A battery cross-life stage aging trend prediction method and system

Abstract

This invention discloses a method and system for predicting the aging trend of a battery across its life stages. The method includes: S1, acquiring early aging characteristic data of the first life stage of the battery to be predicted and corresponding load condition information; S2, constructing a condition-aware model based on a cross-attention mechanism, using battery data from no-load switching as the main input and corresponding load condition information as the condition input to train the condition-aware model; S3, using the condition-aware model to output the evolution sequence of subsequent aging characteristics of the battery to be predicted from the early stage of the first life stage in a time-series prediction manner, the evolution sequence of subsequent aging characteristics covering the remaining lifespan of the first life stage and the aging characteristic changes of the second life stage after load condition switching; S4, estimating the battery health state based on the evolution sequence of subsequent aging characteristics to generate a cross-life stage aging trajectory prediction curve from the first life stage to the second life stage, thereby improving the robustness and practicality of the prediction.

Description

Technical Field

[0001] This invention relates to the field of battery health management technology, and in particular to a method for predicting battery aging trends based on data-driven and machine learning models. Specifically, it relates to a method and system that can use early battery operating data to predict its aging trajectory across life stages before and after load condition switching. Background Technology

[0002] With the rapid development of the new energy industry, the large-scale application of power batteries has resulted in a large number of retired batteries. These batteries still retain a considerable amount of remaining capacity and can be reused in scenarios with relatively low performance requirements, such as energy storage systems and backup power supplies. This is of great significance for improving resource utilization efficiency, reducing environmental pollution, and promoting the development of a circular economy.

[0003] The core challenge of battery reuse lies in accurately assessing the State of Health (SOH) of retired batteries and predicting their remaining lifespan under new operating conditions. When a battery transitions from its first life stage (e.g., automotive applications) to its second life stage (e.g., energy storage applications), its operating conditions, such as operating current and load characteristics, typically change significantly. This shift in operating conditions alters the battery's aging rate and degradation mechanism, rendering traditional predictive models built on first-life stage data ineffective.

[0004] Currently, battery health status prediction methods can be mainly divided into model-based methods and data-driven methods. Model-based methods rely on electrochemical mechanisms, require precise battery parameters, and have poor generalization ability. Data-driven methods, such as recurrent neural networks using Long Short-Term Memory (LSTM) networks or Gated Recurrent Units (GRUs), can learn aging patterns from historical data. However, these methods usually implicitly assume that the battery will continue operating under past conditions in the future, resulting in significant prediction errors when operating conditions change abruptly. Some studies have attempted to improve the generalization ability of models under different individual batteries or slight changes in operating conditions through transfer learning. However, this essentially addresses the problem of differences between samples in the spatial dimension. For the aging trajectory changes of the same battery over time due to sudden changes in operating conditions, existing methods lack effective explicit modeling means, making it difficult to accurately predict the long-term aging behavior of batteries under unknown operating conditions with only early data.

[0005] Therefore, there is an urgent need in this field for a method that can explicitly model the effects of operating current and achieve long-term, high-precision prediction of aging trends across life stages and operating conditions using only early battery aging data. Summary of the Invention

[0006] This invention aims to address the fundamental technical challenge of inaccurate aging trajectory prediction across battery life stages caused by the inability of existing battery health prediction technologies to cope with abrupt changes in battery operating conditions. Specifically, under the constraint of "using only early battery data," it solves the following two key intermediate problems: first, it struggles to handle abrupt changes in the aging trajectory of the same battery during its life cycle due to abrupt changes in operating conditions; second, it cannot accurately predict the long-term aging behavior of a battery under unknown operating conditions in the second life stage with only early data. The core of this invention is to solve how to construct a prediction model that explicitly models operating current information as the core driving variable, thereby accurately predicting the remaining lifespan of the first life stage and the aging trend of the second life stage after the operating condition change, based solely on limited data from the early stages of the battery's first life stage.

[0007] The technical problem of this invention is solved by the following technical solution: In a first aspect, the present invention provides a method for predicting the aging trend of batteries across life stages, comprising the following steps: S1. Data preparation: Obtain aging characteristic data and corresponding load condition information of the early stage of the first life of the battery to be predicted. S2. Model Training: Construct a working condition perception model based on a cross-attention mechanism, using no-load switching battery data as the main input and corresponding load working condition information as the condition input, and train the working condition perception model. S3. Aging Feature Prediction: Using the operating condition perception model, the subsequent aging feature evolution sequence of the battery to be predicted from the early stage of the first life is output in a time-series prediction manner. The subsequent aging feature evolution sequence covers the remaining life of the first life stage and the aging feature changes of the second life stage after the load condition switch. S4. Health Status Estimation: Based on the subsequent aging feature evolution sequence, estimate the health status of the battery to generate a cross-life stage aging trajectory prediction curve from the first life stage to the second life stage.

[0008] In some embodiments, the following technical features are also included: Before step S1, there is also a feature selection step: based on feature contribution, correlation and prediction accuracy evaluation, a feature set for health status prediction is selected from the multidimensional aging feature data of the battery; the aging feature data obtained in step S1 is the data in the feature set; the multidimensional aging feature data includes at least one of time feature, incremental capacity feature, capacity feature, temperature feature and current change feature.

[0009] In some embodiments, after acquiring the data in step S1, a preprocessing step is further included: aggregating the aging characteristic data of the early stage of the first life of the battery to be predicted and the corresponding load condition information at a predetermined time step interval; the time step interval is configured such that the change in health status within a single time step does not exceed 2%; the time step interval is one time step for every 10 charge-discharge cycles.

[0010] In some embodiments, in step S2, the working condition perception model is a Transformer model, and the training process based on the cross-attention mechanism includes the following sub-steps: S21. Perform linear mapping or feature transformation on the main input to construct a query vector; perform feature encoding on the conditional input to construct a key vector and a value vector; S22. Calculate the similarity between the query vector and the key vector, and obtain the attention weight after normalization; S23. The value vector is weighted and summed or weighted and combined according to the attention weight to generate an aging feature representation that integrates load condition information.

[0011] In some embodiments, in step S3, a recursive time-series prediction method is adopted, and before the time-series prediction begins, the operating condition perception model is fine-tuned using real observation data from the early stage of the first life phase of the battery to be predicted; in step S4, the estimation of the battery's health status specifically involves estimating it using a Gaussian process regression model constrained by load operating conditions.

[0012] In some embodiments, the estimation using a Gaussian process regression model constrained by load conditions includes the following sub-steps: S41. For different load conditions, establish corresponding Gaussian process regression sub-models based on complete no-load switching battery data to form a model library. S42. For each time step in the subsequent aging feature evolution sequence, select a matching Gaussian process regression sub-model from the model library based on its corresponding load condition information. S43. Using the selected sub-model, estimate the aging characteristics of this time step as a healthy state.

[0013] In some embodiments, the load condition information includes at least one of charge / discharge current, voltage, or power.

[0014] In a second aspect, the present invention provides a battery management system, comprising: a processor and a memory; the memory storing a computer program; the processor being configured to execute the computer program to implement the above-described method.

[0015] Thirdly, the present invention provides a battery health status prediction system, comprising: The data acquisition module is configured to collect battery aging characteristic data and load condition information; The data processing and prediction module is configured to execute the above methods to generate aging trajectory prediction curves across life stages; The result output module is configured to output the predicted curve.

