Lithium-ion battery state of health estimation method based on domain generalization of adversarial learning

Abstract

A lithium ion battery state of health estimation method based on domain generalization of adversarial learning belongs to the technical field of lithium ion battery health management. First, incremental capacity curve features are extracted from a multi-source heterogeneous data set. Structural adaptive alignment processing is performed on the incremental capacity curve, including amplitude normalization and voltage axis normalization, to map the heterogeneous features to a unified representation space. A hybrid neural network of a residual convolutional network and a Transformer encoder is constructed as a feature extractor to jointly capture the local morphological features and global degradation dependencies of the incremental capacity curve. A domain generalization method based on an adversarial mechanism is used to learn the degradation representation that is invariant to the domain through adversarial training of the multi-domain discriminator and the feature extractor. The SOH regression loss and the adversarial domain classification loss are jointly optimized, and the end-to-end adversarial training is realized through the gradient reversal layer. Finally, zero-shot SOH estimation is realized for unobserved heterogeneous battery data sets. The present application realizes high-precision zero-shot state of health estimation.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery health management technology, specifically relating to a domain-generalized lithium-ion battery health status estimation method based on adversarial learning. Background Technology

[0002] In recent years, with the intensification of the global environmental crisis, the transition to clean energy and sustainable development has been accelerating. Lithium-ion batteries, as a key component of modern energy systems, are widely used in electric vehicles, portable electronic devices, energy storage power stations, and smart grids due to their advantages such as high energy density, long cycle life, excellent safety performance, and low self-discharge rate. However, lithium-ion batteries inevitably experience performance degradation during use due to internal physicochemical mechanisms such as electrode material aging, electrolyte decomposition, and structural degradation. This leads to a gradual decrease in energy efficiency and storage capacity, while significantly increasing the risk of failures such as thermal runaway, electrolyte leakage, and internal short circuits. To ensure the safe and reliable operation of lithium-ion batteries, accurate state of health (SOH) estimation must be performed throughout their entire lifespan.

[0003] Existing SOH estimation methods are mainly classified into three categories: direct measurement methods, model-based methods, and data-driven methods. Direct measurement methods estimate SOH through physical testing and direct acquisition of electrochemical parameters, providing high-precision results. However, they typically require stringent experimental conditions, specialized equipment, and lengthy testing times, making them unsuitable for online SOH estimation in practical applications. Model-based methods utilize mathematical models to describe the electrochemical and thermodynamic behavior within the battery. Their accuracy and applicability depend on accurate parameter identification and extensive domain knowledge, limiting their scalability under diverse real-world conditions. Data-driven methods utilize neural networks to learn the complex nonlinear mapping relationship between measurable data and SOH from historical data, and have become the mainstream method for SOH estimation. However, their effectiveness largely depends on a large amount of high-quality labeled battery data.

[0004] While existing SOH estimation methods have achieved considerable success in specific scenarios, they still face significant limitations in generalization ability. Most methods are designed for specific battery chemistry systems, operating conditions, or datasets, and often exhibit significant performance degradation when applied to batteries with different aging mechanisms, temperature distributions, or usage histories. In recent years, some researchers have explored strategies such as transfer learning and domain adaptation to enhance the adaptability of SOH estimation models, but these methods typically focus on one-to-one domain adaptation and lack the ability to generalize to more complex, multi-source, and multi-condition battery environments, hindering the widespread deployment of battery management systems in heterogeneous real-world applications.

[0005] Ma et al. proposed a CNN SOH estimation method based on an improved maximum mean difference domain adaptation strategy to reduce distributional differences between different battery datasets (Ma G, Xu S, Yang T, et al. A TransferLearning-Based Method for Personalized State of Health Estimation of Lithium-Ion Batteries[J]. IEEE Trans Neural Netw Learn Syst, 2022.). Kim et al. developed a deep learning method combining transfer learning and variational inference to address the distribution mismatch problem between different battery types (Kim S, Choi YY, Kim KJ, et al. Forecasting state-of-health of lithium-ionbatteries using variational long short-term memory with transfer learning[J]. Journal of Energy Storage, 2021, 41.). However, all of the above methods suffer from limited generalization.

