A method for estimating the state of health of a lithium battery for vehicles and predicting future degradation

Abstract

The application discloses a kind of vehicle lithium battery health state cross-domain estimation and future degradation prediction method, belong to battery management technical field.The method includes: obtaining source domain with label data, target domain unlabelled data and to be predicted data;Each data is preprocessed, and the discrete characteristic sequence of each data is obtained;The characteristic sequence of source domain and target domain is input to evaluation module, and the gradient inversion layer is combined with domain discriminator and SOH evaluator to carry out confrontation training, so that feature extractor I learns domain invariant aging feature, and outputs health state estimation value;Then the health state estimation value is input as pseudo label into prediction module, to iteratively predict future health state trajectory in autoregressive mode.The overall framework of health state cross-domain estimation and future degradation prediction integration is adopted in the application, and the evaluation module learns domain invariant aging feature based on domain confrontation training, which significantly improves the accuracy of lithium battery health state estimation and future degradation trajectory prediction.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, and in particular to a method for cross-domain estimation of the health status of automotive lithium batteries and prediction of future degradation. Background Technology

[0002] With the rapid development of new energy vehicles and energy storage systems, automotive lithium-ion batteries, as the core energy carrier, play a crucial role in ensuring safe vehicle operation, improving energy management, and extending battery life. Existing technologies for estimating SOH in lithium batteries include ampere-hour integration methods, internal resistance analysis methods, methods based on equivalent circuit models, methods based on electrochemical models, and data-driven methods based on machine learning and deep learning. Among these, data-driven methods are widely used in battery aging assessment research due to their strong nonlinear fitting ability, low dependence on mechanistic modeling, and online application capabilities, and have achieved significant progress in specific application scenarios.

[0003] However, existing technologies still have the following shortcomings in real-world automotive scenarios: First, models are trained based on data from specific battery types or operating conditions. When battery types or operating conditions change, the estimation accuracy of existing methods drops significantly. Although the fusion strategy of signal analysis and data-driven approaches has alleviated this problem to some extent in recent years, real-world engineering applications involve diverse battery types and complex, variable operating conditions. Modeling each specific scenario individually is costly and time-consuming, making it difficult to meet practical engineering needs, and the cross-domain adaptability problem remains unresolved. Second, existing methods generally rely on complete battery cycle data, while batteries in actual operation are mostly in a shallow charge-discharge state, making it difficult to obtain complete charge-discharge cycles and accurate capacity labels. A large amount of unlabeled data is not effectively utilized. Finally, most existing prediction methods can only achieve real-time assessment of historical SOH and cannot predict long-term future SOH based on known historical SOH. This fitting of the "current state" rather than inference of the "future degradation process" is difficult to meet the engineering needs of proactive health management and safety warnings throughout the battery's entire life cycle in practical engineering.

[0004] Therefore, there is an urgent need to design a cross-domain estimation method for the state of health (SOH) of automotive lithium batteries and a method for predicting future degradation to improve the accuracy of SOH estimation and maintain a high accuracy of future degradation prediction under limited observation data conditions. This would better meet the application needs of automotive lithium batteries for health assessment and long-term state prediction in complex service environments. Summary of the Invention

[0005] To address the shortcomings of the prior art, this invention provides a method for cross-domain estimation of the health status of automotive lithium batteries and prediction of future degradation. It aims to solve the problems of weak cross-domain generalization ability, insufficient utilization of unlabeled data, and inability to effectively predict future degradation trends in existing automotive lithium battery health status estimation and degradation prediction technologies.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for cross-domain estimation of the state of health and prediction of future degradation of automotive lithium batteries includes the following steps: Step 1: Obtain the source domain labeled data, target domain unlabeled data, and target domain unlabeled data to be predicted of the lithium battery. Preprocess the obtained data to obtain the discrete feature sequence of each data. Step 2: Using the discrete feature sequences of the source domain labeled data and the target domain unlabeled data, a pre-built evaluation model is jointly trained. The evaluation model includes a feature extractor I, a domain discriminator, and a SOH estimator. The domain discriminator distinguishes feature sources and constructs a domain classification loss. The SOH estimator calculates a regression loss based on the source domain labels and introduces a gradient inversion layer to achieve adversarial training, enabling the feature extractor I to learn unified aging features across domains. The trained evaluation model is used to evaluate the discrete feature sequences of the target domain unlabeled data to be predicted, outputting the corresponding current and historical health status SOH estimation sequences. Step 3: The current and historical SOH estimation sequences are used as pseudo-labels and input into the prediction model. The prediction model uses feature extractor II to extract time-series features from the pseudo-labels, and the SOH predictor outputs the future SOH trajectory in an autoregressive manner.

