Lithium ion battery health state estimation method and system

Abstract

The invention relates to the technical field of battery detection, and discloses a lithium ion battery health state estimation method and a lithium ion battery health state estimation system. Carrying out charge-discharge cross-working-condition fusion based on causal convolution and a sparse attention mechanism and correction based on a historical degradation track of the battery to obtain a health state prediction value of the lithium ion battery; if it is judged that the lithium ion battery has the high-value calibration segment in the current cycle, collecting multi-dimensional evaluation data in the current cycle and inputting the multi-dimensional evaluation data into a preset multivariate nonlinear physical mapping model to obtain a health state reference value of the lithium ion battery; and finally, carrying out dynamic fusion calibration on the health state prediction value and the health state reference value to obtain a final health state estimation value of the lithium ion battery, so that high-precision online estimation of the SOH of the lithium ion battery is realized, and the problem of accumulative drift of a pure data driving method under a complex working condition is eliminated.

Description

Technical Field

[0001] This invention relates to the field of battery testing technology, and in particular to a method and system for estimating the health status of lithium-ion batteries. Background Technology

[0002] With the continuous expansion of new energy grid connection, energy storage systems have become a key means to balance grid fluctuations and improve energy utilization efficiency. As the core energy storage carrier, the state of health (SOH) of lithium-ion batteries directly affects the safety and lifespan of the system.

[0003] Currently, data-driven methods are commonly used to estimate the health status of lithium-ion batteries. Although this method avoids complex modeling, feature extraction methods based on feature factors often extract a single health factor by reducing dimensionality, losing dynamic details in the original sampled waveform and making it difficult to sensitively detect early minor losses. Under actual operating conditions, the discharge process is easily affected by random loads and sensor noise, and time-series features are difficult to characterize the statistical distribution of battery operating characteristics.

[0004] Therefore, how to solve the problem that data-driven methods cannot capture all the underlying details and lack robustness under complex operating conditions when estimating the health status of lithium-ion batteries has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This invention provides a method and system for estimating the health status of lithium-ion batteries, which solves the problems of incomplete capture of low-level details and insufficient robustness under complex operating conditions when using data-driven methods to estimate the health status of lithium-ion batteries.

[0006] To address the aforementioned technical problems, the first aspect of this invention provides a method for estimating the state of health of a lithium-ion battery, comprising: Acquire real-time operating data of lithium-ion batteries during the full charging and discharging phases, and extract multi-dimensional operating features from the real-time operating data; The multidimensional operating features are fused across charging and discharging conditions based on causal convolution and sparse attention mechanisms to obtain charging and discharging features. The charging and discharging features are then corrected based on the battery's historical degradation trajectory to obtain the predicted health status of the lithium-ion battery. If, based on the real-time operating data, it is determined that the lithium-ion battery has a high-value calibration segment in the current cycle, then the multidimensional evaluation data of the current cycle is collected and input into a preset multivariate nonlinear physical mapping model, and the health status reference value of the lithium-ion battery is output. The final health status estimate of the lithium-ion battery is determined based on the predicted health status value and the reference health status value.

[0007] A second aspect of the present invention provides a lithium-ion battery health status estimation system, comprising: A multi-dimensional feature extraction module is used to acquire real-time operating data of lithium-ion batteries during the full charging and discharging process, and extract multi-dimensional operating features from the real-time operating data. The health status prediction module is used to perform charge and discharge cross-condition fusion based on causal convolution and sparse attention mechanism on the multi-dimensional operating features to obtain charge and discharge features, and to correct the charge and discharge features based on the battery's historical degradation trajectory to obtain the predicted health status value of the lithium-ion battery. The reference value generation module is used to collect multidimensional evaluation data of the current cycle and input it into a preset multivariate nonlinear physical mapping model if it is determined based on the real-time running data that the lithium-ion battery has a high-value calibration segment in the current cycle, and output the health status reference value of the lithium-ion battery. The dynamic fusion calibration module is used to determine the final health state estimate of the lithium-ion battery based on the predicted health state value and the health state reference value.

[0008] Compared with the prior art, the beneficial effects of the embodiments of the present invention are as follows: By leveraging the horizontal collaboration of charge and discharge data, this invention addresses the incomplete information problem caused by relying solely on single operating conditions in existing technologies, thereby improving the utilization rate of cross-stage data. Through the fusion of charge and discharge operating conditions and the introduction of historical degradation trajectories as a reference, the model can distinguish between sudden fluctuations and trend-based aging, significantly reducing the drift rate of the evaluation results. The identification of high-value calibration segments makes the data more accurate and lays the foundation for the physical anchoring stage of pure data-driven estimation. The dynamic fusion and calibration of predicted and estimated values ​​of lithium-ion battery health status ensures the robustness of SOH estimation. This invention achieves high-precision online estimation of lithium-ion battery SOH through a collaborative correction mechanism of normalized data-driven estimation and intermittent physical calibration, combined with horizontal cross-operating condition data collaboration of charge and discharge and vertical constraints of battery historical aging trajectories, while eliminating the cumulative drift problem of pure data-driven methods under complex operating conditions. Attached Figure Description

[0009] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of a lithium-ion battery health status estimation method provided in a certain embodiment of the present invention; Figure 2This is a structural diagram of a lithium-ion battery health status estimation system provided in a certain embodiment of the present invention; Figure label: Among them, 10 is the multi-dimensional feature extraction module; 20 is the health status prediction module; 30 is the reference value generation module; and 40 is the dynamic fusion calibration module. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings and examples. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0012] In the description of this application, the terms "first," "second," "third," 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," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art will be able to understand the specific meaning of the above terms in this application according to the specific circumstances.

