Method and system for testing and predicting capacity attenuation of power battery

Abstract

The invention discloses a power battery capacity attenuation test and prediction method and system, and relates to the technical field of battery capacities, and the method comprises the steps: constructing a multi-dimensional state characterization system through fusing electrochemical impedance related characteristics, internal physical field characteristics and macroscopic cycle data, thereby achieving the more precise fitting and prediction of a complex attenuation mode; through a dual-drive self-adaptive test decision mechanism, a time-consuming information enrichment test is triggered only when the model prediction uncertainty is high or the internal state is abnormal, so that the optimal balance between the test efficiency and the model performance maintenance is realized; the model not only outputs capacity prediction, but also can qualitatively or semi-quantitatively reveal dominant attenuation mechanisms of different aging stages through internal feature importance analysis; and meanwhile, the abnormal state of the battery can be found earlier than macroscopic capacity fading based on real-time monitoring of the abnormal degree of the internal physical field characteristics, and a prospective basis is provided for safety early warning and maintenance decision making.

Description

Technical Field

[0001] This application relates to the field of battery capacity technology, specifically to a method and system for testing and predicting the capacity degradation of power batteries. Background Technology

[0002] With the rapid development of electric vehicles and energy storage systems, the life prediction and health management of lithium-ion batteries have become key challenges. Traditional capacity degradation prediction methods mainly rely on empirical model fitting or simple machine learning modeling based on macroscopic cycling data (such as discharge capacity and cycle number). Their prediction accuracy and robustness are limited, especially when dealing with complex aging paths and sudden failures.

[0003] Existing technology, such as the invention application patent with publication number CN117686933A, discloses a battery capacity degradation analysis method and system, relating to the field of battery capacity analysis. This includes: obtaining the non-polarized charge-discharge curves of an aged battery and a non-degraded battery; obtaining the differential voltage curves of the aged battery and a fresh battery based on these curves; plotting and aligning the differential voltage curves of the aged and non-degraded batteries on the same coordinate system; and determining the degradation mode and degradation amount of the aged battery based on the correspondence between the degradation mode and the differential voltage. This correspondence is used to quantitatively analyze the degradation mode and degradation amount of the aged battery by comparing the full-cell differential voltage curve of the non-degraded battery with the full-cell differential voltage curve under the degradation mode. This invention, by applying the differential voltage curve, can determine the specific degradation mode and accurately quantify it.

[0004] Regarding the above-mentioned solutions, the inventors of this application have found that the above-mentioned technologies have at least the following technical problems: 1. Currently, there is a lack of systematic acquisition and construction of multi-source heterogeneous characteristic data, which cannot lay a comprehensive and in-depth data foundation for battery health status assessment; 2. The limitations of traditional methods relying on single macroscopic performance data (such as capacity) have not been broken, and the external electrochemical response, internal physical field state and macroscopic cycle performance have not been integrated across scales; 3. The ability of subsequent models to learn complex aging laws from data cannot be enhanced, and key data support cannot be provided for solving the problem of multi-factor coupling and nonlinearity of battery degradation.

[0005] 2. Currently, there is a lack of a dual-drive adaptive testing decision-making mechanism, which hinders efficient allocation of testing resources and adaptive management of model risks. The lack of "data-driven" decision-making prevents the quantification of model prediction uncertainty and the intelligent triggering of costly information enrichment tests when model confidence is insufficient. Consequently, it fails to acquire the most effective "valuable data" for improving prediction accuracy with the fewest possible tests, thus failing to improve testing economics. The lack of "mechanism-driven" decision-making also prevents timely intervention at critical junctures where potential safety hazards or aging mechanisms of the battery occur, failing to enhance the method's early warning and safety assurance capabilities by monitoring abnormal abrupt changes in internal physical field characteristics. Summary of the Invention

[0006] To address the aforementioned technical shortcomings, the purpose of this application is to provide a method and system for testing and predicting the capacity degradation of power batteries.

[0007] To solve the above-mentioned technical problems, this application adopts the following technical solution: In the first aspect, this application provides a method for testing and predicting the capacity degradation of a power battery, which includes the following steps: S1, acquiring multi-source feature data of the battery and integrating it into a structured feature dataset.

[0008] S2. Based on the multi-source feature data, construct and train a capacity decay prediction model.

[0009] S3. During the battery aging process, execute dual-drive adaptive test decisions.

[0010] S4. When the dual-drive adaptive test decision is triggered, update the capacity decay prediction model.

[0011] S5. Based on the updated capacity decay model, output prediction results and analyze the battery capacity decay mechanism.

[0012] Preferably, the multi-source feature data includes the electrochemical response characteristics and internal physical field characteristics of the battery.

[0013] Preferably, the acquisition of multi-source feature data of the battery and integration into a structured feature dataset includes: A1. Applying a current excitation containing several frequency components to the target battery based on a preset excitation current waveform, and simultaneously acquiring its terminal voltage response, calculating the battery impedance value at each frequency through a frequency domain signal processing algorithm to obtain the battery impedance spectrum; calculating relevant parameters of the battery state through a feature extraction algorithm to obtain an electrochemical response feature vector.

[0014] A2. During the target battery aging test, physical field parameters characterizing the internal state of the battery are obtained to obtain the internal physical field feature vector.

[0015] A3. At the end of each charge-discharge cycle of the target battery, record the current cycle number and measure its actual discharge capacity to obtain a macroscopic performance data sequence.

[0016] A4. Based on a predetermined standardization method, the electrochemical response feature vector, the internal physical field feature vector, and the macroscopic performance data sequence are respectively processed to be dimensionless, and then combined into a structured feature dataset through feature fusion operation.

[0017] Preferably, the step of constructing and training the capacity decay prediction model based on the multi-source feature data includes: defining the number of cycles, historical capacity sequence, electrochemical response characteristics of the battery, and internal physical field characteristics as input feature sets based on the structured feature dataset, and defining the current discharge capacity as the output target, so as to constitute the model training samples.

