MAML-based battery health state estimation method

Abstract

The invention discloses a battery health state estimation method based on MAML, and the method specifically comprises the following steps: collecting data sets of different types of batteries, carrying out the standardized preprocessing and SOH labeling of the data, dividing the existing data into a Support set and a Query set, and respectively carrying out the training and verification of an MAML model, a universal initialization model M and a model parameter theta are learned for different types of battery data sets, so that from m (theta), the model can be efficiently used for a new data set task through a small number of gradient steps, the SOH rapid estimation capability of MAML is evaluated on the new data set, a support set is used as data input, and training fine adjustment is performed on the basic model M; and performing verification evaluation through the query set to obtain an optimal model parameter theta * aiming at the new data set and task, and performing health state estimation on the target battery data set by using the model M (theta *). According to the method, a collaborative framework combining a traditional deep learning model and the MAML is provided, so that the model can quickly and accurately complete SOH estimation of a new task.

Description

Technical Field

[0001] The present invention relates to the technical field of battery state of health estimation, and specifically to a method for estimating the state of health of a battery based on MAML. Background Technique

[0002] As a core component of electric vehicles, electrochemical energy storage, electronic devices, etc., the accurate estimation of the state of health (SOH) of lithium batteries is directly related to the energy management of the system, and is also closely related to the state estimation, safety assessment, charge and discharge control, etc. after the attenuation of the battery. The two main parameters characterizing SOH are generally parameters such as capacity and internal resistance, but in actual applications, it is affected by factors such as the dynamic working conditions of battery charge and discharge, battery nonlinear aging, insufficient data volume or data quality.

[0003] Traditional SOH estimation methods (such as experience-based and mechanism-based models) are facing challenges: experience-based SOH estimation is not accurate enough and cannot meet the requirements of real-time online monitoring; although the mechanism model has physical interpretability, parameter identification is complex and the adaptability is poor. At present, many SOH estimations adopt data-driven methods to improve the prediction ability by mining data features, but they rely on a large amount of labeled data. In actual scenarios, the cost of obtaining battery aging data is high, and the generalization performance drops sharply when crossing battery models and working conditions. At present, there is still a lack of a method for estimating the state of health of a battery that is practical-oriented and applicable to actual electric vehicles, energy storage, and other scenarios.

[0004] Existing methods for estimating the state of health of a battery are mainly divided into three categories:

[0005] Empirical model: With the Arrhenius formula as the core, a model is constructed by fitting the capacity decay curve (such as a semi-empirical model combined with attenuation mechanisms such as LLI and LAM). Although the calculation amount is small, parameter identification depends on a large amount of cyclic data, and the adaptability to complex working conditions is poor. For example, the method of extracting charging curve features based on incremental capacity analysis (ICA) can correlate the attenuation mechanism, but still requires accurate laboratory data support and is difficult to generalize to actual dynamic scenarios.

[0006] Mechanism model: Based on an electrochemical model, superimposing side reactions such as lithium plating and SEI film growth to simulate the attenuation path. Although it has strong physical interpretability, the calculation complexity is high, and a large number of parameters (such as the loss rate of active materials and the impedance growth coefficient) need to be accurately calibrated, which is prone to overfitting due to parameter errors.

[0007] Data-driven model: Using deep learning (such as LSTM) or transfer learning to mine the capacity decay law from data. Although it avoids strong dependence on mechanisms, traditional methods require a large amount of labeled data for training, and the generalization ability is insufficient in scenarios with cross-battery models or small samples (such as transfer learning fine-tuning requires more than 20% of the target data), making it difficult to meet actual usage requirements. Summary of the Invention

[0008] (1) Technical problems to be solved

[0009] Aiming at the deficiencies of the prior art, the present invention provides a method for estimating the state of health of a battery based on MAML, which has the advantages of being practical-oriented and applicable to actual scenarios such as electric vehicle energy storage. It solves the challenges faced by traditional SOH estimation methods (such as experience-based and mechanism-model-based methods): Experience-based SOH estimation is not accurate enough and cannot meet the requirements of real-time online monitoring; although the mechanism model has physical interpretability, parameter identification is complex and its adaptability is poor. At present, many SOH estimations adopt data-driven methods to improve prediction ability by mining data features, but they rely on a large amount of labeled data. In actual scenarios, the cost of obtaining battery aging data is high, and the generalization performance drops sharply when crossing battery models and operating conditions. At present, there is still a lack of a method for estimating the state of health of a battery that is practical-oriented and applicable to actual scenarios such as electric vehicles and energy storage.