[0013] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is merely for describing specific embodiments and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0014] In one embodiment, such as Figure 1 As shown, the first aspect of the present invention provides a method for estimating the health status of a lithium-ion battery, comprising: S1. Acquire real-time operating data of the lithium-ion battery during the full charge and discharge cycle, and extract multi-dimensional operating features from the real-time operating data; specifically, continuously collect real-time operating data of the battery during the full charge and discharge cycle at a sampling frequency of 1Hz, including: current, voltage, temperature, and cumulative ampere-hours, and perform preprocessing such as cleaning on these data to improve the accuracy of the data.

[0015] In one embodiment, extracting multi-dimensional operational features from the real-time operational data includes: Extract multi-dimensional health factors of the lithium-ion battery during the charging phase from the real-time operating data, and construct a time-series feature of the health factors based on the multi-dimensional health factors; Based on the real-time operating data, the voltage, current and temperature of the lithium-ion battery during the voltage rise range in the constant current charging stage are extracted and transformed into a two-dimensional characterization matrix with spatiotemporal correlation characteristics to obtain two-dimensional spatiotemporal features. Random load data of the lithium-ion battery during the discharge phase is collected from the real-time operating data, and a three-dimensional operating statistical distribution feature reflecting the correlation between voltage, current and temperature is established based on the random load data. The multidimensional operational features are obtained by combining the temporal characteristics of the health factors, the two-dimensional spatiotemporal characteristics, and the three-dimensional operational statistical distribution characteristics.

[0016] Specifically, based on real-time operating data, this invention extracts multi-dimensional health factors from the three stages of charging operation (fixed operating conditions and low interference) of constant current, constant voltage and relaxation. At the same time, it extracts the electrochemical characteristics in this operating condition through improved incremental capacity (IC) analysis, and constructs the time-series characteristics of health factors.

[0017] Since the current remains relatively constant during the constant current charging phase, but the voltage and temperature change over time, this invention extracts the voltage, current, and temperature of the lithium-ion battery during the voltage rise interval (e.g., from 3.2V to 3.5V, to avoid the nonlinear segments at the beginning and end of charging) from real-time operating data. The scale differences in these data are eliminated through time-domain alignment and sampling standardization. Subsequently, the time-series data of these three variables are transformed into a two-dimensional image matrix using methods such as Gram angle field (GAF) transformation or Markov transition field (MTF). Finally, the three two-dimensional matrices corresponding to voltage, current, and temperature are stacked along the channel dimension to form a two-dimensional spatiotemporal feature. This feature is equivalent to a "three-channel" image, clearly depicting the spatiotemporal evolution of voltage, current, and temperature within a specific voltage rise interval during constant current charging. Simultaneously, this feature reveals the underlying physical details of the charging waveform.

[0018] To address the random fluctuations in current, voltage, and temperature during discharge, this invention collects real-time operational data on the current, voltage, and temperature of lithium-ion batteries during the discharge phase as random load data. Based on this data, a three-dimensional histogram statistical model is established. The voltage range is divided into 20 bins, the current range into 20 bins, and the temperature range into 10 bins. The [V, I, T] values ​​of each sampling point are iterated, and the number of data points falling within each three-dimensional small cube (bin) is counted. After the statistics are completed, a three-dimensional histogram of size [20, 20, 10] is formed, representing the three-dimensional operational statistical distribution characteristics. This feature intuitively reflects the joint distribution probability of voltage, current, and temperature under random discharge conditions, characterizing the dynamic operating condition distribution law during the discharge phase and compensating for the shortcomings of single-time-point features.

[0019] Finally, the time-series characteristics of health factors during the charging phase, the two-dimensional spatiotemporal characteristics, and the three-dimensional operational statistical distribution characteristics during the discharging phase are encapsulated and combined to form multi-dimensional operational characteristics, providing a comprehensive and detailed input data foundation for subsequent estimation models.

[0020] This invention extracts different types of features from the charging and discharging stages and organically integrates these heterogeneous features to achieve a full-dimensional characterization of battery aging status, avoiding information bias. At the same time, it enhances the sensitivity of features to early signs of aging, improves the ability to model battery behavior under complex operating conditions, and provides more comprehensive, refined, and physically meaningful input information for subsequent health status estimation models.

[0021] In one embodiment, extracting multi-dimensional health factors of the lithium-ion battery during the charging phase from the real-time operational data includes: Based on the real-time operating data, the isovoltage rise time and voltage-time integral of the lithium-ion battery during the constant current charging phase, and the constant voltage phase duration and current-time integral of the lithium-ion battery during the constant voltage charging phase are extracted as basic health factors. The voltage curve of the lithium-ion battery in the relaxation stage is determined based on the real-time operating data, and the voltage curve is processed by nonlinear least squares fitting to obtain the polarization time constant and polarization impedance as relaxation health factors. Based on the real-time operating data, the capacity-voltage curve of the lithium-ion battery during the constant current charging stage is constructed, and the peak value, voltage corresponding to the peak value, skewness corresponding to the peak value, kurtosis corresponding to the peak value, and area of ​​the curve are extracted from the capacity-voltage curve as incremental capacity health factors. The multidimensional health factors are obtained by combining the basic health factors, the relaxed health factors, and the incremental capacity health factors.

[0022] Specifically, lithium batteries typically follow a preset control protocol during the charging phase, with relatively fixed operating conditions and minimal interference from load fluctuations, making it an ideal window for extracting baseline health factors. However, during constant current charging, current fluctuations are small, while voltage changes are significant. Considering the unavoidable measurement errors in the original data, this invention extracts the voltage rise time and its integral over time from real-time operating data during the constant current charging phase—a process with simple calculation logic and minimal impact from noise and initial conditions—as the basic health factor. Similarly, the constant voltage duration and the integral of current over time during the constant voltage charging phase are extracted from real-time operating data as the basic health factor. Furthermore, research has shown that the time constant during constant voltage charging can be used to predict battery state of equilibrium (SOH). Therefore, this invention can also fit the current curve during constant voltage charging using a nonlinear least squares fitting algorithm and use the resulting fitting coefficients as the basic health factor.