[0016] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0017] The beneficial effects of this invention compared to the prior art include: This invention provides a method for predicting battery aging trends across life stages based on a condition-aware mechanism. By constructing a cross-attention model with load condition information as input, it achieves explicit modeling of the coupling relationship between aging feature evolution and operating conditions. The technical solution first trains the model using battery data from no-load switching, enabling it to learn the general aging patterns under different load conditions. Then, during prediction, early data of the target battery and future target load condition information are simultaneously input into the model. Through the cross-attention mechanism, the model can dynamically modulate its attention distribution to historical aging features based on future operating condition information, thereby inferring a specific aging trajectory under the constraint of the target load condition. This mechanism allows the model to extrapolate not only time series but also operating conditions, ultimately achieving the beneficial effect of accurately predicting the long-term aging behavior of batteries after experiencing unknown operating condition switching using only early data. This synergistically achieves the top-level goal of early, accurate, and low-cost assessment of the value of retired batteries for secondary use, significantly reducing the dependence of battery secondary use value assessment on full life-cycle test data and improving the robustness and practicality of prediction.

[0018] In addition, some embodiments also have the following beneficial effects: By first predicting aging characteristic sequences that are more closely related to the mechanism and then mapping them to health status values ​​in a two-stage decoupling design, the problem of excessively rapid error accumulation when directly predicting SOH is avoided, and the stability of long-term prediction is enhanced.

[0019] Furthermore, by employing a comprehensive screening strategy based on feature contribution, correlation, and prediction accuracy, an aging feature set was obtained, reducing the interference of redundant features and improving the training efficiency and generalization ability of the model while ensuring prediction accuracy.

[0020] Furthermore, by employing recursive time-series prediction and fine-tuning the model using early real data from the target battery at the beginning, the starting point of the predicted trajectory is anchored to the actual aging path, effectively compressing the initial prediction error.

[0021] Furthermore, by establishing independent Gaussian process regression sub-model libraries for different load conditions and matching and selecting them based on load information during prediction, the health status estimation can be precisely adapted to the mapping relationship under different operating conditions, further improving the accuracy of the final SOH prediction curve.

[0022] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0023] Figure 1 This is a flowchart of a battery aging trend prediction method across life stages provided by an embodiment of the present invention.

[0024] Figure 2 This is a flowchart of a method for predicting battery aging trends across life stages based on a working condition sensing mechanism, provided in another embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of prediction error results under different feature configurations in one embodiment of the present invention.

[0026] Figures 4a to 4f This is a comparison chart of aging trajectory prediction curves of different prediction methods on a test set in one embodiment of the present invention. Detailed Implementation

[0027] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0029] The core of this invention lies in providing a solution for predicting battery aging trends across life stages. This solution first preprocesses and filters the multidimensional aging features of the battery to obtain key features that effectively characterize the evolution of the battery's health state. Then, the core is the construction of a Current-Aware Cross-Attention Transformer (CACA-T) model. This model uses a cross-attention mechanism to deeply fuse and train the aging feature sequence (main input) of a battery switching from no-load conditions with the corresponding load condition information (conditional input), thereby learning the mapping relationship between aging patterns and operating conditions. For a variable-condition battery to be predicted, its early aging features of the first life stage and the target load condition information of the second life stage are input into the trained CACA-T model to predict its subsequent aging feature evolution sequence in a recursive or non-recursive temporal manner. Finally, using a pre-established Gaussian process regression (GPR) model library matched with load condition conditions, the predicted aging feature sequence is mapped to a state of health (SOH) sequence, ultimately generating a smooth and continuous cross-life stage aging trajectory prediction curve. The load condition refers to the external load conditions of the battery, the changes of which are reflected by one or more measurable operating parameters, including but not limited to charging and discharging current, voltage, and power. The core of this invention lies in explicitly sensing the changes in these operating condition parameters through a model and modeling their impact on the battery aging trajectory.

[0030] Example 1 This embodiment provides a method for predicting the aging trend of batteries across their life stages, such as... Figure 1 As shown, it includes the following steps: A1. Data Preparation and Preprocessing: Acquire aging characteristic data and corresponding load condition information of the early stage of the first life phase of the battery to be predicted; aggregate the original cycle data (i.e., aging characteristic data and corresponding load condition information of the early stage of the first life phase of the battery to be predicted) at predetermined time step intervals; the time step interval is configured such that the change in health status within a single time step does not exceed 2%; specifically, preprocess the multidimensional aging characteristic data of the battery to be predicted, with each time step being 10 charge-discharge cycles. Select batteries without load condition switching, and based on feature contribution analysis, correlation analysis, and prediction accuracy evaluation, select the optimal feature set for battery health status prediction from the multidimensional aging characteristic data of the battery; the aging characteristic data obtained in step A1 is the data in the optimal feature set. Specifically, the load condition information includes at least one of charge-discharge current, voltage, or power.

[0031] A2. Model Training: Construct a condition-aware Transformer model based on cross-attention mechanism, using no-load switching battery data as the main input and corresponding load condition information as the condition input to train the condition-aware Transformer model. A3. Aging Feature Prediction: Using the operating condition perception model, the subsequent aging feature evolution sequence of the battery to be predicted is output in a time-series prediction manner from the early stage of the first life phase. The subsequent aging feature evolution sequence covers the remaining lifespan of the first life phase and the aging feature changes in the second life phase after load condition switching. Specifically, the aging features of the early stage of the first life phase of the variable operating condition battery are fused with load condition information to predict the subsequent aging feature evolution process from the early stage of the first life phase. The evolution process includes the remaining lifespan range of the first life phase and the aging feature changes in the second life phase after load condition switching.

[0032] A4. Health Status Estimation: Based on the key aging characteristics predicted in step A3, the battery health status is estimated using a Gaussian process regression model constrained by load conditions, in order to generate a cross-life-stage aging trajectory prediction curve from the first life stage to the second life stage.

[0033] The raw data is aggregated every 10 cycles as a time step. Under the aging rate conditions of the dataset used in this embodiment, the SOH change within a single time step is typically within 1%, meeting the requirement that the aging behavior within a time step can be linearly approximated. The time step interval is an adjustable hyperparameter, which can be adjusted accordingly by those skilled in the art based on the aging rate characteristics of the target battery. It is recommended that the SOH change within a single time step not exceed 2% as a reference standard. Step A1 includes the following sub-steps:

[0034] A11. Denoise, normalize and unify the time scale of the multidimensional aging characteristic data of the battery. A12. Use the characteristic contribution analysis method to assess the degree of influence of each aging characteristic on the prediction results of health status; A13. Calculate the correlation index between each aging characteristic and the battery health status; A14. Based on the prediction accuracy evaluation results of aging features, the optimal feature set for subsequent model training and health status estimation is selected.

[0035] The multidimensional aging characteristic data includes at least one of the following: time characteristics, incremental capacity (IC) characteristics, capacity characteristics, temperature characteristics, and current change characteristics.

[0036] Step A2 includes the following sub-steps: A21. Select the aging characteristic data of the battery switching under no-load conditions as the main input, and select the corresponding load condition information, temperature characteristics and load condition change characteristics as condition inputs. A22. Input the main input and conditional input into the corresponding network structure of the working condition perception Transformer model, and introduce a cross attention mechanism into the model to perform information interaction and feature fusion on the main input and conditional input. A23. Based on the fused feature representation, the operating condition-aware Transformer model is trained to learn the mapping relationship between the aging feature evolution process and the load conditions, providing a model foundation for subsequent cross-life stage aging feature prediction.

[0037] The training process based on the cross-attention mechanism in sub-step A22 includes the following steps: A221. Perform linear mapping or feature transformation on aging feature data to construct a query vector for characterizing aging state evolution information; at the same time, perform feature encoding on load condition information to construct corresponding key vectors and value vectors, so that the key vectors and value vectors can reflect the operating characteristics under different load conditions. A222. Based on the similarity relationship between the query vector and the key vector, calculate the attention weight between the query vector and the key vector, and normalize the attention weight to obtain a weight distribution that reflects the correlation between aging characteristics and load conditions. A223. Based on the normalized attention weights, the value vectors are weighted and summed or weighted and combined to generate an aging feature representation that integrates load condition information, thereby enhancing the model's ability to perceive the evolution law of aging features under different load conditions.