[0006] Therefore, there is an urgent need for a lithium-ion battery state of health estimation method with strong generalization ability, which can achieve accurate and robust zero-sample SOH estimation under different chemical systems, usage scenarios and operating conditions, and provide reliable technical support for battery management systems in heterogeneous practical applications. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, the present invention aims to provide a domain-generalized lithium-ion battery state of health estimation method based on adversarial learning, which constructs a basic model for battery SOH estimation across chemical systems, usage scenarios, and operating conditions, and achieves high-precision zero-sample state of health estimation on unseen heterogeneous battery data.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A domain-generalized lithium-ion battery health state estimation method based on adversarial learning includes the following steps: Step 1: Obtain the voltage-capacity curve of the charging process from a multi-source heterogeneous dataset of lithium-ion batteries, which covers battery data with different chemical systems, rated capacity, voltage windows, operating temperatures and cycle protocols; Step 2: Based on the voltage-capacity curve of the charging process, calculate the incremental capacity curve as the core feature representation. The incremental capacity curve is defined as the differential of the charging capacity to the terminal voltage. Step 3: Perform structure-adaptive alignment processing on the incremental capacity curves of each dataset obtained in Step 2, including amplitude normalization and voltage axis normalization, to map the heterogeneous incremental capacity features to a unified representation space; Step 4: Construct a hybrid neural network consisting of a residual convolution module and a Transformer encoder module as a feature extractor to jointly extract local morphological features and global degradation dependencies of the incremental capacity curve; Step 5: Construct the SOH regression module and the multi-domain discriminator module, which are used for regression prediction of health status values ​​and classification and recognition of domain labels, respectively; Step 6: Adopt an adversarial generalization method, implement adversarial training between the feature extractor and the multi-domain discriminator module through a gradient reversal layer, jointly optimize the SOH regression loss and the adversarial domain classification loss, and learn a domain-invariant degradation representation. Step 7: Deploy the trained model for inference, remove the multi-domain discriminator module, and retain only the feature extractor and SOH regression module to achieve zero-sample health status estimation for unseen battery data.

[0009] Step 1 specifically involves: acquiring charging process data from multiple lithium-ion battery datasets from different sources, covering nickel-cobalt-manganese, nickel-cobalt-aluminum, lithium cobalt oxide, lithium iron phosphate, and mixed chemical systems; battery rated capacity ranging from 0.74 Ah to 3.5 Ah, voltage window ranging from 2.0 V to 4.3 V, and operating temperature ranging from 15°C to 45°C; charging protocols including standard constant current-constant voltage charging and various customized cycling strategies.

[0010] Step 2 specifically involves: the incremental capacity curve is derived from the voltage-capacity curve of the charging process, and is defined as follows: IC=dQdV Where Q represents the charging capacity and V represents the terminal voltage; the incremental capacity curve serves as a general representation of degradation characteristics across different battery types and usage conditions.

[0011] Step 3, the adaptive alignment process, includes two steps: (a) Amplitude normalization: Minimum-maximum normalization is performed independently on the amplitude of the incremental capacity curve in each dataset, mapping the amplitude to the [0, 1] interval, eliminating amplitude deviations caused by differences in battery capacity, internal resistance and chemical system, and preserving the relative shape and local trend of the incremental capacity curve; (b) Voltage axis normalization: Mini-maximum normalization is performed on the voltage axis of the incremental capacity curves of each dataset to scale the original voltage range to a fixed interval of [0, 1], ensuring a uniform voltage domain while preserving the internal structure of each curve.