[0023] Furthermore, this invention introduces relaxation stage data as an important supplementary feature reflecting the electrochemical evolution state of the battery. Based on real-time operational data, it identifies the resting event after charging is completed and extracts the relaxation voltage sequence during the voltage drop process to construct a voltage curve. Since the relaxation behavior of a single cell follows a multi-order RC equivalent physical model, this invention uses a double exponential decay model to analytically fit the voltage curve, and its fitting equation is expressed as: In the formula, This represents the current terminal voltage. This is the open-circuit voltage balance value; The current at the moment before entering the relaxation phase; , These are the polarization impedances corresponding to the polarization process; , These represent the polarization time constants at different time scales.

[0024] Subsequently, the voltage curves of the observed relaxation segments were parameterized using the nonlinear least squares method. Given... and This invention can keenly capture changes in electrochemical impedance caused by the loss of active materials inside the battery. The increase in impedance has a significant physical mapping relationship with the decline in battery health. The present invention uses the fitted polarization time constant... , and polarization impedance , The relaxation health factor was extracted to describe the degree of battery aging, and this factor is also used to reflect changes in battery electrochemical impedance.

[0025] Battery incremental capacity analysis involves mathematically processing the battery constant current data (voltage-capacity curve) extracted from real-time operating data to reflect the internal electrochemical processes of the battery. The expression is as follows: In the formula, IC represents the incremental capacity; Q represents the capacity.

[0026] A common approach is to directly calculate using battery voltage and current data, and then filter the results (such as Savitzky-Golay filtering). However, numerical differentiation methods are highly susceptible to outliers and sensitive to noise. Therefore, this invention adopts a method of filtering before calculation. Specifically, smooth spline interpolation is first used to fit the capacity-voltage curve (i.e., the Q(V) curve) to obtain a spline function. Then, since the spline function is an explicit function, the incremental capacity data can be directly derived. The specific steps are as follows: The constant current stage capacity sequence is obtained by the ampere-hour integration method: Fit capacity-voltage data points using a cubic spline function. To minimize the objective function: In the formula, It is a cubic spline function; For function The second derivative; This is the penalty coefficient; , These represent the upper and lower limits of the voltage during the constant current phase.

[0027] and In each voltage range The expression is: In the formula, a j b j c j d j Let be the constant term coefficient, linear term coefficient, quadratic term coefficient, and cubic term coefficient of the cubic spline for the j-th interval.

[0028] For voltage interpolation, calculate each interpolation point. corresponding function The first derivative is used to obtain the point set of the incremental capacity curve. .

[0029] Because interpolation algorithms suffer from edge effects, the incremental capacity curve point set... Further processing is required. This invention takes into account the physical processes of the battery during the constant current phase. At the beginning and end of the phases, the battery voltage polarizes and changes rapidly; therefore, the incremental capacity data... The size is relatively small, therefore the point set of the incremental capacity curve is small. Select incremental capacity data Curve points with an absolute value greater than 1 are considered as valid data for the final feature extraction.

[0030] Features are extracted from the curve as incremental capacity health factors, such as curve peaks. Peak voltage The peak value corresponds to skewness SK, and the peak value corresponds to kurtosis Kur, satisfying the condition that the absolute value of the incremental capacity data is greater than 1 for the minimum or maximum voltage value, and the area under the IC curve, etc. The expressions for calculating skewness and kurtosis are as follows: In the formula, μ is the mean of the incremental capacity data; σ is the standard deviation of the incremental capacity data.

[0031] Finally, the three types of health factors mentioned above are combined to obtain a multi-dimensional health factor vector of the lithium-ion battery in the current cycle, which comprehensively describes the health status information of the battery in the current charging cycle.

[0032] This invention proposes a multi-level, multi-physics feature construction method for extracting health factors during the charging stage of lithium-ion batteries. By extracting health factors with clear physical meaning from four dimensions—constant current charging, constant voltage charging, relaxation stage, and incremental capacity analysis—and combining them into multi-dimensional health factors, this method can not only comprehensively capture the changes in electrochemical characteristics during battery aging, enhance the feature's ability to identify specific aging modes, and improve the feature's robustness to noise and operating condition fluctuations, but also provide richer, more refined, and robust input information for subsequent health status estimation.

[0033] S2. Perform charge-discharge cross-condition fusion based on causal convolution and sparse attention mechanism on the multi-dimensional operating features to obtain charge-discharge features, and perform correction on the charge-discharge features based on the battery's historical degradation trajectory to obtain the predicted health status value of the lithium-ion battery. In one embodiment, step S2 includes: The causal convolution integrates the temporal features of the health factors and the two-dimensional spatiotemporal features to form a comprehensive charging characterization feature; Calculate the correlation contribution between each element contained in the comprehensive charging characterization feature, so as to strengthen the information expression strongly related to battery aging through feature recombination and obtain the charging enhancement feature; The sparse attention mechanism is used to fuse and calibrate the charging enhancement features and the three-dimensional operational statistical distribution features to obtain the charging and discharging features; The charging and discharging characteristics are processed by a feedforward neural network to obtain the charging and discharging transient characteristics; Multiple historical health state estimates of the lithium-ion battery are obtained to extract a state vector reflecting the battery's historical degradation trajectory. The charge-discharge transient characteristics are then corrected using the state vector to obtain the current state characteristics. A linear mapping is performed on the current state characteristics to obtain the predicted health status value of the lithium-ion battery.

[0034] Specifically, this invention utilizes causal convolution to capture the underlying physical details in multidimensional operational features, and fuses them through a sparse attention mechanism. Combined with global trend constraints provided by long-term aging path guidance, it outputs a data-driven health status prediction. This process includes the following steps: (I) Evolution and Integration of Multidimensional Features This invention employs a model constructed using causal convolution to integrate the operational characteristics of the charging phase, thereby obtaining a unified comprehensive charging characterization feature. In one embodiment, the integration of the temporal features of the health factors and the two-dimensional spatiotemporal features through causal convolution to form the comprehensive charging characterization feature includes: The temporal trend of the health factors is extracted by constructing a long-range evolutionary correlation model based on the causal convolution, and long-term degradation features are obtained. The two-dimensional spatiotemporal feature matrix is ​​input into the local perception model constructed based on the causal convolution to extract the local physical details of the charging waveform and obtain the local waveform detail features. The long-period degradation features and the local waveform detail features are integrated to obtain the comprehensive charging characterization features.

