Small sample lithium battery SOH estimation method based on domain similar element learning

Abstract

The invention discloses a small sample lithium battery SOH estimation method based on domain similar element learning, and the method comprises the following steps: S1, designing a dynamic-static feature extraction module, extracting dynamic features and static features from a battery data set through the dynamic-static feature extraction module, and connecting the obtained features through a gray correlation coefficient; s2, designing a task sorting module based on domain similarity, calculating domain similarity among the training tasks by the task sorting module by utilizing a maximum mean difference MMD algorithm, and sorting the training tasks; and S3, adding an LSTM meta learning module into the meta learning framework, taking the LSTM model as a learning device, capturing a dependency relationship between the SOH and the voltage time sequence, and embedding the sorted training task into a meta training process. According to the small sample lithium battery SOH estimation method based on domain similar meta learning, a mathematical tool and a meta learning framework are combined, and the method is used for accurately estimating the SOH of the lithium battery under different working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of development and testing of lithium battery health state estimation methods, and in particular to a small sample lithium battery SOH estimation method based on domain similarity meta-learning. Background Art

[0002] In recent years, with the rapid development of the global economy, the demand for energy has continued to increase, and energy shortages and environmental pollution have gradually attracted close attention from countries around the world. All countries and regions have put forward the goal of energy conservation, emission reduction and environmental protection. Compared with lead-acid batteries, lithium batteries have about 5-7 times higher storage energy density, 2-6 times longer cycle life, and more efficient and stable operation. And because lead-acid batteries contain a variety of heavy metals and lead compounds in electrolyte solutions, they cause serious environmental pollution, while lithium batteries are relatively less polluting. Therefore, lithium batteries are currently one of the key directions in the energy field of various countries. Since 2019, the global demand for lithium batteries has continued to rise, and they have been widely used in electric vehicles, mobile communications, power grids, aerospace, and military industries.

[0003] The SOH of a battery is defined as the loss of capacity or the increase of resistance. It is usually impossible to measure directly and needs to be estimated by certain methods to provide effective information for the battery management system. Since the internal chemical reactions of lithium batteries are very complex and the aging of batteries is also affected by factors such as production process, operating conditions, and environment, it is difficult to establish an accurate physical model to estimate the SOH of lithium batteries.

[0004] In the prior art, the idea of ​​using data-driven methods to estimate the SOH of lithium batteries has attracted widespread attention from scholars at home and abroad. The data-driven method only needs to find certain rules based on the historical degradation data of the battery and compare it with the online operation data to estimate the SOH, without accurately describing its degradation mechanism, so it shows good application potential. However, since the aging of lithium batteries is a long process, it is relatively complex and difficult to obtain data through aging experiments, so there is relatively little data available for research. In addition, there are inconsistencies even between batteries from the same batch, and the storage conditions and charging and discharging conditions cannot be exactly the same. Therefore, the data used to establish data-driven modeling have distribution differences from the historical operation data, which affects the estimation effect of traditional methods. Due to the limited amount of data and large differences, in order to effectively estimate the SOH of lithium batteries with limited data in practical applications, it is particularly important to study the health status estimation method of lithium batteries under small sample conditions.

[0005] The use of meta-learning algorithms for lithium battery SOH estimation is effective to a certain extent. However, it also has some problems that need to be solved. When the working state of the battery changes, the entire network of the above method needs to be retrained, which limits the adaptability of the learning method to spacecraft battery SOH estimation. Summary of the invention

[0006] The purpose of this invention is to provide a small sample lithium battery SOH estimation method based on domain similarity meta-learning, which combines mathematical tools with the meta-learning framework to achieve accurate learning of meta-SOH in limited spacecraft data scenarios and different working conditions, and is used to accurately estimate the SOH of lithium batteries under different working conditions.

[0007] To achieve the above object, the present invention provides a small sample lithium battery SOH estimation method based on domain similarity meta-learning, comprising the following steps:

[0008] S1. Design a dynamic-static feature extraction module, use the dynamic-static feature extraction module to extract dynamic features and static features from the battery data set, and connect the obtained features through the grey correlation coefficient to collect sufficient information;

[0009] S2. Design a task sorting module based on domain similarity. The task sorting module uses the maximum mean difference (MMD) to calculate the domain similarity between training tasks, sort the training tasks, and reduce the retraining time.

[0010] S3. Add an LSTM meta-learning module to the meta-learning framework. The LSTM meta-learning module uses the LSTM model as a learner. The LSTM meta-learning module captures the long-term dependency between SOH and voltage time series and embeds the post-sorting training task into the meta-learning process.

