Characteristic decoupling-based passive domain lithium battery health state estimation method

Abstract

The invention provides a passive domain lithium battery SOH (state of health) estimation method based on feature decoupling, which is used for solving the problem of model performance degradation in lithium battery cross-working condition SOH estimation under privacy protection constraint, and the privacy protection constraint means that source domain data cannot be accessed in a target domain adaptive stage. The method comprises the following three stages: (1) a source domain pre-training stage: constructing a heterogeneous dual-predictor network, training on source domain data, and isolating the source domain data after storing a model weight; and (2) a target domain self-adaption stage: pre-training network parameters are frozen, label-free target domain data are input, a self-supervision signal is constructed by using a dual-prediction head output difference, a dual-head decoupling module is driven to explicitly separate target domain features into domain invariant features and domain private features, and the domain invariant features and the domain private features are extracted. The feature collapse is prevented through the generative reconstruction constraint and the feature orthogonal constraint, and the feature purity is ensured; and (3) a target domain inference stage: inputting label-free actual measurement data, inputting domain invariant features after model purification into two frozen heterogeneous predictors in parallel, and taking arithmetic average as a final SOH estimation result.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, and specifically to a method for estimating the state of health of passive domain lithium batteries based on feature decoupling. Background Technology

[0002] Lithium-ion batteries, due to their high energy density, high power density, and long cycle life, have been widely used in electric vehicles, power tools, and large-scale energy storage systems. However, lithium batteries undergo irreversible performance degradation during repeated charge-discharge cycles. If the State of Health (SOH) cannot be estimated in a timely and accurate manner, the battery may experience serious safety accidents such as thermal runaway, fire, or even explosion during continuous degradation. Therefore, accurate and reliable SOH estimation plays a crucial role in the Battery Management System (BMS).

[0003] The existing lithium battery capacity estimation methods are mainly divided into two categories: mechanistic model-based methods and data-driven methods. Mechanism-based methods mainly rely on electrochemical models, equivalent circuit models or empirical models, and establish mathematical descriptions through expert knowledge. Although such methods have a certain degree of interpretability, they have the following limitations: First, there are many model parameters, the calibration process is complicated, and they rely heavily on expert knowledge in specific fields; second, it is difficult to accurately characterize the nonlinear degradation characteristics of batteries under complex working conditions, resulting in limited estimation accuracy. Data-driven methods use machine learning models to automatically learn feature representations from data without explicit mechanistic modeling, and have received widespread attention in the field of battery health monitoring in recent years. For example, Zhang et al. [1] proposed a dual-channel deep learning method for estimating the state of charge under dynamic working conditions. However, such methods usually assume that the training data and the test data follow the same data distribution. In practical applications, due to changes in working conditions such as temperature, internal resistance, and charge / discharge rate, battery data often show significant distribution differences, i.e., there is a domain shift problem. When the training data and the test data do not satisfy the independent and identically distributed assumption, the performance of the model obtained from the training will be significantly reduced when applied directly to the test scenario.

[0004] Transfer learning techniques provide an effective way to solve the domain shift problem by transferring knowledge from a well-labeled source domain to a relevant but poorly labeled target domain. For example, Wen et al. [2] used unsupervised subdomain adversarial networks to predict the remaining useful life under different working conditions. However, the above transfer learning methods all require access to both source domain data and target domain data during the training process. In actual engineering scenarios, due to considerations such as data privacy protection, control of commercial secrets, or data transmission costs, the original source domain data is often unavailable, and only pre-trained models can be provided. This lack of source domain constraints makes it difficult to directly apply traditional transfer learning methods.

[0005] Therefore, there is an urgent need for a passive domain adaptive method under strict privacy protection constraints, which can achieve high-precision estimation of SOH of lithium batteries across operating conditions without accessing the original source domain data.