[0035] This invention utilizes a long-range evolutionary correlation model (multi-level dilated sampling causal convolution) constructed by causal convolution to extract the temporal trend of health factors, capture their capacity degradation trajectory over long periods, and obtain long-term degradation features. This process is achieved through multi-level dilated sampling convolution operators, as shown in the following formula: In the formula, This represents the long-period degradation feature; t is the current time step; i is the index of the element within the convolution kernel; K is the kernel length; d is the inflation factor, which grows exponentially (e.g., 1, 2, 4). The input time series value at time td·i is used. The results of processing with different dilation factors are concatenated as the final input. Alternatively, the long-range evolutionary association model can also adopt a temporal convolutional network (TCN) structure, including an input layer, a causal convolutional layer, a global average pooling layer, and an output layer; wherein the causal convolutional layer is stacked with 4 dilated causal convolutional layers, with dilation factors of 1, 2, 4, and 8 respectively, and the kernel size of each layer is 3, and each layer is followed by a ReLU activation function and Dropout (0.2).

[0036] The two-dimensional spatiotemporal feature matrix is ​​then input into the local perception model for processing to extract local physical details of the charging waveform, obtaining local waveform detail features, which show the local spatiotemporal patterns of voltage, current, and temperature within a specific range during the charging process. This model uses a 2D causal convolutional network, maintaining causality in both the image height and width. However, considering that the time axis of the spatiotemporal feature matrix corresponds to the height direction of the image, causal convolution is applied along the height direction, including an input layer, a causal convolutional layer (stacked with 3 layers of 2D causal convolutional layers, each with a kernel size of [3,3] (3 in the height direction and 3 in the width direction), and causal padding along the height direction, i.e., when calculating the output of the i-th row, it only depends on the i-th row and the rows before it, with each layer followed by BatchNormalization and ReLU activation functions), global average pooling (performing global average pooling on the output of the last convolutional layer in the spatial dimensions, i.e., height and width), and an output layer.

[0037] Finally, the long-term degradation features and local waveform detail features are spliced ​​together to achieve dimensional integration and form a comprehensive characterization under charging conditions, which is the comprehensive charging characterization feature. It includes both the long-term evolution law of battery aging and the instantaneous physical details of the current charging process.

[0038] This invention addresses the problem of multi-source heterogeneous feature fusion during the charging stage of lithium-ion batteries by proposing a dual-path feature extraction and integration method based on causal convolution. Through a long-range evolutionary correlation model and a local perception model, the method specifically processes the temporal features of health factors reflecting long-term degradation trends and the two-dimensional spatiotemporal features reflecting instantaneous physical characteristics, respectively. These two features are then dimensionally integrated to form a comprehensive charging characterization feature, achieving a multi-scale refined expression of charging features. Furthermore, it fully leverages the unique advantages of causal convolution in temporal modeling and improves the targeting and efficiency of feature extraction through a divide-and-conquer strategy, providing high-quality input for subsequent feature enhancement and fusion.

[0039] (ii) Enhanced internal connections This invention strengthens the expression of key aging information by calculating the correlation contribution between elements within the comprehensive charging characterization feature, and weakens irrelevant / noisy features, making the features input to subsequent networks more representative: Using the comprehensive charging characterization feature as row vectors, an M×N feature matrix is ​​constructed (M is the feature dimension, N is the number of sampling points); the Pearson correlation coefficient matrix R (dimension M×M) of the feature matrix is ​​calculated, where each element R... ij This represents the linear correlation between the i-th feature and the j-th feature (values ​​range from -1 to 1); a significance test is performed on the correlation coefficient matrix R (with a significance level of α=0.05), retaining only significantly correlated elements with p<0.05 and filtering out weak correlations without statistical significance; based on the significant correlation matrix, the contribution score of each feature to SOH is calculated: the contribution of each feature is obtained by multiplying the absolute value of the univariate regression coefficient between the feature and SOH by the mean of the significantly correlated elements between the features; the contribution of all features is normalized, and features with higher contribution are assigned greater weights, and the original features are recombined according to the weights; residual connections are performed on the recombined features to retain the basic information of the original features, while the weights after association enhancement are superimposed, finally outputting a "denoised + enhanced" comprehensive charging feature.

[0040] (III) Cross-operating condition collaborative calibration and feature selection This invention processes the three-dimensional statistical histogram of the discharge condition through nonlinear mapping; then it introduces a cross-condition collaborative calibration mechanism, using the comprehensive charging characteristics as a reference benchmark, and achieves the fusion and calibration of charging and discharging characteristics through a sparse cross-attention mechanism, automatically identifying and retaining the key statistical feature points that are most relevant to the current state.

[0041] In one embodiment, the step of fusing and calibrating the charging enhancement features and the three-dimensional operational statistical distribution features using the sparse attention mechanism to obtain charging and discharging features includes: The charging enhancement features are mapped into a key / value vector matrix, and the three-dimensional running statistical distribution features are non-linearly mapped into a query vector matrix. Calculate the relevance score between the key / value vector matrix and the query vector matrix, and introduce a sparsity factor to retain key-value pairs whose relevance scores meet the preset requirements; The Softmax function is used to normalize the relevant scores after screening to obtain a sparse attention distribution, which is then weighted and summed with the value vector matrix in the key-value pair to obtain a sparse aggregated feature. All sparse aggregated features are then concatenated to form the charge-discharge feature.

[0042] This invention inputs the charging enhancement features into two parallel fully connected layers (without activation functions) to generate a key vector matrix K and a value vector matrix V, respectively. The three-dimensional running statistical distribution features are flattened into a one-dimensional vector, and then nonlinearly mapped through a two-layer multilayer perceptron (MLP) to obtain the query vector matrix Q.