[0018] Using the input feature set and the output target as training data, the parameters of the preset gradient boosting decision tree algorithm are initialized to obtain the initialized gradient boosting decision tree model. The initialized gradient boosting decision tree model is trained based on the iterative additive modeling process of the gradient boosting decision tree algorithm, including calculating residuals, fitting a new decision tree, and updating the model, so as to obtain the finally trained gradient boosting decision tree model as the initial capacity decay prediction model.

[0019] Preferably, the dual-drive adaptive test decision includes a first drive and a second drive, wherein the first drive is based on the prediction uncertainty of the capacity decay prediction model, and the second drive is based on the real-time anomaly degree of the internal physical field characteristics.

[0020] Preferably, the first driving force is based on the prediction uncertainty of the capacity decay prediction model, and the second driving force is based on the real-time change anomaly degree of the internal physical field characteristics, including: B1. At the current aging cycle time of the battery, obtaining the input feature set of the current prediction model, the input feature set including the current cycle number, historical capacity sequence, electrochemical response feature vector at the current time, and internal physical field feature vector; processing the input feature set based on the current prediction model, predicting future critical lifetime nodes, and then processing the predicted values ​​of future critical lifetime nodes to calculate the prediction uncertainty metric.

[0021] B2. Based on the internal physical field feature vectors at the current and historical times, perform time series analysis on the key parameters to calculate the internal physical field anomaly index.

[0022] Preferably, the execution of dual-drive adaptive test decision includes: setting a first decision threshold and a second decision threshold; performing logical judgment on the prediction uncertainty measure and the internal physical field anomaly index; comparing the prediction uncertainty measure with the first decision threshold, and simultaneously comparing the internal physical field anomaly index with the second decision threshold; when "the prediction uncertainty measure is greater than the first decision threshold" or "the internal physical field anomaly index is greater than the second decision threshold", generating an instruction to trigger the next information enrichment test, determining that new high-dimensional test data needs to be added to reduce model uncertainty, and confirming physical field anomalies.

[0023] Preferably, updating the capacity decay prediction model when the dual-drive adaptive test decision is triggered includes: when an instruction to trigger the next information enrichment test is generated, based on the current battery cycle state, performing an information enrichment test on the battery in a subsequent preset charge-discharge cycle to collect new electrochemical response data and new internal physical field data, so as to obtain a new electrochemical response feature vector and a new internal physical field feature vector; and obtaining the new actual discharge capacity value corresponding to this information enrichment test; and performing feature fusion of the new electrochemical response feature vector, the new internal physical field feature vector and the new actual discharge capacity value to form a new data point.

[0024] The newly added data points are added to the structured feature dataset to obtain the updated feature dataset. Incremental learning is then performed on the constructed initial capacity decay prediction model to obtain the updated capacity decay prediction model.

[0025] Preferably, the step of outputting prediction results and analyzing battery capacity decay mechanisms based on the updated capacity decay model includes: using the updated capacity decay prediction model as the current model based on the updated capacity decay prediction model and the current cycle state, returning to the execution of the dual-drive adaptive test decision process, and performing the next round of monitoring and decision-making for subsequent battery aging cycles to form a closed-loop loop path of model iteration and test execution; based on the current cycle state and preset iteration termination conditions, performing a termination judgment on the execution of the closed-loop loop path, and terminating the iteration when the battery capacity decays to the end-of-life threshold or the number of cycles reaches the preset total test cycle to obtain the iteration termination judgment result; when the iteration terminates, using the latest capacity decay prediction model at the time of termination, calculating and outputting a high-confidence capacity decay curve, the remaining service life prediction value and its confidence interval, and combining the feature importance analysis results of the latest capacity decay prediction model, performing a qualitative or semi-quantitative analysis of the dominant mechanism of battery decay to obtain a comprehensive battery life assessment report.

[0026] In a second aspect, this application provides a system for testing and predicting the capacity degradation of a power battery, comprising: preferably, a structured feature dataset generation module for acquiring multi-source feature data of the battery and integrating it into a structured feature dataset.

[0027] The capacity decay prediction model construction module constructs and trains a capacity decay prediction model based on the multi-source feature data.

[0028] The decision module is used to perform dual-drive adaptive test decisions during battery aging.

[0029] The model update module is used to update the capacity decay prediction model when the dual-drive adaptive test decision is triggered.

[0030] The results output module outputs prediction results and battery capacity decay mechanism analysis based on the updated capacity decay model.

[0031] The beneficial effects of this application are as follows: 1. The power battery capacity degradation testing and prediction method and system provided in this application, by integrating electrochemical impedance related characteristics, internal physical field characteristics and macroscopic cycling data, constructs a multi-dimensional state characterization system, thereby achieving more accurate fitting and prediction of complex degradation modes; through the "dual-drive adaptive test decision" mechanism, the time-consuming information enrichment test is triggered only when the model prediction uncertainty is high or the internal state is abnormal, achieving the best balance between test efficiency and model performance maintenance; the model not only outputs capacity prediction, but its internal feature importance analysis can qualitatively or semi-quantitatively reveal the dominant degradation mechanism of different aging stages; at the same time, based on the real-time monitoring of the anomaly degree of internal physical field characteristics, the abnormal state of the battery can be detected earlier than the macroscopic capacity degradation, providing a forward-looking basis for safety warning and maintenance decision-making.

[0032] 2. This application systematically acquires and constructs multi-source heterogeneous feature data, laying a comprehensive and in-depth data foundation for battery health status assessment. Its core benefit lies in breaking the limitations of traditional methods that rely on single macroscopic performance data (such as capacity), and for the first time, fusing external electrochemical response, internal physical field state, and macroscopic cycle performance across scales. This fusion can comprehensively characterize the battery's aging state from multiple dimensions of "signal-physics-performance," especially with the introduction of internal physical field characteristics, which allows for the quantitative characterization of previously unobservable internal thermal and mechanical state changes. The constructed structured dataset includes richer information on degradation correlations and clues to potential failure mechanisms, significantly enhancing the ability of subsequent models to learn complex aging patterns from the data, and providing crucial data support for solving the challenges of multi-factor coupling and nonlinearity in battery degradation.

