Lithium battery SOH estimation method fusing time sequence reconstruction and test training

Abstract

The invention discloses a lithium battery SOH estimation method fusing time sequence reconstruction and test training. According to the method, the complex features in the lithium battery aging process can be captured more comprehensively by fusing the multi-scale features. The battery pack level statistical characteristics can reflect the overall performance trend, the PCA dimension reduction characteristics of the single voltage can reveal the local subtle change, and the multi-dimensional characteristic construction mode is beneficial to improving the accuracy and robustness of SOH estimation, so that the model can understand the aging state of the battery from different levels, and the accuracy and robustness of SOH estimation are improved. A generalization enhancement mechanism effectively solves the problem of scarcity of supervision data in a traditional method. Pre-training is carried out by means of a Transform self-supervised voltage reconstruction task, the model can autonomously learn the general rule of battery aging under the condition that no SOH label exists, the model can keep good adaptability when facing different types of lithium batteries in different use environments due to the universal aging characterization, and the application range of the method is expanded.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of battery management, and specifically relates to a lithium battery SOH estimation method combining time sequence reconstruction and test training. BACKGROUND

[0002] In recent years, with the large-scale popularization of new energy vehicles and the rapid development of renewable energy storage systems, the performance, service life and safety of lithium ion battery packs as the core energy storage units are directly related to the operation efficiency and reliability of the entire system. Accurate assessment and prediction of the state of health (SOH) of the battery pack has become a key task of the battery management system (BMS). SOH is a core indicator representing the degradation of the current actual capacity, internal resistance and other key performance parameters of the battery relative to a new battery, and directly determines the effective available energy range of the system, the formulation of the operation strategy and the assessment of the remaining useful life (RUL). However, the accurate online estimation of the SOH of the battery pack in the actual complex operating environment and long service life service process faces many severe challenges.

[0003] The currently widely used SOH estimation methods mainly rely on offline trained models, including electrochemical characteristic analysis based on equivalent circuit models (ECM), feature engineering methods based on statistics and machine learning (such as extracting constant current charging voltage curve features, incremental capacity analysis ICA, etc.), and end-to-end modeling methods based on deep learning (such as CNN, LSTM) that have emerged in recent years.

[0004] Although deep learning methods exhibit advantages in automatic feature extraction and nonlinear modeling, they are still limited by the inherent limitations of static models: on the one hand, they are highly dependent on offline features that are carefully designed and selected, and the robustness and generalization of these features under actual complex operating conditions may not be sufficient; on the other hand, model parameters are usually frozen after deployment and lack the ability to adapt to dynamic changes in the distribution of actual operating data, resulting in a decline in model performance over time. In particular, when faced with new operating conditions, abnormal events or continuous aging trends that have not been trained, static models often perform poorly or even fail.

[0005] Therefore, developing a novel SOH estimation method that can adaptively adapt online, effectively address deviations in actual operating conditions, and integrate multi-dimensional aging characterization information is particularly urgent. This method not only needs to accurately extract key features that comprehensively describe battery aging characteristics, but also needs the ability to continuously sense environmental changes and self-adjust and optimize during actual operation, thereby ensuring high accuracy, robustness, and reliability of SOH estimation under complex dynamic conditions throughout the entire battery lifecycle. This technological innovation is of vital importance for improving the intelligence level of battery management systems, optimizing energy utilization efficiency, ensuring safe system operation, and maximizing battery lifespan. Summary of the Invention

[0006] The purpose of this invention is to provide a lithium battery SOH estimation method that integrates time-series reconstruction and test training in order to solve the problems mentioned above.

[0007] The technical solution adopted in this invention is as follows: a lithium battery SOH estimation method integrating time-series reconstruction and test training, the method comprising the following steps:

[0008] S1: First, the data is cleaned, and invalid data such as abnormal charging and discharging segments are deleted and analyzed. Charging segments covering 50% to 100% SOC are selected for subsequent processing. Then, hierarchical feature extraction is performed from two levels: battery pack and battery cell for subsequent model pre-training.

[0009] S2: Perform model pre-training. During the model pre-training stage, construct a dual-branch, dual-task learning framework. The network utilizes existing labeled data to perform multi-task learning, including SOH regression and self-supervised voltage sequence reconstruction tasks. The input to the SOH regression task is the hierarchical features extracted from the charging segments in step one, and the input to the voltage sequence reconstruction task is the voltage sequence of the charging segments.

[0010] S3: Fine-tuning during testing. In order to further improve the model's adaptability to external environmental shifts during actual operation, a test-time training strategy was introduced during the testing phase.