[0012] Step 4 specifically involves the following: The hybrid feature extraction network consists of three ResNet modules and one Transformer encoder module connected in series, serving as a feature extractor; The Residual Convolutional Network (ResNet) module is constructed by stacking multiple convolutional layers. Each convolutional layer is followed by batch normalization and a non-linear activation function, and identity shortcut connections are introduced to enable the network to learn the residual function. The output of the Residual Convolutional Network (ResNet) module is represented as follows: y=Fx+x Where x is the input of the block, and Fx represents the non-linear transformation performed by the stacked convolutional layers; The Transformer encoder module first adds positional encoding to the feature embeddings. Positional encoding is defined as: PEp,2i=sinp100002id PEp,2i+1=cosp100002id Where d represents the embedding dimension of the feature, p represents the position, and i represents the dimension index; Then, a self-attention mechanism is used to calculate the attention score between all positions in the sequence. Self-attention is represented as: Attention Q,K,V = softmaxQKTdkV Where Q, K, and V are the query, key, and value matrices obtained from the input sequence X through learnable linear projection, respectively; dk is the dimension of the key; KT represents the transpose operation; and the softmax(·) function is used to normalize the attention score to the probability weights at all positions. Furthermore, a multi-head attention mechanism is adopted to execute multiple self-attention operations in parallel on different representation subspaces to capture diverse relational patterns.

[0013] Step 5 specifically includes: The SOH regression module is designed as a fully connected feedforward network, consisting of multiple dense layers with nonlinear activation and a linear output layer. It maps the extracted latent features to continuous SOH values, quantifying the degree of battery degradation. The multi-domain discriminator module can identify features from multiple heterogeneous source domains and interact adversarially with the feature extractor, enabling the network to learn domain-invariant representations shared across heterogeneous battery datasets.

[0014] Step 6 is specifically as follows: Let Gf·, Gd·, and Gy· represent the feature extractor, the multi-domain discriminator module, and the SOH regression module, respectively; given N source domains D1, D2, ..., DN, the discriminator Gd is trained to predict the domain label of each feature characterizing Gfx, while the feature extractor Gf is optimized to generate features that can confuse the discriminator. The overall learning objective of domain generalization methods can be expressed as a minimax optimization problem: minGf,Gy maxGd LSOHX,Y-λd·LdomainX,D Where LSOH represents the SOH regression loss, Ldomain represents the domain classification loss, λd is the weight coefficient that controls the balance between regression accuracy and domain invariance, and X, Y, and D represent the input features, SOH labels, and domain labels, respectively. A gradient inversion layer is inserted between the feature extractor and the multi-domain discriminator. During backpropagation, the gradient inversion layer reverses the gradient direction and scales with a negative coefficient, enabling the discriminator to learn to minimize the domain classification loss, while the feature extractor learns to maximize the domain classification loss. Through adversarial competition, the feature extractor gradually generates cross-domain indistinguishable features that are highly sensitive to degradation processes, achieving effective domain generalization and robust SOH estimation.

[0015] Step 7 specifically involves: using only the feature extractor and SOH regression module during the inference phase, and removing the multi-domain discriminator module; for unseen battery datasets, first extracting incremental capacity features from the charging data, performing adaptive structural alignment, and then inputting them into the trained hybrid feature extraction network, outputting continuous health state estimates through the SOH regression module to achieve zero-shot generalization.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a domain-generalized lithium-ion battery state of health (SOH) estimation method based on adversarial learning. It eliminates distributional differences between multi-source heterogeneous datasets through structure-adaptive incremental capacity feature alignment, jointly captures local morphological features and global degradation dependencies using a hybrid CNN-Transformer architecture, and learns deep domain-invariant representations through an adversarial domain generalization strategy. This invention fundamentally overcomes the bottleneck of existing SOH estimation methods in terms of generalization capability, constructing a basic battery health state estimation model capable of generalizing across chemical systems, usage scenarios, and operating conditions. Comprehensive evaluation on 14 multi-source heterogeneous datasets shows that this method consistently outperforms representative baseline methods across different tasks, providing reliable technical support for battery management systems in large-scale energy storage, electric vehicles, and other applications. Attached Figure Description

[0017] Figure 1 This is an overall flowchart of the method proposed in the embodiments of the present invention.

