A system, method, and medium for estimating the health status of an electrochemical battery.

Abstract

This invention discloses an electrochemical battery health state estimation system, method, and medium, belonging to the field of electric vehicle electrochemical battery management and state monitoring technology. The method acquires operational data during the cyclic charging and discharging process of the electrochemical battery, extracts discharge voltage drop time interval features, incremental capacity features, and temperature-voltage differential features, and divides them into multiple feature groups according to their source. Through group-level gated feature fusion, the weights of each feature group are learned, and multi-source health factors are weighted and fused to obtain fused features. Then, a sliding time window sequence is constructed based on the fused features from multiple consecutive cycles, and input into a long short-term memory network for time-series modeling, outputting the health state estimation result. This invention can improve the accuracy, stability, and engineering application capability of health state estimation.

Description

Technical Field

[0001] This invention belongs to the field of electric vehicle electrochemical battery management and status monitoring technology, and in particular relates to an electrochemical battery health status estimation system, method and medium. Background Technology

[0002] With the rapid development of the electric vehicle industry, electrochemical batteries, as the core power source for vehicle drive and energy storage systems, are widely used in various types of electric vehicles. Under long-term charge-discharge cycles and complex operating conditions, electrochemical batteries gradually experience capacity decay and internal resistance increase, leading to reduced driving range, weakened power output, and increased safety risks during operation. Therefore, accurately estimating the health status of electrochemical batteries is a crucial foundation for battery management systems to achieve safe operation, lifespan management, and maintenance decisions.

[0003] Existing methods for estimating the state of health of electrochemical batteries mainly include mechanistic model-based methods, methods based on equivalent circuits and filtering estimation, and data-driven learning methods. Among these, data-driven methods can directly learn battery degradation patterns from operational data without requiring precise electrochemical parameters, thus showing promising prospects for engineering applications.

[0004] However, existing technologies still have the following shortcomings: First, many methods rely on a small number of health factors from a single source, making it difficult to fully characterize the multidimensional changes of electrochemical batteries at different degradation stages, resulting in limited adaptability of the model to operating condition disturbances and individual battery differences; Second, although some methods introduce multi-source feature fusion, the sensitivity of different features is not consistent under different battery individuals and different degradation stages, and they are easily affected by measurement noise and abnormal data. Traditional fixed weights or simple splicing methods are prone to introducing redundant information, leading to unstable fusion results and reduced generalization performance of the model; Third, the degradation process of electrochemical batteries has obvious time correlation. If the model cannot effectively characterize the long-term and short-term time-series dependencies, the accuracy and stability of health state estimation will still be difficult to further improve.

[0005] Therefore, there is an urgent need for an electrochemical battery health state estimation scheme that can fully utilize multi-source complementary information, has adaptive fusion capabilities, and can effectively model degradation time-series characteristics, so as to improve the estimation accuracy, robustness, and usability of different individual batteries and different degradation stages. Summary of the Invention

[0006] In view of this, the present invention aims to provide an electrochemical battery health status estimation system, method and medium to at least solve one of the problems in the background art.

[0007] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0008] Firstly, this scheme discloses a method for estimating the health status of an electrochemical battery, including the following steps:

[0009] The operation data of the electrochemical battery during the cyclic charge and discharge process is acquired, and the operation data is preprocessed to obtain effective data segments divided by cycle. The operation data includes voltage data, capacity data and temperature data.

[0010] Based on the effective data segments corresponding to each preprocessed cycle, multi-source health factors are extracted. These multi-source health factors include discharge voltage drop time interval characteristics, incremental capacity characteristics, and temperature-voltage differential characteristics.

[0011] The various multi-source health factors extracted in the same cycle are divided into multiple feature groups according to their source, and each feature group is uniformly processed to form the feature vector corresponding to the cycle.

[0012] The feature vector is input into the group-level gated feature fusion module, which generates group-level weights based on each feature group. According to the correspondence between each feature and its respective feature group, the group-level weights are mapped to the gate coefficients corresponding to each feature. The feature vector is then weighted according to the gate coefficients to obtain the fused feature corresponding to the loop.

[0013] A sliding time window sequence is constructed based on the fusion features corresponding to multiple consecutive cycles arranged in chronological order. The sliding time window sequence is then input into a long short-term memory network for time-series modeling, and the corresponding electrochemical cell health state estimation results are output.

[0014] Furthermore, the preprocessing includes: cyclically segmenting the running data to obtain charge / discharge data segments corresponding to each cycle; and performing at least one of outlier handling, missing value handling, smoothing, resampling, and unification processing on the cyclically segmented running data.

[0015] The preprocessed running data is used to support the extraction of discharge voltage drop time interval features, incremental capacity features, and temperature-voltage differential features, respectively, so that health factors from different sources correspond to data segments under the same cycle number.

[0016] Furthermore, the extraction of the discharge equal voltage drop time interval feature includes:

[0017] Within the preset effective discharge voltage range, multiple adjacent voltage intervals are divided according to the preset voltage step size;

[0018] For each cycle, record the time difference on the discharge voltage curve as the voltage drops from the upper boundary to the lower boundary of each voltage range;

[0019] The time difference mentioned above is used as the discharge equal voltage drop time interval characteristic corresponding to this cycle;

[0020] The multiple time differences within the same cycle are grouped into a first type of feature group in a preset order to characterize the changes in discharge duration within the cycle as capacity decays and internal resistance increases.

[0021] Furthermore, the extraction of the incremental capacity features includes:

[0022] Based on capacity and voltage data in the same cycle, an incremental capacity curve is constructed using a cross-point differential method.

[0023] The incremental capacity curve is smoothed to suppress noise and stabilize the curve shape;

[0024] Extract at least one of the main peak position, main peak amplitude, and area of ​​the corresponding interval of the main peak from the smoothed incremental capacity curve as the incremental capacity feature corresponding to the cycle.

[0025] The extracted incremental capacity features are then grouped into a second type of feature group, wherein the area of ​​the interval corresponding to the main peak is determined by the curve integral result within a preset voltage bandwidth interval centered on the main peak position.

[0026] Furthermore, the extraction of the temperature-voltage differential features includes:

[0027] Based on temperature and voltage data in the same cycle, a temperature-voltage differential curve is constructed using a cross-point differential method.

[0028] The temperature-voltage differential curve is smoothed.

[0029] Extract at least one of the peak position, peak amplitude, valley position, and valley amplitude from the smoothed temperature-voltage differential curve as the temperature-voltage differential feature corresponding to the cycle.

[0030] The extracted temperature-voltage differential features were then grouped into a third type of feature set to characterize the impact of temperature and terminal voltage coupling changes on the health status during battery aging.