[0038] Preferably, in step A3, the prediction of the subsequent aging feature evolution process adopts a recursive temporal prediction method, generating aging feature prediction results step by step according to time steps, thereby forming a continuous aging feature sequence covering the remaining lifespan of the first life stage and the second life stage. Before the recursive temporal prediction begins, the operating condition perception model is fine-tuned using real observation data from the early stage of the first life stage of the target battery, anchoring the recursive starting state to the measured aging trajectory of the target battery, thus compressing the initial error. The cross-attention mechanism performs global modeling based on a complete historical window, which has a dilution and smoothing effect on single-step prediction deviations and suppresses long-range drift of local errors. Experimental results demonstrate that error accumulation is controllable in practice.

[0039] Step A4 includes the following sub-steps: A41. Based on the predicted aging characteristics, construct a Gaussian process regression model with the aging characteristics as input and the battery health status as output; A42. For different load conditions, establish corresponding Gaussian process regression sub-models constrained by load conditions based on the complete data of battery switching under no-load conditions, and form a model library. A43. For each time step in the subsequent aging feature evolution sequence, based on its corresponding load condition information, a matching Gaussian process regression sub-model is selected from the model library, and the aging feature of that time step is estimated as a healthy state in order to estimate the battery health state. A44. Combine the estimated battery health status over time to form a cross-life stage aging trajectory prediction curve from the first life stage to the second life stage.

[0040] The battery health status is referred to as the battery's state of health (SOH).

[0041] The condition-aware Transformer model based on the cross-attention mechanism is a multi-layer encoder structure. Each layer of the encoder contains a self-attention module, a cross-attention module, and a feedforward neural network module, which are used to extract the temporal dependencies of aging features and their coupling relationship with load conditions layer by layer, thereby improving the accuracy and stability of aging feature prediction across life stages.

[0042] The method uses only aging characteristic data collected during the early stage of the battery's first life phase, i.e., when the SOH decreases by no more than 5%, to train the operating condition-aware Transformer model. Without introducing measured aging data from the second life phase, it can predict the battery's aging characteristics and health status under different load conditions in the second life phase.

[0043] In this embodiment, a battery with no-load switching refers to a battery whose load condition setting is unique and constant throughout its entire first life stage, without any abrupt switching between different load condition levels. The determination of load condition switching is based on whether the current command level of the charging / discharging device undergoes a step change (e.g., switching from 16.0 A to 12.9 A). Instantaneous current fluctuations caused by measurement noise or the CC-CV charging protocol during charging and discharging are not considered load condition switching. During screening, it is only necessary to check whether the discharge current setting level for each cycle is unique to classify batteries into two categories: no-load current switching (for training) and variable load condition (for prediction and verification). The load current range covered by the training data of the no-load current switching battery includes the target load current value for the second life stage of the variable load condition battery. The batteries are lithium-ion batteries of the same chemical system and specification. The load current switching occurs within the normal operating current range of the battery and does not involve extreme conditions that lead to the emergence of novel degradation mechanisms.

[0044] Example 2 This invention provides a preferred embodiment using charge / discharge current as the operating condition parameter. Those skilled in the art will understand that replacing the current with other equivalent operating condition parameters (such as voltage or power) and making corresponding adjustments during model training and prediction also falls within the scope of this invention. This embodiment addresses the problems of changing operating conditions, limited available early lifespan data, and the difficulty of accurately predicting aging trajectories using traditional models during battery reuse. It proposes a cross-lifespan battery aging trend prediction method based on a current sensing mechanism, enabling aging trajectory prediction across lifespans from the first to the second life, providing reliable technical support for forward-looking health prediction and application scenario planning for batteries in reuse scenarios. Specifically, this embodiment provides a cross-lifespan battery aging trend prediction method based on a current sensing mechanism, including the following steps:

[0045] B1. Preprocess the multidimensional aging characteristic data of the battery, with each cycle of 10 iterations serving as a time step. Select batteries without operating current switching, and based on feature contribution analysis, correlation analysis, and prediction accuracy evaluation, select the optimal feature set for battery health status prediction;

[0046] B2. Construct a current-sensing Transformer model based on the cross-attention mechanism and train the model using battery data with no-operation current switching. B3. The early aging characteristics of the first life stage of the variable operating condition battery are fused with the operating current information to predict the evolution process of subsequent aging characteristics from the early stage. The evolution process includes the remaining life range of the first life stage and the aging characteristics changes of the second life stage after the operating current is switched. B4. Based on the key aging characteristics predicted in step B3, the battery health status is estimated using a Gaussian process regression model constrained by current conditions, thereby realizing the prediction curve of the aging trajectory across life stages from the first life to the second life.

[0047] Further: Step B1 includes the following sub-steps: B1. Denoise, normalize, and unify the time scale of the multidimensional aging characteristic data of the battery. B2. Use the characteristic contribution analysis method to assess the degree of influence of each aging characteristic on the health status prediction results; B3. Calculate the correlation indicators between various aging characteristics and battery health status; B4. Based on the prediction accuracy evaluation results of aging features, the optimal feature set for subsequent model training and health status estimation is selected.

[0048] Step B2 includes the following sub-steps: B21. Select the aging characteristic data of the battery with no operating current switching as the main input, and select the corresponding operating current, temperature characteristics and current change characteristics as condition inputs. B22. Input the main input and conditional input into the corresponding network structure of the current sensing Transformer model, and introduce a cross-attention mechanism into the model to perform information interaction and feature fusion on the main input and conditional input. B23. Based on the fused feature representation, the current-sensing Transformer model is trained to learn the mapping relationship between the aging feature evolution process and the operating current, providing a model basis for subsequent cross-life stage aging feature prediction.

[0049] The cross-attention mechanism in step B22 includes the following sub-steps: B221. Perform linear mapping or feature transformation on the input aging feature data to construct a query vector to characterize the evolution information of aging state; at the same time, perform feature encoding on the operating current information to construct corresponding key vectors and value vectors, so that the key vectors and value vectors can reflect the operating characteristics under different operating current conditions. B222. Based on the similarity relationship between the query vector and the key vector, calculate the attention weight between the query vector and the key vector, and normalize the attention weight to obtain a weight distribution that reflects the correlation between aging characteristics and operating current. B223. Based on the normalized attention weights, the value vectors are weighted and summed or weighted and combined to generate an aging feature representation that integrates the operating current information, thereby enhancing the model's ability to perceive the evolution law of aging features under different operating current conditions.

[0050] In step B3, the prediction of the subsequent aging feature evolution process adopts a recursive prediction method, generating aging feature prediction results step by step according to time steps, thereby forming a continuous aging feature sequence covering the remaining lifespan of the first life stage and the second life stage.

[0051] Step B4 includes the following sub-steps: B41. Based on the predicted aging characteristics, construct a Gaussian process regression model with the aging characteristics as input and the battery health status as output; B42. For different operating conditions and current conditions, establish corresponding Gaussian process regression sub-models with current condition constraints. B43. Based on the operating current information corresponding to the prediction stage, select a matching Gaussian process regression sub-model to estimate the battery health status. B44. Combine the estimated battery health status over time to form a cross-life stage aging trajectory prediction curve from the first life stage to the second life stage.

[0052] The multidimensional aging characteristic data includes at least one of time characteristics, incremental capacity (IC) characteristics, capacity characteristics, temperature characteristics, and current change characteristics. The operating current information is used as a conditional input and does not participate in the main feature sequence modeling, but is used to modulate the response of the aging characteristics to different load stresses.

[0053] The training process of the Transformer model includes: using the aging feature data of the battery in the first stage of its life cycle without operating current switching as the main input, using the corresponding operating current data as the conditional input, and introducing a cross-attention mechanism into the model to perform information interaction and feature fusion between the main input and the conditional input, thereby training the current-sensing Transformer model to establish the mapping relationship between the aging feature evolution process and the operating current, providing a model foundation for subsequent cross-life-stage aging feature prediction.