[0018] Figure 2 This is an overall architecture diagram of the hybrid neural network model in an embodiment of the present invention.

[0019] Figure 3 This is a comparison chart of SOH prediction results for ablation experiments on the Tongji-NCM dataset in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the embodiments and accompanying drawings.

[0021] Reference Figure 1 A domain-generalized lithium-ion battery health state estimation method based on adversarial learning includes the following steps: Step 1: Acquisition of Multi-Source Heterogeneous Data. Voltage-capacity curves of the charging process are obtained from a multi-source heterogeneous dataset of lithium-ion batteries. This dataset covers six battery chemistry systems, including nickel-cobalt-manganese (NCM), nickel-cobalt-aluminum (NCA), lithium cobalt oxide (LCO), lithium iron phosphate (LFP), and mixed chemistry systems such as NCM-LCO and NCM-NCA. The rated battery capacity ranges from 0.74 Ah to 3.5 Ah, the voltage window from 2.0 V to 4.3 V, and the operating temperature from 15°C to 45°C. The charging protocols include standard constant current-constant voltage (CC-CV) charging and various customized cycling strategies, designed to simulate diverse application scenarios. This embodiment uses a total of 14 heterogeneous datasets, fully covering the diversity of batteries in practical applications. Step 2: Incremental Capacity Curve Feature Extraction. Based on the voltage-capacity curve during the charging process, the incremental capacity (IC) curve is calculated as the core feature representation. The incremental capacity curve is defined as the differential of the charging capacity with respect to the terminal voltage, expressed as: IC=dQdV Where Q represents the charging capacity and V represents the terminal voltage; the incremental capacity curve can highlight the phase transition process that occurs in the electrode material and sensitively reflect the changes caused by aging. These characteristics make the incremental capacity feature particularly suitable as a general degradation feature representation across different battery types and usage conditions, providing informative and physically interpretable input for generalizable SOH estimation.

[0022] Step 3: Adaptive Structure Feature Alignment. Adaptive structure alignment is performed on the incremental capacity curves of each dataset obtained in Step 2, comprising the following two sub-steps: (a) Amplitude normalization: The amplitude of the incremental capacity curve may vary significantly between different datasets due to differences in battery capacity, internal resistance and chemical system. These variations may cause the model to overfit to the amplitude scale of a specific dataset and ignore general degradation features. To address this, min-max normalization is performed independently on the incremental capacity amplitude within each dataset, mapping it to the [0, 1] interval. This operation preserves the relative shape and local trend of each incremental capacity curve while eliminating amplitude differences across datasets, enabling the model to learn degradation features that are independent of amplitude.

[0023] (b) Voltage Axis Normalization: Different cutoff voltages, charging protocols, and battery chemistry systems often lead to inconsistent voltage ranges across datasets. This misalignment can shift the position of key features on the incremental capacity curves (such as the peak corresponding to the phase transition), thereby impairing the model's ability to identify common degradation features. To address this, min-max normalization is performed on the voltage axis of the incremental capacity curves for each dataset, scaling the original voltage range to a fixed [0, 1] interval. This normalization ensures a uniform voltage domain while preserving the internal structure of each curve, enabling the model to compare and extract consistent features across datasets with different voltage ranges.

[0024] Through the above normalization steps, the incremental capacity features extracted from the heterogeneous battery dataset are mapped to a unified representation space with consistent amplitude and voltage structure. This standardized feature format significantly reduces the distribution differences between datasets while preserving key degradation patterns.