[0031] Furthermore, multi-source health factors are divided into multiple characteristic groups according to their source, and each characteristic group is standardized, including:

[0032] The discharge equal voltage drop time interval characteristics are divided into the first characteristic group, the incremental capacity characteristics are divided into the second characteristic group, and the temperature-voltage differential characteristics are divided into the third characteristic group.

[0033] Normalization and consistency processing are performed on each feature group separately;

[0034] During the standardization process, missing values ​​are filled or removed, and redundant dimensions are suppressed to obtain the standardized feature vectors corresponding to each loop.

[0035] In this standardized feature vector, each feature dimension retains its corresponding feature group identifier, which is used for subsequent group-level gating weight mapping.

[0036] Furthermore, the group-level gating feature fusion includes:

[0037] The feature vectors are input into a gating network, which then scores each feature group at the group level.

[0038] The group-level scoring results are normalized to obtain the group-level weights corresponding to each feature group;

[0039] Based on the correspondence between each feature and its corresponding feature group, the group-level weights of each feature group are mapped to the gating coefficients corresponding to each feature;

[0040] The feature vector is weighted element-wise according to the gating coefficient to obtain the fused feature.

[0041] The fusion features serve as input for the construction of subsequent sliding time window sequences.

[0042] Furthermore, the gated network includes a hidden layer and an output layer connected in sequence;

[0043] The hidden layer is used to learn intermediate representations of the differences in the contribution of different feature groups to health status estimation;

[0044] The output layer is used to output the group-level scoring results corresponding to each feature group;

[0045] The scoring results of the output layer are normalized by softmax to obtain the group-level weights corresponding to each feature group;

[0046] Furthermore, when the group-level weights are mapped to gating coefficients, a lower gating limit parameter is introduced to ensure that each feature group retains its basic contribution during the fusion process, thus preventing any feature group from being completely suppressed.

[0047] Furthermore, the sliding time window sequence is composed of multiple consecutive cyclic fusion features stacked in chronological order;

[0048] After setting the window length, the fused features corresponding to multiple consecutive cycles are combined into a time window sample;

[0049] The time window samples are input into a unidirectional long short-term memory network for sequence modeling.

[0050] Based on the hidden state corresponding to the end of the window, the health status estimation result is output through the regression layer;

[0051] The health status estimation results are used to characterize the health status of the electrochemical cell corresponding to the current cycle.

[0052] Furthermore, the Long Short-Term Memory network performs gated updates on the fused features of the inputs at each time step within the sliding time window to learn the long-term dependencies and short-term changes in the electrochemical cell degradation process.

[0053] The health status estimation result is obtained from the hidden state at the end of the window through a fully connected regression layer;

[0054] Furthermore, after obtaining the health status estimation result, the output result is smoothed before being output in real time or offline for use in battery management or life assessment.

[0055] Furthermore, the method also includes model training, parameter management, and online inference steps, which include:

[0056] During the training phase, the parameters of the group-level gated feature fusion module and the long short-term memory network are trained based on historical battery data.

[0057] During the inference phase, the trained model parameters are loaded, and the newly arrived recurrent data are processed sequentially to extract health factors, construct feature groups, perform group-level gated feature fusion, construct sliding time windows, and output health status.

[0058] The normalization parameters, model version information, and inference logs are stored in the memory.

[0059] Furthermore, a battery-level cross-validation strategy is used to evaluate model performance during the training phase;

[0060] The battery-level cross-validation strategy includes selecting one battery dataset as the test set in turn and using the remaining battery datasets as the training set.

[0061] At least one of the mean absolute error, mean absolute percentage error and root mean square error is used as the evaluation index.

[0062] The model parameters for online inference are determined based on the evaluation results.

[0063] Secondly, this solution discloses an electrochemical battery health status estimation system, including a processor, a memory, and a program stored in the memory and executable by the processor. When the program is executed by the processor, it implements the electrochemical battery health status estimation method.

[0064] The system includes:

[0065] The data acquisition and preprocessing module is used to acquire the operating data of the electrochemical cell and perform preprocessing to output effective data segments divided by cycle.

[0066] The multi-source health factor extraction module is used to extract discharge voltage drop time interval features, incremental capacity features, and temperature-voltage differential features based on the effective data fragments.

[0067] The feature group construction module is used to divide the multi-source health factors corresponding to the same cycle into multiple feature groups according to their source, and perform unified processing to form feature vectors;

[0068] The group-level gated feature fusion module is used to generate group-level weights based on each feature group, and to perform weighted fusion of the feature vectors according to the group-level weights to obtain fused features;

[0069] A sliding time window construction module is used to construct a sliding time window sequence based on the fused features of multiple consecutive cycles;

[0070] The health status regression module is used to perform time-series modeling on the sliding time window sequence based on a long short-term memory network and output the health status estimation results.

[0071] The model training, parameter management, and online inference module is used to complete model parameter training during the training phase and load the trained model parameters during the inference phase to perform health status estimation on newly arrived recurrent data.

[0072] The results output and interface module is used to output the health status estimate to the battery management system control strategy module, life management module, alarm module, or visualization terminal.

[0073] Furthermore, the group-level gating feature fusion module includes a gating network and a mapping unit;

[0074] The gated network is used to output the group-level weights corresponding to each feature group;

[0075] The mapping unit is used to generate feature dimension gating coefficients based on the correspondence between group-level weights and feature group identifiers;

[0076] The health status regression module includes a sliding time window long short-term memory network and a regression output layer, which are used to convert fused features into health status estimation results.

[0077] Thirdly, this solution discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the electrochemical battery health state estimation method.

[0078] Compared with existing technologies, the electrochemical battery health status estimation system, method, and medium described in this invention have the following advantages:

[0079] (1) This invention extracts the discharge voltage drop time interval feature, incremental capacity feature and temperature-voltage differential feature to characterize the degradation state of electrochemical battery from multiple perspectives such as discharge kinetics change, electrochemical reaction feature change and thermal behavior coupling change, making the health state characterization more comprehensive and reducing the impact of single feature failure on the estimation results;

[0080] (2) This invention divides multi-source health factors into multiple feature groups according to their sources, learns the weights of each feature group by using a group-level gated feature fusion method, and then maps the group-level weights to the gate coefficients of the feature dimensions to perform weighted fusion of the feature vectors. This can enhance the role of high-contribution feature groups in health status estimation and suppress the interference of noise and redundant information, thereby improving the discriminative ability of the fused features and the robustness of the model.