[0033] 3. The dual-drive adaptive testing decision-making mechanism proposed in this application achieves efficient allocation of testing resources and adaptive management of model risks. Its benefits are reflected in two aspects: First, "data-driven" decision-making, by quantifying the uncertainty of model predictions, intelligently triggers costly information enrichment tests when model confidence is insufficient, thereby obtaining the most effective "valuable data" for improving prediction accuracy with the fewest number of tests, greatly improving testing economy. Second, "mechanism-driven" decision-making, by monitoring abnormal changes in internal physical field characteristics, can intervene in a timely manner at critical nodes where potential safety hazards or aging mechanisms of the battery occur, enhancing the method's early warning and safety assurance capabilities. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart illustrating the steps involved in implementing the method described in this application.

[0036] Figure 2 This is a schematic diagram of the system structure connection of this application. Detailed Implementation

[0037] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0038] Please see Figure 1 As shown, this application provides a method for testing and predicting the capacity degradation of a power battery in a first aspect, including: S1, acquiring multi-source feature data of the battery and integrating it into a structured feature dataset.

[0039] In one specific instance, the multi-source feature data includes the electrochemical response characteristics and internal physical field characteristics of the battery.

[0040] In a specific example, the acquisition of multi-source feature data of the battery and integration into a structured feature dataset includes: A1. Applying a current excitation containing several frequency components to the target battery based on a preset excitation current waveform, and simultaneously acquiring its terminal voltage response, calculating the battery impedance value at each frequency through a frequency domain signal processing algorithm to obtain the battery impedance spectrum; calculating relevant parameters of the battery state through a feature extraction algorithm to obtain the electrochemical response feature vector.

[0041] A2. During the target battery aging test, physical field parameters characterizing the internal state of the battery are obtained to obtain the internal physical field feature vector.

[0042] A3. At the end of each charge-discharge cycle of the target battery, record the current cycle number and measure its actual discharge capacity to obtain a macroscopic performance data sequence.

[0043] A4. Based on a predetermined standardization method, the electrochemical response feature vector, the internal physical field feature vector, and the macroscopic performance data sequence are respectively processed to be dimensionless, and then combined into a structured feature dataset through feature fusion operation.

[0044] It should be noted that applying current excitation containing multiple frequency components means inputting a complex waveform synthesized excitation signal into the battery, whose spectrum covers a wide frequency range from millihertz to kilohertz, aiming to excite multiple timescale electrochemical processes inside the battery at once. This testing method is called information-enriched testing.

[0045] It should be noted that the simultaneous acquisition of its terminal voltage response is a discrete sampling sequence of the current excitation signal and the battery terminal voltage response signal, which are synchronized and time-aligned, and is recorded as the original current-voltage time series data (the basis data for subsequent frequency domain analysis and feature extraction).

[0046] It should be noted that frequency domain signal processing algorithms typically include Fast Fourier Transform or correlation analysis. Their core is to transform the time-domain current-voltage pair to the frequency domain (decoupling the voltage and current relationships at different frequency components from the mixed time-domain response). The battery impedance value at a specific frequency is obtained by calculating the complex ratio of the voltage response to the current excitation at that frequency. Specifically, this process involves using the calculation formula... The battery impedance spectrum was obtained. ,in Angular frequency, Represents the Fast Fourier Transform. and These are represented as discrete sampling sequences of the time-aligned current excitation signal and battery terminal voltage response signal, respectively. Further, the battery impedance spectrum is a set of complex data describing the battery's impedance values ​​at different test frequencies, including the real part (typically representing ohmic resistance and charge transfer resistance) and the imaginary part (typically related to double-layer capacitance and diffusion processes). This spectrum is a key fingerprint reflecting the battery's internal electrochemical state, such as SEI film growth and changes in the lithium-ion diffusion coefficient.

[0047] It should be noted that relevant parameters of the battery state are calculated through feature extraction algorithms to obtain the electrochemical response feature vector; based on the battery impedance spectrum, a series of quantitative feature parameters are calculated, such as, but not limited to, the ohmic internal resistance, charge transfer resistance and Warburg diffusion impedance coefficient obtained by fitting the equivalent circuit model; these parameters together constitute the electrochemical response feature vector (condensing complex impedance spectrum information into low-dimensional numerical features strongly correlated with the capacity decay mechanism).

[0048] It should be noted that the methods for obtaining the internal physical field feature vector include: directly measuring the temperature distribution (thermal field) on or inside the battery surface through embedded or attached temperature sensors; measuring the deformation of the battery during charging and discharging (stress-strain field) through strain gauges or acoustic sensors; and estimating the potential distribution (potential field) at the electrode / electrolyte interface through reference electrodes or simulation methods.

[0049] It should be noted that the macroscopic performance data sequence is a set of measured discharge capacity values ​​indexed by the number of cycles. It is a direct quantitative indicator of the degradation of the battery's macroscopic performance and is used as a supervised learning target value for machine learning models.

[0050] It should be noted that dimensionless processing is used to eliminate the impact of differences in the dimensions and numerical ranges of different features on subsequent model training. This is typically achieved using z-score standardization. The distribution of each feature is adjusted to a standard normal distribution with a mean of 0 and a standard deviation of 1, ensuring all features are on the same scale, which facilitates model learning.

[0051] It should be noted that feature fusion operation concatenates standardized features from different sources (electrochemical, physical field, macroscopic cycle) into a high-dimensional vector in a predetermined order, forming a structured data sample. For a data sample with a cycle count of... The fusion process at the test point can be represented as follows: ,in This represents a vector concatenation operation. The number of iterations is the standardized number of iterations (which is usually also normalized). Represented as the standardized first The electrochemical response characteristic vector of the next cycle, This is represented as the standardized internal physical field eigenvector of the i-th cycle. For the standardized first The discharge capacity of each cycle (data from the macroscopic performance data sequence) is used to construct a complete feature vector containing multi-dimensional state information, which serves as the unified input for subsequent capacity decay prediction models.