[0025] Step 4: Construct the hybrid feature extraction network. (Refer to...) Figure 2 To effectively capture the complex degradation features embedded in the incremental capacity curve and enhance the model's generalization ability across heterogeneous battery datasets, a hybrid neural network consisting of a Residual Convolutional Network (ResNet) module and a Transformer encoder module is constructed as a feature extractor. This aims to simultaneously capture local features and global dependencies, both crucial for accurate SOH estimation across heterogeneous datasets. Specifically, the model architecture consists of three residual blocks and a Transformer encoder module connected in series, followed by an SOH regression module and a multi-domain discriminator module, jointly achieving robust feature extraction, reliable SOH estimation, and domain-invariant representation learning.

[0026] The Residual Convolutional Network (ResNet) module is constructed by stacking multiple convolutional layers, each followed by batch normalization and a non-linear activation function. Unlike traditional feedforward convolutional networks that directly learn the mapping from input to output, the Residual Convolutional Network (ResNet) module introduces identity shortcut connections, enabling the network to learn a residual function referenced to the input. Formally, the output of the Residual Convolutional Network (ResNet) module can be represented as: y=Fx+x Where x represents the input of the block, and Fx represents the nonlinear transformation performed by the stacked convolutional layers. This design effectively alleviates the gradient vanishing problem, enabling deeper architectures to be trained with stable convergence, and ensuring that both low-level information and high-level abstract features are preserved during the learning process.

[0027] After extracting local features through the ResNet module, these representations are further fed into the Transformer encoder module to capture long-range dependencies. To preserve sequence structure and enable the model to distinguish positional relationships, positional encoding is first added to the feature embeddings; positional encoding is defined as: PEp,2i=sinp100002id PEp,2i+1=cosp100002id Where d represents the embedding dimension of the feature, after adding the position encoding to the embedding feature, the Transformer uses a self-attention mechanism to calculate the attention score between all positions in the sequence; self-attention can be represented as: Attention Q,K,V = softmaxQKTdkV Where Q, K, and V are the query, key, and value matrices obtained from the input sequence X through learnable linear projection, respectively; dk is the dimension of the key; KT represents the transpose operation; and the softmax(·) function is used to normalize the attention score to the probability weights at all positions.

[0028] To further enhance the model's ability to capture long-range dependencies, Transformer employs a multi-head attention mechanism. Multi-head attention executes multiple self-attention operations in parallel across different representation subspaces, rather than relying on a single self-attention operation. This design allows the model to simultaneously focus on diverse relational patterns, capturing both fine-grained and high-level dependencies within sequences. By integrating information from multiple attention heads, Transformer effectively enhances the expressive power and robustness of feature representations.

[0029] Step 5: Construction of the SOH Regression Module and Multi-Domain Discriminator. Features extracted by ResNet and Transformer are simultaneously fed into two task-specific modules: the SOH Regression Module and the Multi-Domain Discriminator Module.

[0030] The SOH regression module is designed as a fully connected feedforward network consisting of multiple dense layers with nonlinear activation and a linear output layer. This module maps the extracted latent features to continuous SOH values, thereby quantifying the degree of battery degradation.

[0031] As a core component of the adversarial-based domain generalization framework, the multi-domain discriminator module interacts adversarially with the feature extractor. The multi-domain discriminator module can identify features from multiple heterogeneous source domains, enabling the network to learn domain-invariant representations shared across heterogeneous battery datasets.

[0032] Step 6: Domain Generalization Training Based on Adversarial Mechanisms. Traditional adversarial transfer learning methods typically assume a predefined target domain, focusing on aligning a source domain with a target domain. However, this one-to-one alignment limits its applicability in real-world scenarios. The domain generalization method based on adversarial mechanisms (DG-AM) proposed in this invention learns domain-invariant features from multiple heterogeneous source domains without assuming a target domain, thereby improving the model's generalization ability and robustness to unseen battery conditions.

