fluctuations were evaluated.

[0080] like Figure 7 As shown, Figure 7 A comprehensive comparison chart of the battery SOH estimation accuracy and error distribution under the traditional cyclic domain and the proposed SOC reconstruction method is presented. Figure 7 (a)~(d) and (e)~(h) in the figure show the global deviation between the predicted value and the actual value; Figure 7 In (i)~(l), the distribution characteristics of point-by-point absolute error (AE) and absolute percentage error (APE) are analyzed.

[0081] observe Figure 7 As shown in (a) to (d), under the traditional paradigm, the predicted scatter points of the four datasets all exhibit varying degrees of detachment and distortion. Local predicted scatter point clusters deviate significantly downwards from the diagonal, forming obvious trajectory discontinuities. Furthermore, in these severely distorted regions, the color of the scatter points degrades from dark blue to cyan, green, and even yellow, indicating a significant increase in the local absolute error (Absolute Error) of the traditional SOH estimation method. As described in the preceding mechanistic analysis, the root cause of this structural distortion in the predicted trajectory lies in the unavoidable characteristic phase shift under the traditional cyclic domain. The pseudo-fluctuations generated with increasing aging contaminate the physical consistency of the input sequence, preventing the GRU state transition operator from establishing a stable dynamical mapping.

[0082] In comparison, Figure 7 Figures (e) to (h) illustrate the prediction performance using the method proposed in this invention. After removing the interference of time-scale scaling, the predicted scatter points of the four battery groups are extremely tightly "attracted" to both sides of the ideal diagonal, achieving good convergence and linearity. Originally in Figure 7In (a) to (d), the large areas of high-error cyan-green scattering points were effectively removed, and the overall color returned to dark blue, representing extremely low error. This proves that the intrinsic degradation features extracted by SOC reconstruction provide a smooth evolution trajectory with clear physical direction for the lightweight state transition model, enabling the model to effectively suppress the influence of local transient perturbations and extract irreversible capacity decay trends.

[0083] Figure 7 Figures (i) to (l) show a back-to-back violin plot comparison of the error indices Absolute Error (AE) and Absolute Percentage Error (APE). The width of the violin plot reflects the probability density of the error within a specific interval, the thick black line segment inside represents the interquartile range (IQR), the white dot represents the median, and the line running through the belly of the plot precisely indicates the arithmetic mean (MAE) and MAPE of the distribution. The superiority of the proposed method can be seen from the morphological evolution of the error distribution. In the blue violin plot representing the traditional method (Cycle-based), the error distribution generally exhibits obvious "fat belly" and "long tail" characteristics.

[0084] like Figure 8 As shown, Figure 8 Ablation analysis and multidimensional performance evaluation diagram for decoupling the physical mechanism of the proposed framework, wherein, Figure 8 In the diagram, (a) to (d) represent the scatter regression results of the SOC Trend-Only model on four typical battery cells. Specifically, the ablation model is implemented as follows: when constructing the state-aware input vector at the network input layer, reversible transient disturbance terms affected by external test conditions and environment (such as local temperature, mean current, etc.) are removed, and only intrinsic degradation trend features (such as voltage minimum, feature quantile, etc.) are retained, and these are input into the GRU-MLP network for health state estimation.

[0085] and Figure 7 Compared to the traditional cycle-based benchmark in (a), the model's prediction scatter points are more concentrated in... Near the diagonal, the trajectory discontinuity phenomenon of the former is effectively alleviated. The results show that SOC phase alignment helps reduce capacity prediction bias; even without the introduction of temperature and operating condition compensation, the intrinsic degradation characteristics alone can provide a stable global decay trend for the GRU network. However, in the local nonlinear accelerated aging range, the prediction results of this ablation model still have slight fluctuation bias.