[0054] The battery health state is referred to as the battery's State of Health (SOH). In practical applications during the second life stage, current information under different operating conditions is input, and the prediction model iteratively predicts the battery's aging characteristics and corresponding health state, thereby achieving aging evolution prediction across life stages.

[0055] Compared with the prior art, this embodiment has the following advantages: (1) The cross-life stage aging prediction method proposed in this embodiment can accurately predict the aging behavior under different operating conditions in the second life stage by using only the early operating data of the first life stage of the battery, reducing the dependence on full life cycle test data. The early data of the first life stage referred to in this embodiment refers to the aging characteristic data collected from the battery from its initial state until the SOH decrease does not exceed 5%, which corresponds to the first 30 charge-discharge cycles in this embodiment. In this embodiment, "first life stage" refers to the period when the battery operates under the first set of operating current conditions. "Second life stage" refers to the period when the battery operates under the second set of operating current conditions, which are different from the first life stage. "Cross-life stage" specifically refers to the evolution and prediction of the aging law of the battery after switching from the first set of operating conditions to the second set of operating conditions. The method in this embodiment only involves the prediction of aging trends based on early data, and does not involve the physical maintenance links such as detection and reorganization that may be carried out after the battery ends the first life stage. (2) By introducing a cross-attention mechanism with operating current information as the input condition, explicit modeling between aging characteristics and operating conditions is realized, improving the model's prediction robustness and generalization ability in operating condition transition scenarios. (3) Through the decoupling design of aging characteristic prediction and health status mapping, the embodiments of the present invention enable the model to flexibly adapt to different second-life application scenarios, enhancing the practicality of the method in tiered utilization. (4) The embodiments of the present invention provide a high-precision, low-data-dependency technical solution for the application matching, operation management, and life assessment of retired batteries in the second life stage, which helps to improve the economy and safety of tiered utilization of retired batteries and promote the sustainable recycling and safe use of power batteries.

[0056] Example 3 See Figure 2 This embodiment provides a method for predicting battery aging trends across life stages based on a current sensing mechanism. This method models early aging data from the first stage of battery life to predict battery aging characteristics and health status under different operating conditions in the second stage of battery life. Specifically, it includes the following steps: Step G1: Preprocessing and feature selection of multidimensional aging feature data: In this embodiment, multidimensional aging characteristic data of the battery during the charge-discharge cycle is first obtained, and the aging characteristic data is time-series processed with each 10 cycles as a time step.

[0057] In a preferred embodiment, the multidimensional aging characteristic data includes, but is not limited to, at least one of time features, incremental capacity (IC) features, capacity features, temperature features, and current change features. Feature selection is an independent preprocessing step, completed before CACA-T model training and GPR health state estimation model construction. It is performed using complete first-life stage measured data of the battery with no-condition current switching, and the selection results are used as fixed feature configurations for all subsequent steps. Further, step G1 includes the following sub-steps:

[0058] G11. The acquired multidimensional aging feature data is denoised, normalized, and time-scale unified to eliminate the impact of dimensional differences and noise on model training. G12. Select battery samples without operating current switching and use the feature contribution analysis method to evaluate the degree of influence of each aging feature on the prediction of its own evolution law. G13. Calculate the correlation index between each aging characteristic and the battery health status to measure the statistical association between different characteristics and the evolution of health status; G14. Based on the accuracy evaluation of the prediction results of aging features, the optimal feature set for subsequent model training and health status estimation is selected.

[0059] Step G2: Construction and training of the current-sensing Transformer model: In this embodiment, a current-sensing Transformer model based on a cross-attention mechanism is constructed to explicitly model the coupling relationship between the battery aging characteristic evolution process and the operating current. In a preferred embodiment, step G2 includes the following sub-steps:

[0060] G21. Select the aging characteristic data of the battery without operating current switching as the main input, and select the corresponding operating current, temperature characteristics and current change characteristics as conditional inputs. The operating current information does not participate in the main feature sequence modeling, but is used as a conditional input to modulate the response of the aging characteristics to different load stresses. G22. Input the main input and conditional input into the corresponding network structure of the current sensing Transformer model respectively, and introduce a cross attention mechanism inside the model to realize information interaction and feature fusion between the main input features and the conditional input features. G23. Based on the fused feature representation, the current-sensing Transformer model is trained to learn the mapping relationship between the aging feature evolution process and the operating current, providing a model basis for subsequent cross-life stage aging feature prediction.

[0061] The cross-attention mechanism introduced in step G22 includes the following processing steps: G221. Perform linear mapping or feature transformation on the input aging feature data to construct a query vector for characterizing the evolution information of aging state; at the same time, perform feature encoding on the operating current information to construct corresponding key vectors and value vectors, so that the key vectors and value vectors can reflect the operating characteristics under different operating current conditions. G222. Based on the similarity relationship between the query vector and the key vector, calculate the attention weight between the query vector and the key vector, and normalize the attention weight to obtain a weight distribution that reflects the degree of correlation between aging characteristics and operating current. G223. Based on the normalized attention weights, the value vectors are weighted and summed or weighted and combined to generate an aging feature representation that integrates the operating current information, thereby enhancing the model's ability to perceive the evolution law of aging features under different operating current conditions.

[0062] Step G3: Prediction of the evolution of aging characteristics across life stages: In this embodiment, the early aging characteristics of the first life stage of the variable operating condition battery and the corresponding operating current information are input into a trained current-sensing Transformer model to predict the subsequent aging characteristic evolution process. In a preferred embodiment, step G3 adopts a recursive prediction method, generating aging characteristic prediction results step by step according to time steps, thereby forming a continuous aging characteristic sequence covering the remaining lifespan of the first life stage and the second life stage after the operating current switch.

[0063] Step G4: Health state estimation based on current condition constraints: In this embodiment, based on the key aging characteristics predicted in step G3, a battery health status estimation model is constructed to achieve aging trajectory prediction across life stages. In a preferred embodiment, step G4 includes the following sub-steps:

[0064] G41. Based on the predicted aging characteristics, construct a Gaussian process regression model with the aging characteristics as input and the battery health status as output. G42. For each operating current level in the training set, using the complete first-life data of a battery operating at a constant operating current without operating condition switching as training samples, establish a Gaussian process regression sub-model with current constraints, forming a GPR sub-model library. The size of the model library is consistent with the number of operating current levels covered by the training set. For operating current values ​​encountered in the prediction stage, if they are already included in the model library, the corresponding sub-model is directly selected; if they are intermediate current values, nearest neighbor sub-model matching or adjacent sub-model hyperparameter interpolation is used. Both of these implementation methods are within the protection scope of this invention.

[0065] G43. Based on the operating current information corresponding to the prediction stage, select the matching Gaussian process regression sub-model to estimate the battery health status. G44. Combine the estimated battery health status over time to form a cross-life stage aging trajectory prediction curve from the first life stage to the second life stage.

[0066] In practical applications, by inputting current information under different operating conditions, the prediction model is used to iteratively predict the aging characteristics and corresponding health status of the battery, thereby realizing aging evolution analysis across life stages. It should be noted that the long-term prediction capability across life stages in this embodiment of the invention is based on the following assumptions, which users should verify before application: (1) The dominant aging mechanism type of the battery remains consistent throughout the entire prediction cycle, and there is no fundamental mechanism change such as lithium plating caused by out-of-specification operating conditions; (2) The target operating current value of the second life stage is included in the operating range covered by the training set; (3) The target battery and the training battery belong to the same chemical system and specification model. Under the condition that the above conditions are met, this method has been experimentally verified to achieve a mean absolute percentage error (MAPE) accuracy of less than 1.31% within the complete cross-life stage prediction interval. For application scenarios that do not meet the above conditions, or where the prediction cycle significantly exceeds the experimental verification range of this embodiment of the invention, the prediction results are uncertain. It is recommended to regularly update and correct the model in conjunction with online observation data of the second life stage to maintain the reliability of the prediction accuracy.