[0006] [1] Zhang Q, Wan G, Li C, et al. State of charge estimation for Li-ionbattery during dynamic driving process based on dual-channel deep learning methods and conditional judgment[J]. Energy, 2024, 294: 130948.

[0007] [2] Wen Z, Lyu Y, Chen A. An unsupervised subdomain adversarial network for remaining useful life estimation under various conditions[J]. Quality and Reliability Engineering International, 2024, 40(4): 1652-1671. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention proposes a passive domain lithium battery health state estimation method based on feature decoupling, which includes the following three stages and a total of nine steps:

[0009] Phase 1: Source Domain Pre-training Phase, including:

[0010] S1, Source Domain Data Acquisition: Acquire the degradation data of lithium batteries under typical operating conditions and the corresponding capacity tags as source domain data. ,in and These represent data samples from the source domain and their corresponding capacity labels;

[0011] S2, Source Domain Pre-training: Constructing independent predictors consisting of a feature extractor E and two heterogeneous structures. , The network is composed of labeled source domain data, and supervised training is performed on the network to obtain a pre-trained model and save the model weights.

[0012] Phase Two: Target Domain Adaptation Phase, including:

[0013] S3, Target Domain Adaptive Initialization: Acquire degraded data without capacity labels under target operating conditions. Freeze the feature extractor E and the predictor , The parameters are set, and a trainable dual-head decoupling module is set after the feature extractor E, wherein the dual-head decoupling module contains two linear heads. and These are used to generate domain-invariant features and domain-private features, respectively; the maximum number of training rounds is set.

[0014] S4, Feature Decoupling: Input the target domain data into the frozen feature extractor E to obtain basic features, and then input the basic features in parallel into the linear head. and The domain-invariant characteristics after decoupling are obtained. and domain private features ;

[0015] S5, Self-supervised signal generation: Extracting purified domain-invariant features Parallel input to frozen predictor and Calculate the difference between the outputs of the two predictors and construct the difference loss. , serving as a self-supervised signal driving feature decoupling;

[0016] S6, Feature Quality Constraint: Domain Invariant Features and domain private features The features are concatenated along the feature dimension to form a combined feature vector. Input the generative reconstructor R to obtain reconstructed data , Then calculate the reconstruction loss. Simultaneously, the feature orthogonality loss between domain-invariant features and domain-private features is calculated. ;

[0017] S7, Model Optimization: Based on the Total Loss Function The dual-head decoupling module and the generative refactorer R are optimized and trained, where , For non-negative hyperparameters, Used to control the strength of generative reconstruction loss in order to maintain information integrity. Used to control the strength of orthogonal loss to ensure the purity of domain-invariant features;

[0018] S8, Iteration judgment: Determine whether the training has reached the maximum number of iterations. If not, return to S4.

[0019] Phase Three: Target Domain Reasoning Phase, including:

[0020] S9, Target Domain Inference: Input the target domain test sample into the trained feature decoupling encoder, extract the purified domain-invariant features, and feed the domain-invariant features into the frozen predictors P1 and P2 in parallel. The arithmetic mean of the outputs of the two predictors is used as the final SOH estimation result.

[0021] Specifically, the feature extractor E in step S2 uses a bidirectional long short-term memory network to map the input time series and extract bidirectional temporal features from the lithium battery degradation data.

[0022] Specifically, in step S2, predictor P1 is a deep and narrow multilayer perceptron structure, using a deep neural network with four linear layers and ReLU activation function; predictor P2 is a shallow and wide multilayer perceptron structure, using a shallow structure with only a single wide hidden layer and LeakyReLU activation function.

[0023] Specifically, the joint objective function for end-to-end training used in step S2 is:

[0024] ,

[0025] in The total number of samples in the source domain dataset. For the i-th training sample in the source domain dataset, This is the corresponding capacity label.

[0026] Specifically, the difference loss in step S5 Adopting a combination norm ( )and norm ( The mixed objective function is:

[0027] ,

[0028] in, The dynamic penalty coefficient follows a linear warm-up strategy in the early stages of adaptive training, increasing linearly from 0 to the maximum value η_max within a set number of warm-up rounds.

