A lithium battery state of health estimation method based on meta-transfer learning

Abstract

This invention relates to a lithium battery state of health estimation method based on meta-transfer learning, belonging to the interdisciplinary field of battery management systems and signal processing. This method enhances feature mining capabilities through a cascaded structure of a 1-D CNN and a Transformer encoder. Simultaneously, it integrates meta-learning and transfer learning mechanisms. Meta-learning optimizes model initialization parameters across multiple source tasks, reducing the dependence on the similarity between the source and target domain distributions; transfer learning, through a pre-training-fine-tuning paradigm, reduces the requirements for the number and diversity of source tasks. This joint mechanism overcomes the limitations of traditional single learning methods in cross-domain generalization, maintaining high SOH estimation accuracy even when there are significant differences in distributions between domains.

Description

Technical Field

[0001] This invention relates to a lithium battery health state estimation method based on meta-transfer learning, which belongs to the interdisciplinary field of battery management system and signal processing. Background Technology

[0002] Lithium-ion batteries, with their high energy density and long cycle life, have become the core energy storage device for electric vehicles. However, during repeated charge and discharge cycles, irreversible electrochemical reactions occur inside the battery, leading to a gradual decline in battery performance—a process known as battery aging. This aging phenomenon not only reduces the driving range and power performance of electric vehicles but may also cause safety issues. To represent the degree of battery aging, the ratio of the current maximum usable capacity to the initial rated capacity is typically used as a quantitative indicator; this indicator is the State of Health (SOH). Therefore, the accuracy of SOH estimation is crucial for the safe operation of electric vehicles.

[0003] Existing methods for estimating the state of harm (SOH) of lithium-ion batteries are mainly divided into two categories: model-based methods and data-driven methods. Among them, data-driven methods based on deep learning have shown significant advantages in SOH estimation due to their powerful feature extraction and nonlinear modeling capabilities, and have become a research hotspot in recent years. However, these methods are all based on the assumption that the distributions of training data and test data are consistent. In reality, lithium-ion batteries have diverse material systems, and even with the same material system and manufacturing process, there are inconsistencies among individual cells. Furthermore, battery operating conditions vary, including differences in charge / discharge rates and ambient temperature. These factors all lead to a domain shift in the distribution of test data compared to training data. This domain shift severely restricts the generalization ability of existing methods, making it difficult to accurately characterize the complex nonlinear degradation process of batteries in the target domain, a problem that is particularly prominent when monitoring data is limited.

[0004] To address the poor generalization ability of deep learning-based methods, some studies have attempted to leverage transfer learning (TL) techniques to transfer knowledge learned in the source domain to the target domain, thereby improving model generalization. Dong et al., in "Deep transfer learning enabled state of health estimation of lithium-ion battery using voltage sample entropy under fast charging profiles," proposed a parameter transfer method across charging strategies. By freezing some layers of the pre-trained network and fine-tuning the remaining structure, they achieved the transfer of the state of health (SOH) estimation model between different fast charging strategies. Song et al., in "A novel transferlearning model for battery state of health prediction based on driving behavior classification," developed a TL framework based on pre-trained model parameters, enabling cross-domain estimation of battery SOH between vehicles with similar driving behaviors. Amogne ZE et al., in "Transfer learning based on transferability measures for state of health prediction of lithium-ion batteries," proposed a cross-lifetime domain parameter transfer framework based on transferability measures. This framework filters transferable battery samples by quantifying the similarity of degradation trajectories and freezes the network backbone layers to retain general features. While the aforementioned time-travel (TL) method improves the generalization performance of the model to some extent in specific cross-domain scenarios, in real-world battery applications, complex coupling effects exist between multiple factors such as temperature, charge / discharge rate, and material system, leading to significant differences in data distribution under different operating conditions. The performance of TL largely depends on the similarity of data distributions between the source and target domains. When the distribution differences are large, negative migration can easily occur, making it difficult to accurately estimate the battery's state of equilibrium (SOH) under conditions of significant distribution shift.