[0052] This application systematically acquires and constructs multi-source heterogeneous feature data, laying a comprehensive and in-depth data foundation for battery health status assessment. Its core benefit lies in breaking the limitations of traditional methods that rely on single macroscopic performance data (such as capacity), and for the first time, fusing external electrochemical response, internal physical field state, and macroscopic cycle performance across scales. This fusion can comprehensively characterize the battery's aging state from multiple dimensions of "signal-physics-performance," especially with the introduction of internal physical field characteristics, which allows for the quantitative characterization of previously unobservable internal thermal and mechanical state changes. The constructed structured dataset includes richer information on degradation correlations and clues to potential failure mechanisms, significantly enhancing the ability of subsequent models to learn complex aging patterns from the data, and providing crucial data support for solving the challenges of multi-factor coupling and nonlinearity in battery degradation.

[0053] S2. Based on the multi-source feature data, construct and train a capacity decay prediction model.

[0054] In a specific example, the construction and training of the capacity decay prediction model based on the multi-source feature data includes: defining the number of cycles, historical capacity sequence, electrochemical response characteristics of the battery, and internal physical field characteristics as input feature set based on the structured feature dataset, and defining the current discharge capacity as output target, so as to constitute model training samples.

[0055] Using the input feature set and the output target as training data, the parameters of the preset gradient boosting decision tree algorithm are initialized to obtain the initialized gradient boosting decision tree model. The initialized gradient boosting decision tree model is trained based on the iterative additive modeling process of the gradient boosting decision tree algorithm, including calculating residuals, fitting a new decision tree, and updating the model, so as to obtain the finally trained gradient boosting decision tree model as the initial capacity decay prediction model.

[0056] It should be noted that the cycle count, historical capacity sequence, electrochemical response characteristics, and internal physical field characteristics of the battery are defined as the input feature set. This means selecting the values ​​of the cycle count, historical capacity sequence, electrochemical response feature vector, and internal physical field feature vector from the structured feature dataset as model input variables. The cycle count characterizes the aging process of the battery; the historical capacity sequence reflects the capacity decay trend; the electrochemical response feature vector contains key parameters extracted from the impedance spectrum that reflect the internal electrochemical state of the battery (such as ohmic resistance and charge transfer resistance); the internal physical field feature vector contains key parameters characterizing the internal thermal field and stress-strain field of the battery. These features together constitute a multidimensional state feature vector, representing the comprehensive state of the battery at a specific cycle time, providing an information basis for the model to learn the complex nonlinear relationship between state and capacity.

[0057] It should be noted that defining the current discharge capacity as the output target means using the standardized actual discharge capacity value, measured under the same number of cycles, corresponding to the input features in the structured feature dataset, as the target variable that the model expects to predict. This output target is a direct quantitative indicator of the degradation of the battery's macroscopic performance, serving as the target value for supervised learning to guide the model in learning the mapping relationship from multidimensional input features to capacity values.

[0058] It should be noted that constructing the model training samples means pairing the input feature set (including the number of cycles, historical capacity sequence, electrochemical response feature vector, and internal physical field feature vector) for each cycle with the corresponding output target (current discharge capacity) to form a complete training data point. The set of data points from all cycles constitutes the training sample set used to train the capacity decay prediction model.

[0059] It should be noted that parameter initialization of the preset gradient boosting decision tree algorithm refers to setting initial hyperparameters for the gradient boosting decision tree model, including but not limited to: the maximum depth of the base learners (i.e., decision trees), the learning rate (or shrinkage coefficient), the subsampling rate, and the number of iterations for model training (i.e., the total number of decision trees). The initialization process sets a starting point for model training. Furthermore, the initialized gradient boosting decision tree model refers to a gradient boosting decision tree framework with initial parameter values ​​that has not yet begun learning data patterns; it awaits receiving training data to execute subsequent iterative optimization processes.

[0060] It should be noted that the iterative addition modeling process of the gradient boosting decision tree algorithm is mathematically described as follows: Assume the training sample set is... ,in For the first The input feature vector of each sample (i.e., the aforementioned input feature set). This corresponds to the true capacity value (i.e., the output target). The gradient boosting decision tree model aims to find a function... To minimize the loss function For example, mean square error Furthermore, the final output of the model is a weighted sum of a series of regression decision trees: in, For the model in the first The prediction function after iterative regression decision tree For the first A regression decision tree, This is represented by the number corresponding to the regression decision tree. , This represents the total number of regression decision trees. This is expressed as the learning rate (shrinkage coefficient). Further, the training process for each iteration is as follows: S401, Calculate the negative gradient of the current model (for squared loss, i.e., the residual): S402, with residuals To achieve the goal, fit a new regression decision tree. S403, Update Model: .

[0061] It's important to note that the residual vector reflects the error in the current model's predictions, representing the patterns the model hasn't yet learned. Training the decision tree with the residuals as the new target aims to ensure the new tree focuses on correcting the current model's errors. In S402, fitting the new decision tree means training a new regression decision tree using the residuals calculated in the current iteration as the target variable and the original input feature set as input. This decision tree recursively partitions the feature space, searching for optimal split points and leaf node values ​​to fit the current residual distribution as well as possible. In S403, updating the model involves multiplying the prediction results of the newly fitted decision tree by a learning rate less than 1 and then accumulating this into the model's prediction results from the previous iteration. The learning rate controls the contribution of each tree, preventing overfitting due to excessively large update steps. This gradual accumulation allows the model to approximate complex nonlinear functions in a progressive and ensemble manner.

[0062] It should be noted that the final trained gradient boosting decision tree model, serving as the initial capacity decay prediction model (establishing a nonlinear mapping relationship from high-dimensional, multi-source battery state characteristics to macroscopic performance indicators (capacity), which forms the basis for subsequent capacity decay prediction and closed-loop optimization), refers to a model that integrates a series of regression decision trees after training for a predetermined number of iterations. The final model The model can take a new, multi-dimensional state feature vector as input, which includes the number of cycles, historical capacity sequence, electrochemical response characteristics, and internal physical field characteristics, and output a predicted current capacity value. .

[0063] S3. During the battery aging process, execute dual-drive adaptive test decisions.