[0029] Specifically, the generative reconstruction loss in step S6 Calculated using mean square error:

[0030] ,

[0031] in, Where L is the batch size, L is the sequence length, and d is the input dimension. The original input data, The data restored by the generative reconstructor This represents the Frobenius norm.

[0032] Specifically, the feature orthogonal loss in step S6 The steps include:

[0033] (1) The field-invariant characteristic matrix and domain private feature matrix Centralization and The calculation method is as follows , ;

[0034] (2) Calculate the normalized Frobenius norm of the cross-covariance matrix between the centered eigenma matrices: ,in, and These are the dimensions of domain-invariant features and domain-private features, respectively. This represents the Frobenius norm.

[0035] Specifically, the final SOH estimation result in step S9 The calculation formula is as follows:

[0036] . Attached Figure Description

[0037] Figure 1 This is a flowchart of a passive domain lithium battery health state estimation method based on feature decoupling provided in an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of the network structure of a passive domain lithium battery capacity estimation method based on feature decoupling provided in an embodiment of the present invention. Detailed Implementation

[0039] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0040] refer to Figure 1This embodiment provides a passive domain lithium battery health state estimation method based on feature decoupling, which includes three stages: the first stage is the source domain pre-training stage, which uses source domain data with battery capacity labels; the second stage is the target domain adaptation stage, which isolates the source domain data and uses only the data under the target operating condition to perform passive domain adaptation of the model; and the third stage is the field inference stage after implementation and deployment, which directly inputs the measured data to obtain the estimated value of battery SOH.

[0041] Phase 1: Source Domain Pre-training Phase

[0042] S1, acquire degradation data collected under typical operating conditions and the corresponding capacity labels as the source domain. ,in and These represent data samples from the source domain and their corresponding capacity labels;

[0043] The degradation data under typical operating conditions in this embodiment refers to battery state information under preset target operating conditions, such as temperature, charging / discharging current, battery voltage, etc. Among them, charging / discharging current and battery voltage are essential information. For example, it can be charging current and battery voltage, or discharging current and battery voltage, or charging / discharging current, battery voltage and battery temperature information.

[0044] The battery degradation data in this embodiment has a corresponding capacity label, and the corresponding capacity can be detected by a capacity testing instrument.

[0045] Furthermore, step S1 also includes: normalizing the source domain data.

[0046] This embodiment uses a sliding time window technique and a data normalization method to normalize the degraded data. The calculation formula is as follows.

[0047]

[0048] in, This represents the j-th reading of the k-th selected sensor in the data sample. , These are the maximum and minimum values ​​among the readings of the k-th selected sensor, respectively. for Normalized data sample.

[0049] S2, this step is the source domain pre-training stage. Its core objective is to train a prediction model that performs well under preset conditions using the number of labeled source domains D𝑠. The subsequent source domain adaptation is based on this pre-trained model, and optimizes and adapts the model under the target conditions without accessing the source domain data, relying only on the unlabeled target domain data, thus solving the problems of data privacy protection and continuous model optimization.

[0050] The model consists of a feature extractor E and two heterogeneous independent predictors. , The neural network is composed of [various components]. In this embodiment, the feature extractor E uses a bidirectional long short-term memory network (BiLSTM) to map the input time-series data and extract temporal features. Predictor (Deep Narrow Type): Employs a deep neural network with four linear layers and ReLU activation function, designed to capture subtle fluctuations and high-frequency components during battery degradation. Predictor (Shallow Wide Type): Employs a shallow structure containing only a single wide hidden layer, with LeakyReLU as the activation function. It focuses on fitting global smooth trends and enhancing robustness to outliers.