[0010] (2) Technical solutions

[0011] To achieve the above-mentioned objectives of strong few-shot learning ability, strong generalization across battery types and operating conditions, and high computational efficiency, the present invention provides the following technical solutions: A method for estimating the state of health of a battery based on MAML, to achieve accurate and rapid SOH estimation with robustness, generalization, and few-shot estimation ability for single cells, electric vehicles, and energy storage application scenarios, specifically including the following steps:

[0012] Step 1: Collection and preprocessing of existing data: Collect datasets of different types (generally ≥2) of batteries and perform standardized preprocessing and SOH labeling on the data.

[0013] a) The information of the existing dataset can be mainly divided into time series data and labeled data. Among them, the time series data mainly includes but is not limited to: time, current, voltage, maximum voltage of a single cell, minimum voltage of a single cell, temperature, maximum temperature, minimum temperature, etc., and the labeled data mainly includes but is not limited to: number of cycles, mileage, operating condition label (charging, discharging, stationary), etc.

[0014] b) The first step of data preprocessing mainly includes: data cleaning, abnormal sample removal, data interpolation, regularization, etc.; the second step mainly includes: calculating its variance, skewness, differential, etc. based on the current, temperature, voltage, etc. data of the battery, and using them together with the original time series data as input.

[0015] c) If each cycle of the data itself has an accurate SOH (capacity) label, there is no need to perform additional tagging; if not, identify and extract the data during the battery's stationary, charging, or discharging segments, and calculate the SOH capacity corresponding to the number of cycles through methods such as open-circuit voltage calibration and ampere-hour integration, which will be used as the battery's label.

[0016] Step 2: Divide the existing data into a Support set and a Query set; generally, the Support set accounts for a relatively large proportion of the total number of batteries (about 80% generally), and the remaining data is the Query set, which are respectively used for training and validating the MAML model.

[0017] Step 3: Meta-training: Learn a general initial model M and model parameters θ from different types of battery datasets, so that starting from m(θ), the model can be efficiently applied to new dataset tasks after a small number of gradient steps. The optimization of meta-training is divided into an inner loop and an outer loop.

[0018] a) Inner loop: Select a basic neural network model and perform a small number of gradient updates on a single battery dataset.

[0019] b) Outer loop: Train and evaluate the model parameters on different datasets to obtain a basic model M and its parameters θ that are comprehensively optimal for all input data. This can make the comprehensive performance of the model M(θ) optimal on different datasets, learning the general characteristics of battery aging rather than overfitting on a certain dataset.

[0020] Step 4: Meta-testing: Evaluate the SOH rapid estimation ability of MAML on a new dataset. Similarly, preprocess the new data samples and standardize them. At this time, select a small number of samples (5 - 20%) as the Support set, and the rest as the Query set.

[0021] Step 5: Use the Support set as the data input to perform training fine-tuning on the basic model M, and perform validation and evaluation through the Query set to obtain the optimal model parameters θ* for the new dataset and task.

[0022] Step 6: Use the model M(θ*) to estimate the health state of the target battery dataset.

[0023] Preferably, after preprocessing multiple input datasets, train to obtain a comprehensively optimal basic model; when facing a new dataset and task, perform rapid fine-tuning on the basic model to obtain a new optimal model for SOH estimation.

[0024] Preferably, for the new dataset, this method only requires a small number of samples for training to adapt to new scenario tasks, and the generalization and small-sample learning ability of the method are significantly superior.