[0064] In a specific example, the dual-drive adaptive test decision includes a first drive and a second drive, wherein the first drive is based on the prediction uncertainty of the capacity decay prediction model, and the second drive is based on the real-time anomaly degree of the internal physical field characteristics.

[0065] In a specific example, the first driving force is based on the prediction uncertainty of the capacity decay prediction model, and the second driving force is based on the real-time change anomaly degree of the internal physical field characteristics, including: B1, at the current aging cycle time of the battery, obtaining the input feature set of the current prediction model, the input feature set including the current cycle number, historical capacity sequence, electrochemical response feature vector at the current time, and internal physical field feature vector; processing the input feature set based on the current prediction model, predicting future critical lifetime nodes, and then processing the predicted values ​​of future critical lifetime nodes to calculate the prediction uncertainty metric.

[0066] B2. Based on the internal physical field feature vectors at the current and historical times, perform time series analysis on the key parameters to calculate the internal physical field anomaly index.

[0067] It should be noted that processing the input feature set based on the current prediction model refers to inputting a multi-dimensional vector containing cycle count, historical capacity, electrochemical characteristics, and physical field characteristics into the finally trained gradient boosting decision tree model (i.e., the current prediction model). This model maps complex nonlinear relationships through its integrated series of regression decision trees, outputting a predicted value for a specific future performance degradation event. Furthermore, a future critical lifetime node refers to a predicted time point or cycle count, such as the future cycle count corresponding to the first degradation of the battery's discharge capacity to 80% of its rated capacity. This predicted value is the core output of "Driver One" in the dual-drive decision-making process, used to evaluate the model's understanding of future degradation states.

[0068] It should be noted that the prediction uncertainty metric is calculated as follows: For the gradient boosting decision tree ensemble model, the set of prediction results from all its base learners (i.e., each regression decision tree) for the future key lifespan node is recorded when performing S02 prediction. The statistical variance of this prediction result set is then calculated, and this variance value is used as the prediction uncertainty metric. Furthermore, the prediction uncertainty metric (quantifies the dispersion between the prediction results of each base learner within the ensemble prediction model); the larger the variance, the worse the model's consistency in predicting the future event, and the lower the confidence level of the model's prediction, i.e., the higher the "cognitive uncertainty." The uncertainty metric is a key indicator for evaluating the predictive reliability of the model itself.

[0069] It should be noted that the calculation of the internal physical field anomaly index is achieved as follows: From the internal physical field feature vector, parameters strongly correlated with key aging mechanisms such as battery thermal runaway, lithium plating, or mechanical failure (e.g., temperature at a specific measurement point and shell strain values) are selected as key parameters. For each key parameter, its values ​​at the current moment and several consecutive aging cycles prior to it are extracted to form a time series. The information entropy of this time series is calculated, and this information entropy value is used as the current anomaly index of the key parameter. Further, firstly, the frequency of each unique value in the time series is counted, and its frequency is calculated as a probability. Then, the product of each probability value and the base-2 logarithm is calculated. Finally, the sum of the above products of all unique values ​​is taken as a negative value, thus obtaining the information entropy of the time series. Information entropy is a dimensionless scalar. In information theory, information entropy is used to measure the uncertainty or information content of a random variable (by calculating the information entropy of the time series of key parameters of the internal physical field, the degree of disorder in the data distribution of the parameter is quantified). If the internal state of the battery undergoes a sudden change (such as local overheating and a surge in stress), the temporal regularity of its physical field parameters will be broken, resulting in a significant change (usually an increase) in the calculated information entropy value, thus indicating a potential abnormal state.

[0070] In a specific example, the execution of dual-drive adaptive test decision includes: setting a first decision threshold and a second decision threshold; performing logical judgment on the prediction uncertainty measure and the internal physical field anomaly index; comparing the prediction uncertainty measure with the first decision threshold, and simultaneously comparing the internal physical field anomaly index with the second decision threshold; when "the prediction uncertainty measure is greater than the first decision threshold" or "the internal physical field anomaly index is greater than the second decision threshold", generating an instruction to trigger the next information enrichment test, determining that new high-dimensional test data needs to be added to reduce model uncertainty, and confirming physical field anomalies.

[0071] It should be noted that the first decision threshold is a pre-set variance threshold based on the historical model's predictive performance, used to determine whether the model's predictions have become unreliable due to insufficient data. The second decision threshold is a threshold set based on the baseline level of information entropy of the internal physical field characteristics during normal aging, used to determine whether the physical field state deviates from the normal degradation path. This decision logic ensures the sensitivity of the decision; whether the model's own confidence is insufficient or an actual physical anomaly is detected, it can promptly trigger information enrichment testing, achieving a dual-engine drive of data-driven adaptive optimization and physical state-driven anomaly response.

[0072] The dual-drive adaptive test decision mechanism proposed in this application achieves efficient allocation of test resources and adaptive management of model risks. Its benefits are twofold: First, "data-driven" decision-making, by quantifying the uncertainty of model predictions, intelligently triggers costly information enrichment tests when model confidence is insufficient, thereby obtaining the most effective "valuable data" for improving prediction accuracy with the fewest number of tests, greatly improving test economy. Second, "mechanism-driven" decision-making, by monitoring abnormal abrupt changes in internal physical field characteristics, can intervene in a timely manner at critical nodes where potential safety hazards or aging mechanisms of the battery occur, enhancing the method's early warning and safety assurance capabilities.

[0073] S4. When the dual-drive adaptive test decision is triggered, update the capacity decay prediction model.

[0074] In a specific example, updating the capacity decay prediction model when the dual-drive adaptive test decision is triggered includes: when an instruction to trigger the next information enrichment test is generated, based on the current battery cycle state, performing an information enrichment test on the battery in a subsequent preset charge-discharge cycle to collect new electrochemical response data and new internal physical field data, so as to obtain a new electrochemical response feature vector and a new internal physical field feature vector; and obtaining the new actual discharge capacity value corresponding to this information enrichment test; and performing feature fusion of the new electrochemical response feature vector, the new internal physical field feature vector and the new actual discharge capacity value to form a new data point.