[0005] To address this challenge, researchers have begun to introduce meta-learning (ML) methods to tackle cross-domain generalization. Meta-learning, by extracting shared prior knowledge from multiple source tasks during the training phase, can quickly adapt to new target tasks with limited samples. Its performance is less dependent on the similarity between the source and target domains, making it particularly suitable for battery SOH estimation scenarios with limited data and varied operating conditions. Miao et al., in "Multi-source Domain Meta-learning Network for Battery State-of-Health Estimation under Multi-target Working Conditions," proposed a multi-source domain meta-learning framework based on a meta-fusion module to characterize the diversity of battery degradation behavior and improve the model's cross-domain generalization ability under multiple target conditions. Wu et al., in "Domain similarity meta-learning for lithium-ion battery state-of-health estimation of spacecraft systems," further introduced a task similarity measurement mechanism to construct a domain similarity meta-learning framework to improve the cross-task adaptability of lithium-ion battery SOH estimation. However, the effectiveness of meta-learning methods still heavily depends on the number and diversity of available source tasks.

[0006] Existing cross-domain SOH estimation methods still have the following limitations: (1) Dong et al., Song et al., and Amogne et al. used the TL method to achieve inter-domain knowledge transfer under specific variable differences, which improved the model's adaptability under limited target data. However, most of these methods are for single variables, and their performance is highly dependent on the distribution similarity between the source domain and the target domain. Therefore, such methods are difficult to effectively deal with complex distribution differences caused by the coupling of multiple variables such as ambient temperature, charge / discharge rate, and material system, and their generalization ability is limited; (2) The meta-learning method proposed by Miao et al. and Wu et al. can learn generalizable knowledge from multiple source tasks and improve cross-domain generalization ability, but its performance depends on a large-scale and diverse set of source tasks. Therefore, insufficient number or lack of diversity of source tasks will limit the model's ability to capture complex multi-factor degradation patterns, resulting in a decrease in generalization performance in unknown target domains. Summary of the Invention

[0007] The purpose of this invention is to address the limitations of existing cross-domain SOH estimation methods by proposing a domain-generalized SOH estimation method based on meta-transfer learning. This method reduces the dependence on high similarity between the source and target domains, alleviates the requirements for a large number of diverse source tasks, and enhances adaptability under complex multi-factor operating conditions. The core idea of ​​this method is to integrate transfer learning (TL) and machine learning (ML) to leverage their complementary advantages. The framework employs a two-stage learning mechanism: in the pre-training stage, meta-learning is performed on multi-source tasks to extract domain-independent degradation features and obtain model initialization parameters with high generalization ability; in the fine-tuning stage, the parameters learned in the pre-training stage are used for initialization, and the model is fine-tuned using limited target domain data to capture task-specific degradation features. By combining TL and ML, this method exhibits stronger generalization ability when dealing with multivariate coupled conditions. In terms of model architecture, a feature extraction structure is adopted that combines a one-dimensional convolutional neural network (1D-CNN) with a Transformer encoder. Among them, 1D-CNN is responsible for capturing local features in temporal data, while the Transformer encoder focuses on extracting global features of the sequence. The two work together to significantly enhance the ability to represent degraded temporal features, thereby comprehensively improving the accuracy of SOH estimation.

[0008] The above-mentioned objectives of the present invention are mainly achieved through the following technical solutions:

[0009] A lithium battery health state estimation method based on meta-transfer learning, considering different ambient temperatures, charge / discharge rates, and material systems involved in actual battery service scenarios, includes the following steps: (1) Collection of full life cycle data of lithium batteries with different material systems under different working conditions; This method focuses on the domain-generalized State of Harmony (SOH) estimation problem under different ambient temperatures, charge / discharge rates, and material systems. Therefore, in this step, a series of charge-discharge cycle tests are conducted on lithium batteries at various ambient temperatures and charge / discharge rates. During the tests, key operating data such as battery voltage, current, and capacity are collected in real time to provide a data foundation for subsequent model training and validation.