[0081] (3) This invention constructs a sliding time window sequence based on the fusion features of multiple consecutive cycles and uses a long short-term memory network for time series modeling, which can characterize the dynamic changes and time series dependencies between cycles in the degradation process of electrochemical cells, thereby improving the accuracy and stability of the health status estimation results;

[0082] (4) The present invention adopts a modular structure to realize data acquisition and preprocessing, multi-source health factor extraction, feature group construction, group-level gating fusion, sliding time window construction, health status regression and result output, which is easy to deploy in battery management system, vehicle computing unit or cloud server, and has good engineering implementation and system integration capabilities.

[0083] (5) In the training and testing process of different battery individuals, the present invention can improve the adaptability across battery scenarios through multi-source health factor fusion and time-series modeling, which is conducive to improving the generalization performance and application value of the health status estimation model in practical applications. Attached Figure Description

[0084] 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 undue limitation of the invention. In the drawings:

[0085] Figure 1 This is a schematic diagram of the method described in an embodiment of the present invention;

[0086] Figure 2 This is a comparison chart of SOH estimation results for electrochemical cells under different feature input schemes as described in the embodiments of the present invention;

[0087] Figure 3This is a comparison chart of SOH error results of electrochemical cells under different feature input schemes described in the embodiments of the present invention;

[0088] Figure 4 This is a comparison chart of SOH estimation results for electrochemical cells under different prediction models described in the embodiments of the present invention;

[0089] Figure 5 This is a comparison chart of SOH error results of electrochemical cells under different prediction models described in the embodiments of the present invention. Detailed Implementation

[0090] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0091] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0092] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0093] The electrochemical batteries described in this invention include, but are not limited to, lithium-ion batteries and sodium-ion batteries. The following embodiments use lithium-ion batteries as an example to illustrate the technical solution of this invention. However, the processing procedures of this invention, such as data acquisition, preprocessing, multi-source health factor extraction, group-level gated feature fusion, sliding time window construction, and long short-term memory network time-series modeling, are also applicable to other electrochemical batteries such as sodium-ion batteries. For different battery systems, the effective voltage range, voltage step size, cross-point differential interval, smoothing parameters, peak-valley extraction range, sliding time window length, and model training parameters in the feature extraction process can be adjusted accordingly based on their charge-discharge platform, voltage range, temperature response characteristics, and degradation patterns, without affecting the overall technical concept of this invention.

[0094] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0095] Figure 1 This is a schematic diagram of the method described in an embodiment of the present invention. The diagram illustrates the complete execution logic of the method from bottom-level data processing to top-level model output. Specifically, the process shows the entire process from acquiring and preprocessing lithium battery operating data, through the extraction and grouping of multi-source health factors, to dynamic weighted fusion using a group-level gated feature fusion module, and finally combining the sliding time window sequence input to a long short-term memory network for time-series modeling, thereby accurately outputting the electrochemical battery health state estimation result.

[0096] Figure 2 This is a comparison chart of SOH estimation results for electrochemical cells under different feature input schemes described in this invention. The chart shows that, in five sub-charts using Cell1, Cell3, Cell4, Cell7, and Cell8 as test sets, the fusion feature scheme proposed in this invention consistently matches the true SOH value curve compared to the single feature input scheme. This directly verifies that by fusing multiple health factors, this invention can effectively overcome the fluctuations and biases that easily occur with single features in the later stages of battery aging, exhibiting higher SOH estimation accuracy and better robustness across different individual batteries.

[0097] Figure 3This is a comparison of the SOH error results of electrochemical cells under different feature input schemes described in this invention. The figure clearly shows that in the five sub-figures using Cell1, Cell3, Cell4, Cell7, and Cell8 as test sets, the SOH estimation error of the multi-source feature fusion scheme proposed in this invention consistently remains within a very small range, closest to 0. In contrast, the traditional single-feature input scheme exhibits varying degrees of drastic fluctuations and significant deviations throughout the entire cycle. This further verifies that the feature fusion method of this invention can significantly suppress uncertainties in the prediction process, greatly reduce estimation errors, and improve the overall prediction accuracy and stability of the model.

[0098] Figure 4 This is a comparison chart of the SOH estimation results of electrochemical cells under different prediction models described in the embodiments of the present invention. The chart clearly shows that, in the five sub-charts with Cell1, Cell3, Cell4, Cell7, and Cell8 as test sets, the GFF-LSTM model proposed in this invention consistently maintains a close fit to the true SOH value compared to traditional BP, TCN, GRU, and basic LSTM models. This indicates that the model constructed in this invention has high estimation accuracy and stronger cross-individual generalization ability when dealing with the nonlinear dynamic characteristics of battery degradation.

[0099] Figure 5 This figure compares the State of Health (SOH) error results of electrochemical cells under different prediction models described in this invention. The figure clearly shows that in the tests of five different battery cells, the SOH estimation error of the GFF-LSTM model proposed in this invention consistently remained within an extremely narrow range, closest to zero. Compared to the large error shifts that occur in the prediction process of traditional BP, LSTM, GRU, and TCN models, the GFF-LSTM model of this invention exhibits better error control and anti-interference capabilities, further confirming the high accuracy and robustness advantages of this model in battery health estimation.

[0100] This embodiment provides a method for estimating the health status of electrochemical batteries, used to predict the health status of electric vehicle electrochemical batteries during cyclic charge-discharge processes. This method can be deployed in a battery management system, on-board computing unit, or cloud server. It utilizes operational data from the electrochemical battery during cyclic charge-discharge processes to extract multi-source health factors, fuse group-level gated features, and perform time-series modeling, thereby outputting the electrochemical battery's health status estimation results. This provides a basis for battery life management, maintenance decisions, and safety alarms.

[0101] In this embodiment, operational data of the electrochemical battery during cyclic charging and discharging is first acquired. This operational data includes at least voltage, capacity, and temperature data, and in some embodiments, may further include current data. The operational data is then preprocessed, including at least one of the following: cycle segmentation, outlier handling, missing value handling, smoothing, resampling, and standardization. The standardization process transforms features from different sources, with different dimensions, and different sampling scales into a unified form that can be jointly input into the same model, including but not limited to normalization, consistency alignment, and unified data format processing. Through this preprocessing, the original operational data meets the requirements for subsequent health factor extraction and model input.

[0102] After pretreatment, multi-source health factors are extracted from each cycle. These multi-source health factors include at least discharge isobaric time interval characteristics, incremental capacity characteristics, and temperature-voltage differential characteristics. Specifically, the discharge isobaric time interval characteristics characterize the changes in voltage maintenance capability of the electrochemical battery during discharge, the incremental capacity characteristics characterize the changes in electrochemical reaction characteristics, and the temperature-voltage differential characteristics characterize the coupled changes in thermal behavior. These three types of health factors originate from different physical phenomena and can characterize the degradation state of the electrochemical battery from different perspectives, thus exhibiting strong complementarity.