[0025] Preferably, L1 and L2 regularization are two common regularization methods. For L1 regularization, its regularization formula is as follows:

[0026] R(w) = λ∑|w|

[0027] where R(w) represents the regularization term, w represents the weight parameters of the model, and λ is the regularization parameter.

[0028] Preferably, for L2 regularization, its regularization formula is as follows:

[0029] R(w) = λ∑(w 2 )

[0030] where R(w) represents the regularization term, w represents the weight parameters of the model, and λ is the regularization parameter.

[0031] Preferably, the data interpolation calculation formula is as follows:

[0032] f(x) = ∑[i = 0,n]yi*Li(x)

[0033] where f(x) represents the value of the unknown data point to be estimated, yi represents the value of the known data point, Li(x) represents the Lagrange interpolation polynomial, and n represents the number of known data points.

[0034] Preferably, the formula of the Lagrange interpolation polynomial is as follows:

[0035] Li(x) = Π[j = 0,n,j≠i](x - xj) / (xi - xj)

[0036] where i represents the subscript of the known data point currently being calculated, j represents the subscripts of other known data points, xj represents the abscissa of other known data points, and xi represents the abscissa of the known data point currently being calculated.

[0037] Preferably, the basic model M uses a total of two convolutional layers, and the rectified linear unit (ReLU) and hyperbolic tangent (tanh) are used as activation functions respectively. Among them, ReLU only takes the positive part of the input, while tanh maps the positive and negative parts of the input to the interval [-1,1] at the same time.

[0038] Preferably, the calibration formula of the open circuit voltage (Open Circuit Voltage, OCV)

[0039] V_open = Ф+ - Ф-

[0040] where Ф+ and Ф- are the electrode potentials of the positive and negative electrodes of the battery respectively.

[0041] (3) Beneficial effects

[0042] Compared with the prior art, the present invention provides a method for estimating the state of health of a battery based on MAML, which has the following beneficial effects:

[0043] 1. The method for estimating the state of health of a battery based on MAML, a battery SOH estimation framework based on MAML: a collaborative framework combining a traditional deep learning model and MAML is proposed, including meta-training and meta-testing. In the meta-training stage: perform co-training on multiple datasets across models and operating conditions to find the optimal parameters of the basic model; in the meta-testing stage: input a small number of samples of the new dataset for model fine-tuning, so that the model can quickly and accurately complete the SOH estimation of the new task.

[0044] 2. The method for estimating the state of health of a battery based on MAML, a small-sample fast fine-tuning mechanism: for a new target battery, only the historical cycle numbers of a small number of its samples need to be input to complete the parameter fine-tuning of the basic model, breaking through the dependence of traditional data-driven methods on a large amount of labeled data and meeting the fast deployment requirements of the real vehicle scenario.

[0045] 3. The method for estimating the state of health of a battery based on MAML, an enhanced design of cross-scenario generalization ability: by fusing the commonalities (such as non-linearity, diving, etc.) and personalized differences of battery degradation of different models and operating conditions through meta-training, the model maintains high robustness in scenarios of cross-battery types, aging modes, and noise interference. Description of the drawings

[0046] Figure 1 : A framework of a method for estimating the state of health of a battery based on MAML;

[0047] Figure 2 : Data of a certain charging segment of a battery dataset in an embodiment of the present invention;

[0048] Figure 3 : Curves of battery capacity decay with life for 5 datasets in an embodiment of the present invention;

[0049] Figure 4 : A framework of an SOH estimation model combining MAML and a neural network in an embodiment of the present invention;

[0050] Figure 5 : Comparison of root mean square errors of battery capacity estimation based on GCNN and combined with MAML (GCNNs-MAML) in an embodiment of the present invention. Detailed implementation manners

[0051] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0052] Please refer to Figures 1-5 , a method for estimating the state of health of a battery based on MAML, to achieve accurate and rapid SOH estimation with robustness, generalization, and small-sample estimation capabilities for single cells, electric vehicles, and energy storage application scenarios, specifically including the following steps:

[0053] 1. Collection and preprocessing of existing data: Collect datasets of different types (generally ≥2) of batteries and perform standardized preprocessing and SOH labeling on the data.