[0075] The newly added data points are added to the structured feature dataset to obtain the updated feature dataset. Incremental learning is then performed on the constructed initial capacity decay prediction model to obtain the updated capacity decay prediction model.

[0076] It should be noted that information enrichment testing refers to a comprehensive test specifically executed after receiving an instruction to trigger the next information enrichment test. This test involves current excitation with multiple frequency components and is specifically designed to obtain richer battery state information to reduce model uncertainty or confirm physical field anomalies. It proactively supplements high-dimensional state data, providing new and high-quality data sources for iterative optimization of the model.

[0077] It should be noted that the "new electrochemical response data" and "new internal physical field data" refer to the current-voltage time-series data pairs synchronously acquired during the information enrichment test triggered in this instance, as well as the real-time physical field parameters such as temperature, stress-strain, and potential distribution estimated through sensor measurements or simulations. These data constitute the most original information reflecting the current internal electrochemical and physical state of the battery (i.e., at a certain cycle moment after receiving the trigger command).

[0078] It should be noted that feature fusion refers to combining the new electrochemical response feature vector, the new internal physical field feature vector, and the new actual discharge capacitance value into a unified structured data sample, using the same standardized methods and splicing rules as the original feature dataset. This fusion process can be formally represented as adding new data points. in This represents the standardized number of loops required for this test. , and These are the electrochemical response eigenvectors after feature fusion and normalization, the new internal physical field eigenvectors, and the new actual discharge capacitance value, respectively. This operation ensures that the newly added data points... Compared with the original dataset The samples in the model have the same feature dimensions and data scale, and can be directly used for model updates.

[0079] It should be noted that the "add" operation refers to adding newly created data points. The sample set is appended to the existing structured feature dataset. The updated feature dataset. The expression is ,in This represents the union operation of sets. This operation expands the data samples available for the model to learn from, incorporating information reflecting the latest state of the battery.

[0080] It's important to note that "incremental learning" is a model update strategy. Its core principle is to fine-tune model parameters using only new data points or small batches of new data, while retaining the knowledge already learned by the original model, without needing to completely retrain using all historical data. For gradient boosting decision tree models, incremental learning can be achieved by fitting and adding new decision trees based on the residuals calculated from the new data, building upon existing model ensembles. The advantages of incremental learning are high update efficiency, low computational resource consumption, and suitability for online, continuous model optimization scenarios.

[0081] It should be noted that the "updated capacity degradation prediction model" refers to a new prediction model with optimized parameters obtained after the aforementioned incremental learning operation. Compared to the original model, this model incorporates the latest battery state and performance degradation information reflected from the newly added data points into its internal mapping relationships. Therefore, theoretically, it will have higher accuracy and lower prediction uncertainty in predicting subsequent battery capacity degradation, achieving adaptive improvement and iterative evolution of model performance.

[0082] S5. Based on the updated capacity decay model, output prediction results and analyze the battery capacity decay mechanism.

[0083] In a specific example, the output prediction results and battery capacity decay mechanism analysis based on the updated capacity decay model include: using the updated capacity decay prediction model as the current model based on the updated capacity decay prediction model and the current cycle state, returning to the execution of the dual-drive adaptive test decision process, and performing the next round of monitoring and decision-making for the subsequent aging cycles of the battery to form a closed-loop loop path of model iteration and test execution; based on the current cycle state and the preset iteration termination condition, performing a termination judgment on the execution of the closed-loop loop path, and terminating the iteration when the battery capacity decays to the end-of-life threshold or the number of cycles reaches the preset total test cycle to obtain the iteration termination judgment result; when the iteration terminates, using the latest capacity decay prediction model at the time of termination, calculating and outputting a high-confidence capacity decay curve, the remaining service life prediction value and its confidence interval, and combining the feature importance analysis results of the latest capacity decay prediction model, performing a qualitative or semi-quantitative analysis of the dominant mechanism of battery decay to obtain a comprehensive battery life assessment report.

[0084] It should be noted that "executing the next round of monitoring and decision-making" means using the capacity decay model optimized in the previous round as the starting point for the new cycle and restarting the process described in step three. Specifically, in each (or specific) aging cycle, based on the current battery state (electrochemical response characteristics and internal physical field characteristics), the updated capacity decay model is used to predict the capacity and calculate its prediction uncertainty (first driver). At the same time, the anomaly of the real-time physical field characteristics is analyzed (second driver), and the decision-making logic of the dual drivers determines whether the next information enrichment test is needed to supplement the data and may trigger another model update. This process realizes a closed-loop feedback mechanism of "testing - acquiring new data - model updating - improving decision-making," enabling the testing strategy and prediction model to adaptively evolve as the battery ages.

[0085] It's important to note the closed-loop path of model iteration and test execution (which constructs a dynamic, self-learning battery health management framework). This path is not a one-off, unidirectional testing and modeling process, but rather a continuous iterative optimization process. Each capacity degradation model update means that the model's understanding of the battery's current and future aging behavior is corrected and enhanced based on new measured data. This makes subsequent testing decisions (determining when to conduct information enrichment testing) more intelligent and efficient, aiming to obtain the most accurate lifespan prediction possible with the least possible testing cost.

[0086] It's important to note that the termination check means that after each decision in step three or model update in step four, the system checks whether two preset iteration termination conditions are met. The first condition is that the battery's actual discharge capacity (from the macroscopic performance data sequence) has decayed to a preset threshold (e.g., 70% or 80% of the rated capacity, i.e., the end-of-life threshold). The second condition is that the battery has undergone a preset total number of test cycles (preset test period). These two conditions are judged by an "OR" logic relationship; if either condition is met, the entire closed-loop iteration process stops. This check ensures that the testing and prediction process automatically terminates when the battery's lifespan naturally ends or when the predetermined study period is completed, avoiding meaningless continuous operation.

[0087] It should be noted that the iteration termination judgment result is a binary status flag used to indicate whether the closed-loop iteration process should continue. If the judgment result is "continue," the process returns to the previous step; if it is "terminate," the process proceeds to the next step for the final synthesis output. This result controls the flow branches of the entire method execution.