[0103] In this embodiment, the discharge voltage drop time interval feature is extracted as follows: During the discharge process, an effective discharge voltage range of 3.4V to 4.2V is selected, and divided into 8 adjacent voltage intervals with a step size of 0.1V; for each cycle, the time difference experienced by the voltage dropping from the upper boundary to the lower boundary of each voltage interval is recorded on the discharge voltage curve, resulting in 8 time interval features, denoted as F1 to F8. The above time interval features can characterize the phenomenon of shortened discharge duration caused by capacity decay and internal resistance increase.

[0104] Specifically, the discharge equal voltage drop time interval (DEVDT) characteristic is as follows: during the discharge process, the effective discharge voltage range of 3.4V-4.2V is selected, and 8 adjacent voltage intervals t are divided with a step size of 0.1V. V For each cycle, the time difference that the voltage takes from the upper boundary to the lower boundary of the interval is recorded on the discharge voltage curve, thus obtaining 8 time interval characteristics, which are denoted as F1-F8, and are used to characterize the phenomenon of shortened discharge duration caused by capacity decay and internal resistance growth.

[0105] For lithium-ion batteries, multiple adjacent voltage ranges can be divided in 0.1V increments within the effective discharge voltage range of 3.4V–4.2V. For sodium-ion batteries or other electrochemical batteries, the voltage range and increment are determined based on the effective discharge platform, voltage operating range, and sampling resolution of the corresponding battery system.

[0106] In this embodiment, the incremental capacity feature is extracted as follows: an incremental capacity curve is constructed based on capacity and voltage data. To reduce the impact of sampling noise on differential operations, a cross-point differential method is used to calculate the approximate incremental capacity value at each sampling point. That is, two sampling points separated by a preset sampling interval are selected, and the ratio of the capacity difference to the voltage difference between the two sampling points is used as the approximate incremental capacity value of the current sampling point. In a specific implementation, the sampling interval is 20. The obtained discrete incremental capacity sequence is further smoothed using a Gaussian kernel to suppress residual noise and stabilize the curve shape. Then, the peak voltage and peak height corresponding to the main peak are extracted from the smoothed incremental capacity curve, and a fixed bandwidth of 0.03V is taken on both sides of the main peak position. The area under the curve in this interval is calculated, and the obtained peak position, peak value, and area are used as incremental capacity features.

[0107] Specifically, in the incremental capacity (IC) curve characteristics, the discharge voltage and capacity sequence are discretized. To reduce the influence of noise, the incremental capacity at the kth sampling point is approximated by cross-point difference calculation.

[0108]

[0109] Where Q is the capacity, V is the discharge voltage, L is the sampling interval, and L = 20.

[0110] Gaussian kernel smoothing filter is applied to the discrete IC sequence to obtain a smoothed IC curve, thereby suppressing residual noise and stabilizing the curve shape.

[0111]

[0112] Where μ is the mean, which is used to make the Gaussian kernel expand symmetrically around the current sampling point. The standard deviation is used to control the width of the smoothing window.

[0113] The IC curve after GS is denoted as

[0114]

[0115] Extract the peak voltage and peak height corresponding to the main peak from the smoothed IC curve, and take a fixed bandwidth of 0.03V to the left and right of the peak as the center. Calculate the area under the curve in this interval as the main peak area to form the IC feature group.

[0116] In this embodiment, the temperature-voltage differential feature is extracted as follows: a temperature-voltage differential curve is constructed based on temperature and voltage data. Considering that direct differentiation between adjacent sampling points is easily affected by noise, this embodiment uses a cross-point differentiation method to calculate the approximate temperature-voltage differential value. That is, two sampling points separated by a preset differential interval are selected, and the ratio of the temperature difference to the voltage difference between the two sampling points is used as the approximate temperature-voltage differential value of the current sampling point. In a specific implementation, the differential interval is taken as 1 / 15 of the total number of sampling points in this discharge process. The obtained temperature-voltage differential sequence is also smoothed using a Gaussian kernel. Then, the peak position, peak amplitude, valley position, and valley amplitude are extracted from the smoothed temperature-voltage differential curve. The peak and valley amplitudes can be combined to further construct a temperature-voltage differential feature group, which is denoted as F9 to F14 in sequence.

[0117] Specifically, in the temperature-voltage differential (DTV) characteristic, electrochemical cell aging leads to increased internal resistance and polarization, resulting in increased irreversible heat generation. Although the internal heat source is difficult to measure directly, the coupling change between the terminal voltage and surface temperature under constant current conditions is strongly correlated with capacity degradation. Therefore, the temperature-voltage differential (DTV) characteristic is introduced.

[0118]

[0119] The actual measured T and V are discrete sampling sequences. Directly differentiating adjacent points would be subject to severe noise interference. Therefore, this paper uses a cross-point differencing method to calculate the approximate DTV value at the k-th sampling point, that is, selecting two points on the temperature-voltage curve separated by a certain sampling interval L for differencing.

[0120]

[0121] Where T is temperature, V is voltage, t is sampling time, and N is the total number of sampling points in this discharge process, which is empirically taken as L=N / 15.

[0122] The DTV sequence was also Gaussian smoothed, and the peak and trough values ​​were extracted from the smoothed DTV curve as health factors. The voltage position and amplitude corresponding to the highest and lowest points were recorded to form four basic quantities. A set of DTV features was constructed by combining the peak and trough amplitudes, and denoted as F9–F9. 14 .

[0123] After extracting the three types of health factors mentioned above, the discharge isobaric time interval feature, incremental capacity feature, and temperature-voltage differential feature are divided into multiple feature groups according to their sources, and each feature group is normalized and consistency processed. In some embodiments, redundancy suppression and missing value processing can also be combined to obtain multi-source health feature inputs that can be used for fusion. The division into multiple feature groups according to their sources means grouping features from the same type of physical representation method or the same type of calculation method into the same group. For example, the discharge isobaric time interval feature is grouped into one group, the incremental capacity feature into one group, and the temperature-voltage differential feature into one group.