[0054] a) The information of the existing dataset can be mainly divided into time series data and labeled data. Among them, the time series data mainly includes but is not limited to: time, current, voltage, maximum cell voltage, minimum cell voltage, temperature, maximum temperature, minimum temperature, etc. The labeled data mainly includes but is not limited to: number of cycles, mileage, operating condition labels (charging, discharging, stationary), etc.

[0055] b) The first step of data preprocessing mainly includes: data cleaning, abnormal sample removal, data interpolation, regularization, etc.; the second step mainly includes: calculating its variance, skewness, differential, etc. based on the current, temperature, voltage, etc. data of the battery, and using them together with the original time series data as input.

[0056] c) If each cycle of the data itself has an accurate SOH (capacity) label, there is no need to do the labeling work again; if not, identify and extract the data of the battery in the stationary, charging, or discharging section, and calculate the SOH capacity corresponding to the number of cycles through methods such as open-circuit voltage calibration and ampere-hour integration as the label of the battery.

[0057] 2. Divide the existing data into a Support set and a Query set; among them, the support (Support) set generally accounts for a relatively large proportion of the total number of batteries (generally about 80%), and the remaining data is the query (Query) set, which are respectively used for training and validating the MAML model.

[0058] 3. Meta-training: Learn a general initialization model M and model parameters θ from different types of battery datasets, so that starting from m(θ), the model can be efficiently used for new dataset tasks after a small number of gradient steps. The optimization of meta-training is divided into an inner loop and an outer loop.

[0059] a) Inner loop: Select a basic neural network model and perform a small number of gradient updates on a single battery dataset.

[0060] b) Outer loop: Train and evaluate the model parameters on different datasets to obtain a basic model M that is comprehensively optimal for all input data and its parameters θ.

[0061] Doing so can make the comprehensive performance of the model M(θ) optimal on different datasets, learn the general characteristics of battery aging, rather than overfitting on a certain dataset.

[0062] 4. Meta-test: Evaluate the SOH rapid estimation ability of MAML on a new dataset. Similarly, preprocess the new data samples and standardize them. At this time, select a small number of samples (5 - 20%) as the support set, and the rest as the query set.

[0063] 5. Use the support set as the data input to perform training fine-tuning on the basic model M, and perform verification and evaluation through the query set to obtain the optimal model parameters θ* for the new dataset and task.

[0064] 6. Use the model M(θ*) to estimate the state of health of the target battery dataset.

[0065] The above is to achieve an accurate estimation of the current capacity of the battery based on MAML by inputting the historical cycle data of the already operating battery. This method has good accuracy, robustness, and generalization ability when facing different types of batteries and data samples. When the model is migrated to a new battery dataset and task, even when the amount of new data samples is very small, it can well complete the SOH estimation. This method is expected to be applied to actual scenarios to solve problems such as the state of health estimation of electric vehicles and energy storage.

[0066] Such as Figure 1 The SOH estimation framework and steps for battery cells based on MAML in an embodiment of the present invention are as follows:

[0067] 1. S1: Collect the cycle data of different types of batteries. In this embodiment, 5 different battery datasets are collected; 4 of them are used to train the basic model, and 1 is used for model fine-tuning and SOH estimation. Extract the charging segments according to the data information, and calculate the charging amount Q of each charging segment and its differential with voltage V (dQ / dV) according to information such as time, current, and voltage. Such as Figure 2 A 2000s-long charging segment of a certain battery data, including current, voltage, temperature, charging amount, and dQ / dV.

[0068] 2. S2: The dataset comes with the capacity information of the battery after each cycle. Such as Figure 3As shown, it is the change of the capacity of single battery cells in different datasets. It can be seen that under different charge and discharge cycle conditions such as different temperatures and currents, the attenuation paths and speeds of the batteries are different.