[0033] The core idea of ​​DG-AM is to enable the feature extractor to learn a domain-invariant representation that is discriminative to the SOH estimate but invariant across multiple source domains. To achieve this, an adversarial learning strategy is employed, with the feature extractor and multi-domain discriminator trained in a minimax game. The multi-domain discriminator attempts to correctly identify the domain label for each feature, while the feature extractor focuses on generating features that can confuse the discriminator. Through this adversarial interaction, the model aligns the feature distributions of different domains within a shared latent space, thereby learning domain-invariant representations without requiring any target domain data.

[0034] Let Gf·, Gd·, and Gy· represent the feature extractor, multi-domain discriminator module, and SOH regression module, respectively. Given N source domains D1, D2, ..., DN, the discriminator Gd is trained to predict the domain label of each feature representing Gfx, while the feature extractor Gf is optimized to generate features that can confuse the discriminator. The overall learning objective of DG-AM can be expressed as a minimax optimization problem: minGf,Gy maxGd LSOHX,Y-λd·LdomainX,D Where LSOH represents the SOH regression loss, Ldomain represents the domain classification loss, λd is the weight coefficient that controls the balance between regression accuracy and domain invariance, and X, Y, and D represent the input features, SOH labels, and domain labels, respectively.

[0035] To achieve adversarial training, a gradient inversion layer (GRL) is inserted between the feature extractor and the multi-domain discriminator. During backpropagation, the GRL inverts the gradient direction and scales it with a negative coefficient. As a result, the multi-domain discriminator learns to minimize the domain classification loss, while the feature extractor learns to maximize the domain classification loss. Through this adversarial competition, the feature extractor gradually generates cross-domain indistinguishable features that are highly sensitive to the degradation process, achieving effective domain generalization and robust SOH estimation.

[0036] Step 7: Model Inference and Zero-Shot SOH Estimation. After training, the model is deployed for inference. During the inference phase, only the feature extractor and SOH regression module are used, removing the multi-domain discriminator module. For unseen battery datasets, incremental capacity features are first extracted from the charging data, and adaptive structural alignment is performed. The aligned features are then input into the trained hybrid feature extraction network, and the SOH regression module outputs continuous health state estimates, achieving zero-shot generalization.

[0037] This embodiment uses 14 heterogeneous lithium-ion battery datasets for comprehensive evaluation, employing a unified zero-shot testing strategy: for each target dataset, the model is trained on the remaining 13 heterogeneous battery datasets, completely excluding the target dataset from training and using it only for testing, without any fine-tuning or domain-specific calibration. Mean absolute error (MAE) and root mean square error (RMSE) are used as evaluation metrics, defined as follows: MAE=1ki=1kyi-yi RMSE = 1ki = 1kyi - yi2 Where yi represents the true SOH value of the i-th sample, yi represents the predicted SOH value of the i-th sample, and k represents the total number of samples in the dataset.

[0038] Experimental verification – Case 1: Zero-shot generalization on the Tongji-NCM dataset.

[0039] The Tongji-NCM dataset contains 18650-specification NCM batteries with a nominal capacity of 3.5 Ah, providing degradation data collected at three operating temperatures: 25°C, 35°C, and 45°C. This dataset is used to evaluate the model's generalization ability under temperature-dependent degradation patterns.

[0040] Reference Figure 3 Ablation experiments show that the complete method proposed in this invention achieves the lowest estimation error, with an average MAE of 0.786% and an RMSE of 0.839%, and exhibits stability under all temperature conditions. Using raw charging data instead of incremental capacity features leads to a significant increase in error (MAE 10.92%, RMSE 11.22%), indicating that the raw data is sensitive to protocol and measurement variations. Using incremental capacity features without performing adaptive structural alignment also produces high errors (MAE 11.17%, RMSE 11.31%), demonstrating that inconsistencies in amplitude and voltage scales across datasets are a major obstacle to learning transferable degradation representations. Removing the DG-AM module produces moderate performance (MAE 1.378%, RMSE 1.424%), with significant deviations in SOH prediction at different cycling stages. These results collectively demonstrate that incremental capacity feature extraction, adaptive structural alignment, and DG-AM are complementary and indispensable components.