[0007] Furthermore, in step 1, the acquired data is preprocessed, including: The source domain labeled data, target domain unlabeled data, and target domain unlabeled data of the acquired lithium battery are cleaned. The cleaning includes extracting charging segment data, removing abnormal charge and discharge cycles, repairing outliers, and filling in missing values. A capacity increment curve was plotted based on the cleaned data; Within a preset voltage range, discrete values ​​on the capacity increment curve are sampled at equal voltage intervals to form the discrete feature sequence.

[0008] Furthermore, in step 2, feature extractor I is a multi-layer Transformer encoder used to extract common features from battery input data of different systems and operating conditions.

[0009] Furthermore, during the adversarial training process, the gradient reversal layer is passed identically during forward propagation, and during backward propagation, the gradient of the discriminator part is multiplied by a negative gradient reversal intensity factor -λ(t) so that the parameters of the discriminator and feature extractor I are updated simultaneously in one backward propagation.

[0010] Furthermore, step 2 also includes: during the evaluation model training process, introducing the Beluga optimization algorithm to optimize the hyperparameters of the evaluation model.

[0011] Furthermore, in step 3, the feature extractor II is an LSTM network with an embedded multi-head self-attention mechanism, which is used to extract the hidden state sequence after linear mapping and position encoding of the input sequence, and enhance the global perception capability through the self-attention mechanism.

[0012] Furthermore, the autoregressive method includes: predicting the future single-step SOH value in the previous time window, and iteratively feeding the predicted value back to the input sequence to achieve autoregressive prediction of the future capacity trajectory until the SOH drops to a preset threshold.

[0013] The present invention also provides a cross-domain estimation and future degradation prediction system for automotive lithium battery health status, for performing the above prediction method, including: a data preprocessing module for preprocessing input multi-source battery data and extracting discrete feature sequences with equal voltage intervals; The health status assessment module includes a feature extractor I, a domain discriminator, and a SOH estimator. The domain discriminator distinguishes feature sources and constructs a domain classification loss. The SOH estimator calculates a regression loss based on the source domain labels. The health status assessment module also introduces a gradient inversion layer to achieve adversarial training, enabling the feature extractor I to learn unified aging features across domains. Finally, the trained health status assessment module evaluates the discrete feature sequences of the unlabeled data in the target domain to be predicted, outputting the corresponding current and historical health status SOH estimation sequences. The future degradation prediction module includes a feature extractor II and a SOH predictor. The future degradation prediction module is used to receive the current and historical health state SOH estimation sequences as pseudo-labels, extract the temporal features of the pseudo-labels through the feature extractor II, and predict the future SOH trajectory by the SOH predictor in an autoregressive manner.