[0124] Specifically, for the group-level gated feature fusion module GFF, to reflect the differences in the contribution of features from different sources to SOH and to suppress redundancy and noise interference, this invention introduces a group-level gated fusion module at the feature layer. Let the t-th cycle-normalized feature vector be... ∈ d represents the dimension of the feature vector, which is divided into G feature groups based on its source. The gated network uses a two-layer fully connected structure to learn the weights of the groups. The hidden layer is calculated first:

[0125]

[0126] Where ReLU(·) is the linear rectified activation function, W1 is the first-layer weight matrix, b1 is the first-layer bias vector, and H is the gated hidden layer dimension. The group-level scoring vector is then obtained:

[0127]

[0128] Where W2 is the first-layer weight matrix, b2 is the first-layer bias vector, and u t Applying softmax yields the group weights:

[0129]

[0130] in, Let M be the group-level weight corresponding to the g-th feature group. To map the group weights to the feature dimension, construct a group-to-feature mapping matrix M∈{0,1}. d×G :

[0131]

[0132] Among them, M j,g Let Ω be the element in the j-th row and g-th column of M, where j = 1, 2, ..., d are the feature dimension numbers. g Let g be the set of feature dimension indices contained in the g-th feature group. Further, by introducing a lower gate bound parameter ε∈(0,1), we obtain the feature dimension gate vector:

[0133]

[0134] Where, 1∈ The original features are d-dimensional vectors of all ones. Finally, element-wise gated weighting is applied to the original features to obtain the fused features:

[0135]

[0136] Where ⊙ represents the Hadamard product. Let z t The fused features are then used as input to the subsequent time series model for SOH prediction.

[0137] In this embodiment, the feature vectors corresponding to each loop are input into the group-level gated feature fusion module. The group-level gated feature fusion module first divides all features into G feature groups according to their source, and then generates corresponding group-level weights based on each feature group. The group-level weights are weight parameters used to characterize the contribution of each feature group to the current health status estimation. Further, based on the correspondence between each feature and its corresponding feature group, the group-level weights of each feature group are mapped to the corresponding gate coefficients. The gate coefficients are used to characterize the weighting degree of each feature dimension during the fusion process. Then, the feature vectors are weighted according to the gate coefficients to obtain the fused features. Through the above processing, the influence of high-contribution feature groups on health status estimation can be enhanced, and the interference of noisy and redundant features on the estimation results can be suppressed.

[0138] In a more specific implementation, the group-level weights are generated through a gating network, which includes a hidden layer and an output layer connected in sequence. The hidden layer is used to learn intermediate representations for each feature group, and the output layer is used to output the scoring results for each feature group. The results of the output layer are normalized to obtain the group-level weights corresponding to each feature group. In some embodiments, the gating network uses a two-layer fully connected structure to learn the group-level weights. To avoid completely suppressing a certain group of features, a lower gating bound parameter can be introduced during the gating mapping process to ensure that each group of features retains its basic contribution during the fusion process.

[0139] After obtaining the fusion features, a sliding time window sequence is constructed based on the fusion features of multiple consecutive cycles, and the sliding time window sequence is input into a long short-term memory network for time series modeling.

[0140] In GFF-LSTM, battery degradation is time-dependent. This invention uses a sliding time window to construct sequence samples and LSTM to model the inter-cycle degradation dynamics. Let the fused feature be z. t Given a window length P, construct the sliding time window input sequence:

[0141]

[0142] Z tInput a unidirectional LSTM. For the first [unit] within the window... Input at each time step The LSTM gating is updated as follows:

[0143]

[0144] in, Number the time steps within the sliding time window. This is a feature of fusion. For the Sigmoid function, tanh It is the hyperbolic tangent function. , and These are the forget gate, input gate, and output gate, respectively. and These are the hidden state and the cellular state, respectively. Candidate cell state. W and These are the weight matrix and the bias vector, respectively. The hidden state corresponding to the end of the window is then retrieved. As a sequence representation, the predicted value is obtained through a fully connected regression layer:

[0145]

[0146] The sliding time window sequence is an input sequence composed of fused features from multiple consecutive cycles stacked in chronological order, used to characterize the degradation dynamics of electrochemical cells during continuous cycling. In a specific implementation, a sliding time window sequence can be constructed using fused features from eight consecutive cycles. This sliding time window sequence is input into a unidirectional long short-term memory network, the hidden state corresponding to the end of the window is taken as the sequence representation, and the corresponding health state estimation result is output through a fully connected regression layer.

[0147] When estimating the health status of an electrochemical battery under test, the operating data of the battery during cyclic charge-discharge processes are first acquired. Then, the preprocessing, multi-source health factor extraction, feature group construction, group-level gating fusion, and sliding time window construction are completed according to the steps described above. Finally, the obtained sliding time window sequence is input into a trained long short-term memory network model to output the health status estimation result of the electrochemical battery under test. The health status estimation result can be smoothed and output in real time, or it can be output offline to a battery management system, life management module, alarm module, or visualization terminal.

[0148] This embodiment further illustrates the training and verification process of the method described in Embodiment 1.

[0149] In this embodiment, the health status (SOH) is defined as the ratio of the current available capacity to the initial rated capacity. By collecting the actual available capacity in each cycle and calculating the ratio with the initial rated capacity, the health status label for the corresponding cycle is obtained, which serves as the target output for model training and prediction.

[0150] During the model training phase, publicly available battery aging datasets can be used as training and validation samples. In one specific implementation, the publicly available electrochemical battery aging dataset from Oxford University is selected (see "Diagnosis and Prognosis of Degradation in Lithium-Ion Batteries", PhD thesis, Department of Engineering Science, University of Oxford, 2017). This dataset includes aging data of multiple Kokam lithium-ion pouch batteries with a rated capacity of 740mAh under dynamic operating conditions at 40℃. To ensure sample quality, individual battery data with abnormal capacity drops are removed during data preprocessing, as well as incomplete lifecycle battery data that has not reached the end of its lifespan. Finally, battery data with a complete capacity decay process and a relatively smooth trajectory are selected as the valid dataset.

[0151] To verify the model's cross-battery generalization ability, a battery-level leave-one-out cross-validation strategy can be adopted. That is, each time, one battery is selected from the effective dataset as the test set and the remaining batteries are used as the training set. This process is repeated multiple times, and the average value of each test result is used as the final evaluation criterion.

[0152] To verify the effectiveness of group-level gated feature fusion, under the same long short-term memory network architecture, schemes were set up with only incremental capacity features, only temperature and voltage differential features, only discharge voltage drop time interval features, and features after group-level gated fusion, and the above schemes were compared. By comparing the health state estimation results under different input schemes, it can be shown that multi-source health factors have better information complementarity after group-level gated fusion.

[0153] To verify the advantages of the health status estimation model in time series modeling, the model using group-level gated fusion and long short-term memory network can be compared with error backpropagation neural network, temporal convolutional network, gated recurrent unit network and basic long short-term memory network, and the prediction performance of each model can be compared under the same data partitioning and evaluation conditions.