[0011] Preferably, in S1, the dynamic-static feature extraction module includes a direct static feature extraction part, a direct dynamic feature extraction part and an internal dynamic feature extraction part. In the feature extraction process of the direct static feature extraction part, a fully connected neural network FCNN with three hidden layers is used. DS Fit the voltage curve, then take 20 points at equal intervals from the voltage range of the battery dataset and feed the obtained points into FCNN DS The corresponding time value obtained is the equal voltage difference discharge time, recorded as characteristic H TUd ;

[0012]

[0013] Among them, U i is the input voltage sequence, variable i is the discrete time, m is the number of cycles, and the fully connected neural network FCNN DSFor the network used to directly extract static features, the input voltage sequence U i Fitted to a continuous voltage curve U(t);

[0014] The direct dynamic feature extraction part extracts the features of the incremental capacity IC curve in the battery data set as the mechanism parameter features that reflect the battery health status, and establishes the following feature set:

[0015]

[0016] The IC curve reflects the battery status information, where H peak1 is the horizontal coordinate of the peak value of the IC curve at the lower right corner, H peak2 is the horizontal coordinate of the peak value of the IC curve at the upper left corner, H peak1 and H peak2 Both change with the number of cycles; H A and H V Respectively represent the mean and variance of the IC curve;

[0017] The internal dynamic feature extraction part uses the physical information neural network PINN to perform feature identification using the linear relationship between the ohmic resistance and the battery identified by the discharge curve in the battery data set to obtain the internal dynamic feature ohmic resistance value H R ;

[0018]

[0019] Among them, PINN is a physical information neural network, MLP is a multi-layer perceptron, and RNN is a recurrent neural network. After connecting PINN, MLP and RNN in series, the input voltage sequence U i and battery capacity Q i Identify and process to identify the internal ohmic resistance characteristics of the battery;

[0020] The extracted dynamic features and static features are processed and fused using the grey correlation coefficient to obtain the feature set Feature Set Used to input into subsequent networks.

[0021] Preferably, in S2, the calculation of the maximum mean difference MMD algorithm is based on the reproducing kernel Hilbert space, and the MMD quantifies the gap between two sample points by calculating the average distance between the two distribution sample points in the feature space. The calculation method is as follows:

[0022]

[0023] Among them, φ(·) is a nonlinear mapping function, H represents the mapping of the reproducing kernel Hilbert space, represents a randomly selected task, n is The total number of data in, m is The total number of data in, i and j are used for summing up the count, representing the i-th and j-th data respectively. Represents the training data during the meta-learning process.

[0024] Preferably, in S3, meta-learning includes two steps: a meta-training process and a meta-testing process. Tasks are introduced into the meta-learning framework and are used as training units.

[0025] First, the task data to be studied is used as the meta-test data set, and the task data of other categories are used as the meta-training data set; the samples in each task in the meta-training data set and the meta-test data set are divided into training data and test data respectively. The training data and test data are called the support set and query set respectively in meta-learning; in the meta-training stage, each task includes the same number of samples in the support set and the same number of samples in the query set;

[0026] When executing the gradient descent in the meta-learning algorithm, several tasks are randomly selected to form a batch. The meta-training stage includes a two-layer learning process, namely the task-level inner layer and the task-level outer layer. Each layer has a gradient update. The purpose of the task-level inner layer is to use the basic learner to find the optimal parameters for each task in the support set. Assume that the training model is f(θ), where θ represents the task parameters in the support set. In the inner loop, the training model is copied into multiple layers and used The tasks in the task layer are used to calculate the copy model loss respectively, and the first gradient descent is performed to calculate the parameter meta-level of the copy model; the outer layer of the task layer learns a common parameter representation across multiple tasks based on the query set, with the aim of quickly adapting to new tasks.

[0027] Preferably, the meta-training process based on task sorting includes the following specific steps:

[0028] Input the meta-training dataset and the meta-test support set to initialize the meta-learning learner;

[0029] Sample a batch of tasks from the meta-training dataset, calculate the domain similarity MMD between the batch of tasks and the meta-test support set, and re-rank the tasks according to the calculation results;

[0030] The reordered tasks use the gradient descent method to calculate the learner parameters in turn and update the system model.

[0031] Preferably, an LSTM model is constructed using Python language, Tensorflow framework and keras library. The LSTM meta-learning module uses the LSTM model as the basic learner for meta-learning. The parameter training of the LSTM network is different from the error back propagation and gradient descent method of ordinary neural networks, but adopts the task ranking meta-learning method.