[0069] 3.S3: Select 4 of the battery datasets for the training of the basic model. In this embodiment, a Gated Convolutional Neural Network (GCNN) is selected as the basic deep learning model for parameter optimization. This implementation case combines GCNN to design the training framework of MAML. Use the 4 datasets to train the basic general model M(θ), and then fine-tune the basic model on the 5th dataset to obtain the optimal SOH estimation model M(θ*) for the new dataset. Its framework is as Figure 4 .

[0070] a) GCNN is a neural network that combines a convolutional layer and a gated layer, and is widely used in fields such as natural language processing and sequence analysis. The input charging segment is normalized as the input data and embedded into two convolutional layers for feature extraction. The output results pass through the gated layer, pooling layer, and fully connected layer respectively, and finally the training results are output. The working principles of the three key layers in GCNN are described in detail:

[0071] (1) Convolutional layer. This layer extracts features from the input sequence through convolution, and consists of N parallel N-dimensional convolutional kernels. The input matrix M is padded with a zero-dimensional vector of dimension (L - 1) / 2 to obtain M input , which is expressed as follows: padding ,

[0072]

[0073] Therefore, the output sequence y (n) of the convolutional layer can be calculated as follows:

[0074]

[0075] In the formula, b(n) is the bias vector, * represents the convolution calculation, and K (n) is the weight coefficient of the nth convolutional kernel, which can be calculated by the following formula:

[0076]

[0077] After the overall calculation, the input y conv :

[0078]

[0079]

[0080] In this implementation case, the model uses a total of two convolutional layers, and the rectified linear unit (ReLU) and hyperbolic tangent (tanh) are used as activation functions respectively. Among them, ReLU only takes the positive part of the input, while tanh maps both the positive and negative parts of the input to the interval [-1, 1]. The application of these two activation functions is very important for maintaining the integrity of features during training.

[0081] (2) Gating layer. This layer is used to filter redundant feature information, and the output is:

[0082]

[0083] In the formula represents the product of matrix elements, which are the outputs from the above two convolutional layers and are integrated and calculated here.

[0084] (3) Linear layer. The y of the pooling layer maxpool is linearly transformed through the weight vector W and the bias coefficient b, and the output after processing is:

[0085]

[0086] The multi-layer input and output structure enables the GCNN to flexibly handle the diversity and complexity of input data, which is beneficial for its battery life estimation in complex working conditions with diverse data.

[0087] Based on the GCNN, MAML uses meta-training to train on a dataset with 4 inputs to obtain the basic model M(θ). The specific steps are as follows:

[0088] Put all the training parameters into the basic model θ. For N training tasks Randomly select a task T i to train the model parameters θ i , and each task T i contains a small number of training samples, which are further divided into a support set and a query set, denoted as as the smallest training unit. Put T i into the inner loop training, and update the model θ i through a parameter search function to Note that is obtained through three gradient descent updates based on task T i . The formula for one gradient descent update is:

[0089]

[0090] where α is a hyperparameter of the step size, and its magnitude is related to the learning rate. For simplicity, only the formula for one-step gradient descent is listed here, and it is also expressed in the same way in the remaining parts of this section that require repeated gradient descent. In fact, the model undergoes multiple gradient updates during the training process, and only the calculation in Equation (3–16) needs to be repeated. After the update is completed, the Query set is used to evaluate the performance of the model . After separate training and optimization for each task, {θ′1, θ′2, …, θ' N} is obtained. The outer-loop training is based on task T S to train θ to minimize the objective loss function, which can be expressed as:

[0091]

[0092] Meta-optimization searches for the optimal θ over all tasks using the stochastic gradient descent method:

[0093]

[0094] where β is another hyperparameter related to the learning rate. In few-shot learning, the same loss function is used for the optimization in both the inner loop and the outer loop:

[0095]

[0096] where x j and y j are an input-output pair of task T i . In the K-shot regression task, each task provides K pairs of samples for the meta-training process.