[0011] In a preferred embodiment, in step S1, the battery pack level features can be selected from the maximum single cell voltage, minimum single cell voltage, highest single cell temperature, lowest single cell temperature, charging current, charging voltage, trickle charging duration, average current of the trickle charging segment, charging energy, etc. For the single cell level features, PCA can be used to reduce the dimensionality of the single cell voltage data.

[0012] In a preferred embodiment, in step S1, the individual cell voltages at each time point are first subjected to zero-mean and standardization processing, so that the mean of each dimension of the data is 0 and the variance is 1, thereby constructing a standardized data matrix X∈R.N×K Where K represents the number of battery cells and N represents the number of samples, the covariance is calculated based on this matrix, as shown in formula (1).

[0013]

[0014] Then, eigenvalue decomposition is performed on the covariance matrix to obtain K pairs of eigenvalues ​​λ. k and the corresponding unit eigenvector v k Where λ k Indicates along v k The magnitude of the variance of the data retained after directional projection. Let λ (1) ≥λ (2) ≥...≥λ (20) If the eigenvalues ​​are arranged in descending order, then the variance ratio explained by the i-th principal component is:

[0015]

[0016] If the cumulative explained variance of the first L principal components Once the set threshold is reached, typically 95%, L principal components are retained, and the original high-dimensional vector is projected onto a subspace of dimension L.

[0017] To verify the correlation between the features extracted by hierarchical analysis and SOH, methods such as Pearson correlation coefficient, Spearman correlation coefficient, and grey correlation coefficient can be used to quantitatively evaluate the correlation.

[0018] In a preferred embodiment, in step S1, in order to verify the correlation between the features extracted hierarchically and SOH, methods such as Pearson correlation coefficient, Spearman correlation coefficient, and grey correlation coefficient can be used to quantitatively evaluate the correlation.

[0019] In a preferred embodiment, in step S2, the network architecture for the SOH regression task can be selected, but is not limited to, CNN + fully connected neural network. CNN can explore local correlations between features through convolutional kernels, while fully connected neural networks are good at learning global nonlinear interactions. For the voltage sequence reconstruction task, the Transformer model is used. The core of the Transformer is the self-attention mechanism, which can effectively capture long-distance dependencies between elements at any position in the sequence, making it very suitable for sequence prediction and reconstruction tasks.

[0020] In a preferred embodiment, in step S2, during the joint pre-training phase, the weighted mean square error loss is used as the loss function, defined as in formula (3):

[0021]

[0022] Where s is the true normalized SOH label. These are the model's predicted values; This is the voltage sequence after true alignment. For reconstructed output; λ recon To determine the weighting of the reconstruction task loss, a suitable value can be obtained through multiple experiments to strike a balance between improving generalization ability and avoiding excessive bias towards the reconstruction task.

[0023] In a preferred embodiment, in step S3, the model weights are first initialized to the optimal parameters obtained in the joint pre-training stage, and the relevant parameters of the SOH regression task head, as well as subsequent layer normalization and pooling operations, are frozen to ensure that the SOH regression features are not affected. Only the Transformer layer and the reconstruction head of the voltage sequence are fine-tuned, that is, only the weights of the Transformer layer and the reconstruction head are allowed to participate in gradient updates.

[0024] In a preferred embodiment, in step S3, for each sample in the actual operation process, forward inference and reconstruction output are performed on its voltage sequence, that is, the voltage sequence is reconstructed using a pre-trained model, and the self-supervised reconstruction loss is calculated, as shown in formula (4):

[0025]

[0026] Then, optimizations are performed on the parameters of the Transformer and the reconstruction head through backpropagation, so that L TTT It reaches its minimum after several iterations. After fine-tuning, forward inference is performed again using the same input, and the normalized SOH prediction value is output through the SOH regression head.

[0027] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0028] 1. In this invention, by fusing multi-scale features, the complex characteristics of the lithium battery aging process can be captured more comprehensively. Battery pack-level statistical features can reflect the overall performance trend, while PCA dimensionality reduction features of individual cell voltages can reveal subtle local changes. This multi-dimensional feature construction method helps improve the accuracy and robustness of SOH estimation, allowing the model to understand the battery aging state from different levels.

[0029] 2. In this invention, the generalization enhancement mechanism effectively solves the problem of scarce supervised data in traditional methods. By using the Transformer self-supervised voltage reconstruction task for pre-training, the model can autonomously learn the general laws of battery aging without SOH labels. This universal aging characterization enables the model to maintain good adaptability when facing different types of lithium batteries and different usage environments, thus expanding the applicability of the method.