[0010] (2) Dataset partitioning Given the inherent differences among individual lithium-ion battery cells, to comprehensively evaluate the performance of this patented technology under different conditions, it categorizes cells into groups based on three factors: material system, ambient temperature, and charge / discharge rate, using various combinations such as single-variable, bivariate, and trivariate combinations. For example, under the condition that the material system and charge / discharge rate are the same, batteries with ambient temperatures of 25℃, 35℃, and 45℃ can be divided into three groups: A, B, and C. Groups A and B are used as the training set, and group C is used as the test set, thereby verifying the estimation accuracy and generalization ability of this patented technology under the single variable of ambient temperature. In the meta-learning framework, the data partitioning is set as follows: In the pre-training stage, the first 70% of the data for each battery in the source domain is used as the support set, and the remaining 30% is used as the query set; in the fine-tuning stage, fine-tuning samples are extracted from the first 20% of the data for each battery in the target domain, and further divided into a 70% support set and a 30% query set; the remaining 80% of the data in the target domain is used for the final test.

[0011] (3) Feature extraction This invention first proposes a feature extractor based on a cascaded 1-D CNN and a Transformer encoder. This feature extractor first extracts local features from battery data through a 1-D CNN layer. These local features not only preserve detailed information about small voltage segments but also assign preliminary semantics to each segment. Subsequently, these semantically informative local features are organized into token sequences and input into the Transformer encoder, making them more consistent with the Transformer's design premise of structured semantic input. Building on this, the Transformer further captures global features using a self-attention mechanism, effectively fusing local details and global trends, significantly enhancing the model's ability to express general battery degradation features across different operating conditions and domain scenarios.

[0012] (4) Learning strategies This learning strategy comprises two phases: pre-training and fine-tuning. The overall structure is as follows: Figure 1 As shown. In the pre-training phase, the model learns from multiple source tasks based on a meta-learning framework. For each source task, a loss function is calculated based on the support set, and gradient updates are performed. The updated model parameters are then used to further calculate the loss on the query set. Finally, the global average loss is calculated by combining the losses from all source task query sets, and the meta-model parameters are updated accordingly. This process enables the model to obtain initial parameters with strong generalization ability. In the fine-tuning phase, fine-tuning is performed using a limited number of samples in the target domain. This phase is also based on the meta-learning framework, further optimizing the model parameters by performing iterative updates on limited data, enabling it to quickly capture the specific degradation features of the target task.

[0013] (5) Test After the aforementioned training steps are completed, a cross-domain SOH estimation network based on meta-transfer learning is obtained, which has optimal model parameters. Subsequently, the network is subjected to a final performance test using test set data from the target domain that was not involved in training and fine-tuning, to obtain the SOH estimates for the corresponding battery samples, thereby verifying the network's generalization ability and estimation accuracy in cross-domain scenarios.

[0014] Beneficial effects 1. This invention relates to a cross-domain lithium battery state of health (SOH) estimation method based on meta-transfer learning. This method integrates meta-learning and transfer learning mechanisms. Through meta-learning, it extracts generalizable knowledge representations from multiple source tasks and obtains model initialization parameters with strong adaptability. Furthermore, by combining the transfer learning mechanism, it uses a small number of samples in the target domain to fine-tune the model and quickly adapt to the target task.

[0015] 2. For the time-series signals unique to lithium batteries, a cascaded structure of 1-D CNN and Transformer encoder is used as a feature extractor to extract local and global features from the sequence data.

[0016] 3. This invention provides a domain-generalized SOH estimation method based on meta-transfer learning. This method enhances feature mining capabilities through a concatenated structure of a 1-D CNN and a Transformer encoder. Simultaneously, it integrates meta-learning and transfer learning mechanisms. Meta-learning optimizes model initialization parameters across multiple source tasks, reducing the dependence on the similarity between the source and target domain distributions; transfer learning, through a pre-training-fine-tuning paradigm, reduces the requirements for the number and diversity of source tasks. This joint mechanism overcomes the limitations of traditional single learning methods in cross-domain generalization, maintaining high SOH estimation accuracy even when there are significant differences in domain distributions.

[0017] 4. The method proposed in this invention exhibits strong generalization in cross-domain scenarios involving multiple coupled variables (including material system, ambient temperature, and charge / discharge rate). It can achieve rapid adaptation regardless of whether it is a single-variable, bivariate, or trivariate domain shift. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a local feature extraction module; Figure 3 This is a global feature extraction module; Figure 4 This is a flowchart of the meta-learning process. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] Example Take the lithium battery lifecycle dataset collected by Tongji University as an example.