[0051] During training, the entire network is trained by minimizing a loss function. End-to-end optimization is performed. For each sample in the source domain, the function calculates the prediction errors of both predictors and sums them. The specific formula for the loss function is as follows:

[0052] , ⑷

[0053] in, Represents the total number of samples in the source domain dataset. This represents the i-th training sample in the source domain dataset. The MSE loss function is... Here, we use the Huber loss function. By minimizing this joint loss, the model is driven to learn deep features that effectively represent the source domain (SOH) and forces the two predictors to converge to local optima with different error sensitivities, providing stable and diverse self-supervised signals for the subsequent passive domain adaptation stage. Once the model's performance converges on the source domain validation set, training is complete, and the model weights of feature extractor E, predictors P1, and P2 are saved for use in the next stage.

[0054] Phase Two: Unsupervised Adaptation in the Target Domain

[0055] S3, Unsupervised Adaptive Initialization of the Target Domain: Acquiring degraded data without capacity labels under target operating conditions. The data is then subjected to the data normalization process described in step S1; the feature extractor E and the predictor are frozen. , The parameters are appended to the feature extractor E as a trainable dual-head decoupling module, the dual-head decoupling module comprising two linear heads. and These are used to generate domain-invariant features and domain-private features, respectively; the maximum number of training epochs is set; the network model and model weights consisting of feature extractor E, predictor P1 and P2 trained in S2 are loaded, the maximum number of training epochs max_epoch and exit conditions are designed, and the iteration counter epoch is started, and the process proceeds to step S4.

[0056] S4, Feature decoupling, converting target domain data The frozen feature extractor E is used to obtain basic features. These basic features are then input into a trainable dual-head decoupling module to obtain decoupled domain-invariant features. and domain private features The dual-head decoupling module consists of a linear head. and composition.

[0057] For data from the target domain First, through the feature extractor with frozen network parameters. Generate a basic feature representation , This fundamental feature It inherits the powerful feature extraction capabilities of the source domain model, but still incorporates core information related to SOH prediction as well as operating noise specific to the target domain. Subsequently, these basic features are input in parallel into two trainable linear heads. and In this module, these two heads together form a dual-head decoupling module. They respectively generate the decoupled feature vectors, namely, domain-invariant features and domain-private features. The formula for this process is as follows:

[0058] ,

[0059] ,

[0060] in, , and , These are the trainable weights and biases of the two linear heads, respectively. These two decoupled feature vectors are assigned explicit and mutually exclusive semantic roles:

[0061] Domain-invariant features: Their design goal is to capture core semantic information that is relevant to the SOH prediction task and is universal across domains. They will serve as the sole carrier of knowledge transfer and the sole basis for subsequent input into the frozen predictor for SOH prediction.

[0062] Domain-private features: These are designed to isolate and encapsulate noise or patterns that are only relevant to specific operating conditions within the target domain (such as unique charge / discharge rates and strategies). While these features may contain some information within the target domain, their domain dependence can interfere with the model's generalization ability, and therefore they must be separated out.

[0063] This step explicitly separates the original, entangled target domain features into two clearly defined sub-features. This lays the foundation for subsequent step S5, which allows for model optimization and driving based on pure domain-invariant features.

[0064] S5, self-supervised signal generation, is the core mechanism in the passive domain adaptation process, namely, difference-driven self-supervised signal generation. In S4, the target domain features have been decoupled into domain-invariant features and domain-private features. The key operation in this step is to input this purified domain-invariant feature in parallel into two predictors P1 and P2, which were frozen in S3, and to calculate the difference between their outputs.

[0065] The basic principle is that P1 and P2, as predictors trained on the source domain, have consistent judgment power regarding features that conform to the data distribution of the source domain. Although their decision boundaries differ slightly due to structural differences, their prediction results should be highly consistent for an ideal, pure, domain-invariant feature, i.e., the difference should be very small. Conversely, if the domain-invariant feature still contains noise or patterns specific to the target domain, the two predictors with different decision boundaries are more likely to diverge in their prediction results when interpreting these unseen patterns, leading to larger differences.