[0014] Compared with the prior art, the beneficial technical effects of the present invention are as follows: This invention employs a semi-supervised learning mechanism while introducing labeled and unlabeled data. Through domain adversarial training, it extracts consistent aging characteristics under different chemical systems and operating conditions, improving the cross-domain adaptability of health status estimation. Secondly, this invention introduces pseudo-labels generated by the health status estimation module into the future degradation prediction process, realizing the effective transfer of assessment results to the prediction model and further improving the continuity and accuracy of extrapolating future capacity decay trajectories. In addition, this invention uses discrete incremental capacity sequences as input and directly performs feature learning, reducing the reliance on manual feature engineering and simplifying the modeling process. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the cross-domain estimation of the health status of automotive lithium batteries and the prediction of future degradation according to the present invention. Figure 2 This is a flowchart illustrating the overall framework of the cross-domain estimation of the health status of automotive lithium batteries and the prediction of future degradation, as presented in this invention. Figure 3 This is a schematic diagram illustrating the IC curve analysis and cycle performance of the lithium-ion battery of the present invention. Figure 4 This is a schematic diagram of the modules of the evaluation model of the present invention; Figure 5 This is a schematic diagram of the modules of the prediction model of the present invention; Figure 6 The image shows the prediction results of the prediction model of this invention for three target domain conditions. Figure 7 This is a schematic diagram of the residual results of the prediction model of this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] refer to Figure 1-2 The present invention provides a method for cross-domain estimation of the health status and prediction of future degradation of automotive lithium batteries, comprising the following steps: Step 1: Data Preprocessing This invention addresses the problem of insufficient available battery data in the target domain during practical engineering by introducing a large amount of unlabeled data for model training. Therefore, the acquired data consists of a small amount of labeled target domain data, unlabeled target domain data to be predicted, and a large amount of labeled source domain data. The acquired data includes battery measurement data, such as temperature and voltage.

[0020] In terms of data preprocessing, the raw data is first cleaned and outlier checked, which mainly includes extracting charging segment data, removing abnormal charge and discharge cycles, repairing outliers, and filling in missing values. After data cleaning, feature curves are plotted. In this embodiment, the incremental capacity curve (IC) is used as the feature curve.

[0021] Capacity increment feature analysis is a commonly used method in existing battery health state prediction research. By converting constant current charge / discharge data into dQ / dV-V curves, it can sensitively identify battery aging mechanisms. The horizontal axis of this curve represents the battery measured voltage (V), and the vertical axis represents the capacity increment (dQ / dV). Existing studies mostly extract features such as peak values ​​and slopes from the IC curve for prediction. However, such manually extracted features have poor universality, are difficult to maintain effectiveness under different battery systems and operating conditions, and additionally increase the complexity of data processing.

[0022] Based on this, this embodiment uses a battery IC sequence with equal voltage intervals as a battery aging characteristic, such as... Figure 3 As shown, Figure 3 (a) is an example of a capacity increment curve, such as Figure 3 (b) is a schematic diagram showing the systematic shift in the IC curve shape as the battery ages (e.g., upward shift of peaks and valleys, and a decrease in dQ / dV values ​​near 3.0V). These changes contain crucial aging information. To avoid tedious manual feature extraction, this embodiment directly selects 20 equally spaced discrete IC values ​​within the 3.0V–3.95V voltage range (covering the commonly used charging range in real vehicles) to construct the battery feature sequence input model. The specific value selection process is as follows: Figure 3 As shown in (a).

[0023] Step 2: Perform cross-domain estimation of the state of health of lithium batteries using an evaluation model. Existing methods can achieve high-precision SOH estimation under the same system and operating conditions, but their performance degrades significantly in cross-domain scenarios due to differences in data distribution. To address this, this invention introduces a domain adversarial network. By introducing a domain discriminator after the feature extractor, the feature extractor is forced to output universal battery aging features across different domains through adversarial interaction, thus mitigating the performance degradation caused by differences in the distribution of the training set (source domain) and the test set (target domain). The core objective of the evaluation model in this invention is to assess the current and historical SOH of the target domain battery. Its training process jointly utilizes labeled data from the source domain and unlabeled data from the target domain to enhance the model's ability to represent universal aging features. The two types of data feature sequences are encoded by feature extractor I and then input into the domain discriminator and the SOH estimator, respectively. The domain discriminator constructs a domain classification loss by distinguishing the feature source, while the SOH estimator calculates a regression loss based on the source domain label. Adversarial training is achieved through a gradient inversion layer, forcing feature extractor I to learn the unified aging features across domains, ultimately achieving the assessment of the target domain battery's SOH.