[0032] Preferably, the method of task ranking meta-learning comprises the following specific steps:

[0033] First, randomly sample a batch of tasks from the meta-training dataset, and update the parameters based on the data of a batch. When using a batch to train the LSTM model, the calculation of each time step is performed in parallel, that is, after processing the first time step, the LSTM model will simultaneously process the second time step of all tasks in the batch, and so on;

[0034] Calculate the MMD between all tasks in a batch and the target test task, and sort the tasks in the batch according to the size of the MMD. The smaller the MMD, the higher the ranking. Then, input the support sets in the sorted tasks into the LSTM network in turn for forward calculation to obtain the output of the LSTM network, calculate the error loss function between the predicted value and the true value, and calculate the parameters of the replication model based on the gradient of the loss function; then input the meta-test data set into the replication model and calculate the loss function value for the second time; after all the tasks in the batch have gone through the above two operations, sum the meta-test data set loss values ​​of all the tasks in the batch, and calculate the gradient value. According to this gradient value, the training model parameters are actually updated to complete the gradient descent of the outer layer of the task;

[0035] After repeating the above process for several epochs, the training model f(θ) is trained. The trained model is called a meta-model. In ordinary neural networks, an epoch refers to a training process in which all data in the training data set are used only once. In an epoch, the training algorithm inputs all training data into the training model f(θ) in a set order for forward propagation, loss calculation, back propagation, and parameter update. In meta-learning, an epoch refers to a process in which all tasks in a batch are used once.

[0036] By inputting battery voltage data into the meta-model, the estimated value of battery SOH can be output, realizing accurate estimation of the health status of a small sample of lithium batteries.

[0037] Therefore, the present invention adopts the above-mentioned small sample lithium battery SOH estimation method based on domain similarity meta-learning, and the beneficial effects are as follows:

[0038] (1) The present invention combines the dynamic-static feature extraction method to solve the problem of insufficient information caused by limited data.

[0039] (2) The present invention uses a task ranking meta-learning method to obtain accurate small sample lithium battery SOH estimation, and enhances the application of data-based methods under variable battery operating conditions, thereby improving the generalization ability under different battery operating conditions.

[0040] (3) The present invention uses the MMD algorithm to calculate domain similarity and then sorts the training tasks. The algorithm sorts the training tasks from easy to difficult, thereby reducing the retraining time. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the overall scheme structure of an embodiment of a small sample lithium battery SOH estimation method based on domain similarity meta-learning of the present invention;

[0042] Figure 2 1 is a graph of the battery SOH estimation performance test results of a NASA data set of an embodiment of a small sample lithium battery SOH estimation method based on domain similarity meta-learning of the present invention, wherein (a) is the SOH estimation curve of Experiment 1, (b) is the SOH estimation error curve of Experiment 1, (c) is the SOH estimation curve of Experiment 2, (d) is the SOH estimation error curve of Experiment 2, (e) is the SOH estimation curve of Experiment 3, and (f) is the SOH estimation error curve of Experiment 3;

[0043] Figure 3 It is a curve chart of the battery SOH estimation performance test results of the MIT data set of an embodiment of the small sample lithium battery SOH estimation method based on domain similarity meta-learning of the present invention, wherein (a) is the SOH estimation curve of experiment one, (b) is the SOH estimation error curve of experiment one, (c) is the SOH estimation curve of experiment two, (d) is the SOH estimation error curve of experiment two, (e) is the SOH estimation curve of experiment three, and (f) is the SOH estimation error curve of experiment three. DETAILED DESCRIPTION

[0044] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0045] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0046] Example

[0047] like Figure 1 As shown, a small sample lithium battery SOH estimation method based on domain similarity meta-learning includes the following steps:

[0048] S1. Design a dynamic-static feature extraction module, use the dynamic-static feature extraction module to extract dynamic features and static features from limited battery data information, and connect the obtained features through the grey correlation coefficient to collect sufficient information.

[0049] The dynamic-static feature extraction module includes a direct static feature extraction part, a direct dynamic feature extraction part, and an internal dynamic feature extraction part. Furthermore, the feature extraction process of the direct static feature extraction part mainly uses a fully connected neural network FCNN with three hidden layers (32 neurons in each layer) DS To fit the voltage curve. Then take 20 points at equal intervals from the voltage range of the data set and bring them into FCNN DS The corresponding time value is the equal voltage difference discharge time, which is recorded as characteristic H TUd ;

[0050]

[0051] Among them, U i is the input voltage sequence, variable i is the discrete time, m is the number of cycles, and the fully connected neural network FCNN DS is a network for directly extracting static features, where the input voltage sequence U i Fit to a continuous function U(t).

[0052] The direct dynamic feature extraction part extracts the features of the incremental capacity IC curve as the mechanism parameter features that reflect the health status of the battery, and establishes the following feature set:

[0053]

[0054] The IC curve reflects the battery status information, where H peak1 is the horizontal coordinate of the peak value of the IC curve at the lower right corner, H peak2 is the horizontal coordinate of the peak value of the IC curve at the upper left corner, H peak1 and H peak2 Both change with the number of cycles; H A and H V Respectively represent the mean and variance of the IC curve;

[0055] The internal dynamic feature extraction part uses the physical information neural model PINN, and uses the linear relationship between the ohmic resistance and the battery identified by the discharge curve to perform feature identification to obtain the internal dynamic feature ohmic resistance value H R ;

[0056]

[0057] Among them, PINN is a physical information neural network, MLP is a multi-layer perceptron, and RNN is a recurrent neural network. After connecting PINN, MLP and RNN in series, the input voltage sequence U i and battery capacity Q i Identify and process to identify the internal ohmic resistance characteristics of the battery.