[0088] It's important to clarify that "calculating and outputting a high-confidence capacity decay curve" refers to using the best-performing capacity decay model at the end of the iteration (i.e., the final model after multiple rounds of closed-loop iteration optimization), inputting a multi-dimensional feature set indexed by the number of iterations from the initial loop to the future prediction endpoint (this may require reasonable estimation of certain features for future loops or use of historical trends), and having the model predict the capacity value corresponding to each loop point. Connecting these predicted "loop number - capacity" data points forms a predicted capacity decay curve. The term "high confidence" stems from the fact that this model is trained and optimized based on continuously supplemented high-value data (triggered by dual-driven decisions) during the closed-loop iteration process, resulting in a significant improvement in predictive performance compared to the initial model.

[0089] It should be noted that the process of "calculating and outputting the predicted remaining lifespan and its confidence interval" is as follows: First, define the "end of life" of the battery, for example, capacity decay to 80% of the rated capacity. Then, based on the predicted capacity decay curve, find the number of cycles corresponding to when the capacity first falls below the capacity threshold. The difference between this number of cycles and the number of cycles at the current termination of the iteration is the predicted remaining lifespan. Furthermore, the calculation of the confidence interval is usually based on the uncertainty of the model prediction. For ensemble models such as gradient boosting decision trees, the statistical distribution can be calculated using the set of predictions of remaining lifespan from all base learners (single trees). For example, the confidence interval of the predicted remaining lifespan can be obtained by calculating the mean (point estimate) of this set and the mean plus or minus a certain number of standard errors, representing the range within which the actual remaining lifespan value may fall at a given confidence level (e.g., 95%).

[0090] It should be noted that "combining the model feature importance analysis results" refers to extracting and analyzing the importance scores of each input feature (such as impedance at a specific frequency, temperature at a certain temperature measurement point, and historical capacity trends) in the latest gradient boosting decision tree model for the predicted capacity target. Feature importance scores are typically measured by calculating the number of times each feature is used to split nodes across all decision trees, or by the total reduction in impurity resulting from its splitting. Features with high importance are considered to contribute significantly to capacity decay prediction.

[0091] It should be noted that "qualitative or semi-quantitative analysis of the dominant battery degradation mechanism" refers to interpreting the results of the aforementioned feature importance analysis in conjunction with professional knowledge. For example, if the feature importance of ohmic internal resistance and charge transfer resistance in a specific mid-frequency range remains high, it may indicate that SEI film growth is the dominant degradation mechanism; if the importance of temperature or casing strain characteristics in high-temperature regions increases abnormally, it may suggest an increased risk of thermal aging or mechanical stress failure. Based on the ranking and relative magnitude of the importance scores, a "qualitative or semi-quantitative analysis report" can be generated, revealing the main contradictions or potential failure modes in the current battery aging process from a data-driven perspective, providing guidance for battery design improvements or usage strategy optimization.

[0092] It should be noted that the "Comprehensive Battery Life Assessment Report" is the final output of the method of this invention. This report integrates high-confidence capacity degradation trend prediction, remaining life estimation with reliability ranges, and data-driven insights into aging mechanisms. It not only provides a quantitative answer to "how much longer can it be used," but also explains the underlying reasons for "why it degrades in this way," achieving a comprehensive assessment from condition monitoring and life prediction to mechanism analysis, and has higher engineering application value.

[0093] Please see Figure 2 As shown, in a second aspect, this application provides a system for testing and predicting the capacity degradation of a power battery.

[0094] The system 100 of the power battery capacity degradation testing and prediction method described in this invention can be installed in an electronic device. Depending on the functions implemented, the system 100 may include a structured feature dataset generation module 101, a capacity degradation prediction model construction module 102, a decision module 103, a model update module 104, and a result output module 105. The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, stored in the memory of the electronic device.

[0095] In this embodiment, the functions of each module / unit are as follows: The structured feature dataset generation module is used to acquire multi-source feature data of the battery and integrate it into a structured feature dataset.

[0096] The capacity decay prediction model construction module constructs and trains a capacity decay prediction model based on the multi-source feature data.

[0097] The decision module is used to perform dual-drive adaptive test decisions during battery aging.

[0098] The model update module is used to update the capacity decay prediction model when the dual-drive adaptive test decision is triggered.

[0099] The results output module outputs prediction results and battery capacity decay mechanism analysis based on the updated capacity decay model.

[0100] This application provides a method and system for testing and predicting the capacity degradation of power batteries. By integrating electrochemical impedance spectroscopy, internal physical field characteristics, and macroscopic cycling data, a multi-dimensional state characterization system is constructed, thereby achieving more accurate fitting and prediction of complex degradation patterns. Through a "dual-drive adaptive test decision" mechanism, a time-consuming information enrichment test is triggered only when the model prediction uncertainty is high or the internal state is abnormal, achieving the best balance between test efficiency and model performance maintenance. The model not only outputs capacity predictions, but its internal feature importance analysis can qualitatively or semi-quantitatively reveal the dominant degradation mechanisms at different aging stages. At the same time, based on real-time monitoring of the anomaly degree of internal physical field characteristics, abnormal battery states can be detected earlier than macroscopic capacity degradation, providing a forward-looking basis for safety warnings and maintenance decisions.

[0101] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0102] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0103] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0104] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0105] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for testing and predicting the capacity degradation of a power battery, characterized in that, include: S1. Obtain multi-source feature data of the battery and integrate it into a structured feature dataset; S2. Based on the multi-source feature data, construct and train a capacity decay prediction model; S3. During the battery aging process, execute dual-drive adaptive test decisions; S4. When the dual-drive adaptive test decision is triggered, update the capacity decay prediction model; S5. Based on the updated capacity decay model, output prediction results and analyze the battery capacity decay mechanism.

2. The method for testing and predicting the capacity degradation of a power battery according to claim 1, characterized in that, The multi-source feature data includes the battery's electrochemical response characteristics and internal physical field characteristics.