[0066] Therefore, by minimizing the difference between the outputs of these two predictors, it is equivalent to providing the model with a strong self-supervised signal even without real labels. This signal will propagate back, forcing the dual-head decoupling module in S4 to continuously adjust its parameters, separating all domain-related features that may cause prediction discrepancies into domain-private features, thus making them increasingly pure and more consistent with the distribution pattern of the source domain features.

[0067] This driving mechanism is quantified as a difference loss, and a hybrid difference loss function is defined. It consists of L1 norm terms ( ) and L2 norm terms with dynamic coefficients ( )composition:

[0068] ,

[0069] in, This is a dynamic penalty coefficient. It is set during the initial stages of adaptive training. Following a linear preheating strategy, i.e. The temperature increases linearly from 0 to the maximum value within the set preheating cycles. middle, This indicates that the expectation is calculated over all target domain data. The L1 term leverages its insensitivity to outliers to ensure stable training when dealing with self-supervised signals with significant noise; while the L2 term provides stronger gradient directions for samples with large prediction discrepancies. Through a dynamic warm-up mechanism, the model can be guided to smoothly transition from the source domain state to the adaptive representation space of the target domain, making the training process more stable and robust.

[0070] Ultimately, this difference loss will serve as the most important component of the total loss function in S7, directly guiding the optimization process in S8, and is key to achieving effective passive domain adaptation.

[0071] Step S6 introduces two complementary key regularization constraints to the core difference-driven mechanism in S5, aiming to ensure the effectiveness and stability of the adaptive process. These two constraints are the generative reconstruction loss and the feature orthogonality loss, which together ensure that the model does not get trapped in trivial solutions during optimization and can achieve thorough feature separation.

[0072] First, to ensure the integrity of the information in the decoupled features, this step introduces a generative reconstruction mechanism. Simply minimizing the predictor difference loss in S5 carries an inherent risk: the model might learn a shortcut solution to achieve the goal, such as outputting a constant or minimally informative domain-invariant feature. While this solution can easily bring the difference loss close to zero, it completely loses the original data information crucial for SOH prediction; this phenomenon is called feature collapse. To address this issue, this step decouples the domain-invariant features from S4... and domain private features The features are concatenated along the feature dimension to form a combined feature vector that contains all the original information. The feature vector is then input into a trainable generative reconstructor R, whose task is to reconstruct the original input data as accurately as possible. This is achieved by reconstructing the data... Between with the original data The error is quantified as a reconstruction loss. The specific formula for calculating the loss is as follows:

[0073] ,

[0074] in, Where L is the batch size, L is the sequence length, and d is the input dimension. The original input data, The data restored by the generative reconstructor This represents the Frobenius norm.

[0075] By minimizing this loss, the model is forced to ensure that both domain-invariant and domain-private features retain enough information to fully reconstruct the original input. This provides a strong regularization signal that effectively prevents feature collapse.

[0076] Secondly, to ensure the purity of the decoupled features, this step introduces a feature orthogonality loss. Its goal is to force domain-invariant features and domain-private features to be statistically independent, thereby achieving more thorough feature decoupling and completely isolating target domain-specific noise information into the domain-private features. This loss directly applies to both the domain-invariant and domain-private feature subspaces. In specific calculations, the domain-invariant features corresponding to a batch of data are first... and domain private features Centering the eigenma matrix yields... and Then, the cross-covariance matrix between the two centered feature matrices is calculated. This orthogonal loss... The goal is to make this covariance matrix as close as possible to a zero matrix, and its size is measured by the normalized Frobenius norm. The specific calculation formula is as follows:

[0077] ,

[0078] ,

[0079] ,

[0080] in, and These are the dimensions of domain-invariant features and domain-private features, respectively. Let Frobenius norm be denoted. By minimizing the orthogonal loss, the model is motivated to learn two mutually orthogonal (i.e., uncorrelated) feature subspaces. This effectively isolates target domain-specific noise and artifacts into domain-private feature subspaces, thereby refining domain-invariant features so that they can more purely carry cross-domain general core information strongly relevant to the SOH prediction task.