[0024] The module diagram of the evaluation model in this embodiment is shown below. Figure 4 As shown, the evaluation model takes labeled and unlabeled sample data as parallel inputs. After passing through multiple transformer encoders, high-dimensional features of the labeled and unlabeled samples are obtained respectively. These high-dimensional features are then fed into a discriminator to perform domain classification. Before feature discrimination, the high-dimensional features of the labeled and unlabeled samples need to be labeled with distinguishing tags, denoted as 1 and 0 respectively. The cross-entropy loss between the discriminator output and the domain label is calculated. The discriminator error is calculated using the following formula: ; in, Representative feature extractor, Represents the discriminator. , These represent the parameters of the feature extractor and discriminator, respectively. x i Representative sample, d i Represents the actual domain label; The sum of the discrimination errors of labeled and unlabeled data is the overall discrimination error of the discriminator, i.e.: ; in, N and M These represent the cycle number for labeled and unlabeled data, respectively.

[0025] While calculating the discriminator error, the general features of the labeled data are fed into the SOH evaluator for SOH evaluation. The prediction result is then compared with the true label to obtain the prediction error. ; in For capacity predictor, These are the parameters for the capacity predictor. After calculating the discrimination error and prediction error, they need to be combined into a total error. Simultaneously, unified parameter learning is performed on the SOH estimator, feature discriminator, and feature generator. During optimization, it is desirable to reduce the SOH error, meaning the SOH estimator's calculation result is closer to the true capacity label; conversely, to increase the feature discriminator error, meaning the features extracted by the feature generator are closer to general features, making it difficult for the feature discriminator to identify the feature source. To reduce prediction error while increasing discrimination error, the feature extractor needs to be frozen first, and the discriminator parameters updated during model parameter updates; then the discriminator needs to be frozen, and the feature extractor parameters updated. This invention inserts a gradient inversion layer (GRL) between the feature extractor and the discriminator. Its forward propagation is identical, and during backpropagation, a portion of the discriminator's gradient is multiplied by -λ(t), thus simultaneously updating the discriminator and feature extractor parameters in one backpropagation. Here, λ(t) is a gradient inversion strength factor that varies with the training process, used to adjust the degree of influence of the discriminator loss on the feature extractor parameter updates. During forward propagation, the gradient inversion layer keeps the input features unchanged; during backpropagation, the gradient returned from the discriminator to the feature extractor is multiplied by -λ(t), causing the feature extractor to update in the direction of increasing discrimination error, thereby enhancing the domain invariance or robustness of the features. The backpropagation process is as follows: ; In the formula, I It is an identity matrix.

[0026] The overall error function of the model can be expressed as: ; in, D s Represents the source domain. This represents adjustable weights.

[0027] To prevent model oscillations caused by excessively strong adversarial constraints in the early stages of training, this embodiment employs a method of gradually introducing adversarial training, specifically as follows: Where epochs represents the total number of training epochs, epoch represents the current training epoch, and λ tThis is an intermediate scheduling factor that varies with the training rounds, and is used to... t Calculate the inversion strength coefficient λ of the gradient inversion layer. In the first 10 epochs of training, let λ=0, that is, do not introduce adversarial training, so that the feature extractor learns the basic prediction features first; when the epoch>10, λ smoothly increases from 0 to 1 during the training process, which is used to gradually enhance the inverse constraint effect of the discriminator loss on the update of the feature extractor parameters, thereby avoiding model oscillation caused by excessive adversarial constraints in the early stage of training.