[0021] (1) Collection of full life cycle data of lithium batteries with different material systems under different operating conditions The Tongji dataset contains long-cycle data of 18650 lithium-ion batteries with various material systems, including NCA, NCM, and NCM+NCA hybrid systems. This dataset records the complete aging process of the batteries, from an initial capacity of 100% to below 80%, with cycle counts ranging from 50 to 800, fully demonstrating the different degradation trajectories of lithium-ion batteries. All battery experiments were conducted under strictly controlled temperature conditions, with ambient temperatures set at 25°C, 35°C, and 45°C. The current rate during charge and discharge was systematically controlled, covering multiple rate conditions from 0.25C to 4C, thus constructing a comprehensive battery aging dataset encompassing multiple material systems, multiple ambient temperatures, and multiple charge / discharge rates.

[0022] (2) Dataset partitioning In a specific embodiment, 45 batteries from the Tongji dataset were selected to construct an experimental dataset, covering 4 charge / discharge rates, 3 ambient temperatures, and 3 material systems. Based on the differences between different batteries in the three dimensions of material system, ambient temperature, and charge / discharge rate, they were divided into various experimental groups, including single-variable, bivariate, and trivariate groups, for the construction of training and test sets.

[0023] Taking a single-variable experiment across material systems as an example, samples numbered CY25-05_1-#1~#4, #8, #11~#13, and #15~#16 in the NCM battery were used as the training set, and samples numbered CY25-05_1-#1~#3 in the NCM+NCA hybrid battery were used as the test set.

[0024] Taking a single-variable experiment across ambient temperature as an example, samples numbered CY35-05_1-#1~#2 in the NCM battery are used as the training set, and samples numbered CY45-05_1-#1~#2 in the NCM battery are used as the test set.

[0025] Taking the single-variable experiment across charge-discharge rates as an example, samples numbered CY25-05_1-#1~#3 and CY25-05_2-#1~#3 in the NCM+NCA hybrid system battery were used as the training set, and samples numbered CY25-05_4-#1~#3 in the NCM+NCA hybrid system battery were used as the test set.

[0026] Taking the bivariate experiment across environmental temperature and material system as an example, the samples numbered CY25-025_1-#3~#5 in the NCA battery are used as the training set, and the samples numbered CY25-05_1-#1~#4, #8, #11~#13, #15~#16 in the NCM battery are used as the test set.

[0027] Taking the bivariate experiment across ambient temperature and charge / discharge rate as an example, the samples numbered CY25-025_1-#3 to #5 in the NCA battery were used as the training set, and the samples numbered CY35-05_1-#1 to #2 in the NCA battery were used as the test set.

[0028] Taking the bivariate experiment across material systems and charge / discharge rates as an example, the samples numbered CY25-025_1-#3~#5 in the NCA battery are used as the training set, and the samples numbered CY25-05_1-#1~#4, #8, #11~#13, #15~#16 in the NCM are used as the test set.

[0029] Taking a three-variable experiment involving ambient temperature, material system, and charge / discharge rate as an example, G uses samples numbered CY25-025_1-#3 to #5 from NCA batteries as the training set and samples numbered CY35-05_1-#1 to #4 from NCM batteries as the test set. Similarly, based on the above principles, battery data under different material systems and operating conditions are systematically divided, constructing multiple training-test set combinations covering single-variable, bivariate, and three-variable scenarios to comprehensively verify the performance of the proposed method in different cross-domain scenarios.

[0030] (3) Feature extraction This feature extraction module concatenates a 1-D CNN with a Transformer encoder to extract local and global features from battery data, respectively.

[0031] Local feature extraction module To address the characteristics of lithium battery voltage time series data, this invention constructs a 1-D CNN structure for extracting local features, such as... Figure 2 As shown. The input data is... The network consists of five sequentially connected convolutional blocks, each including a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. This design controls computational complexity while maintaining the model's representational power, making it suitable for embedded or real-time processing environments. Finally, a fully connected layer transforms the feature map output of the last convolutional block to the target dimension. Specifically, its input dimension equals the number of feature channels of the last convolutional layer, and its output dimension matches the embedding dimension of subsequent Transformer modules, achieving feature space alignment.