[0067] The following further illustrates the implementation method of specific embodiments of the present invention.

[0068] Feature filtering: In the problem of predicting battery aging across its lifespan, different discharge conditions, current levels, and cycling stages significantly affect battery voltage, capacity, and their derived characteristics. This invention addresses the problem of predicting battery health status across its lifespan under changing operating conditions. To effectively characterize the gradual aging process of the battery while reducing computational complexity, this embodiment discretizes the original cycling data, dividing each 10 consecutive charge-discharge cycles into a time step. The operating data sequences for the first and second lifespan stages of the battery can be represented as follows:

[0069] (1) (2) in, This represents the principal eigenvector used to characterize the intrinsic aging properties of a battery. This represents a conditional feature vector that reflects external operating conditions. and These represent the total number of time steps corresponding to the first and second life stages of the battery, respectively. This indicates the index of the current time step.

[0070] The main features include: time features F0–F3 measured at 5.0 V, 5.5 V, 6.0 V, and 6.5 V during the discharge phase; IC curve features F4–F6, corresponding to the IC value at 6.0 V, the peak value of the IC curve, and the voltage corresponding to the peak value of IC, respectively; and discharge capacity features F7–F10, which are the capacity values ​​at 5.0 V, 5.5 V, 6.0 V, and 6.5 V, respectively.

[0071] The conditional features include: discharge current and cycle number representing instantaneous load level and operation process; temperature-related features F11–F15, including maximum discharge temperature, time taken to reach maximum discharge temperature, initial temperature rise rate, battery temperature at the end of discharge, and cumulative heat generation; and current change features F16–F17, including maximum current change rate and its corresponding time.

[0072] By utilizing early operational data from the first stage of a battery's lifespan, it is possible to extrapolate and predict operational data from the later stages of the first and second lifespans. Therefore, the cross-lifespan SOH prediction problem can be formalized as learning a mapping function. To satisfy:

[0073] (3) in, Indicates at time step The predicted battery health status covers the period from the first life stage to the second life stage. This indicates the starting time step for the prediction.

[0074] To reduce the interference of redundant features on model training and improve the model's ability to represent key aging mechanisms, this invention introduces a feature screening step before model training to effectively extract and screen multi-dimensional aging features. Considering the varying contributions of different aging features to the evolution of battery health status, this invention comprehensively evaluates input features based on three aspects: feature contribution, feature relevance, and prediction accuracy. A method based on Shapley values ​​or feature importance assessment is used to quantify the contribution of each aging feature in the health status prediction process. By statistically analyzing the average contribution values ​​of different features across multiple battery samples and different cycle stages, a subset of key features with stable influence on the SOH evolution trend is selected. The screened aging features are used for subsequent cross-lifecycle prediction model training, thereby reducing redundant feature interference and improving model generalization performance.

[0075] Feature contribution was quantified using SHapley Additive exPlanations (SHAP) analysis, while the statistical correlation between features and battery health status was measured using the Pearson correlation coefficient, thereby identifying the key input features that are most informative for predicting health status.

[0076] The SHAP method originates from the Shapley value in game theory. Its core idea is to assign an importance score to each input feature to characterize its marginal contribution to the model's output. For an initial baseline prediction model trained on full-lifespan data of a battery with no-current switching capability, using all candidate features as input... and input samples , No. The SHAP value of each feature is defined as follows:

[0077] (4) in, Represents the set of all features. A set of feature indices, Indicates using only a subset of features The prediction results of the time model.

[0078] feature The global importance of a sample is measured by the mean of its absolute SHAP values ​​across all samples, and is defined as: (5) in, Features To further characterize the statistical correlation between each feature and the SOH, and to assess the global importance of each feature's time series, we will analyze the time series data of each feature. With the corresponding SOH sequence Calculate the Pearson correlation coefficient , is represented as: (6) in, and These represent the mean values ​​of the characteristic time series and the SOH series, respectively.

[0079] By combining the SHAP contribution analysis results, Pearson correlation coefficient, and the predictive performance of features in the iterative prediction process, the input feature structure is optimized and filtered to improve the stability and robustness of the model in prediction scenarios across different working conditions and life stages while ensuring prediction accuracy.

[0080] Current sensing for predicting battery aging characteristics across life stages: After completing feature screening, this embodiment of the invention further proposes a cross-life-stage aging trend prediction method (CACA-T) for current-sensing Transformers, which is used to predict the subsequent aging process and the evolution trend of the second life stage using early data of the first life stage of the battery.

[0081] Cross-attention is a mechanism in neural networks that allows a model to focus on a specific part of an input based on information from another input. In practical applications, this mechanism is particularly effective when the relevance of certain features in one data stream depends on additional information from another data stream. To explicitly characterize the impact of operating current on the evolution of aging features, this embodiment introduces a cross-attention mechanism into the Transformer model to achieve information interaction between the main input aging features and the conditional input operating current.

[0082] The CACA-T model introduces a current modulation mechanism. By explicitly embedding real-time operating current information during attention calculation, the model can achieve adaptive feature interaction under different load conditions, thereby effectively improving its cross-operating condition aging modeling capability. This embodiment of the invention uses MAPE as the training loss function; in other embodiments, mean squared error (MSE) or other alternative loss functions can be used. The optimizer employs the Adam optimizer with weight decay. In a specific embodiment of the invention, the learning rate is set to the order of 0.001, and the weight decay coefficient is 1 × 10⁻⁶. -5 The total number of training epochs ranges from several hundred to one thousand, with a hidden dimension of 128, 8 attention heads, and 3 encoder layers. All of these hyperparameters can be adjusted according to the specific application scenario and dataset characteristics. A cosine annealing (Cosine AnnealingLR) learning rate scheduling strategy is introduced during training to smoothly decay the learning rate from its initial value to near zero during the training cycle, thereby improving model convergence stability. Simultaneously, a gradient pruning strategy is employed to limit the upper bound of the gradient norm, preventing gradient explosion during training. The input to this network consists of the main input features. With conditional input features Together they constitute the first dimension of the conditional characteristics, which is the real-time current. The main input refers to the aging feature sequence collected in the early stages of the battery's first life phase. It describes the evolution of the battery's internal aging state and is a sequence of shapes... The tensor, in which For sequence length, The primary feature dimension is defined as follows. Conditional input refers to the operating current information for the corresponding time period or prediction phase, used to characterize the operating environment under different load conditions. It is a shape-based input... The tensor, in which This is the conditional feature dimension. All input features are first projected onto a unified hidden dimension through a linear mapping. :

[0083] (7) (8) in, The hidden representation of the main input after linear mapping. Let W be the hidden representation of the conditional input after linear mapping, W be the learnable weight matrix, and b be the learnable bias vector.

[0084] Extracting real-time current scalars from conditional features: (9) in, For conditional input tensors The real-time operating current scalar extracted is used, taking the value of the first dimension (index 0) of the conditional feature. This current is nonlinearly mapped via a lightweight fully connected network to generate modulation parameters for the query vector and key vector. The fully connected network consists of a linear transformation layer (input dimension 1, output dimension...). In the embodiments of the present invention The output dimension is 512) and the GELU activation function layers are stacked sequentially, containing a total of weight matrices. and corresponding bias vector The fully connected network is jointly trained end-to-end with the linear mapping layer, self-attention encoder, cross-attention layer, and output layer of the CACA-T model, sharing the same loss function, optimizer, and learning rate scheduling strategy, without requiring a separate pre-training stage or additional supervision signals. The modulation parameter is divided into query scaling coefficients. Query translation items Key scaling factor and key translation terms It operates on both the primary feature and the conditional features in a feature-by-feature manner:

[0085] (10) (11) (12) in, This represents element-wise multiplication. K is the query vector, K is the key vector, and V is the value vector.