[0097] 4.S4: After meta-training, use M(θ) to perform fine-tuning on a new dataset for meta-testing. Based on the GCNN, perform model fine-tuning on the support set used for testing. The original parameters of the model are already in a relatively optimal state. Facing a new task, only a small amount of fine-tuning is required to quickly update it into a model M(θ*) that adapts to the new task. Then, verify the SOH estimation effect on the query set of the test task .

[0098] 5.S5: Apply the model M(θ*) to perform SOH estimation on the third battery dataset. If there are other new datasets and battery health state estimation tasks, only repeat step S4 to obtain the new optimal model parameters for the corresponding scenario.

[0099] For example Figure 5 ​In the implementation case of the present invention, based on the comparison of the battery state of health estimation results of the traditional GCNN deep learning model and the model combined with MAML in dataset 5, it can be seen that the root mean square error (RMSE) of the estimation of the method proposed by the present invention is significantly smaller than that of the traditional method.

[0100] In summary, the battery state of health estimation method based on MAML has the following advantages:

[0101] 1. Significantly improved few-shot learning ability

[0102] Existing technology: Traditional machine learning models rely on a large amount of labeled data for training. In the estimation of battery SOH, especially in the case of new battery types or new working condition scenarios, the performance drops sharply when the data volume is insufficient.

[0103] Advantages of the present invention: Through the MAML meta-learning framework, a general initialization model is generated using multi-type battery data. Only a small number of samples (5% - 20%) are required for fine-tuning to adapt to new tasks, solving the problem of SOH estimation in data-scarce scenarios. The combined training of the inner loop (single-task fast adaptation) and the outer loop (multi-task global optimization) ensures that the model efficiently learns degradation features from a small number of samples and avoids overfitting.

[0104] 2. Strong generalization across battery types and working conditions

[0105] Existing technology: Traditional models are trained and tested on the same dataset, and due to differences in battery types, temperatures, and charge-discharge protocols, the generalization ability is poor, and re-training is required when migrating to new scenarios.

[0106] Advantages of the present invention: Through the training of multi-type battery data, the model learns the general aging features of batteries (such as voltage differential and capacity decay mode), does not overfit in a single dataset, and can be quickly migrated to new battery types or energy storage scenarios.

[0107] 3. Computational efficiency and real-time advantages

[0108] Existing technology: New scenarios require re-training the data from scratch, with high computational costs, making it difficult to meet the health state estimation requirements of different types of electric vehicles and energy storage.

[0109] Advantages of the present invention: Based on the pre-trained general model of MAML, only the model needs to be fine-tuned, and it can converge using a small number of gradient steps, greatly shortening the training time compared with traditional methods.

[0110] It should be noted that, in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variation thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the process, method, article or device comprising said element.