[0041] Comparative experiments with different machine learning methods show that the hybrid model proposed in this invention achieves the highest estimation accuracy (average MAE 0.786%, RMSE 0.839%); the CNN model achieves moderate accuracy with MAE 1.776% and RMSE 1.842%; and the Transformer model records MAE 1.985% and RMSE 2.015%. The proposed architecture integrates residual convolution and Transformer attention, enabling joint learning of fine-grained local degradation features and long-range relational dependencies, generating smoother, more accurate, and more consistent SOH trajectories.

[0042] Experimental Validation – Case 2: Zero-shot generalization on the IECON-2022 dataset.

[0043] The IECON-2022 dataset, derived from long-term non-accelerated cycling experiments, contains significant fluctuations, diurnal variations, and measurement noise in the SOH trajectory, making it more representative of real-world degradation behavior. On this dataset, the proposed method still achieves state-of-the-art performance, with an average MAE of 6.027% and an RMSE of 6.184%. These results further demonstrate the robustness and generalizability of the proposed method in real-world degradation scenarios.

Claims

1. A domain-generalized lithium-ion battery health state estimation method based on adversarial learning, characterized in that, Includes the following steps: Step 1: Obtain the voltage-capacity curve of the charging process from a multi-source heterogeneous dataset of lithium-ion batteries, which covers battery data with different chemical systems, rated capacity, voltage windows, operating temperatures and cycle protocols; Step 2: Based on the voltage-capacity curve of the charging process, calculate the incremental capacity curve as the core feature representation. The incremental capacity curve is defined as the differential of the charging capacity to the terminal voltage. Step 3: Perform structure-adaptive alignment processing on the incremental capacity curves of each dataset obtained in Step 2, including amplitude normalization and voltage axis normalization, to map the heterogeneous incremental capacity features to a unified representation space; Step 4: Construct a hybrid neural network consisting of a residual convolution module and a Transformer encoder module as a feature extractor to jointly extract local morphological features and global degradation dependencies of the incremental capacity curve; Step 5: Construct the SOH regression module and the multi-domain discriminator module, which are used for regression prediction of health status values ​​and classification and recognition of domain labels, respectively; Step 6: Adopt an adversarial generalization method, implement adversarial training between the feature extractor and the multi-domain discriminator module through a gradient reversal layer, jointly optimize the SOH regression loss and the adversarial domain classification loss, and learn a domain-invariant degradation representation. Step 7: Deploy the trained model for inference, remove the multi-domain discriminator module, and retain only the feature extractor and SOH regression module to achieve zero-sample health status estimation for unseen battery data.

2. The estimation method according to claim 1, characterized in that, Step 1 specifically includes: Charging process data were obtained from multiple lithium-ion battery datasets from various sources, covering nickel-cobalt-manganese, nickel-cobalt-aluminum, lithium cobalt oxide, lithium iron phosphate, and mixed chemistry systems; battery rated capacities ranged from 0.74 Ah to 3.5 Ah, voltage windows from 2.0 V to 4.3 V, and operating temperatures from 15°C to 45°C; charging protocols included standard constant current-constant voltage charging and various custom cycling strategies.

3. The estimation method according to claim 1, characterized in that, Step 2 specifically includes: The incremental capacity curve is derived from the voltage-capacity curve during the charging process and is defined as follows: IC=dQdV Where Q represents the charging capacity and V represents the terminal voltage; the incremental capacity curve serves as a general representation of degradation characteristics across different battery types and usage conditions.