[0043] For each query vector Calculate its relationship with all key vectors correlation score (d) k Let k be the key vector. j The dimension size is determined by introducing a sparse factor k to select key-value pairs whose relevance scores meet the preset requirements. In this embodiment, k=10, that is, for each query, only the scores corresponding to the 10 keys with the highest scores are retained, and the scores of the remaining positions are set to negative infinity (-∞), so that their weights in the subsequent Softmax are close to zero.

[0044] The filtered score matrix is ​​subjected to a Softmax operation row-wise (i.e., for each query) to obtain a sparse attention distribution, where the weights of the masked positions are 0, and the weights of the remaining positions are summed to 1. The attention distribution is then weighted and summed with the value matrix to obtain a sparse aggregated feature vector. Concatenating all the sparse aggregated features yields the charging and discharging features. Alternatively, feature representation can be optimized using residual connections and normalization.

[0045] This invention employs a sparse attention mechanism to selectively focus on the most critical parts of charging features for the current discharging features, achieving efficient fusion and mutual calibration of features across operating conditions (charge and discharge). It introduces a sparsity factor to reduce computational complexity, avoiding the high computational overhead of the full attention mechanism, while retaining the most important information interactions, making the model more suitable for real-time embedded systems. By using charging enhancement features as keys and discharging features as queries, it reflects the retrieval of information related to charging features guided by the discharging operating condition, thereby achieving calibration of charging features by discharging features. The final charge and discharge features integrate complementary information from both operating conditions, providing a more comprehensive input for subsequent health status prediction.

[0046] (iv) Feedforward unit processing The fused charge and discharge features are input into a feedforward neural network (using a two-layer fully connected network) to perform nonlinear transformation and dimension unification of the features, generating charge and discharge transient features. These features further refine the dynamic response information in the current charge and discharge cycle, preparing for subsequent time-series correction.

[0047] (v) Global constraints guided by long-term aging paths This invention introduces the battery's historical degradation trajectory as a global logical constraint, maps the fusion features at the current moment to the aging trajectory of the entire life cycle, and uses historical evolution trends to correct transient evaluation biases, ensuring the continuity and logical consistency of the estimation results on the time scale.

[0048] In one embodiment, the step of extracting a state vector reflecting the battery's historical degradation trajectory and using the state vector to correct the charge-discharge transient features to obtain the current state features includes: Convolutional encoding is performed on each of the historical health state estimates to obtain a state vector reflecting the historical degradation trajectory of the battery. Using the state vector as the query vector and the charging enhancement feature as the key / value vector, the mapping weight of the historical trend to the current feature is calculated. The key / value vector is weighted and aggregated based on the mapping weights to obtain trend correlation features that reflect the battery under long-term aging background; The current state features are obtained by nonlinearly fusing the trend correlation features and the charging / discharging transient features.

[0049] This invention first collects the SOH (State of Health) estimates of lithium-ion batteries over N historical cycles and inputs them into a one-dimensional convolutional encoder (a three-layer causal convolutional network with global max pooling) for convolutional encoding to extract a state vector reflecting the long-term aging trend of the battery. Then, using the state vector as the query vector and charging enhancement features as the source of the key and value vectors, the mapping weights of historical evolution trends to current features are calculated, i.e., the dot product similarity between the query vector and the key vector is calculated to obtain a correlation score matrix, which is then Softmax normalized to obtain a mapping weight vector. Based on the mapping weights, the value vectors are weighted and aggregated to obtain trend-related features reflecting the battery's long-term aging background. These features represent the aggregation of the most noteworthy feature information during the charging phase against the backdrop of historical degradation trends. Finally, the trend-related features and charge / discharge transient features are globally averaged and concatenated before being input into a fully connected network with a Sigmoid activation layer for nonlinear fusion, resulting in current state features that deeply integrate the long-term degradation trend information of the battery over multiple past cycles and the charge / discharge transient response information of the current cycle.

[0050] This invention addresses the time-series dependency problem in lithium-ion battery health state estimation by proposing a correction method based on historical degradation trajectories. The method encodes historical health state estimates into state vectors, which are then used as queries to interact with charging enhancement features. Finally, the extracted trend-related features are nonlinearly fused with current charge / discharge transient features to form the current state feature. This achieves dynamic correction of the current state by historical trends, enhancing the continuity and stability of the estimation. Furthermore, an attention mechanism enables refined interaction between historical trends and charging features, and nonlinear fusion improves the expressive power and adaptability of the features. Ultimately, this results in a current state feature that combines historical memory and real-time response, laying the foundation for accurate prediction.

[0051] (vi) Linear mapping output The fused current state features are input into a linear mapping unit (a fully connected layer without an activation function), and the output is a pure data-driven prediction of the health status of the lithium-ion battery in the current cycle.

[0052] This invention addresses the problem of multi-source feature fusion and temporal modeling in lithium-ion battery health state estimation. It proposes a multi-level, multi-stage information processing workflow: integrating charging stage features through causal convolution, strengthening key information through an attention mechanism, fusing charging and discharging features through sparse attention, and correcting them by incorporating historical degradation trajectories. The final output is a high-precision health state prediction, achieving deep integration and enhancement of features within the charging stage, as well as efficient fusion and mutual calibration of features across charging and discharging conditions. Simultaneously, the introduction of historical degradation trajectories for temporal correction improves the continuity and trend consistency of the prediction. This decomposes the complex feature processing into multiple steps with clearly defined functions, forming a hierarchical feature processing pipeline that enhances the interpretability and modularity of the model.