[0030] 3. In this invention, the dynamic online adaptability ensures the reliability of the model in practical applications. The test-time training framework can locally fine-tune the Transformer parameters by reconstructing the loss based on the actual operating voltage sequence, enabling the model to autonomously adapt to the shift in operating condition distribution. Even when encountering complex situations such as temperature changes and load fluctuations during battery use, it can continuously and accurately estimate SOH, thus enhancing the practical value of the method in real-world scenarios. Attached Figure Description

[0031] Figure 1 This is a schematic diagram illustrating the process principle of the present invention;

[0032] Figure 2 This is a schematic diagram of the pre-trained model of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0034] Example:

[0035] Reference Figures 1-2 A novel lithium battery SOH estimation method integrating time-series reconstruction and test training is proposed. This method achieves highly robust online estimation through the following core innovations:

[0036] A hierarchical feature extraction system covering both battery pack and individual cell layers was constructed to solve the problem of multi-scale aging characterization.

[0037] A joint pre-training mechanism combining self-supervised voltage sequence reconstruction and SOH regression tasks is introduced to significantly improve feature generalization.

[0038] This invention designs a time-to-test (TTT) freeze-time fine-tuning strategy to achieve real-time adaptation of the model to dynamic operating condition distribution shifts. The technical solution adopted by this invention to solve its technical problem mainly includes the following parts (feature selection, model pre-training, and time-to-test fine-tuning), and the overall flowchart of the technical solution is shown below. Figure 1 As shown.

[0039] Specifically, the following steps are included:

[0040] Step 1: First, clean the data, delete faulty data such as abnormal charging and discharging segments, and analyze and process them. Select charging segments covering 50% to 100% SOC range for subsequent processing, and extract hierarchical features from two levels: battery pack and battery cell for subsequent model pre-training.

[0041] The battery pack level features can be selected from the maximum and minimum single-cell voltage, highest and lowest single-cell temperature, charging current, charging voltage, trickle charging duration, average current during trickle charging, and charging energy, etc. For the single-cell level features, PCA can be used to reduce the dimensionality of the single-cell voltage data.

[0042] For each individual cell voltage at any given time, the data is first normalized to zero mean, ensuring that the mean of each dimension is 0 and the variance is 1. This results in the construction of a standardized data matrix X∈R. N×K Where K represents the number of battery cells and N represents the number of samples, the covariance is calculated based on this matrix, as shown in formula (1).

[0043]

[0044] Then, eigenvalue decomposition is performed on the covariance matrix to obtain K pairs of eigenvalues ​​λ. k and the corresponding unit eigenvector v k Where λ k Indicates along v k The magnitude of the variance of the data retained after directional projection. Let λ (1) ≥λ (2) ≥...≥λ (20) If the eigenvalues ​​are arranged in descending order, then the variance ratio explained by the i-th principal component is:

[0045]

[0046] If the cumulative explained variance of the first L principal components Once the set threshold is reached, typically 95%, L principal components are retained, and the original high-dimensional vector is projected onto a subspace of dimension L.

[0047] To verify the correlation between the features extracted by hierarchical analysis and SOH, methods such as Pearson correlation coefficient, Spearman correlation coefficient, and grey correlation coefficient can be used to quantitatively evaluate the correlation.

[0048] Step 2: Model pre-training, the specific steps are as follows:

[0049] During the model pre-training phase, a dual-branch, dual-task learning framework is constructed. The network utilizes labeled data to perform multi-task learning, including SOH regression and self-supervised voltage sequence reconstruction tasks. The input to the SOH regression task is the hierarchical features extracted from the charging segments in the first step mentioned above, while the input to the voltage sequence reconstruction task is the voltage sequence of the charging segments.

[0050] For the SOH regression task, the network architecture can be selected, but is not limited to, CNN + fully connected neural network. CNN can explore the local correlations between features through convolutional kernels, while fully connected neural networks are good at learning global nonlinear interactions. For the voltage sequence reconstruction task, the Transformer model is used. The core of the Transformer is the self-attention mechanism, which can effectively capture the long-distance dependencies between elements at any position in the sequence, making it very suitable for sequence prediction and reconstruction tasks.

[0051] In the joint pre-training phase, the weighted mean squared error loss is used as the loss function, as defined in formula (3).

[0052]

[0053] Where s is the true normalized SOH label. These are the model's predicted values; This is the voltage sequence after true alignment. For reconstructed output; λ recon To determine the weighting of the reconstruction task loss, a suitable value can be obtained through multiple experiments to strike a balance between improving generalization ability and avoiding excessive bias towards the reconstruction task.