[0154] During the evaluation phase, mean absolute error, mean absolute percentage error, and root mean square error can be used as evaluation indicators to quantitatively analyze the health status estimation results on different test objects. Through these evaluation methods, the prediction accuracy and stability of the health status estimation model on different individual batteries can be verified.

[0155] Specifically, in this embodiment, the SOH estimation process is divided into four stages: data acquisition, data preprocessing and health factor preparation, model training, and SOH estimation. Detailed implementation and verification are as follows:

[0156] (1) Acquisition and selection of training dataset

[0157] In the model training and validation phases, this invention is based on the battery aging dataset publicly available from the University of Oxford. The original dataset contains dynamic aging data of eight Kokam lithium-ion pouch batteries (Cell1-Cell8) with a rated capacity of 740mAh at 40℃. During the data preprocessing phase, to ensure sample quality and avoid interference from abnormal degradation on LSTM time-series modeling, this invention removed Cell2 and Cell5, which exhibited unexplained capacity drops, and Cell6, which did not reach the end of its lifespan. Finally, five batteries—Cell1, Cell3, Cell4, Cell7, and Cell8—with complete capacity decay processes and relatively smooth trajectories were selected as the valid dataset.

[0158] State of Operation (SOH) is typically defined as the ratio of current available capacity to initial rated capacity, calculated using the following formula:

[0159]

[0160] Among them, C t C0 represents the actual usable capacity of the battery obtained in this test; C0 represents the initial rated capacity of the battery.

[0161] The training dataset described above is merely an exemplary source of training and validation data for this invention, used to illustrate the feasibility and effectiveness of the method. In practical applications, cyclic aging data collected from sodium-ion batteries or other electrochemical batteries under corresponding operating conditions can also be used to train, validate, and deploy the method of this invention.

[0162] (2) Cross-validation strategy

[0163] To verify the model's generalization ability across individual batteries, this invention employs a battery-level "leave-one-out cross-validation" method. This involves selecting one battery from five valid batteries in turn as the test set, and the remaining four as the training set, repeating this process five times. The average of the evaluation results from each test set is used as the final evaluation criterion.

[0164] (3) To verify the necessity of the group-level gated feature fusion (GFF) module proposed in this invention, this embodiment sets up four sets of contrast input schemes and trains them under the same LSTM network architecture:

[0165] IC solution: Prediction is performed by inputting only the incremental capacity curve characteristics;

[0166] DTV scheme: Prediction is performed by inputting only the temperature-voltage differential characteristics;

[0167] DEVDT scheme: Prediction is performed by only inputting the discharge equal voltage drop time interval characteristics;

[0168] GFF scheme: Input features after group-level gating fusion.

[0169] (4) Performance comparison of different prediction models

[0170] To verify the superiority of the proposed GFF-LSTM model in handling nonlinear degradation time dependencies, this embodiment compares GFF-LSTM with the following four mainstream algorithms:

[0171] BP (Backpropagation Neural Network): a representative of traditional machine learning regression models;

[0172] TCN (Temporal Convolutional Network): a representative of parallel temporal modeling based on convolutional structures;

[0173] GRU (Gated Recurrent Unit Network): a representative of lightweight recurrent neural networks;

[0174] LSTM (Long Short-Term Memory Network): a representative of basic time-series regression models.

[0175] (5) SOH prediction and evaluation indicators

[0176] To quantitatively evaluate the accuracy and stability of the GFF-LSTM model in SOH prediction, this invention uses mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) as core evaluation indicators. The specific calculation formulas are as follows:

[0177] MAE:

[0178]

[0179] MAPE:

[0180]

[0181] RMSE:

[0182]

[0183] In the formula, S is the total number of samples in the test set, and y i This represents the true SOH test value for the i-th sample. The model predicts the SOH value. By evaluating these three indicators, the reliability and advantages of this invention in assessing the health status of different individual batteries can be verified.

[0184] This embodiment provides a system for implementing the above-described electrochemical battery health status estimation method.

[0185] The system includes a processor, a memory, and a program stored in the memory and executable by the processor. When executed by the processor, the program implements the electrochemical battery health state estimation method described in the foregoing embodiments. The system can be deployed in a battery management system, an onboard computing unit, or a cloud server, and is connected to a battery acquisition terminal via a communication interface.

[0186] The system includes a data acquisition and preprocessing module for acquiring and preprocessing the operating data of the electrochemical battery during cyclic charging and discharging. The operating data includes at least voltage data, capacity data, and temperature data, and in some embodiments, it may further include current data. The preprocessing includes at least one of cycle segmentation, missing value processing, outlier processing, smoothing processing, resampling processing, and normalization processing.

[0187] The system includes a multi-source health factor extraction module, used to extract discharge isobaric time interval features, incremental capacity features, and temperature-voltage differential features from each cycle. Specifically, the discharge isobaric time interval feature submodule divides the effective discharge voltage range of 3.4V to 4.2V into multiple adjacent voltage intervals of 0.1V and records the time difference between the voltage drop from the upper boundary to the lower boundary of each interval; the incremental capacity feature submodule calculates the incremental capacity sequence through cross-point difference and performs Gaussian smoothing, extracting the main peak correlation features from the smoothed incremental capacity curve; the temperature-voltage differential feature submodule calculates the temperature-voltage differential sequence through cross-point difference and performs Gaussian smoothing, extracting peak and valley correlation features from the smoothed temperature-voltage differential curve.

[0188] The system includes a feature group construction module for dividing the multi-source health factors into multiple feature groups according to their source and performing unified processing; it also includes a group-level gated feature fusion module for weighted fusion of multiple feature groups to obtain fused features. The group-level gated feature fusion module generates group-level weights corresponding to each feature group through a gating network, and maps the group-level weights to gating coefficients corresponding to each feature based on the correspondence between each feature and its respective feature group. Then, it weights each dimension of the feature vector according to the gating coefficients to obtain the fused features.

[0189] The system also includes a sliding time window construction module for constructing a sliding time window sequence based on the fused features of multiple consecutive cycles; a health status regression module for performing time-series modeling of the sliding time window sequence based on a long short-term memory network and outputting health status estimation results; a model training, parameter management, and online inference module for training model parameters based on historical battery data during the training phase, loading the trained model parameters during the inference phase, and performing health factor extraction, feature fusion, sliding time window construction, and health status output on newly arriving cycle data; and a result output and interface module for outputting the health status estimate to the battery management system control strategy module, life management module, alarm module, or visualization terminal via an onboard bus or network interface.