[0028] During model training, this embodiment also introduces the Beluga Whale Optimization (BWO) algorithm for hyperparameter optimization. BWO is a swarm intelligence optimization algorithm proposed in 2022, inspired by the collective migration, cooperative hunting, and whale fall ecology of beluga whales in the Arctic environment. The algorithm abstracts each beluga whale as a search agent, achieving global-local collaborative search by simulating three natural whale behaviors: swimming, foraging, and falling. The algorithm iteration mainly includes three stages: exploration, development, and fall. The transition from the exploration stage to the development stage depends primarily on the balance coefficient B. f The specific mathematical expression is: ; Where T represents the current iteration number, T max The maximum number of iterations, B 0 represents the balance factor, which is randomly assigned a value in the interval (0,1) during each iteration. When the balance factor is greater than 0.5, the BWO is in the exploration phase, the beluga whale positions are updated in pairs, and the current optimal individual is perturbed using sine and cosine operators. Mathematically, this is described as follows: ; Where T represents the current iteration number. It is the new position of the i-th beluga whale in the j-th dimension. It is a random number selected from d-dimensional space. This represents the position of the i-th beluga whale in dimension pj. , These represent the current position and indicate the i-th and r-th beluga whales (r is a randomly selected beluga whale). r 1 and r 2 is a random number between (0,1). When the balance factor is less than 0.5, BWO is in the development stage, with beluga whales sharing the optimal prey location for refined searching, while a Levy flight strategy is introduced to enhance convergence. in, L FThis represents the Lévy flight perturbation step size, used to introduce random perturbations during parameter updates to enhance the model's global search capability and avoid getting trapped in local optima. is a normally distributed random number, β is the Lévy distribution exponent with a value of 1.5, σ is the scaling factor, and 0.05 in the formula is the step size scaling factor used to control the perturbation amplitude. v A random variable that follows a standard normal distribution is used to generate a Lévy perturbation with long-tailed distribution characteristics together with the random variable μ. During the whale fall phase, some individuals are made to sink to the seabed with a certain probability, i.e., escape the local extremum. The population size is maintained by random reinitialization and the global escape capability is enhanced.

[0029] In this embodiment, the feature generator in the evaluation model is a multi-layer transformer used to extract common features from battery input data of different systems and operating conditions. The transformer is designed specifically for natural language processing tasks, and is particularly adept at processing sequential data. Its core lies in the self-attention mechanism, which enables the model to process sequences in parallel and capture long-range dependencies. This invention only uses the encoder part of the transformer, which mainly consists of two core sub-layers: a multi-head self-attention network and a positional feedforward network, stacked alternately. The encoder's goal is to map the input sequence into a continuous representation. The self-attention mechanism allows the model to consider information from all other words in the sequence while processing each word. Its core formula is as follows: ; In this context, Q, K, and V represent query, key, and value, respectively. d k Let be the dimension of the key vector. This formula calculates the attention score of each word in the sequence to all other words, and then performs a weighted summation of the value vectors based on the calculated attention scores to obtain the final output.

[0030] A feedforward network is a simple two-layer fully connected network, typically using the ReLU activation function, with the following formula: ; in, x It is input. W 1 and W 2 It is a weight matrix. b 1 and b 2 It is a bias term.

[0031] Step 3: Predict the future degradation trajectory of lithium batteries using predictive models. A simple evaluation module can only perform historical or even current state of equilibrium (SOH) and capacity assessment, but cannot predict future capacity trajectories or estimate remaining lifespan. This embodiment integrates an evaluation model and a prediction model: the evaluation model generates a capacity pseudo-label based on available cyclic data; then, this pseudo-label is input into the prediction model, and the future battery aging trajectory is extrapolated through an autoregressive approach.

[0032] The model architecture of the prediction model is as follows Figure 5 As shown, the model uses historical SOH pseudo-labels output by the evaluation module as input. First, a sliding window is used to construct supervised learning samples from the sequence data, using past fixed-step data as features and the SOH of the next step as the ground truth, to train the feature extractor II and the SOH predictor. After the model is trained, the SOH value of the next single step is predicted using the previous time window, and this prediction is iteratively fed back to the input sequence to achieve autoregressive prediction of the future capacity trajectory until the SOH drops to the 0.8 threshold.

[0033] Feature Extractor II employs an LSTM network with an embedded multi-head self-attention mechanism. After linear mapping and positional encoding of the input sequence, the LSTM extracts the hidden state sequence, and then enhances global perception capabilities through a self-attention mechanism to alleviate the long-term dependency limitations of the LSTM. Finally, the final state after residual connections is input into the SOH predictor (a two-layer fully connected network) to complete SOH prediction.