[0081] Step S7 aims to integrate the various independent loss terms defined in the previous steps into a unified total loss function to guide model optimization. This total loss function cleverly balances the three core objectives of the adaptive process through weighted summation: driving the learning of the core task, ensuring information integrity, and ensuring feature purity. The total loss is composed of the core difference loss from step S5, and two key regularization terms from step S6: generative reconstruction loss and feature orthogonality loss.

[0082] These three loss terms are passed through two non-negative hyperparameters. and Weighted combinations are then performed to obtain the final total loss function. Among these, the hyperparameters... Hyperparameters are used to control the strength of the reconstruction loss, i.e., the importance of the information integrity constraint; This is used to control the strength of the orthogonal loss, i.e., the importance of the feature purity constraint. The specific calculation formula is as follows:

[0083] .

[0084] In the next S8 step, the model will update its parameters by minimizing the total loss function, thereby collaboratively completing the tasks of feature decoupling and domain adaptation. This ensures that while learning cross-domain invariance, key information is not lost and thorough feature separation is achieved.

[0085] S8: Train the system by optimizing the total loss function using stochastic gradient descent; check if the counter has reached the set maximum number; if so, exit; otherwise, return to S4.

[0086] Phase 3: Target Domain Reasoning

[0087] After completing passive domain adaptive training, for a given target domain test sample The final health status (SOH) prediction result is obtained through the following steps: [Sample data is missing from the original text.] Input the trained feature decoupling encoder to extract purified domain-invariant features. .Will Two heterogeneous predictors, saved and frozen during the pre-training phase, are fed in parallel. and Two predicted output values ​​were obtained respectively. and An integrated prediction strategy is employed, and the arithmetic mean of the two output values ​​is calculated as the final SOH estimate. The calculation formula is as follows:

[0088] .

[0089] This strategy significantly enhances the model's noise resistance and estimation accuracy when migrating between different battery domains by combining the sensitivity to detail of the deep and narrow structure model with the trend-capturing ability of the shallow and wide structure model.

[0090] This invention proposes a passive domain feature decoupling method for transferable estimation of the State of Health (SOH) of lithium batteries, aiming to address the performance degradation of models under different operating conditions while meeting strict data privacy protection requirements. First, the core of this method lies in utilizing the output difference between the two predictors in the pre-trained model as a unique self-supervised signal. This signal drives the model to explicitly decouple the target domain features into orthogonal domain-invariant features and domain-private features. Then, generative reconstruction constraints are introduced to ensure the information integrity of the decoupled features and prevent feature collapse, while simultaneously applying orthogonal constraints to ensure the purity of the domain-invariant features. Finally, only the purified domain-invariant features are fed into the frozen predictor for SOH prediction, thereby achieving robust cross-domain estimation. This method can effectively decouple the core features relevant to the task under passive domain conditions, solving the feature entanglement problem in traditional methods and improving the prediction accuracy and stability of the model in the target domain. This method requires no access to source domain data during the entire model adaptation phase, only a pre-trained model and unlabeled target domain data, possessing extremely high practical application value and data security.