[0032] Global feature extraction module Transformer encoders can effectively capture global features, thus overcoming the limitations of traditional CNNs in global representation capabilities, such as... Figure 3 As shown in the diagram, the core of this encoder consists of four encoder layers. Each layer comprises a multi-head attention structure and a feed-forward network (FFN), and embeds residual connections and layer normalization. Its structure is as follows: Figure 3 As shown, the results from the CNN are first positionally encoded before being input into the Transformer encoder. This design ensures efficient gradient propagation and stability during training even with deep layer stacking.

[0033] Lithium-ion battery health status estimation model based on meta-transfer learning This model employs a learning strategy that integrates training (TL) and machine learning (ML), consisting of two stages: pre-training and fine-tuning, each using a meta-learning framework. Let the parameterization function of the CNN-Transformer encoder concatenated model be denoted as... .

[0034] During the source domain phase, the source domain contains One task, denoted as Each task Divided into support sets and query set This serves as the basic unit for model training. Let's consider a random task... The corresponding model parameters are For the lithium battery SOH estimation problem, the task... The support set and query set can be represented as:

[0035] Where V represents voltage data and SOH represents health status label.

[0036] ML employs a two-layer optimization framework to improve the model's generalization ability in cross-task scenarios, with the specific structure as follows: Figure 4 As shown, the input data matrix is ​​divided into multiple tasks, and the inner optimization process is tailored to each independent task. From initial parameters Starting with the first 70% of the data in each task, we use it as the support set and the remaining 30% as the query set. Based on the support set of this task... Perform forward propagation and calculate the mean squared error loss. The task-specific parameters are then updated using the gradient descent algorithm. :

[0037] in The inner learning rate. Updated task-specific parameters. In query set Calculate the loss of a single task :

[0038] in To query the number of samples in the set, and They represent the first The true value and the estimated value of each sample.

[0039] The outer meta-optimization process aims to learn a generalized model representation with cross-task generalization capabilities. This mechanism enhances the model's ability to quickly adapt to diverse scenarios by integrating common knowledge from multiple tasks. (In the query set for each task...) The parameters obtained above were evaluated. The performance of meta-optimization is evaluated, and the average task loss within the domain is calculated. The core objective of meta-optimization is to minimize this average task loss. During the pre-training phase, the objective function of meta-learning is defined as the performance of the source domain. Average query loss for each task:

[0040] The meta-optimization process operates on global initialization parameters. Its objective function is based on the parameters of each task after being updated by the inner layer. The calculations show that the optimal global initialization parameters can be found by minimizing this meta-objective function. The model parameters are then updated accordingly, and the update process is as follows:

[0041] in The outer learning rate, The average task loss within the domain eventually converges. These are meta-learning parameters that have the ability to adapt quickly.

[0042] During the fine-tuning phase, the parameters obtained from training in the source domain are... As initialization, the model is fine-tuned using a small sample of 20% of the target domain data to further optimize the network weights and enhance their adaptability to the target domain, ultimately obtaining model parameters adapted to the target domain. Fine-tuning data includes One task, denoted as Each task Divided into support sets and query set ,Right now Similar to the pre-training step, the support set is used for task... Perform a single-step gradient update to obtain task-specific parameters. The query set is then used to evaluate task performance and calculate the average loss across all tasks. :

[0043] Similarly, this objective function Minimum Parameter This is the optimal parameter. The test set is used for subsequent SOH estimation. At this stage, to avoid the risk of overfitting with small samples, the fine-tuning training period should be strictly controlled.

[0044] After completing the above pre-training and fine-tuning process, the optimal model parameters with cross-domain adaptability are obtained. This parameter This enables the model to quickly adapt to the data distribution in the target domain and has the ability to extract domain-invariant health features and achieve high-precision cross-domain SOH estimation.

[0045] During the testing phase, performance was validated using the remaining 80% of data in the target domain that was not used for training. First, discriminative local and global features were extracted from the target domain running data using a cascaded structure of a 1-D CNN and a Transformer encoder. Then, based on the obtained optimal model parameters, an accurate mapping function from multi-dimensional features to SOH values ​​was constructed, ultimately outputting the SOH estimation results of the target domain battery.