4. The estimation method according to claim 1, characterized in that, Step 3, the adaptive alignment process, includes two steps: (a) Amplitude normalization: Minimum-maximum normalization is performed independently on the amplitude of the incremental capacity curve in each dataset, mapping the amplitude to the [0, 1] interval, eliminating amplitude deviations caused by differences in battery capacity, internal resistance and chemical system, and preserving the relative shape and local trend of the incremental capacity curve; (b) Voltage axis normalization: Mini-maximum normalization is performed on the voltage axis of the incremental capacity curves of each dataset to scale the original voltage range to a fixed interval of [0, 1], ensuring a uniform voltage domain while preserving the internal structure of each curve.

5. The estimation method according to claim 1, characterized in that, Step 4 specifically includes: The hybrid feature extraction network consists of three ResNet modules and a Transformer encoder module connected in series, serving as a feature extractor; The Residual Convolutional Network (ResNet) module is constructed by stacking multiple convolutional layers. Each convolutional layer is followed by batch normalization and a non-linear activation function, and identity shortcut connections are introduced to enable the network to learn the residual function. The output of the Residual Convolutional Network (ResNet) module is represented as follows: y=Fx+x Where x is the input of the block, and Fx represents the non-linear transformation performed by the stacked convolutional layers; The Transformer encoder module first adds positional encoding to the feature embeddings. Positional encoding is defined as: PEp,2i=sinp100002id PEp,2i+1=cosp100002id Where d represents the embedding dimension of the feature, p represents the position, and i represents the dimension index; Then, a self-attention mechanism is used to calculate the attention score between all positions in the sequence. Self-attention is represented as: Attention Q,K,V = softmaxQKTdkV Where Q, K, and V are the query, key, and value matrices obtained from the input sequence X through learnable linear projection, respectively; dk is the dimension of the key; KT represents the transpose operation; and the softmax(·) function is used to normalize the attention score to the probability weights at all positions. Furthermore, a multi-head attention mechanism is adopted to execute multiple self-attention operations in parallel on different representation subspaces to capture diverse relational patterns.

6. The estimation method according to claim 1, characterized in that, Step 5 specifically includes: The SOH regression module is designed as a fully connected feedforward network, consisting of multiple dense layers with nonlinear activation and a linear output layer. It maps the extracted latent features to continuous SOH values, quantifying the degree of battery degradation. The multi-domain discriminator module can identify features from multiple heterogeneous source domains and interact adversarially with the feature extractor, enabling the network to learn domain-invariant representations shared across heterogeneous battery datasets.

7. The estimation method according to claim 1, characterized in that, Step 6 specifically includes: Let Gf·, Gd·, and Gy· represent the feature extractor, the multi-domain discriminator module, and the SOH regression module, respectively. Given N source domains D1, D2, ..., DN, the discriminator Gd is trained to predict the domain label of each feature characterizing Gfx, while the feature extractor Gf is optimized to generate features that can confuse the discriminator. The overall learning objective of domain generalization methods can be expressed as a minimax optimization problem: minGf,Gy maxGd LSOHX,Y-λd·LdomainX,D Where LSOH represents the SOH regression loss, Ldomain represents the domain classification loss, λd is the weight coefficient that controls the balance between regression accuracy and domain invariance, and X, Y, and D represent the input features, SOH labels, and domain labels, respectively. A gradient inversion layer is inserted between the feature extractor and the multi-domain discriminator. During backpropagation, the gradient inversion layer reverses the gradient direction and scales with a negative coefficient, enabling the discriminator to learn to minimize the domain classification loss, while the feature extractor learns to maximize the domain classification loss. Through adversarial competition, the feature extractor gradually generates cross-domain indistinguishable features that are highly sensitive to degradation processes, achieving effective domain generalization and robust SOH estimation.

8. The estimation method according to claim 1, characterized in that, Step 7 specifically includes: During the inference phase, only the feature extractor and the SOH regression module are used, and the multi-domain discriminator module is removed. For unseen battery datasets, incremental capacity features are first extracted from the charging data, and structural adaptive alignment is performed. Then, the feature extraction network is input into the trained hybrid feature extraction network, and the SOH regression module outputs continuous health state estimates to achieve zero-shot generalization.

Images