[0054] Step 3: Fine-tuning during testing

[0055] To further improve the model's adaptability to external environmental shifts during actual operation, a test-time training strategy was introduced during the testing phase. The specific process is as follows:

[0056] First, the model weights are initialized to the optimal parameters obtained during the joint pre-training phase, and the relevant parameters of the SOH regression task head, as well as subsequent layer normalization and pooling operations, are frozen to ensure that the SOH regression features are not affected. Only the Transformer layer and the reconstruction head of the voltage sequence are fine-tuned, that is, only the weights of the Transformer layer and the reconstruction head are allowed to participate in gradient updates.

[0057] For each sample in the actual operation process, its voltage sequence is forward inferenced and reconstructed, that is, the voltage sequence is reconstructed using the pre-trained model, and the self-supervised reconstruction loss is calculated, as shown in formula (4):

[0058]

[0059] Then, optimizations are performed on the parameters of the Transformer and the reconstruction head through backpropagation, so that L TTT It reaches its minimum after several iterations. After fine-tuning, forward inference is performed again using the same input, and the normalized SOH prediction value is output through the SOH regression head.

[0060] From the above, we can conclude that:

[0061] This method achieves accurate capture of multi-scale aging features by constructing a hierarchical feature extraction system covering both the battery pack and individual cell levels. At the battery pack level, key statistical features such as maximum and minimum cell voltage, maximum and minimum cell temperature, charging current, charging voltage, trickle charging duration, average current during trickle charging, and charging energy are extracted to comprehensively reflect the overall performance change trend of the battery pack. At the individual cell level, PCA is used to reduce the dimensionality of high-dimensional cell voltage data. A data matrix is ​​constructed through zero-mean and standardization, the covariance matrix is ​​calculated, and eigenvalue decomposition is performed. Principal components that explain up to 95% of the cumulative variance are selected to effectively extract subtle aging features contained in the cell voltage sequence. This hierarchical feature fusion approach not only preserves the macroscopic operating state information of the battery pack but also captures the differences between cells and local aging details, enabling the model to understand the complex physicochemical process of battery aging from different dimensions and providing more comprehensive and accurate feature inputs for SOH estimation.

[0062] The generalization enhancement mechanism significantly improves the model's feature learning ability and generalization performance by introducing joint pre-training of a self-supervised voltage sequence reconstruction task and a SOH regression task. During the model pre-training stage, leveraging the powerful sequence modeling capabilities of the Transformer, voltage sequence reconstruction is treated as a self-supervised task. Without relying on explicit SOH labels, the model learns the deep intrinsic representations of battery aging by reconstructing the input voltage sequence data. Simultaneously, a weighted mean squared error loss function is used to balance the training weights of the reconstruction and SOH regression tasks, achieving a balance between improving generalization ability and avoiding excessive bias towards the reconstruction task. This joint pre-training mechanism enables the model to mine implicit patterns of battery aging from a large amount of unlabeled data, reducing dependence on scarce SOH labeled data. This not only lowers data labeling costs but also enhances the model's adaptability to different types of lithium batteries and different usage environments, maintaining good estimation performance even when facing new battery samples or unknown operating conditions.

[0063] Dynamic online adaptability is achieved through a frozen test-time (TTT) fine-tuning strategy, enabling the model to accurately adapt to distribution shifts under actual operating conditions in real time. During the testing phase, model weights are initialized to the optimal parameters from the joint pre-training stage, freezing the parameters of the SOH regression task head and subsequent layer normalization and pooling operations. Fine-tuning is performed only on the Transformer layer and the reconstruction head of the voltage sequence reconstruction. For each sample in actual operation, the voltage sequence is reconstructed using the pre-trained model, the self-supervised reconstruction loss is calculated, and the weights of the Transformer layer and the reconstruction head are optimized through backpropagation to minimize the reconstruction loss. This fine-tuning method allows the model to autonomously adjust parameters based on real-time operating data while ensuring that the SOH regression features remain unaffected, effectively addressing distribution shifts caused by complex operating conditions such as temperature changes, load fluctuations, and voltage / current fluctuations. Even during the long-term use of the battery throughout its entire lifespan, the model can continuously learn new operating condition features, maintaining high accuracy and strong robustness in SOH estimation, significantly improving the reliability and practical value of the method in real-world applications.