[0046] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for estimating the state of health of lithium batteries based on meta-transfer learning, characterized in that... Includes the following steps: (1) Collect full life cycle data of lithium batteries with different material systems under different working conditions; (2) Divide the full lifecycle data collected in step (1) into source domain data and target domain data; (3) Use a feature extractor to extract local and global features of the voltage data collected in step (1) of the whole life cycle data; (4) Construct a lithium battery health status estimation model based on meta-transfer learning; (5) Use the source domain data in step (2) to pre-train the lithium battery health status estimation model constructed in step (4), update the parameters of the trainable part in the lithium battery health status estimation model, and obtain the trained lithium battery health status estimation model. Use the target domain data in step (2) to fine-tune the trained lithium battery health status estimation model, update the parameters of the trainable part in the lithium battery health status estimation model, and obtain the final model architecture. (6) Input the data in the target domain data in step (2) that did not participate in the fine-tuning in step (5) into the final model architecture obtained in step (5) to obtain the health status estimate of the corresponding battery sample.

2. The lithium battery health state estimation method based on meta-transfer learning according to claim 1, characterized in that: In step (1), the lithium batteries with different material systems include NCA, NCM and NCM+NCA. The different operating conditions are different ambient temperatures and different charge / discharge rates. The ambient temperatures include 25℃, 35℃ and 45℃, and the charge / discharge rates include 0.25C, 0.5C, 1C, 2C and 4C.

3. The lithium battery health state estimation method based on meta-transfer learning according to claim 1, characterized in that: In step (1), the full life cycle data refers to the data of the lithium battery aging from 100% healthy state to 80% healthy state.

4. The lithium battery health state estimation method based on meta-transfer learning according to claim 1, characterized in that: In step (1), when collecting data, multiple charge and discharge rates are used to conduct charge and discharge cycle tests on lithium batteries with different material systems under different ambient temperatures. During the test, the voltage, current and capacity of the battery are collected in real time.

5. The lithium battery health state estimation method based on meta-transfer learning according to claim 1, characterized in that: In step (2), dataset partitioning refers to partitioning across single variables, bivariates, and trivariates.

6. The lithium battery health state estimation method based on meta-transfer learning according to claim 5, characterized in that: Under the same material system and charge / discharge rate, batteries with ambient temperatures of 25℃, 35℃, and 45℃ were divided into three groups: A, B, and C. Groups A and B were used as source domain data, and group C was used as target domain data. The data partitioning was set as follows: In the pre-training stage, the first 70% of the data for each battery in the source domain was used as the support set, and the remaining 30% was used as the query set. In the fine-tuning stage, fine-tuning samples were extracted from the first 20% of the data for each battery in the target domain, and these samples were further divided into a 70% support set and a 30% query set. The remaining 80% of the data in the target domain was used for testing.

7. The lithium battery health state estimation method based on meta-transfer learning according to claim 1, characterized in that: In step (3), the feature extractor is based on a 1-D CNN and a Transformer encoder connected in series. The feature extractor first extracts local features of the battery data through the 1-D CNN layer. Then, the local features are organized into a token sequence and input into the Transformer encoder. The Transformer captures global features with the help of a self-attention mechanism.

8. The lithium battery health state estimation method based on meta-transfer learning according to claim 1, characterized in that: In step (4), the lithium battery health status estimation model based on meta-transfer learning is constructed. In the pre-training stage, it learns from multiple source tasks based on the meta-learning framework. Each source task calculates the loss function based on the support set and performs gradient updates. The updated model parameters are used to further calculate the loss on the query set. Finally, the global average loss is calculated by combining the losses of all source task query sets, and the meta-model parameters are updated accordingly. In the fine-tuning stage, it is fine-tuned using limited samples in the target domain. In this stage, based on the meta-learning framework, the model parameters are further optimized by performing cyclic updates on limited data so that it can quickly capture the specific degradation features of the target task.

9. The lithium battery health state estimation method based on meta-transfer learning according to claim 1, characterized in that: In step (6), the network is finally tested using test set data from the target domain that has not been used for training and fine-tuning, and the SOH estimate of the corresponding battery sample is obtained, thereby verifying the network's generalization ability and estimation accuracy in cross-domain scenarios.

Images