[0086] After completing the current-sensing modulation, the cross-attention mechanism is calculated as follows: (13) in, , , All shapes are , To hide dimensions (in the embodiments of this invention) ); The dimension of a single attention head is defined as follows: ,in Number of attention heads (in this embodiment of the invention) ,therefore ); This is a scaling factor used to prevent the softmax gradient from vanishing due to an excessively large dot product result. The cross-attention mechanism is implemented using multi-head attention, with the query vector... Key vector and value vector Before inputting into the multi-head attention module, it passes through the following paths respectively: The learnable linear projection weight matrix is ​​mapped to the subspace of each attention head. Both the projection weight matrix and the output projection matrix are trainable parameters of the model, automatically optimized during training via backpropagation. The multi-head attention output shape is... Take the output of the last time step. .

[0087] During the training phase, the model is trained using only the first-stage aging data without any load switching to learn the basic laws governing the evolution of aging characteristics. During the prediction phase, the operating current from the prediction phase is used as a conditional input to guide the model in extrapolating aging trends under different load conditions.

[0088] In this embodiment of the invention, the model adopts a stepwise recursive prediction method, that is, it takes the actual or predicted aging characteristics of the previous stage and the operating current of the next stage as inputs and outputs the aging characteristics of the next stage, so as to realize the continuous prediction of aging characteristics across life stages from the first life stage to the second life stage.

[0089] Current-constrained Gaussian process regression estimation of battery health status: By constructing a Gaussian Process Regression (GPR) model library for different current conditions and incorporating a trend correction mechanism, this method ensures that the health status prediction results across life stages are physically reasonable and numerically stable and reliable. The input variables for the GPR model are:

[0090] (14) in, The input feature vector of the GPR model is the filtered feature vector. It consists of several key aging characteristics; For feature index, The features selected from all candidate features The set of indices of the optimal features; This represents the number of features used for GPR modeling after filtering. At each time step... The input sample is composed of the aging feature values ​​corresponding to this feature space. ,in express A 3D real space, meaning each input sample is a 3D real space. .

[0091] The corresponding output target variable is the battery health state (SOH). (15) in, For the complete prediction interval across life stages, the SOH sequence is provided. The SOH values ​​corresponding to each time step; This is the starting time step index for the recursive prediction, corresponding to the end time of the early actual observation data of the target battery; This is the time step index corresponding to the end of the second life phase; Indicates the time step from the start time Towards the end of the second life stage A complete SOH prediction sequence. For different current operating conditions, corresponding GPR model sets are constructed. :

[0092] (16) in, This represents the current under current conditions and belongs to a predefined set of currents. The current collection It consists of all operating current ranges covered by the battery with no operating current switching in the training set, as described in this embodiment of the invention. The operating current value encountered during the prediction phase. Already included in the collection When in the middle, directly select the corresponding... The sub-model performs health status estimation; when encountering a set during the prediction phase... When an intermediate current value is not included in the set, the nearest neighbor matching strategy is adopted, that is, a set is selected. The GPR sub-model corresponding to the operating current with the smallest absolute difference from the target current value is used for estimation, or the hyperparameters of two adjacent sub-models are linearly interpolated according to the current value to generate an interpolated GPR model corresponding to the intermediate current. This processing strategy ensures that the method has operable health state estimation capability for any current value within the operating current range of the training set. Each GPR model follows the Gaussian process assumption:

[0093] (17) in, The operating current condition is The Gaussian process regression function under current conditions represents the regression function of the process under current conditions. Constrained by input features The mapping relationship to SOH. The distribution of a Gaussian process is completely determined by the mean function and the covariance function. It is a mean function. The covariance function is set to 0, and the mean function is set to 0. The covariance function can be either a radial basis function kernel (RBF kernel) or a Matérn kernel (smoothness parameter). (The value can be 1.5 or 2.5). The specific kernel function type is automatically determined by a random search method that maximizes the log-marginal likelihood function on the validation set.

[0094] By maximizing the log-marginal likelihood function for the hyperparameters of the Gaussian process Optimize: (18) in, Operating current conditions The optimal hyperparameter vector of the corresponding GPR sub-model includes learnable parameters such as kernel function length scale, signal variance, and noise variance. Indicates hyperparameters Find the parameter values ​​that maximize the objective function. The value of the log-marginal likelihood function. The operating current condition is The measured SOH sequence corresponding to the non-switching battery, For the corresponding input aging feature matrix, for the test battery At the moment of switching of operating current : (19) (20) (twenty one) in, This refers to the operating current value corresponding to the first life stage before the operating current switch, which belongs to the current set. ; This refers to the target operating current value corresponding to the second life stage after the operating current switch, which belongs to the current set. ; To cover from the start time of recursive prediction Towards the end of the second life stage Complete predicted SOH sequences across life stages; Indicates the time step before current switching. Cause The predicted SOH obtained from the sub-model estimation; Indicates the time step after current switching. Cause The predicted SOH obtained from the sub-model estimation.

[0095] Experimental Example (1) Description of experimental data In this embodiment, the battery aging experimental data used for model validation comes from a publicly published lithium-ion battery accelerated life testing dataset, which was released by K. Fricke et al. in the paper "Accelerated Life Testing Dataset for Lithium-Ion Batteries with Constant and Variable Loading Conditions" (International Journal of Prognostics and Health Management, 2023). This dataset contains multiple sets of full life-cycle test data for battery modules operating under constant and variable load conditions, and can be used for aging behavior analysis across operating conditions and life stages.

[0096] In the experimental test, 18650 type lithium-ion battery cells with a rated capacity of 2.5Ah were selected, and two battery cells were assembled in series to form a battery module. A total of six battery modules were selected for the experimental test. The charging conditions and resting time were set as follows: constant current charging was used, and the charging current was 3A; a 10-minute resting period was set before each charging and discharging process.

[0097] In this embodiment, the experimental subjects are strictly divided into two parts: a training set and a test set. The two sets are completely isolated in terms of data usage. The test set data is not visible throughout the entire process of model training, feature selection, and hyperparameter tuning. It is only used for final performance evaluation after all the above steps are completed.

[0098] The training set includes eight battery modules without operating current switching. Throughout the first lifespan, a single constant discharge condition is used, with discharge currents of 9.3A (modules 1 and 11), 12.9A (modules 22 and 31), 14.3A (modules 23 and 52), and 16.0A (modules 0 and 10), each current level containing two modules. The training set data is used for the following three tasks: (a) SHAP contribution analysis, Pearson correlation calculation, and feature autoregression prediction accuracy evaluation during the feature selection phase; (b) training and hyperparameter tuning of the CACA-T model; and (c) training of the GPR sub-models under various current constraints. Feature selection and hyperparameter tuning are performed within the training set and do not involve any test set information.

[0099] The test set includes six battery modules with varying operating conditions (modules 2, 12, 24, 32, 33, and 53), which undergo operating current switching throughout their lifecycle. The specific load operating current parameters for each battery module in the test set are shown in Table 1. In the second lifecycle application, except for module 53, the operating current of the above battery modules is generally lower than the discharge condition of their corresponding first lifecycle. The test set data is only used to evaluate the model's cross-lifecycle prediction performance on completely unseen batteries and is not used in any training or tuning steps.

[0100] It should be noted that the dataset used in this experiment is limited in size (14 modules in total), and the battery modules in both the training and test sets are from the same batch. To strictly avoid data leakage, the following measures are taken in this embodiment of the invention: feature selection, model training, and hyperparameter optimization all use data from the 8 non-switching batteries in the training set; data from the 6 variable-condition batteries in the test set are accessed only after the above steps are completed, and are only used for the final performance evaluation; the prediction errors reported in the experimental results (Table 2) are the model's true generalization performance on the test set, not the fitting error of the training set. The above data partitioning rules ensure the objectivity and reliability of the experimental evaluation results.