[0086] The true value benchmark represents the ideal reference standard for the actual measured state of harmonics (SOH) of the battery. In the Taylor plot, this benchmark point is fixed on the horizontal axis, i.e., the correlation coefficient. (Normalized standard deviation is 1.0, and centered root mean square error is 0). The closer the data points of each prediction model are to the benchmark point in the graph, the higher the consistency between the prediction results and the actual situation, and the better the overall performance. Cycle-based models represent existing benchmark models that do not adopt the SOC domain reconstruction mechanism of this invention and directly extract features based on traditional time series or cyclic periods. Due to the interference of dynamic operating conditions and environmental fluctuations, the feature data has phase shifts. The data points of this model are farthest from the true benchmark in the graph, showing a low correlation coefficient and a large standard deviation, making it difficult to cope with complex and ever-changing actual operating conditions. SOC trend-only ablation represents the feature ablation experimental model of this invention. This model implements SOC domain reconstruction to align the physical phase, but actively removes the transient perturbation term representing short-term polarization at the input layer, only inputting the irreversible intrinsic degradation trend feature into the network. As shown in the figure, the data points of this model have converged significantly to the true value benchmark compared to the traditional cyclic domain model, proving the effectiveness of reconstructing the intrinsic trend features. However, due to the lack of detailed capture of local nonlinear fluctuations, a certain gap remains between its position and the perfect benchmark point. The method proposed in this invention represents the complete S²-HD architecture model proposed in this invention. This method not only reconstructs the data within a fixed SOC domain but also integrates the "intrinsic degradation trend features" and "reversible transient perturbation terms" in the hidden state space through a lightweight GRU network. In the test results of all four cells, the diamond-shaped marked point best matches the true value benchmark asterisk point, confirming that the feature decoupling and state-aware fusion mechanism proposed in this invention can restore the real battery degradation dynamics to the greatest extent and improve the accuracy and robustness of SOH estimation.

[0087] In order to quantify this evolutionary difference in a multidimensional statistical space, Figure 8 (e)~(h) introduces Taylor diagrams for three different models. These diagrams integrate three metrics—correlation coefficient, normalized standard deviation, and centered root mean square error (CRMSE)—within a two-dimensional coordinate system to comprehensively evaluate the predictive performance of the models.

[0088] Taylor diagrams cleverly utilize geometric relationships to integrate three high-order statistical dimensions—correlation coefficient, normalized standard deviation, and centered root mean square error—within a two-dimensional polar coordinate system. In the Taylor subplots of each cell, the trajectory of predicted performance transitions from blue circles to orange squares and then to red diamonds. Figure 8As shown in (e)~(h), traditional cycle-based coordinate points (blue circles) are generally detached from the periphery, and their standard deviation is often greater than 1.15, indicating a large error and significant oscillations in the prediction results. After stripping away the time domain and applying SOC reconstruction, the ablation model (orange square) takes a decisive step towards the true observation benchmark (black pentagram) in the multidimensional statistical space. The significant improvement indicates that the model has successfully identified the mainstream form of irreversible decline.

[0089] After further introducing the transient perturbation module, the complete model is almost in line with the observation reference point, and its coordinates are close to the reference arc with a normalized standard deviation of 1.0. This shows that the S²-HD framework not only approximates the real SOH, but also replicates the dynamic fluctuation amplitude of the real degenerate envelope.

[0090] The above results demonstrate that the "intrinsic degradation trend + transient operation disturbance" architecture proposed in this invention is an effective method for filtering out operating noise and achieving high-fidelity battery health state estimation.

[0091] Table 2. Statistical evaluation indicators for different methods.

[0092]

[0093] Table 2 compares the statistical evaluation indicators under different methods. As shown in Table 2, under the traditional cycle-based architecture, due to the phase shift in feature evolution, the absolute error of the model is generally high and fluctuates significantly, with MAE ranging from 0.0133 to 0.0277. After introducing the ablation experiment SOC Trend-Only model, the absolute errors of all four battery groups achieved a significant decrease for the first time. For example, the RMSE of B0006 dropped sharply from 0.0354 to 0.0215. This phenomenon indicates that simple SOC domain reconstruction can "filter" the time-domain pseudo-fluctuation noise caused by capacity decay. Further, after adding a complete architecture with transient polarization compensation, the error achieved another significant reduction. Taking B0007 as an example, its MAE was finally reduced to 0.0047, a reduction of 64.7% compared to the baseline model, and the RMSE dropped to 0.0071. On B0006, which exhibits a relatively rapid degradation trend in the early stages and is accompanied by significant nonlinear capacity fluctuations, the proposed method also demonstrates good robustness against interference, further reducing its RMSE to 0.0135, a total reduction of 61.9%. The results show that the lightweight GRU network combined with the multidimensional physical decoupling mechanism can focus the weight parameters on the irreversible intrinsic degradation trend.