[0053] S3. If it is determined based on the real-time operating data that the lithium-ion battery has a high-value calibration segment in the current cycle, then the multidimensional evaluation data of the current cycle is collected and input into the preset multivariate nonlinear physical mapping model, and the health status reference value of the lithium-ion battery is output. In one embodiment, step S3 includes: Based on the real-time operating data, when it is determined that the current cycle is in a low current constant current environment or a complete voltage monitoring environment, it is determined that the lithium-ion battery has a high-value calibration segment in the current cycle; the low current constant current environment is when the current rate is continuously within a preset range and the current fluctuation rate is lower than a preset fluctuation threshold; the complete voltage monitoring environment is when the absolute value of the voltage change rate is continuously within the voltage plateau period and the duration covers a preset complete voltage window. The average current ratio and the peak value of the curve are extracted from the real-time operating data and combined with the ambient temperature of the lithium-ion battery to form the multidimensional evaluation data. The multidimensional evaluation data is input into the multivariate nonlinear physical mapping model for processing, and the health status reference value of the lithium-ion battery is output.

[0054] Specifically, this invention identifies high-value calibration segments of lithium-ion batteries based on real-time operating data. Only when the current cycle meets any of the following high physical confidence conditions is the corresponding real-time operating data determined to contain a high-value calibration segment, and the current data stream is marked as "calibrable." Otherwise, it is "uncalibrable": a low-current constant-current environment, i.e., the current rate remains within a preset low-rate range (e.g., 0.1C to 0.2C) for a preset time (at least 10 minutes), and the current fluctuation rate is lower than a preset threshold (preferably 5%), with weak polarization effect under low current, and the observed data closely approximating the intrinsic change in battery electromotive force; or complete voltage plateau monitoring, i.e., combined with SOC range determination, the absolute value of the voltage change rate remains within a stable low value range (e.g., 0.1mV / s) and is in the voltage plateau period (e.g., [3.25V, 3.35V]), and the segment covers a preset complete voltage window (from the start of the voltage entering the plateau period to the end of the voltage leaving the plateau period, covering the entire plateau period range), with high accuracy in aging feature mapping. If the collected real-time operating data is not calibrable, the health status prediction value is subjected to reasonable constraints (such as limiting the fluctuation range of the single SOH estimate based on the physical laws of battery aging) and then used as the final health status estimate of the lithium-ion battery for smooth output, so as to ensure the physical rationality of the result.

[0055] Identify and extract the peak values ​​of the IC curve from the cached constant current charging segments. This is used as a physical health indicator reflecting battery capacity degradation, and compared with the current average current rating calculated based on real-time operating data. I and ambient temperature T The multidimensional evaluation data is input into a pre-set multivariate nonlinear physical mapping model for processing, outputting a reference value for the health status of the lithium-ion battery. This multivariate nonlinear physical mapping model is fitted based on the full life-cycle degradation data of the same batch of batteries under controlled experimental conditions, and is expressed by the following formula: In the formula, These are reference values ​​for health status. f For a specific battery system, a multivariate nonlinear mapping operator can be constructed based on the Arrhenius formula.

[0056] This invention addresses the online calibration problem in lithium-ion battery health status estimation by proposing a physical model calibration method based on high-value calibration segments. This method intelligently identifies specific segments in the charging process (low-current constant-current environment or full voltage monitoring environment), collects key multidimensional evaluation data, and inputs it into a multivariate nonlinear physical mapping model. Finally, it outputs a high-precision health status reference value, achieving low-cost, high-precision online calibration opportunity identification. This improves the adaptability and accuracy of the physical model, leverages its advantages under specific conditions, and compensates for the shortcomings of data-driven models.

[0057] S4. Determine the final health status estimate of the lithium-ion battery based on the predicted health status value and the reference health status value; In one embodiment, step S4 includes: Calculate the absolute residual between the predicted health status value and the reference health status value; When the absolute residual exceeds a preset threshold, the health status reference value is output as the final health status estimate of the lithium-ion battery. When the absolute residual does not exceed the preset threshold, the weighted fusion value of the predicted health status value and the reference health status value is output as the final health status estimate of the lithium-ion battery.

[0058] Specifically, this invention compares the data-driven health status prediction value with the physical reference value of the health status, and calculates the absolute residual between the two as the criterion for determining model drift. Based on the relationship between the absolute residual and a preset threshold, and combined with the operating condition confidence level, it performs case-by-case processing: when the residual does not exceed the preset threshold, the model has no obvious drift, and the weighted fusion value of the health status prediction value and the health status reference value is output as the final health status estimate of the lithium-ion battery, so as to achieve a smooth fusion of the data-driven result and the physical reference value. The weight coefficients are preset according to the operating condition confidence level. When the absolute residual exceeds the preset threshold, it is determined that the model has drifted, and the physical reference value—health status reference value—is forcibly used as the final health status estimate of the lithium-ion battery. This value is then used to reverse-correct the hidden state of the deep learning model to achieve adaptive fine-tuning. In addition, if the absolute residual exceeds a preset outlier threshold (the outlier threshold is greater than the preset threshold), it is determined that the model has drifted significantly. The weight coefficient values ​​are reduced (e.g., reduced by 10%) to increase the weight of the physical reference value, and the model is fine-tuned.

[0059] This invention addresses the problem of multi-source information fusion in lithium-ion battery health state estimation by proposing an adaptive fusion strategy based on residual judgment. The strategy calculates the absolute residual between the predicted value of the data-driven model and the reference value of the physical model, and determines whether to use a single reference value or a weighted fusion value as the final output based on the magnitude of the residual. This achieves complementary advantages and safe switching between data-driven and physical models, establishes an error monitoring mechanism, improves the robustness of the estimation system, avoids invalid fusion, reduces the risk of introducing erroneous information, and achieves a balance between smooth transition and accuracy improvement.