[0101] Table 1

[0102] (2) Analysis of the prediction results of battery aging trends across life stages Based on the proposed feature selection and cross-life stage battery aging trend prediction method, aging trajectory prediction results were obtained using different feature inputs. Furthermore, the prediction curves obtained by the cross-life stage method proposed in this embodiment of the invention are compared and analyzed with those obtained by other methods to evaluate the prediction performance of different feature combinations and different prediction models in operating condition switching scenarios.

[0103] First, the role of input features in the iterative prediction process is quantitatively evaluated using feature contribution analysis. The results show significant differences in the contribution of different feature groups to model prediction. Capacity, temperature, and current features exhibit high importance in the prediction process, while time and incremental capacity features have relatively low contributions. The contribution values ​​of some features are close to zero, indicating limited impact on the prediction results and making them feasible for removal in subsequent feature selection. Further correlation analysis between features and health status reveals that although the correlation coefficient between temperature features and health status is not ideal, it significantly affects the model output during feature prediction, effectively improving the model's prediction stability under complex conditions. In contrast, time and capacity features demonstrate both high prediction accuracy and strong correlation with health status, making them suitable as key input features for the subsequent health status mapping module. The correlation distribution of incremental capacity features is relatively discrete, resulting in insufficient mapping stability and a small contribution to overall prediction performance.

[0104] Based on the above analysis results, this embodiment retains capacity, time, and temperature features in the feature configuration, while eliminating incremental capacity features, thereby reducing model complexity and computational cost while ensuring prediction accuracy. Figure 3 As shown, the horizontal axis represents the battery module number, and the vertical axis represents MAPE (%). Each group, from left to right (dark blue, light blue, gray, red), represents four types of feature inputs: unconditional feature input, all feature inputs, no IC feature input, and no F11, F14, or F15 feature inputs. The estimation error of the model is measured by MAPE. The four feature input schemes for comparison are described in detail below: (a) All feature input: all candidate features including time features, incremental capacity features, capacity features, temperature features, and current change features are used as model input; (b) Unconditional feature input: only static aging features without conditional features such as operating current, temperature, and current change are used as input to verify the contribution of conditional features to cross-operating condition prediction; (c) No IC feature input: incremental capacity features are removed from all features, and all other features are retained; (d) No F11, F14, and F15 feature input: the maximum discharge temperature (F11), battery temperature at the end of discharge (F14), and accumulated heat generation (F15) features with large prediction errors are removed to form the optimal feature set finally adopted in the embodiments of the present invention. The comparative experimental results show that in the cross-operating condition prediction scenario, the introduction of operating condition-related features can significantly improve the prediction accuracy; at the same time, after removing the incremental capacity features with low contribution, the model prediction performance is further improved, verifying the rationality and effectiveness of the feature selection strategy.

[0105] In the model comparison experiment, the predictive capabilities of various sequence prediction models for battery aging trends under cross-life-stage operating condition switching conditions were compared and analyzed. To ensure the fairness of the comparison experiment, all comparison models in Table 2 (including T-GPR, LSTM-GPR, GRU-GPR, CACA-T-RF, T-RF, LSTM-RF, and GRU-RF) used the exact same optimal feature set as the CACA-T-GPR model in the embodiment of this invention as input. Since Gaussian process regression does not significantly improve the prediction performance compared to Random Forest (RF), and the prediction trends of the two methods are highly consistent, only the prediction curve based on GPR is plotted for ease of demonstration. Figures 4a to 4f As shown, where, Figures 4a to 4f The battery modules represented are numbered 2, 12, 32, 33, 24, and 53, respectively, with current changes of 16.0A→12.9A, 16.0A→9.3A, 16.0A→14.3A, 16.0A→12.9A, 16.0A→9.3A, and 14.3A→16.0A. Figures 4a to 4f The horizontal axis represents the number of cycles, the vertical axis represents SOH, the vertical purple dashed line represents the prediction starting point, the vertical gray dashed line represents the current switching point, the dark blue dotted line represents the actual SOH value, the green dotted line represents T-GPR, the light blue dotted line represents GRU-GPR, the red dotted line represents the CACA-T-GPR model used in the method of this embodiment, and the yellow dotted line represents LSTM-GPR. The comparison results of the prediction error MAPE (%) of different model methods are listed in Table 2. The results show that by introducing the current-sensing cross-attention mechanism and the recursive prediction strategy, the method of this embodiment ( Figures 4a to 4f The red dotted line (indicated by the current) accurately depicts the impact of operating current variations on the aging rate, effectively addressing the shortcomings of existing methods in terms of accuracy and stability during cross-operating condition and cross-lifecycle predictions. The current-sensing cross-attention mechanism introduces an operating current condition modulation module into the standard cross-attention framework. This module uses a lightweight fully connected network to map the operating current scalar into scaling and translation modulation parameters for the query and key vectors, enabling the attention weight calculation to adaptively respond to the current operating current conditions. This modulation module is jointly trained end-to-end with the rest of the model. Ablation experiments show that, compared to the standard cross-attention baseline without the modulation module, the prediction accuracy is improved by 46% to 82% after introducing the current-sensing modulation mechanism, validating its effectiveness.

[0106] Table 2

[0107] The results show that Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs) can fit the battery aging trend well during the first stage of its lifespan. However, after the operating conditions change, the prediction results generally lag or deviate from the aging rate, making it difficult to reflect the impact of the operating condition changes on the aging process in a timely manner. This is mainly because the above models do not explicitly model current changes, but only treat them as ordinary input features, resulting in insufficient perception of long-term evolution characteristics across operating conditions. Models based on the Transformer structure improve the modeling ability of time-series features to some extent, but there are still prediction biases near the operating condition switching points, making it difficult to accurately capture the sudden changes in aging rate caused by changes in operating current. In contrast, the cross-attention prediction model with current sensing capability proposed in this embodiment can closely track the actual health status change trend throughout the entire lifespan prediction process, achieve a smooth transition at the operating condition switching points, and quickly adjust the aging slope to match the new operating conditions.

[0108] Experimental results show that the cross-life-stage battery aging trend prediction method proposed in this invention can effectively adapt to sudden changes in operating conditions, maintain high prediction accuracy and stability over a long prediction period, and has good generalization ability. It is especially suitable for battery health status prediction scenarios where the operating conditions of the second life stage are significantly different from those of the first life stage.

[0109] Advantages of the embodiments of the present invention compared with traditional methods: (1) The present invention introduces a current sensing Transformer structure based on cross-attention mechanism, which explicitly models the operating current as conditional information, realizes the deep correlation between the aging feature evolution process and the operating condition change, effectively overcomes the problem that the traditional method only regards the current as a normal input feature and is not sensitive enough to sudden changes in operating conditions, and can still maintain stable prediction performance across life stages and operating conditions.

[0110] (2) The embodiments of the present invention construct an optimized set of aging features by combining feature contribution analysis, correlation analysis and prediction accuracy evaluation based on the CACA-T model. This reduces model complexity and computational cost while improving the accuracy and robustness of feature prediction and health status estimation.

[0111] (3) The embodiments of the present invention use a recursive method to predict aging characteristics across life stages, and combine a Gaussian process regression model constrained by current conditions to map health status, so as to realize continuous and smooth aging trend prediction from the first life to the second life, effectively avoiding prediction jumps at the working condition switching point.

[0112] The embodiments of the present invention are superior to the prior art in terms of adaptability to complex working conditions, long-term prediction stability, and generalization ability across life stages.