[0094] MAPE is a key evaluation metric. Towards the end of battery life, the actual usable capacity decreases, and even small absolute deviations can cause a sharp increase in APE, leading to serious inaccuracies in remaining life assessment. As shown in Table 2, in the traditional cycle-based method, the MAPE of B0006 reaches 3.9095%, posing a high application risk. The SOC Trend-Only model, which removes transient features, reduces it to 2.0919%, preliminarily validating the effectiveness of the reconstructed features. The method of this invention stably controls the MAPE of all four battery groups below 1.8%. Specifically, the MAPE of B0005 and B0007 reaches extremely low levels of 0.6021% and 0.5343%, respectively. This demonstrates that the state transition dynamics model, which includes a complete compensation mechanism, maintains a highly consistent prediction confidence level at all stages of battery aging.

[0095] Limited by the limitations of static empirical mapping, the traditional method of dynamic decoupling fails due to the shift in the B0018 charge / discharge protocol distribution. The goodness-of-fit score was only 0.8814. The ablation model SOC Trend-Only slightly improved it to 0.8971, indicating that trend reconstruction alone is insufficient to completely overcome the heterogeneity of operating conditions. However, the complete S²-HD framework of this invention, with its physical interpretability including operating condition compensation, effectively overcomes the distribution shift problem under heterogeneous operating conditions, significantly improving the goodness-of-fit score of B0018 to 0.9399. Furthermore, on the B0005 and B0007 datasets, the complete model... The values ​​reached 0.9938 and 0.9930 respectively, which effectively reconstructed the complex electrochemical degradation trajectory inside the lithium-ion battery.

[0096] It should be noted that the cycle-based baseline model set in this experiment already includes complete transient features. However, the results in Table 2 show that even under multi-dimensional feature-driven conditions, the error of the traditional time-domain model is still significantly higher than that of the intrinsic degradation trend ablation model that relies on only a single feature. For example, the RMSE of B0006 decreased from 0.0354 to 0.0215. This result further confirms that if the phase misalignment problem of feature evolution is not addressed in advance through SOC domain reconstruction, simply increasing the dimensionality of the input variables cannot fundamentally correct the deviation of the dynamic mapping. The unification of the physical domain is a prerequisite for the effective representation of features.

[0097] Based on this, the state transition S²-HD architecture proposed in this invention, which combines "intrinsic trend reconstruction + transient disturbance compensation," effectively solves the prediction bias caused by differences in different operating conditions and early nonlinear fluctuations by physically decoupling the battery aging mechanism. The results of ablation experiments also demonstrate that this feature reconstruction combined with physical knowledge is key to improving the model's global prediction accuracy and robustness.

[0098] (iv) SOH estimation and analysis in small sample scenarios: To verify the generalization ability of the proposed algorithm under limited training data, this embodiment conducts small-sample prediction experiments based on a self-built LFP battery dataset. As mentioned earlier, due to the use of a large-interval sparse sampling strategy, the effective observation data of a single battery cell during its complete lifespan is only 18 to 31 times. In such sparse and discontinuous data scenarios, traditional deep learning models are prone to severe overfitting due to insufficient sample size; at the same time, due to the lack of continuous historical time-series information, existing models have difficulty accurately predicting the nonlinear "capacity drop" behavior in the later stages of battery life. To address this challenge, this embodiment analyzes the actual effect of the S²-HD architecture in resisting overfitting and restoring the true aging pattern under extremely small sample conditions through comparative experiments.