[0034] This embodiment validated the battery data, using NCA batteries (nominal capacity 3500 mAh, cutoff voltage 2.65-4.2 V) and NCM batteries (nominal capacity 3500 mAh, cutoff voltage 2.5-4.2 V). The experiment was conducted in a constant temperature chamber. All batteries underwent cyclic aging at different temperatures (25℃, 35℃, and 45℃) and different charging rates (0.25C, 0.5C, and 1C), with a uniform discharge rate of 1C. The test conditions were as follows: constant current charging to 4.2V, then constant voltage charging until the current was below 0.05C, followed by 30 minutes of rest, then constant current discharging to the cutoff voltage (NCA 2.65V, NCM 2.5V), and then another 30 minutes of rest. All batteries were aged to approximately 71% of their nominal capacity. To simplify the description, the battery operating conditions are abbreviated, such as "A2505" representing the NCA battery operating at 25℃ and 0.5C charging rate, and "M3505" representing the NMC battery operating at 35℃ and 0.5C charging rate.

[0035] Figure 6 and Figure 7 Based on the overall framework of "evaluation-prediction," the performance of the prediction method proposed in this embodiment under cross-condition scenarios was systematically verified. Figure 6 (a) Figure 6 (b) and Figure 6(c) The SOH evolution curves and corresponding error changes are given under the three source domain conditions of A4505, M2505 and M4505 respectively; Figure 7 The statistical results of residual distribution under five target working conditions were further summarized.

[0036] In another embodiment of the present invention, a cross-domain estimation system for the health status of automotive lithium batteries and a system for predicting future degradation are also provided, for performing the above-described prediction method, including: The data preprocessing module is used to preprocess the input multi-source battery data and extract discrete feature sequences with equal voltage intervals. The health status assessment module includes a feature extractor I, a domain discriminator, and a SOH estimator. The domain discriminator distinguishes feature sources and constructs a domain classification loss. The SOH estimator calculates a regression loss based on the source domain labels. The health status assessment module also introduces a gradient inversion layer to achieve adversarial training, enabling the feature extractor I to learn unified aging features across domains. Finally, the trained health status assessment module evaluates the discrete feature sequences of the unlabeled data in the target domain to be predicted, outputting the corresponding current and historical health status SOH estimation sequences. The future degradation prediction module includes a feature extractor II and a SOH predictor. The future degradation prediction module is used to receive the current and historical health state SOH estimation sequences as pseudo-labels, extract the temporal features of the pseudo-labels through the feature extractor II, and predict the future SOH trajectory by the SOH predictor in an autoregressive manner.

[0037] In summary, the degradation prediction method provided by this invention adopts an integrated overall framework that combines cross-domain health status estimation and future degradation prediction. Instead of treating the health status estimation and future degradation prediction of automotive lithium batteries as two independent tasks, it constructs a unified integrated modeling framework that links "current health status assessment" and "future degradation trajectory prediction" together, thereby enabling the assessment results to directly support and synergize with the prediction process.

[0038] This invention utilizes both tagged and untagged battery data to perform cross-domain learning across different battery chemistry systems, operating conditions, and service scenarios. This allows for better extraction of domain-invariant aging features, thereby enhancing the model's generalization ability across operating conditions and material systems.

[0039] This invention introduces a domain adversarial learning mechanism in the health status estimation stage. Through joint training between the feature extraction module, the health status regression module, and the domain discrimination module, the model can reduce the data distribution differences between the source domain and the target domain and learn aging characteristics that are relatively unrelated to changes in material system and operating conditions.

[0040] This invention does not directly predict future degradation trajectories using historical capacity sequences alone. Instead, it introduces the SOH pseudo-labels output by the health status estimation module into the prediction module, using the pseudo-labels as a bridge connecting assessment and prediction, so that the learned health information can be transferred to the extrapolation process of future degradation trajectories.

[0041] This invention uses discrete incremental capacity sequences obtained by sampling at equal voltage intervals as input to directly perform feature learning. It does not rely on artificially constructed aging features such as peaks, areas, and inflection points, thereby reducing the complexity of artificial feature engineering and improving the engineering applicability of the prediction method.