[0060] This application addresses the issues of incomplete low-level detail capture and insufficient robustness under complex operating conditions when using data-driven methods to estimate the state of health (SOH) of lithium-ion batteries. It proposes a new SOH estimation method that leverages the horizontal collaboration of charge and discharge data to overcome the incomplete information problem caused by relying solely on single operating conditions in existing technologies, thus improving cross-stage data utilization. By fusing charge and discharge operating conditions and introducing historical degradation trajectories as a reference, the model can distinguish between sudden fluctuations and trend-based aging, significantly reducing the drift rate of the evaluation results. The identification of high-value calibration segments makes the data more accurate and lays the foundation for the physical anchoring of pure data-driven estimation. Dynamic fusion and calibration of the predicted and estimated values ​​of the lithium-ion battery SOH ensures the robustness of the SOH estimation. This invention achieves high-precision online SOH estimation of lithium-ion batteries through a collaborative correction mechanism of normalized data-driven estimation and intermittent physical calibration, combined with horizontal cross-operating condition data collaboration of charge and discharge and vertical constraints of battery historical aging trajectories, while eliminating the cumulative drift problem of pure data-driven methods under complex operating conditions.

[0061] It should be noted that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order requirement for the execution of these steps, and they can be executed in other orders.

[0062] In another embodiment, such as Figure 2 As shown, a second aspect of the present invention provides a lithium-ion battery health status estimation system, comprising: The multidimensional feature extraction module 10 is used to acquire real-time operating data of the lithium-ion battery during the full charging and discharging process, and extract multidimensional operating features from the real-time operating data. The health status prediction module 20 is used to perform charge and discharge cross-condition fusion based on causal convolution and sparse attention mechanism on the multi-dimensional operating features to obtain charge and discharge features, and to correct the charge and discharge features based on the battery's historical degradation trajectory to obtain the predicted health status value of the lithium-ion battery. The reference value generation module 30 is used to collect multidimensional evaluation data of the current cycle and input it into a preset multivariate nonlinear physical mapping model if it is determined based on the real-time running data that the lithium-ion battery has a high-value calibration segment in the current cycle, and output the health status reference value of the lithium-ion battery. The dynamic fusion calibration module 40 is used to determine the final health state estimate of the lithium-ion battery based on the predicted health state value and the health state reference value.

[0063] It should be noted that each module in the aforementioned lithium-ion battery health state estimation system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module. For specific limitations regarding a lithium-ion battery health state estimation system, please refer to the limitations of a lithium-ion battery health state estimation method described above; both have the same function and role, and will not be repeated here.

[0064] In summary, this invention relates to the field of battery testing technology and discloses a method and system for estimating the state of health (SOH) of lithium-ion batteries. By extracting multidimensional operational characteristics of the lithium-ion battery in real time across all charging and discharging conditions, and performing cross-condition fusion based on causal convolution and sparse attention mechanisms, as well as correction based on the battery's historical degradation trajectory, a predicted SOH value for the lithium-ion battery is obtained. If a high-value calibration segment is determined to exist in the current cycle, multidimensional evaluation data within the current cycle is collected and input into a pre-set multivariate nonlinear physical mapping model to obtain a reference SOH value for the lithium-ion battery. Finally, the predicted SOH value and the reference SOH value are dynamically fused and calibrated to obtain the final estimated SOH value for the lithium-ion battery. This achieves high-precision online estimation of the SOH of lithium-ion batteries and eliminates the cumulative drift problem of pure data-driven methods under complex operating conditions.

[0065] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0066] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this invention, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A method for estimating the state of health of a lithium-ion battery, characterized in that, include: Acquire real-time operating data of lithium-ion batteries during the full charging and discharging phases, and extract multi-dimensional operating features from the real-time operating data; The multidimensional operating features are fused across charging and discharging conditions based on causal convolution and sparse attention mechanisms to obtain charging and discharging features. The charging and discharging features are then corrected based on the battery's historical degradation trajectory to obtain the predicted health status of the lithium-ion battery. If, based on the real-time operating data, it is determined that the lithium-ion battery has a high-value calibration segment in the current cycle, then the multidimensional evaluation data of the current cycle is collected and input into a preset multivariate nonlinear physical mapping model, and the health status reference value of the lithium-ion battery is output. The final health status estimate of the lithium-ion battery is determined based on the predicted health status value and the reference health status value.

2. The method for estimating the health status of a lithium-ion battery according to claim 1, characterized in that, The extraction of multi-dimensional operational features from the real-time operational data includes: Extract multi-dimensional health factors of the lithium-ion battery during the charging phase from the real-time operating data, and construct a time-series feature of the health factors based on the multi-dimensional health factors; Based on the real-time operating data, the voltage, current and temperature of the lithium-ion battery during the voltage rise range in the constant current charging stage are extracted and transformed into a two-dimensional characterization matrix with spatiotemporal correlation characteristics to obtain two-dimensional spatiotemporal features. Random load data of the lithium-ion battery during the discharge phase is collected from the real-time operating data, and a three-dimensional operating statistical distribution feature reflecting the correlation between voltage, current and temperature is established based on the random load data. The multidimensional operational features are obtained by combining the temporal characteristics of the health factors, the two-dimensional spatiotemporal characteristics, and the three-dimensional operational statistical distribution characteristics.

3. The method for estimating the health status of a lithium-ion battery according to claim 2, characterized in that, The extraction of multi-dimensional health factors of the lithium-ion battery during the charging phase from the real-time operating data includes: Based on the real-time operating data, the isovoltage rise time and voltage-time integral of the lithium-ion battery during the constant current charging phase, and the constant voltage phase duration and current-time integral of the lithium-ion battery during the constant voltage charging phase are extracted as basic health factors. The voltage curve of the lithium-ion battery in the relaxation stage is determined based on the real-time operating data, and the voltage curve is processed by nonlinear least squares fitting to obtain the polarization time constant and polarization impedance as relaxation health factors. Based on the real-time operating data, the capacity-voltage curve of the lithium-ion battery during the constant current charging stage is constructed, and the peak value, voltage corresponding to the peak value, skewness corresponding to the peak value, kurtosis corresponding to the peak value, and area of ​​the curve are extracted from the capacity-voltage curve as incremental capacity health factors. The multidimensional health factors are obtained by combining the basic health factors, the relaxed health factors, and the incremental capacity health factors.