[0099] like Figure 9 As shown, Figure 9 This paper presents the SOH estimation results and corresponding absolute error distribution of three independently tested lithium iron phosphate (LFP) battery cells (denoted as Cell 1, Cell 2, and Cell 3) under the framework proposed in this invention at 0.5C (corresponding to a charging rate of 0.5 times the rated current) and 0.25C (corresponding to a charging rate of 0.25 times the rated current). Wherein: Figure 9 (a), (c), and (e) in the figure show the estimation results of Cell 1, Cell 2, and Cell 3 at a magnification of 0.5C and their absolute error distribution, respectively. Figure 9 Figures (b), (d), and (f) respectively show the estimation results and absolute error distribution of Cell 1, Cell 2, and Cell 3 at a magnification of 0.25C. The black solid dotted line in the figure represents the actual SOH degradation trajectory, and the light orange bar chart at the bottom visually reflects the absolute prediction error of a single lap. From Figure 9 As can be seen, the three independently tested battery cells (Cell 1, Cell 2, and Cell 3) all exhibited nonlinear accelerated aging characteristics within a relatively short cycle life. This is especially true for Cell 2. Figure 9 (c) Figure 9 In (d), a capacity drop occurred around cycle 10. Faced with nonlinear aging abrupt changes, the method proposed in this invention can effectively overcome the influence of the aforementioned transient disturbances and track the true nonlinear degradation trajectory of each test battery.

[0100] For Cell 3 (corresponding to) Figure 9 In (e) and (f) of the model, this model further verifies its generalization robustness under small sample conditions, and further verifies its generalization robustness under extremely small sample conditions. Due to the influence of individual manufacturing differences, Figure 9The early plateau period and accelerated decay rate in (e) and (f) are both related to Figure 9 There are significant differences between (a) and (d) in the model. Even in the absence of any prior historical data for this battery, the model has been able to track the life cycle evolution well under 0.5C and 0.25C conditions, and the predicted curves have a high degree of overlap with the actual trajectory.

[0101] The Abs. Error distribution plot at the bottom of the dual Y-axis architecture shows that the overall prediction error of the model is low. Although the error histogram shows a spike in the SOH range, the model quickly corrects this spike thanks to its high-density electrochemical state-aware features. This error self-suppression mechanism demonstrates the value of introducing intrinsic degradation trends and instantaneous operational disturbance physical features to enhance the model's temporal stability.

[0102] Table 3. Quantitative evaluation of SOH estimation results under different test conditions.

[0103]

[0104] Table 3 details the quantitative evaluation results of SOH estimation for the proposed model under different rates and different single cells. These results were obtained through MAE, RMSE, MAPE, and... The comparison of the four indicators shows that the model exhibits good prediction accuracy and robustness in scenarios with extremely small sample sizes. The error metrics demonstrate that the model maintains excellent accuracy across all test conditions, with maximum errors of 0.0141 and 0.0168 for MAE and RMSE, respectively, and MAPE for all cells remaining below 2%. Particularly noteworthy is Cell 1 under the 0.25C condition, where MAE and RMSE decrease to 0.0057 and 0.0068, respectively, with a MAPE of only 0.69%. This indicates that the model has a strong ability to capture the nonlinear fluctuations in the battery degradation trajectory.

[0105] From the perspective of correlation and goodness of fit, all batteries The values ​​are all above 0.9695, reaching a maximum of 0.9955. It's worth noting that the Cell 3 battery performs well at 0.5C and 0.25C. The values ​​remained at 0.9861 and 0.9695 respectively, with MAPE remaining stable at 1.23% and 1.54%. The high consistency and low error levels exhibited by the three batteries demonstrate that the state-aware model based on physical feature extraction can effectively overcome the temporal drift caused by individual manufacturing and has good generalization robustness in cross-individual SOH tracking.