[0190] Specifically, in this embodiment, the present invention provides a group-level gated fusion and sliding time window LSTM system for predicting the state of health (SOH) of electrochemical batteries. This system can be deployed in a battery management system (BMS), an onboard computing unit, or a cloud server. Its hardware includes a processor, a memory, and a communication interface connected to the battery acquisition terminal. The software functional modules run on the processor and are stored in the memory.

[0191] (1) Data acquisition module

[0192] This is used to acquire battery cycle aging data, including at least sampling sequences of voltage, current, temperature, etc., that change over time or with capacity, and organizes the data by cycle or by charge / discharge segments to provide input for subsequent health factor extraction.

[0193] (2) Data preprocessing and health factor preparation module

[0194] This is used for consistency processing of multi-source raw sequences, including cyclic segmentation, handling of missing and outliers, necessary smoothing or resampling, and unification of different health factors to form a set of health factors that can be used for cross-cell training and inference.

[0195] (3) Multi-source health factor extraction module

[0196] This is used to extract complementary health features from each cycle of data and form feature groups according to their source. The health factors include at least the following three categories.

[0197] The equal voltage drop discharge time characteristic submodule divides the effective discharge voltage range of 3.4V-4.2V into adjacent intervals of 0.1V, records the time difference between the voltage drop from the upper boundary to the lower boundary of the interval, and obtains a set of time characteristics.

[0198] Incremental Capacity IC Feature Submodule: Calculates the IC sequence through cross-point differential calculation and performs Gaussian smoothing. Extracts main peak-related features from the smoothed IC curve, such as peak position, peak value, and area within a fixed voltage bandwidth centered on the peak position.

[0199] Temperature-Voltage Differential DTV Feature Submodule: The DTV is calculated using cross-point differential on the temperature-voltage curve and Gaussian smoothed. The horizontal and vertical axes of the peak and valley values ​​are extracted from the smoothed DTV curve as health factors.

[0200] (4) Group-level gating feature fusion module GFF

[0201] This module is used for adaptive weighted fusion of multi-source feature groups divided by source, while suppressing redundancy and noise. It employs a group-level gating network to learn the weights of each feature group and expands these weights into feature-dimensional gating vectors through a group-to-feature mapping. The original features are then subjected to element-wise gating weighting, outputting a fused feature representation z. t , as input to the time series model.

[0202] (5) Sliding Time Window Construction Module

[0203] This is used to stack the fused features of L consecutive cycles into a sequence sample input to a time-series model to characterize the time dependence of SOH degradation.

[0204] (6) LSTM sequence modeling and SOH regression module

[0205] This module is used to model sliding time window sequences and output SOH prediction values. It takes the sequence as input to a unidirectional LSTM, extracts the hidden state corresponding to the end of the window as the sequence representation, and obtains the SOH estimation result through a fully connected regression layer.

[0206] (7) Model training, parameter management and online inference module

[0207] This system is used to train model parameters based on historical battery data during the training phase, and load the trained GFF and LSTM parameters during the inference phase. It then performs feature extraction, gating fusion, sliding window construction, and SOH output on newly arriving recurrent data. Training and testing can be divided by battery level to evaluate cross-battery generalization ability, and the model version, normalized parameters, and inference log are stored in memory.

[0208] (8) Result Output and Interface Module

[0209] It is used to output SOH estimates and optional intermediate results, and provides them to BMS control strategies, life management strategies, alarm strategies or visualization terminals via vehicle bus or network interface.

[0210] In some implementations, the aforementioned methods and systems can also be used for online health status estimation. Specifically, the vehicle-mounted terminal periodically acquires charge and discharge data of the electrochemical battery during actual operation and extracts valid cycle segments that meet the conditions according to preset rules; then, the extraction and fusion of discharge isobaric time interval features, incremental capacity features, and temperature-voltage differential features are completed in the manner described in Example 1, and the health status estimate corresponding to the current cycle is output through a sliding time window long short-term memory network model; the health status estimate can be further passed to upper-level functional modules such as range estimation, life prediction, maintenance reminders, and safety alarms. Since the discharge isobaric time interval features, incremental capacity features, and temperature-voltage differential features are utilized simultaneously, and a group-level gating fusion method is used to adaptively weight the multi-source features, the accuracy and stability of health status estimation under different degradation stages and different individual batteries can be improved.

[0211] This embodiment provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the electrochemical battery health state estimation method described in any one of Embodiments 1 to 4. The computer-readable storage medium can be a disk, flash memory, solid-state drive, or other tangible storage medium capable of storing program instructions. Since this embodiment corresponds to the implementation of the aforementioned method on a program carrier, its execution process is the same as the method steps described in Embodiment 1, and will not be repeated here.

[0212] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for estimating the health status of an electrochemical battery, characterized in that, Includes the following steps: The operation data of the electrochemical battery during the cyclic charge and discharge process is acquired, and the operation data is preprocessed to obtain effective data segments divided by cycle. The operation data includes voltage data, capacity data and temperature data. Based on the effective data segments corresponding to each preprocessed cycle, multi-source health factors are extracted. These multi-source health factors include discharge voltage drop time interval characteristics, incremental capacity characteristics, and temperature-voltage differential characteristics. The various multi-source health factors extracted in the same cycle are divided into multiple feature groups according to their source, and each feature group is uniformly processed to form the feature vector corresponding to the cycle. The feature vector is input into the group-level gated feature fusion module, which generates group-level weights based on each feature group. According to the correspondence between each feature and its respective feature group, the group-level weights are mapped to the gate coefficients corresponding to each feature. The feature vector is then weighted according to the gate coefficients to obtain the cyclically corresponding fused features. A sliding time window sequence is constructed based on the fusion features corresponding to multiple consecutive cycles arranged in chronological order. The sliding time window sequence is then input into a long short-term memory network for time-series modeling, and the corresponding electrochemical cell health state estimation results are output.

2. The method for estimating the health status of an electrochemical cell according to claim 1, characterized in that, The preprocessing includes: cyclically segmenting the running data to obtain charge / discharge data segments corresponding to each cycle; and performing at least one of the following on the cyclically segmented running data: outlier handling, missing value handling, smoothing, resampling, and unification. The preprocessed running data is used to support the extraction of discharge voltage drop time interval features, incremental capacity features, and temperature-voltage differential features, respectively, so that health factors from different sources correspond to data segments under the same cycle number.