4. The method for estimating the health status of a lithium-ion battery according to claim 2, characterized in that, The process involves fusing the multi-dimensional operational features across charging and discharging conditions using causal convolution and sparse attention mechanisms to obtain charging and discharging features, and then correcting these features based on the battery's historical degradation trajectory to obtain a predicted health status value for the lithium-ion battery. This includes: The causal convolution integrates the temporal features of the health factors and the two-dimensional spatiotemporal features to form a comprehensive charging characterization feature; Calculate the correlation contribution between each element contained in the comprehensive charging characterization feature, so as to strengthen the information expression strongly related to battery aging through feature recombination and obtain the charging enhancement feature; The sparse attention mechanism is used to fuse and calibrate the charging enhancement features and the three-dimensional operational statistical distribution features to obtain the charging and discharging features; The charging and discharging characteristics are processed by a feedforward neural network to obtain the charging and discharging transient characteristics; Multiple historical health state estimates of the lithium-ion battery are obtained to extract a state vector reflecting the battery's historical degradation trajectory. The charge-discharge transient characteristics are then corrected using the state vector to obtain the current state characteristics. A linear mapping is performed on the current state characteristics to obtain the predicted health status value of the lithium-ion battery.

5. The method for estimating the health status of a lithium-ion battery according to claim 4, characterized in that, The integration of the temporal features of the health factors and the two-dimensional spatiotemporal features through the causal convolution to form a comprehensive charging characterization feature includes: The long-term evolutionary correlation model based on the causal convolution is used to extract the temporal trend of the health factors to obtain long-term degradation features. The two-dimensional spatiotemporal feature matrix is ​​input into the local perception model constructed based on the causal convolution to extract the local physical details of the charging waveform and obtain the local waveform detail features. The long-period degradation features and the local waveform detail features are integrated to obtain the comprehensive charging characterization features.

6. The method for estimating the health status of a lithium-ion battery according to claim 4, characterized in that, The sparse attention mechanism is used to fuse and calibrate the charging enhancement features and the three-dimensional operational statistical distribution features to obtain charging and discharging features, including: The charging enhancement features are mapped into a key / value vector matrix, and the three-dimensional operational statistical distribution features are non-linearly mapped into a query vector matrix. Calculate the relevance score between the key / value vector matrix and the query vector matrix, and introduce a sparsity factor to retain key-value pairs whose relevance scores meet the preset requirements; The Softmax function is used to normalize the relevant scores after screening to obtain a sparse attention distribution, which is then weighted and summed with the value vector matrix in the key-value pair to obtain a sparse aggregated feature. All sparse aggregated features are then concatenated to form the charge-discharge feature.

7. The method for estimating the health status of a lithium-ion battery according to claim 4, characterized in that, The process involves extracting a state vector reflecting the battery's historical degradation trajectory and using this state vector to correct the charge-discharge transient characteristics to obtain the current state characteristics, including: Convolutional encoding is performed on each of the historical health state estimates to obtain a state vector reflecting the historical degradation trajectory of the battery. Using the state vector as the query vector and the charging enhancement feature as the key / value vector, the mapping weight of the historical trend to the current feature is calculated. The key / value vector is weighted and aggregated based on the mapping weights to obtain trend correlation features that reflect the battery under long-term aging background; The current state features are obtained by nonlinearly fusing the trend correlation features and the charging / discharging transient features.

8. The method for estimating the health status of a lithium-ion battery according to claim 3, characterized in that, If, based on the real-time operating data, it is determined that the lithium-ion battery has a high-value calibration segment in the current cycle, then the multidimensional evaluation data of the current cycle is collected and input into a preset multivariate nonlinear physical mapping model, and the health status reference value of the lithium-ion battery is output, including: Based on the real-time operating data, when it is determined that the current cycle is in a low current constant current environment or a complete voltage monitoring environment, it is determined that the lithium-ion battery has a high-value calibration segment in the current cycle; the low current constant current environment is when the current rate is continuously within a preset range and the current fluctuation rate is lower than a preset fluctuation threshold; the complete voltage monitoring environment is when the absolute value of the voltage change rate is continuously within the voltage plateau period and the duration covers a preset complete voltage window. The average current ratio and the peak value of the curve are extracted from the real-time operating data and combined with the ambient temperature of the lithium-ion battery to form the multidimensional evaluation data. The multidimensional evaluation data is input into the multivariate nonlinear physical mapping model for processing, and the health status reference value of the lithium-ion battery is output.

9. The method for estimating the health status of a lithium-ion battery according to claim 1, characterized in that, Determining the final health status estimate of the lithium-ion battery based on the predicted health status value and the reference health status value includes: Calculate the absolute residual between the predicted health status value and the reference health status value; When the absolute residual exceeds a preset threshold, the health status reference value is output as the final health status estimate of the lithium-ion battery. When the absolute residual does not exceed the preset threshold, the weighted fusion value of the predicted health status value and the reference health status value is output as the final health status estimate of the lithium-ion battery.

10. A lithium-ion battery health status estimation system, characterized in that, include: A multi-dimensional feature extraction module is used to acquire real-time operating data of lithium-ion batteries during the full charging and discharging process, and extract multi-dimensional operating features from the real-time operating data. The health status prediction module is used to perform charge and discharge cross-condition fusion based on causal convolution and sparse attention mechanism on the multi-dimensional operating features to obtain charge and discharge features, and to correct the charge and discharge features based on the battery's historical degradation trajectory to obtain the predicted health status value of the lithium-ion battery. The reference value generation module is used to collect multidimensional evaluation data of the current cycle and input it into a preset multivariate nonlinear physical mapping model if it is determined based on the real-time running data that the lithium-ion battery has a high-value calibration segment in the current cycle, and output the health status reference value of the lithium-ion battery. The dynamic fusion calibration module is used to determine the final health state estimate of the lithium-ion battery based on the predicted health state value and the health state reference value.

Images