[0106] In summary, this invention addresses two major challenges faced by data-driven models in battery SOH estimation: operating condition phase shift and generalization failure with extremely small samples. It proposes a physical state perception and prediction framework based on SOC reconstruction. The overall framework diagram of Embodiment 1 of this invention is shown below. Figure 10 As shown, this framework decouples the complex degradation process into intrinsic degradation trends and transient operational perturbations under a unified physical baseline, and achieves a collaborative single modeling of long-term aging and transient polarization by combining a lightweight GRU-MLP network. Systematic validation using NASA public datasets and a self-built minimal-sample accelerated aging dataset yields the following conclusions: (1) The phase shift of time-domain features is eliminated, achieving a physically consistent characterization of the degradation law. Mechanism analysis and Taylor diagram evaluation results show that traditional features based on the cyclic domain are prone to temporal phase shifts when available capacity decays, leading to deviations in the predicted trajectory. The SOC spatial reconstruction mechanism proposed in this invention effectively suppresses fluctuation interference caused by operating time scaling. The ablation experiments in Table 2 show that, without introducing transient operating variables, the intrinsic degradation trend extracted solely by SOC reconstruction outperforms the traditional baseline model in all error indices. This indicates that establishing a unified electrochemical reference system is an important prerequisite for ensuring the consistency of SOH estimation.

[0107] (2) Based on the decoupling of dual physical features, the prediction robustness under dynamic battery conditions is improved. In the validation on the NASA dataset, the S²-HD framework overcomes the data distribution shift problem under complex aging conditions by effectively removing and compensating for transient operational disturbances. Evaluation shows that, relying only on the structurally simplified GRU-MLP model, the proposed framework controls the mean absolute percentage error (MAPE) to within 1.8% on four typical single-cell sets. The maximum value can reach 0.9938. Compared with traditional time-domain models with mixed feature inputs, this decoupling mechanism effectively suppresses the long-tailed distribution of prediction errors and improves the reliability and stability of SOH estimation.

[0108] (3) It breaks through the generalization boundary across individuals in scenarios with extremely small sample sizes. For a self-built accelerated aging dataset containing only 18-31 discrete observations within an individual's lifespan, the proposed framework overcomes the deep dependence of traditional time-series models on long-sequence historical data. Even facing the early gradual decay and the nonlinear "capacity plunge" in the later stages, the model can still accurately track the health evolution trajectory across individuals. Table 3 shows that, under sparse data conditions, the MAPE predicted by the model for both 0.5C and 0.25C test conditions remains below 1.81%, and the root mean square error (RMSE) is as low as within 0.0168. The overall stability remained above 0.96. This confirms that the combination of the latent space state transition mechanism and the decoupling of physical features enables the model to maintain good generalization ability even in scenarios with extremely small sample sizes.

[0109] (4) The computational efficiency of the lightweight prediction architecture was verified. Considering the limited computing resources of the battery BMS, the single-stream GRU-MLP network constructed in this invention effectively fused multi-scale degradation features in the hidden state space. Compared with traditional multi-stream parallel or complex attention mechanism networks, this architecture reduces the number of model parameters and computational complexity while maintaining high prediction accuracy. This provides the possibility for lightweight deployment of data-driven algorithms on engineering hardware.

[0110] Example 2 This embodiment provides a battery health state estimation system based on SOC domain reconstruction and feature decoupling, including: The data acquisition module is configured to acquire raw sensing data during the charge-discharge cycle of a lithium-ion battery, including time series of voltage, current, and temperature. The SOC domain reconstruction module is configured to map the raw sensing data to the SOC domain to obtain the sensing data of the SOC domain. The feature decoupling module is configured to extract intrinsic degradation trend features characterizing irreversible aging of the battery from the sensing data in the SOC domain, and simultaneously extract transient operating disturbance features characterizing the reversible polarization intensity under the current operating conditions. The feature fusion module is configured to fuse intrinsic degradation trend features with transient operational disturbance features to construct a joint state-aware vector; The estimation module is configured to input the joint state-aware vector into the temporal neural network model to obtain the battery health state (SOH) estimate.

[0111] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0112] Example 3 This embodiment also provides a computer device, including a computer-readable storage medium, a processor, and a computer program stored on the computer-readable storage medium and executable on the processor. When the processor executes the program, it completes the method described in Embodiment 1.

[0113] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0114] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0115] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0116] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0117] Those skilled in the art will recognize that the units and algorithm steps described in connection with the various examples of this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention.

[0118] Example 4 This embodiment also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the method described in Embodiment 1.