3. The method for testing and predicting the capacity degradation of a power battery according to claim 1, characterized in that, The acquisition of multi-source feature data of the battery and its integration into a structured feature dataset includes: A1. Based on the preset excitation current waveform, apply a current excitation containing several frequency components to the target battery and simultaneously acquire its terminal voltage response. Calculate the battery impedance value at each frequency using a frequency domain signal processing algorithm to obtain the battery impedance spectrum. Calculate the relevant parameters of the battery state using a feature extraction algorithm to obtain the electrochemical response feature vector. A2. During the target battery aging test, physical field parameters characterizing the internal state of the battery are obtained to obtain the internal physical field feature vector. A3. At the end of each charge-discharge cycle of the target battery, record the current cycle number and measure its actual discharge capacity to obtain a macroscopic performance data sequence. A4. Based on a predetermined standardization method, the electrochemical response feature vector, the internal physical field feature vector, and the macroscopic performance data sequence are respectively processed to be dimensionless, and then combined into a structured feature dataset through feature fusion operation.

4. The method for testing and predicting the capacity degradation of a power battery according to claim 1, characterized in that, The construction and training of the capacity decay prediction model based on the multi-source feature data includes: Based on the structured feature dataset, the number of cycles, historical capacity sequence, electrochemical response characteristics of the battery, and internal physical field characteristics are defined as the input feature set, and the current discharge capacity is defined as the output target to form the model training samples. Using the input feature set and the output target as training data, the parameters of the preset gradient boosting decision tree algorithm are initialized to obtain the initialized gradient boosting decision tree model. The initialized gradient boosting decision tree model is trained based on the iterative additive modeling process of the gradient boosting decision tree algorithm, including calculating residuals, fitting a new decision tree, and updating the model, so as to obtain the finally trained gradient boosting decision tree model as the initial capacity decay prediction model.

5. The method for testing and predicting the capacity degradation of a power battery according to claim 1, characterized in that, The dual-drive adaptive test decision includes a first drive and a second drive, wherein the first drive is based on the prediction uncertainty of the capacity decay prediction model, and the second drive is based on the real-time anomaly degree of the internal physical field characteristics.

6. The method for testing and predicting the capacity degradation of a power battery according to claim 1, characterized in that, The first driver is based on the prediction uncertainty of the capacity decay prediction model, and the second driver is based on the real-time anomaly degree of the internal physical field characteristics, including: B1. At the current aging cycle of the battery, obtain the input feature set of the current prediction model. The input feature set includes the current cycle number, historical capacity sequence, electrochemical response feature vector at the current moment, and internal physical field feature vector. Based on the current prediction model, process the input feature set and predict future critical lifetime nodes. Then, process the predicted values ​​of future critical lifetime nodes to calculate the prediction uncertainty metric. B2. Based on the internal physical field feature vectors at the current and historical times, perform time series analysis on the key parameters to calculate the internal physical field anomaly index.

7. The method for testing and predicting the capacity degradation of a power battery according to claim 1, characterized in that, The execution of dual-drive adaptive test decisions includes: A first decision threshold and a second decision threshold are set, and logical judgments are made on the prediction uncertainty measure and the internal physical field anomaly index. The prediction uncertainty measure is compared with the first decision threshold, and the internal physical field anomaly index is compared with the second decision threshold. When the prediction uncertainty measure is greater than the first decision threshold or the internal physical field anomaly index is greater than the second decision threshold, an instruction to trigger the next information enrichment test is generated, it is determined that new high-dimensional test data needs to be added to reduce model uncertainty, and the physical field anomaly is confirmed.

8. The method for testing and predicting the capacity degradation of a power battery according to claim 1, characterized in that, When the dual-drive adaptive test decision is triggered, updating the capacity decay prediction model includes: When a command is generated to trigger the next information enrichment test, based on the current battery cycle state, an information enrichment test is performed on the battery in a subsequent preset charge-discharge cycle to collect new electrochemical response data and new internal physical field data, so as to obtain new electrochemical response feature vectors and new internal physical field feature vectors; and to obtain the new actual discharge capacity value corresponding to this information enrichment test; the new electrochemical response feature vector, the new internal physical field feature vector, and the new actual discharge capacity value are feature fused to form new data points; The newly added data points are added to the structured feature dataset to obtain the updated feature dataset. Incremental learning is then performed on the constructed initial capacity decay prediction model to obtain the updated capacity decay prediction model.

9. The method for testing and predicting the capacity degradation of a power battery according to claim 1, characterized in that, The prediction results and battery capacity degradation mechanism analysis based on the updated capacity degradation model include: Based on the updated capacity decay prediction model and the current cycle state, the updated capacity decay prediction model is used as the current model, and the dual-drive adaptive test decision process is returned to execute. This process performs the next round of monitoring and decision-making for subsequent battery aging cycles, forming a closed-loop cycle path of model iteration and test execution. Based on the current cycle state and preset iteration termination conditions, the execution of the closed-loop cycle path is terminated. When the battery capacity decays to the end-of-life threshold or the number of cycles reaches the preset total test cycle, the iteration is terminated, yielding the iteration termination judgment result. When the iteration terminates, the latest capacity decay prediction model at the time of termination is used to calculate and output a high-confidence capacity decay curve, remaining lifespan prediction value, and its confidence interval. Combined with the feature importance analysis results of the latest capacity decay prediction model, a qualitative or semi-quantitative analysis of the dominant battery decay mechanism is performed to obtain a comprehensive battery life assessment report.

10. A system for implementing the power battery capacity degradation testing and prediction method according to any one of claims 1-9, characterized in that, include: The structured feature dataset generation module is used to acquire multi-source feature data of batteries and integrate it into a structured feature dataset. The capacity decay prediction model construction module constructs and trains a capacity decay prediction model based on the multi-source feature data. The decision module is used to execute dual-drive adaptive test decisions during battery aging. The model update module is used to update the capacity decay prediction model when the dual-drive adaptive test decision is triggered. The results output module outputs prediction results and battery capacity decay mechanism analysis based on the updated capacity decay model.

Images