[0042] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for cross-domain estimation of the state of health and prediction of future degradation of automotive lithium batteries, characterized in that, Includes the following steps: Step 1: Obtain the source domain labeled data, target domain unlabeled data, and target domain unlabeled data to be predicted of the lithium battery. Preprocess the obtained data to obtain the discrete feature sequence of each data. Step 2: Using the discrete feature sequences of the source domain labeled data and the target domain unlabeled data, a pre-built evaluation model is jointly trained. The evaluation model includes a feature extractor I, a domain discriminator, and a SOH estimator. The domain discriminator distinguishes feature sources and constructs a domain classification loss. The SOH estimator calculates a regression loss based on the source domain labels and introduces a gradient inversion layer to achieve adversarial training, enabling the feature extractor I to learn unified aging features across domains. The trained evaluation model is used to evaluate the discrete feature sequences of the target domain unlabeled data to be predicted, outputting the corresponding current and historical health status SOH estimation sequences. Step 3: The current and historical SOH estimation sequences are used as pseudo-labels and input into the prediction model. The prediction model uses feature extractor II to extract time-series features from the pseudo-labels, and the SOH predictor outputs the future SOH trajectory in an autoregressive manner.

2. The method for cross-domain estimation of the health status of automotive lithium batteries and prediction of future degradation according to claim 1, characterized in that, In step 1, the acquired data is preprocessed, including: The source domain tagged data, target domain untagged data, and target domain untagged data to be predicted of the acquired lithium battery are cleaned. The cleaning includes extracting charging segment data, removing abnormal charge and discharge cycles, repairing outliers, and filling in missing values. A capacity increment curve is plotted based on the cleaned data. Within a preset voltage range, discrete values ​​on the capacity increment curve are sampled at equal voltage intervals to form the discrete feature sequence.

3. The method for cross-domain estimation of the health status and prediction of future degradation of automotive lithium batteries according to claim 1, characterized in that, In step 2, feature extractor I is a multi-layer Transformer encoder used to extract common features from battery input data of different systems and operating conditions.

4. The method for cross-domain estimation of the health status of automotive lithium batteries and prediction of future degradation according to claim 3, characterized in that, During adversarial training, the gradient reversal layer is passed identically during forward propagation, and during backward propagation, the gradient of the discriminator part is multiplied by a negative gradient reversal intensity factor -λ(t) so that the parameters of the discriminator and feature extractor I are updated simultaneously in one backward propagation.

5. The method for cross-domain estimation of the health status of automotive lithium batteries and prediction of future degradation according to claim 4, characterized in that, Step 2 further includes: during the evaluation model training process, introducing the Beluga optimization algorithm to optimize the hyperparameters of the evaluation model.

6. The method for cross-domain estimation of the health status of automotive lithium batteries and prediction of future degradation according to claim 1, characterized in that, In step 3, the feature extractor II is an LSTM network with an embedded multi-head self-attention mechanism. It is used to extract the hidden state sequence after linear mapping and position encoding of the input sequence, and enhance the global perception capability through the self-attention mechanism.

7. The method for cross-domain estimation of the health status of automotive lithium batteries and prediction of future degradation according to claim 6, characterized in that, The autoregressive method includes: predicting the future single-step SOH value in the previous time window, and iteratively feeding the predicted value back to the input sequence to achieve autoregressive prediction of the future capacity trajectory until the SOH drops to a preset threshold.

8. A cross-domain estimation system for the state of health of automotive lithium batteries and a system for predicting future degradation, used to execute the prediction method according to any one of claims 1 to 7, characterized in that, include: The data preprocessing module is used to preprocess the input multi-source battery data and extract discrete feature sequences with equal voltage intervals. The health status assessment module includes a feature extractor I, a domain discriminator, and a SOH estimator. The domain discriminator distinguishes feature sources and constructs a domain classification loss. The SOH estimator calculates a regression loss based on the source domain labels. The health status assessment module also introduces a gradient inversion layer to achieve adversarial training, enabling the feature extractor I to learn unified aging features across domains. Finally, the trained health status assessment module evaluates the discrete feature sequences of the unlabeled data in the target domain to be predicted, outputting the corresponding current and historical health status SOH estimation sequences. The future degradation prediction module includes a feature extractor II and a SOH predictor. The future degradation prediction module is used to receive the current and historical health state SOH estimation sequences as pseudo-labels, extract the temporal features of the pseudo-labels through the feature extractor II, and predict the future SOH trajectory by the SOH predictor in an autoregressive manner.

Images