[0119] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A battery health state estimation method based on SOC domain reconstruction and feature decoupling, characterized in that, include: Acquire raw sensing data during the charge-discharge cycle of a lithium-ion battery, wherein the raw sensing data includes time series of voltage, current, and temperature. The original sensing data is mapped to the SOC domain to obtain sensing data in the SOC domain. The intrinsic degradation trend features characterizing irreversible aging of the battery are extracted from the sensing data in the SOC domain, and the transient operating disturbance features characterizing the reversible polarization intensity under the current operating conditions are also extracted. The intrinsic degradation trend features are fused with the transient operational disturbance features to construct a joint state perception vector; The joint state-aware vector is input into a time-series neural network model to obtain the battery health state (SOH) estimate.

2. The method according to claim 1, characterized in that, Mapping the raw sensor data to the SOC domain specifically includes: converting the current time series into an accumulated charge capacity series using the ampere-hour integration method; after monotonicity correction and zero-point shifting, calculating the state of charge (SOC) based on the available capacity at the initial life cycle of the battery; and then mapping the voltage, current, and temperature data to a preset, equally spaced fixed SOC grid through linear interpolation.

3. The method according to claim 1, characterized in that, Extract intrinsic degradation trend features characterizing irreversible aging of the battery from the sensing data in the SOC domain, including: Candidate features are constructed within the SOC domain, including: voltage statistical features, current statistical features, voltage versus time differential statistical features, local capacity increment features, and capacity differential IC curve features within the SOC interval. Pearson correlation coefficient and Spearman rank correlation coefficient were used to calculate the linear and nonlinear correlation between each candidate feature and the actual battery health state (SOH). Calculate the monotonicity score for each candidate feature sequence, which is used to quantify the purity of the evolution direction of the feature sequence within the global lifetime; The absolute values ​​of the Pearson correlation coefficient, the Spearman rank correlation coefficient, and the monotonicity score are normalized, and the three normalized indicators are weighted and summed to obtain the comprehensive score of each candidate feature. Several candidate features with the highest overall scores were selected as intrinsic degradation trend features.

4. The method according to claim 3, characterized in that, The monotonicity score is calculated according to the following formula: ； in, , These represent the candidate feature sequences at the th... Second and third The specific value extracted in the next loop. n The total number of cycles, It is a symbolic function.

5. The method according to claim 1, characterized in that, Extracting transient operation disturbance features specifically includes: defining a target feature interval within the SOC domain; calculating the mean values ​​of voltage, current, and temperature within the target feature interval for each charge-discharge cycle; and using the conditional expectations of the voltage, current, and temperature as transient operation disturbance features.

6. The method according to claim 1, characterized in that, The temporal neural network model is a single-stream GRU-MLP network, where the GRU layer is used to receive the joint state-aware vector at the current time and recursively update the hidden degenerate state, and the MLP layer is connected after the GRU layer to map the hidden state to the SOH estimate; the GRU layer is a single-layer GRU, the MLP layer is a two-layer MLP, and the loss function of the model adopts Huber loss.

7. The method according to claim 1, characterized in that, The intrinsic degradation trend feature and the transient operational disturbance feature are fused together, specifically by directly concatenating the intrinsic degradation trend feature vector and the transient operational disturbance feature vector along the feature dimension to form a joint state perception vector.

8. A battery health state estimation system based on SOC domain reconstruction and feature decoupling, characterized in that, include: The data acquisition module is configured to acquire raw sensing data during the charge-discharge cycle of a lithium-ion battery, wherein the raw sensing data includes time series of voltage, current and temperature. The SOC domain reconstruction module is configured to: map the original sensing data to the SOC domain to obtain sensing data in the SOC domain. The feature decoupling module is configured to extract intrinsic degradation trend features characterizing irreversible aging of the battery from the sensing data in the SOC domain, and simultaneously extract transient operating disturbance features characterizing the reversible polarization intensity under the current operating conditions. The feature fusion module is configured to fuse the intrinsic degradation trend feature with the transient operational disturbance feature to construct a joint state perception vector; The estimation module is configured to input the joint state-aware vector into a temporal neural network model to obtain an estimated battery health state (SOH).

9. A computer device comprising a computer-readable storage medium, a processor, and a computer program stored on the computer-readable storage medium and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

Images