[0111] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A battery state of health (SOH) estimation method based on MAML, aiming to achieve accurate and rapid SOH estimation with robustness, generalization ability, and small-sample estimation ability for single cells, electric vehicles, and energy storage application scenarios. The method specifically includes the following steps: Step 1: Collection and preprocessing of existing data: Collect datasets of different types (generally ≥2) of batteries and perform standardized preprocessing and SOH labeling on the data. a) The information in the existing datasets can be mainly divided into time-series data and labeled data. The time-series data mainly includes, but is not limited to: time, current, voltage, maximum cell voltage, minimum cell voltage, temperature, maximum temperature, minimum temperature, etc. The labeled data mainly includes, but is not limited to: number of cycles, mileage, operating condition labels (charging, discharging, stationary), etc. b) The first step of data preprocessing mainly includes: data cleaning, abnormal sample removal, data interpolation, regularization, etc.; The second step mainly includes: calculating its variance, skewness, differential, etc. based on the current, temperature, voltage, etc. data of the battery, and using them together with the original time-series data as input. c) If each cycle of the data itself has an accurate SOH (capacity) label, there is no need to perform labeling work again; if not, identify and extract the data during the stationary, charging, or discharging segments of the battery, and calculate the SOH capacity corresponding to the number of cycles through methods such as open-circuit voltage calibration and ampere-hour integration as the label of the battery. Step 2: Divide the existing data into a Support set and a Query set; generally, the Support set accounts for a relatively large proportion (about 80% generally) of the total number of batteries, and the remaining data is the Query set, which are respectively used for training and validating the MAML model. Step 3: Meta-training: Learn a general initialization model M and model parameters θ from different types of battery datasets, so that starting from m(θ), the model can be efficiently used for new dataset tasks after a small number of gradient steps. The optimization of meta-training is divided into an inner loop and an outer loop. a) Inner loop: Select a basic neural network model and perform a small number of gradient updates on a single battery dataset. b) Outer loop: Train and evaluate the model parameters on different datasets to obtain a basic model M and its parameters θ that are comprehensively optimal for all input data. This can make the comprehensive performance of the model M(θ) optimal on different datasets, learn the general characteristics of battery aging, rather than overfitting on a certain dataset. Step 4: Meta-testing: Evaluate the SOH rapid estimation ability of MAML on a new dataset. Similarly, perform preprocessing on the new data samples and standardization. At this time, select a small number of samples (5 - 20%) as the Support set, and the remaining as the Query set. Step 5: Use the Support set as the data input to perform training fine-tuning on the basic model M, and perform verification and evaluation through the Query set to obtain the optimal model parameters θ* for the new dataset and task. Step 6: Use the model M(θ*) to estimate the state of health of the target battery dataset.

2. The battery health state estimation method based on MAML according to claim 1, characterized in that: After preprocessing multiple input data sets, a comprehensively optimal basic model is trained; when facing new data sets and tasks, the basic model is quickly fine-tuned to obtain a new optimal model for SOH estimation.

3. A method for estimating the state of health of a battery based on MAML according to claim 1, characterized in that: For the new data set, this method only needs a small number of samples for training to adapt to new scenario tasks, and the generalization and small-sample learning ability of the method are obvious.

4. A method for estimating the state of health of a battery based on MAML according to claim 1, wherein: L1 and L2 regularization are two common regularization methods. For L1 regularization, its regularization formula is as follows: R(w)=λ∑|w| where R(w) represents the regularization term, w represents the weight parameters of the model, and λ is the regularization parameter.

5. The method for estimating the state of health of a battery based on MAML according to claim 4, wherein: For L2 regularization, its regularization formula is as follows: R(w)=λ∑(w²) where R(w) represents the regularization term, w represents the weight parameters of the model, and λ is the regularization parameter.

6. The battery health state estimation method based on MAML according to claim 1, wherein: The calculation formula for data interpolation is as follows: f(x)=∑[i=0,n]yi*Li(x) where f(x) represents the value of the unknown data point to be estimated, yi represents the value of the known data point, Li(x) represents the Lagrange interpolation polynomial, and n represents the number of known data points.

7. A method for estimating the state of health of a battery based on MAML according to claim 1, characterized in that: The formula for the Lagrange interpolation polynomial is as follows: Li(x)=Π[j=0,n,j≠i](x-xj) / (xi-xj) where i represents the subscript of the known data point currently being calculated, j represents the subscripts of other known data points, xj represents the abscissa of other known data points, and xi represents the abscissa of the known data point currently being calculated.

8. The battery health state estimation method based on MAML according to claim 1, characterized in that: The basic model M uses a total of two convolutional layers, and the rectified linear unit (ReLU) and hyperbolic tangent (tanh) are used as activation functions respectively. Among them, ReLU only takes the positive part of the input, while tanh maps the positive and negative parts of the input to the interval [-1,1].

9. A method for estimating the state of health of a battery based on MAML according to claim 1, characterized in that: The calibration formula for the open circuit voltage (Open Circuit Voltage, OCV) Vopen = Ф+-Ф- where Ф+ and Ф- are the electrode potentials of the positive and negative electrodes of the battery respectively.

Images