[0091] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for estimating the state of health of a passive domain lithium battery based on feature decoupling, characterized in that, It includes the following three stages: Phase 1: Source Domain Pre-training Phase, including: S1, Source Domain Data Acquisition: Acquire the degradation data of lithium batteries under typical operating conditions and the corresponding capacity tags as source domain data. ,in and These represent data samples from the source domain and their corresponding capacity labels; S2, Source Domain Pre-training: Constructing independent predictors consisting of a feature extractor E and two heterogeneous structures. , The network is composed of labeled source domain data, and supervised training is performed on the network to obtain a pre-trained model and save the model weights. Phase Two: Target Domain Adaptation Phase, including: S3, Target Domain Adaptive Initialization: Collect degraded data without capacity labels under the target operating conditions as target domain data, and freeze the feature extractor E and the predictor. , The parameters are set, and a trainable dual-head decoupling module is set after the feature extractor E, the dual-head decoupling module comprising two linear heads. and This is used to generate domain-invariant and domain-private features; it also sets the maximum number of training rounds and the termination condition. S4, Feature Decoupling: Input the target domain data into the feature extractor E with frozen parameters to obtain basic features, and then input the basic features in parallel into the linear head. and The domain-invariant characteristics after decoupling are obtained. and domain private features ; S5, Self-supervised signal generation: Extracting purified domain-invariant features Parallel inputs are fed into a predictor with frozen parameters. and Calculate the difference between the outputs of the two predictors and construct the difference loss. , serving as a self-supervised signal driving feature decoupling; S6, Feature Quality Constraint: Domain Invariant Features and domain private features The concatenated data is input into the generative reconstructor R to calculate the reconstruction loss. Simultaneously, the feature orthogonality loss between domain-invariant features and domain-private features is calculated. ; S7, Model Optimization: Based on the Total Loss Function The dual-head decoupling module and the generative reconstructor R are optimized and trained, wherein... , For non-negative hyperparameters, Used to control the strength of generative reconstruction loss in order to maintain information integrity. Used to control the strength of orthogonal loss to ensure the purity of domain-invariant features; S8, Iteration judgment: Determine whether the training has reached the maximum number of iterations. If not, return to S4. Phase Three: Target Domain Reasoning Phase, including: S9, Target Domain Inference: Input the target domain test sample into the trained feature decoupling encoder, extract the purified domain-invariant features, and feed the domain-invariant features into the frozen predictors P1 and P2 in parallel. Calculate the arithmetic mean of the outputs of the two predictors to obtain the final SOH estimation result.

2. The method according to claim 1, characterized in that: The feature extractor E described in step S2 employs a bidirectional long short-term memory network (BiLSTM), and the predictor... , Using a fully connected network, the two linear heads in step S3 and A single-layer fully connected network is used, and the data reconstructor R in step S6 uses a fully connected network.

3. The method according to claim 1, characterized in that: In step S2, the predictor P1 is a deep and narrow multilayer perceptron structure, which uses a deep neural network with four linear layers and the activation function is ReLU; the predictor P2 is a shallow and wide multilayer perceptron structure, which uses a shallow structure with only a single wide hidden layer and the activation function is LeakyReLU.

4. The method according to claim 1, characterized in that, The joint objective function for end-to-end training in step S2 is: ， in The total number of samples in the source domain dataset. For the i-th training sample in the source domain dataset, This is the corresponding capacity label.

5. The method according to claim 1, characterized in that: The difference loss described in step S5 Adopting a combination Norm and Mixed objective function of norm: in, The dynamic penalty coefficient follows a linear warm-up strategy in the early stages of adaptive training, increasing linearly from 0 to the maximum value η_max within a set number of warm-up rounds.

6. The method according to claim 1, characterized in that: The generative reconstruction loss described in step S6 Calculated using mean square error: , in, Where L is the batch size, L is the sequence length, and d is the input dimension. The original input data, The data restored by the generative reconstructor This represents the Frobenius norm.

7. The method according to claim 1, characterized in that: The feature orthogonal loss described in step S6 The calculations include: (1) The field-invariant characteristic matrix and domain private feature matrix Centralization and The calculation method is as follows , ; (2) Calculate the normalized Frobenius norm of the cross-covariance matrix between the centered eigenma matrices: ,in, and These are the dimensions of domain-invariant features and domain-private features, respectively. This represents the Frobenius norm.

8. The method according to claim 1, characterized in that, The final SOH estimation result described in step S9 The calculation formula is: .

Images