[0113] Specific applications of the embodiments of the present invention: The present invention proposes a method for predicting the aging trend of batteries across life stages based on a current sensing mechanism. This method can be widely applied to various fields such as early planning and application adaptation for battery cascade utilization, value assessment of cascade utilization, electric vehicle battery management, and operation and maintenance of energy storage systems. It provides technical support for the early planning, application adaptation, and operation management of power battery cascade utilization and has significant engineering practical value and industrial promotion prospects.

[0114] This invention, without relying on additional accelerated aging tests or pulse tests, can predict the remaining lifespan of a battery under changing operating conditions and the aging trend of its second life stage using only early operational data from the first stage of the battery's life. This provides a scientific basis for the health assessment and application matching of retired batteries. By introducing a current-sensing cross-attention mechanism and a recursive prediction strategy, this invention can accurately characterize the impact of changes in operating current on the aging rate, effectively solving the problems of insufficient accuracy and poor stability in cross-operating condition and cross-life stage predictions of existing methods. The current-sensing cross-attention mechanism introduces an operating current condition modulation module on the basis of the standard cross-attention framework. Through a lightweight fully connected network, the operating current scalar is mapped to scaling and translation modulation parameters of the query vector and key vector, enabling the attention weight calculation to adaptively respond to the current operating current condition. This modulation module is jointly trained end-to-end with the rest of the model. Ablation experiments show that, compared with the standard cross-attention baseline without the modulation module, the prediction accuracy is improved by 46% to 82% after introducing the current-sensing modulation mechanism, verifying the effectiveness of the mechanism. This invention accurately characterizes the impact of operating current variations on aging rates, effectively addressing the shortcomings of existing methods in terms of accuracy and stability during cross-operating condition and cross-life stage predictions. Specifically, it includes:

[0115] 1. Early Planning and Application Adaptation for Battery Second-Life Utilization: This invention provides a method for predicting and planning the aging trends of power batteries across their life stages before retirement. By introducing multi-dimensional aging characteristics and operating current information early in the first life stage of the battery, the evolution of subsequent aging characteristics and changes in health status can be predicted, allowing for the assessment of the battery's potential for secondary use before retirement. This method can identify the degradation rate and remaining usable life of different batteries in the second life stage in advance, providing a basis for the selection, classification, and application planning of second-life utilization targets. By explicitly introducing operating current as a constraint, this invention can predict the aging trend of batteries under different load conditions, solving the problem of drastically shortened lifespan due to mismatched application conditions in existing second-life utilization methods. This method is particularly suitable for application scenario matching decisions in the early stages of second-life utilization, such as communication base station energy storage, distributed energy storage, and grid peak shaving and valley filling, helping to improve the safety, economy, and overall service efficiency of second-life utilization systems.

[0116] 2. Electric Vehicle Sector: As a core component of electric vehicles, the health of the battery directly affects the safety and economy of the entire vehicle. This invention enables long-term prediction of battery aging trends under different charging and discharging currents and usage scenarios during vehicle operation. This provides technical support for battery remaining life assessment, vehicle maintenance decisions, and retirement planning, helping to improve the intelligence level of electric vehicle battery management and extend battery life.

[0117] 3. Applications in Energy Storage Systems and Smart Grids: In smart grids and energy storage power stations, batteries often undergo frequent operating condition switching and complex load changes. This invention can predict the aging behavior of energy storage batteries under different operating strategies, based on limited historical data. This assists in optimizing the operation and lifespan management of energy storage systems, improving system reliability and economy, and solving the problems of insufficient battery health status assessment data and high testing costs in existing energy storage scenarios.

[0118] The embodiments of this invention can guide the prediction of aging trends and health status assessment during the recycling and reuse of retired batteries without increasing additional testing costs, demonstrating good scalability and versatility. With the rapid development of electric vehicles, energy storage systems, and renewable energy, the embodiments of this invention have broad application prospects in fields such as battery recycling and cascade utilization, electric transportation, and smart grids.

[0119] The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0120] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, several equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or purpose, should be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the aging trend of batteries across their life stages, characterized in that, Includes the following steps: S1. Data preparation: Obtain aging characteristic data and corresponding load condition information of the early stage of the first life of the battery to be predicted. S2. Model Training: Construct a working condition perception model based on a cross-attention mechanism, using no-load switching battery data as the main input and corresponding load working condition information as the condition input, and train the working condition perception model. S3. Aging Feature Prediction: Using the operating condition perception model, the subsequent aging feature evolution sequence of the battery to be predicted from the early stage of the first life is output in a time-series prediction manner. The subsequent aging feature evolution sequence covers the remaining life of the first life stage and the aging feature changes of the second life stage after the load condition switch. S4. Health Status Estimation: Based on the subsequent aging feature evolution sequence, estimate the health status of the battery to generate a cross-life stage aging trajectory prediction curve from the first life stage to the second life stage.

2. The battery aging trend prediction method across life stages as described in claim 1, characterized in that, Before step S1, there is also a feature selection step: based on feature contribution, correlation and prediction accuracy evaluation, a feature set for health status prediction is selected from the multidimensional aging feature data of the battery; the aging feature data obtained in step S1 is the data in the feature set; the multidimensional aging feature data includes at least one of time feature, incremental capacity feature, capacity feature, temperature feature and current change feature.

3. The method for predicting battery aging trends across life stages as described in claim 1, characterized in that, After acquiring the data in step S1, a preprocessing step is also included: aggregating the aging characteristic data of the early stage of the first life of the battery to be predicted and the corresponding load condition information at a predetermined time step interval; the time step interval is configured such that the change in health status within a single time step does not exceed 2%; the time step interval is 10 charge-discharge cycles as one time step.

4. The battery aging trend prediction method across life stages as described in claim 1, characterized in that, In step S2, the working condition perception model is a Transformer model; step S2 includes the following sub-steps: S21. Perform linear mapping or feature transformation on the main input to construct a query vector; perform feature encoding on the conditional input to construct a key vector and a value vector; S22. Calculate the similarity between the query vector and the key vector, and obtain the attention weight after normalization; S23. The value vector is weighted and summed or weighted and combined according to the attention weight to generate an aging feature representation that integrates load condition information.

5. The method for predicting battery aging trends across life stages as described in claim 1, characterized in that, In step S3, a recursive time-series prediction method is adopted, and before the time-series prediction begins, the operating condition perception model is fine-tuned using real observation data from the early stage of the first life of the battery to be predicted; in step S4, the estimation of the battery's health status is specifically performed by using a Gaussian process regression model constrained by load operating conditions.

6. The method for predicting the aging trend of batteries across life stages as described in claim 5, characterized in that, The estimation using the Gaussian process regression model constrained by load conditions includes the following sub-steps: S41. For different load conditions, establish corresponding Gaussian process regression sub-models based on complete no-load switching battery data to form a model library. S42. For each time step in the subsequent aging feature evolution sequence, select a matching Gaussian process regression sub-model from the model library based on its corresponding load condition information. S43. Using the selected Gaussian process regression sub-model, estimate the aging characteristics of this time step as a healthy state.

7. The method for predicting the aging trend of batteries across life stages as described in any one of claims 1 to 6, characterized in that, The load condition information includes at least one of charging / discharging current, voltage, or power.

8. A battery management system, characterized in that, include: A processor and a memory; the memory stores a computer program; the processor is configured to execute the computer program to implement the battery life-stage aging trend prediction method as described in any one of claims 1 to 7.

9. A battery health status prediction system, characterized in that, include: The data acquisition module is configured to collect battery aging characteristic data and load condition information; The data processing and prediction module is configured to execute the battery cross-life stage aging trend prediction method as described in any one of claims 1 to 7, and generate a cross-life stage aging trajectory prediction curve. The results output module is configured to output the predicted aging trajectory curve across life stages.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for predicting battery aging trends across life stages as described in any one of claims 1 to 7.

Images