3. The method for estimating the health status of an electrochemical battery according to claim 1, characterized in that, The extraction of the discharge equal voltage drop time interval feature includes: Within the preset effective discharge voltage range, multiple adjacent voltage intervals are divided according to the preset voltage step size; For each cycle, record the time difference on the discharge voltage curve as the voltage drops from the upper boundary to the lower boundary of each voltage range; The time difference mentioned above is used as the time interval characteristic of the discharge equal voltage drop corresponding to the cycle; The multiple time differences under the same cycle are grouped into a first type of feature group in a preset order to characterize the change in discharge sustaining time within the cycle as capacity decays and internal resistance increases. The extraction of the incremental capacity features includes: Based on capacity and voltage data in the same cycle, an incremental capacity curve is constructed using a cross-point differential method. The incremental capacity curve is smoothed to suppress noise and stabilize the curve shape; Extract at least one of the main peak position, main peak amplitude, and area of ​​the corresponding interval of the main peak from the smoothed incremental capacity curve as the incremental capacity feature corresponding to the cycle. The extracted incremental capacity features are then grouped into a second type of feature group, wherein the area of ​​the interval corresponding to the main peak is determined by the curve integral result within a preset voltage bandwidth interval centered on the main peak position.

4. The method for estimating the health status of an electrochemical battery according to claim 1, characterized in that, The extraction of the temperature-voltage differential features includes: Based on temperature and voltage data in the same cycle, a temperature-voltage differential curve is constructed using a cross-point differential method. The temperature-voltage differential curve is smoothed. Extract at least one of the peak position, peak amplitude, valley position, and valley amplitude from the smoothed temperature-voltage differential curve as the temperature-voltage differential feature corresponding to the cycle. The extracted temperature-voltage differential features were then grouped into a third type of feature group to characterize the impact of temperature and terminal voltage coupling changes on the health status during battery aging. This involves dividing multi-source health factors into multiple characteristic groups based on their origin, and then standardizing each characteristic group, including: The discharge equal voltage drop time interval characteristics are divided into the first characteristic group, the incremental capacity characteristics are divided into the second characteristic group, and the temperature-voltage differential characteristics are divided into the third characteristic group. Normalization and consistency processing are performed on each feature group separately; During the standardization process, missing values ​​are filled or removed, and redundant dimensions are suppressed to obtain the standardized feature vectors corresponding to each loop. In this standardized feature vector, each feature dimension retains its corresponding feature group identifier, which is used for subsequent group-level gating weight mapping.

5. The method for estimating the health status of an electrochemical cell according to claim 1, characterized in that, The group-level gating feature fusion includes: The feature vectors are input into a gating network, which then scores each feature group at the group level. The group-level scoring results are normalized to obtain the group-level weights corresponding to each feature group; Based on the correspondence between each feature and its corresponding feature group, the group-level weights of each feature group are mapped to the gating coefficients corresponding to each feature; The feature vector is weighted element-wise according to the gating coefficient to obtain the fused feature. The fusion features serve as input for the subsequent construction of the sliding time window sequence; The gated network includes a hidden layer and an output layer connected in sequence; The hidden layer is used to learn intermediate representations of the differences in the contribution of different feature groups to health status estimation; The output layer is used to output the group-level scoring results corresponding to each feature group; The scoring results of the output layer are normalized by softmax to obtain the group-level weights corresponding to each feature group.

6. The method for estimating the health status of an electrochemical cell according to claim 1, characterized in that, The sliding time window sequence is composed of multiple consecutive cyclic fusion features stacked in chronological order; After setting the window length, the fused features corresponding to multiple consecutive cycles are combined into a time window sample; The time window samples are input into a unidirectional long short-term memory network for sequence modeling. Based on the hidden state corresponding to the end of the window, the health status estimation result is output through the regression layer; The health status estimation results are used to characterize the health status of the electrochemical cell corresponding to the current cycle. The long short-term memory network performs gated updates on the fusion features of the inputs at each time step within the sliding time window to learn the long-term dependencies and short-term changes in the degradation process of the electrochemical battery. The health status estimation result is obtained from the hidden state at the end of the window through a fully connected regression layer; Furthermore, after obtaining the health status estimation result, the output result is smoothed before being output in real time or offline for use in battery management or life assessment.

7. The method for estimating the health status of an electrochemical cell according to claim 1, characterized in that, The method further includes model training, parameter management, and online inference steps, wherein the model training, parameter management, and online inference steps include: During the training phase, the parameters of the group-level gated feature fusion module and the long short-term memory network are trained based on historical battery data. During the inference phase, the trained model parameters are loaded, and the newly arrived recurrent data are processed sequentially to extract health factors, construct feature groups, perform group-level gated feature fusion, construct sliding time windows, and output health status. The normalization parameters, model version information, and inference logs are stored in the memory. Among them, a battery-level cross-validation strategy is used to evaluate model performance during the training phase; The battery-level cross-validation strategy includes selecting one battery dataset as the test set in turn and using the remaining battery datasets as the training set. At least one of the mean absolute error, mean absolute percentage error and root mean square error is used as the evaluation index. The model parameters for online inference are determined based on the evaluation results.

8. A health status estimation system for an electrochemical battery, characterized in that, The system includes a processor, a memory, and a program stored in the memory and executable by the processor, wherein the program, when executed by the processor, implements the electrochemical cell health state estimation method as described in any one of claims 1 to 7. The system includes: The data acquisition and preprocessing module is used to acquire the operating data of the electrochemical cell and perform preprocessing to output effective data segments divided by cycle. The multi-source health factor extraction module is used to extract discharge voltage drop time interval features, incremental capacity features, and temperature-voltage differential features based on the effective data fragments. The feature group construction module is used to divide the multi-source health factors corresponding to the same cycle into multiple feature groups according to their source, and perform unified processing to form feature vectors; The group-level gated feature fusion module is used to generate group-level weights based on each feature group, and to perform weighted fusion of the feature vectors according to the group-level weights to obtain fused features; A sliding time window construction module is used to construct a sliding time window sequence based on the fused features of multiple consecutive cycles; The health status regression module is used to perform time-series modeling on the sliding time window sequence based on a long short-term memory network and output the health status estimation results. The model training, parameter management, and online inference module is used to complete model parameter training during the training phase and load the trained model parameters during the inference phase to perform health status estimation on newly arrived recurrent data. The results output and interface module is used to output the health status estimate to the battery management system control strategy module, life management module, alarm module, or visualization terminal.

9. The electrochemical battery health status estimation system according to claim 8, characterized in that, The group-level gated feature fusion module includes a gated network and a mapping unit; The gated network is used to output the group-level weights corresponding to each feature group; The mapping unit is used to generate feature dimension gating coefficients based on the correspondence between group-level weights and feature group identifiers; The health status regression module includes a sliding time window long short-term memory network and a regression output layer, which are used to convert fused features into health status estimation results.

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

Images