[0064] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0065] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A lithium battery SOH estimation method integrating time-series reconstruction and test training, characterized in that: The method includes the following steps: S1: First, the data is cleaned, abnormal charging and discharging failure data is deleted and analyzed. Charging segments covering 50% to 100% SOC are selected for subsequent processing. Then, hierarchical feature extraction is performed from two levels: battery pack and battery cell for subsequent model pre-training. S2: Perform model pre-training. During the model pre-training stage, construct a dual-branch dual-task learning framework. The network uses existing labeled data to carry out multi-task learning, including SOH regression and self-supervised voltage sequence reconstruction tasks. The input of the SOH regression task is the hierarchical features extracted from the charging segment in the first step above, and the input of the voltage sequence reconstruction task is the voltage sequence of the charging segment. S3: Fine-tuning during testing. In order to further improve the model's adaptability to external environmental shifts during actual operation, a test-time training strategy was introduced during the testing phase.

2. The lithium battery SOH estimation method integrating time-series reconstruction and test training as described in claim 1, characterized in that: In step S1, the battery pack level features can be selected from the maximum single cell voltage, minimum single cell voltage, highest single cell temperature, lowest single cell temperature, charging current, charging voltage, trickle charging duration, average current of the trickle charging segment, and charging energy in the charging segment. For the single cell level features, PCA can be used to reduce the dimensionality of the single cell voltage data.

3. The lithium battery SOH estimation method integrating time-series reconstruction and test training as described in claim 1, characterized in that: In step S1, the individual voltage at each time point is first subjected to zero-mean and standardization processing, so that the mean of each dimension of the data is 0 and the variance is 1, and a standardized data matrix X∈R is constructed. N×K Where K represents the number of battery cells and N represents the number of samples, the covariance is calculated based on this matrix, as shown in formula (1). Then, eigenvalue decomposition is performed on the covariance matrix to obtain K pairs of eigenvalues ​​λ. k and the corresponding unit eigenvector v k ;where λ k Indicates along v k The magnitude of the variance of the data retained after directional projection; let λ (1) ≥λ (2) ≥...≥λ (20) If the eigenvalues ​​are arranged in descending order, then the variance ratio explained by the i-th principal component is: If the cumulative explained variance of the first L principal components Once the set threshold is reached, typically 95%, L principal components are retained, and the original high-dimensional vector is projected onto a subspace of dimension L. To verify the correlation between the features extracted hierarchically and SOH, Pearson correlation coefficient, Spearman correlation coefficient, and grey correlation method can be used to quantitatively evaluate the correlation.

4. The lithium battery SOH estimation method integrating time-series reconstruction and test training as described in claim 1, characterized in that: In step S1, in order to verify the correlation between the features extracted hierarchically and SOH, Pearson correlation coefficient, Spearman correlation coefficient, and grey correlation method can be used to quantitatively evaluate the correlation.

5. The lithium battery SOH estimation method integrating time-series reconstruction and test training as described in claim 1, characterized in that: In step S2, the network architecture for the SOH regression task can be selected, but is not limited to, CNN + fully connected neural network. CNN can explore the local correlation between features through convolutional kernels, while fully connected neural network is good at learning global nonlinear interactions.

6. The lithium battery SOH estimation method integrating time-series reconstruction and test training as described in claim 1, characterized in that: In step S2, during the joint pre-training phase, the weighted mean square error loss is used as the loss function, defined as in formula (3): Where s is the true normalized SOH label. These are the model's predicted values; This is the voltage sequence after true alignment. For reconstructed output; λ recon To determine the weighting of the reconstruction task loss, a suitable value can be obtained through multiple experiments to strike a balance between improving generalization ability and avoiding excessive bias towards the reconstruction task.

7. The lithium battery SOH estimation method integrating time-series reconstruction and test training as described in claim 1, characterized in that: In step S3, the model weights are first initialized to the optimal parameters obtained in the joint pre-training stage, and the relevant parameters of the SOH regression task head, as well as subsequent layer normalization and pooling operations, are frozen to ensure that the SOH regression features are not affected. Only the Transformer layer and the reconstruction head of the voltage sequence are fine-tuned, that is, only the weights of the Transformer layer and the reconstruction head are allowed to participate in gradient updates.

8. The lithium battery SOH estimation method integrating time-series reconstruction and test training as described in claim 1, characterized in that: In step S3, for each sample in the actual operation process, forward inference and reconstruction output are performed on its voltage sequence, that is, the voltage sequence is reconstructed using the pre-trained model, and the self-supervised reconstruction loss is calculated, as shown in formula (4): Then, optimizations are performed on the parameters of the Transformer and the reconstruction head through backpropagation, so that L TTT After several iterations, it reaches its minimum; after fine-tuning, it performs forward inference again using the same input, and outputs the normalized SOH prediction value through the SOH regression head.

Images