[0058] Furthermore, the extracted dynamic and static features are processed and fused using the grey correlation coefficient to obtain the feature set Feature Set Used to input into subsequent networks.

[0059] S2. Design a task ranking module based on domain similarity. The task ranking module uses one of the most widely used domain metrics in domain adaptation, the maximum mean difference (MMD). The calculation of MMD is based on the reproducing kernel Hilbert space (RKHS). The MMD algorithm quantifies the difference between two distributed sample points by calculating the average distance between them in the feature space. The calculation is performed by the following formula:

[0060]

[0061] Among them, φ(·) is a nonlinear mapping function, H represents the mapping of the reproducing kernel Hilbert space, represents a randomly selected task, n is The total number of data in, m is The total number of data in, i and j are used for summing up the count, representing the i-th and j-th data respectively. Represents the training data during the meta-learning process.

[0062] The smaller the MMD between two tasks, the more similar their data distributions are. By calculating the domain similarity between tasks, this method defines the difficulty of the tasks, so that the training tasks can be sorted from easy to difficult. Such an orderly learning strategy trains the outer layer of general knowledge representation, which helps to learn more difficult tasks later.

[0063] For the convenience of calculation, square the above formula to obtain the following calculation method:

[0064]

[0065] Among them, k is the Gaussian kernel function.

[0066] S3. Add an LSTM meta-learning module to the meta-learning framework. The LSTM meta-learning module uses LSTM as a learner to capture the long-term dependency between SOH and voltage time series, and embeds the sorted training task into the meta-training process to improve the adaptability to different working conditions.

[0067] Meta-learning mainly includes two steps: meta-training process and meta-testing process. In addition to data sets, meta-learning also introduces the concept of tasks, taking tasks as training units. First, the task data to be studied is used as meta-testing data, and the task data of other categories are used as meta-training data. Then, the samples in each task in the meta-training and meta-testing data sets are divided into training data and test data, which are called support set and query set in meta-learning. In the meta-training stage, each task includes the same number of samples in the support set and the same number of samples in the query set.

[0068] When performing the gradient descent in the meta-learning algorithm, a number of tasks are randomly selected to form a batch. Meta-training consists of a two-layer learning process, with a gradient update in each layer. The task-level inner layer is similar to traditional machine learning, and its goal is to use the basic learner to find the optimal parameters for each task in the support set. Assume that the model to be trained is f(θ), where θ represents the parameters of the tasks in the support set. In the inner loop, the original model is copied into multiple copies, using The tasks in the task list are used to calculate the copy model loss respectively, and the first gradient descent is performed to calculate the parameter meta-level of the copy model. The outer layer of the task is designed to learn a universal parameter representation across multiple tasks based on the query set, with the goal of quickly adapting to new tasks. The steps of the meta-training process after the above task ranking are as follows:

[0069] Input the meta-training dataset and the meta-test support set to initialize the meta-learning learner;

[0070] Sample a batch of tasks from the meta-training dataset, calculate the domain similarity MMD between the batch of tasks and the meta-test support set, and re-rank the tasks according to the calculation results;

[0071] The reordered tasks are used to calculate the learner parameters using the gradient descent method in sequence and update the system model.

[0072] An LSTM model was built using Python, Tensorflow framework, and keras library as the basic learner for meta-learning. The parameter training of the LSTM network is different from the error back propagation and gradient descent method of ordinary neural networks. Instead, it adopts the task ranking meta-learning method. The specific steps are as follows:

[0073] (1) Extract dynamic and static features from the data set and extract the feature set Then, different data sets can be divided into different training tasks according to different discharge strategies. The task construction process of a complete experiment is as follows: first determine the target task, randomly select a charging strategy containing at least two batteries from the battery data set as the charging strategy of the meta-test set, randomly select a battery data set from it as the support set of the meta-test data set for parameter fine-tuning, and then select another battery data set as the query set for the final verification of the effect of the method. The remaining charging strategies are all used as meta-training data sets.

[0074] (2) Select any two battery datasets from the meta-training dataset as the support set and query set, and take them 20 times in total to form 20 training batches.

[0075] (3) Randomly sample a batch of tasks. Divide the data set into several batches and update the parameters based on the data of a batch each time. When using batch training for LSTM models, gradient descent is performed in batches. Compared with stochastic gradient descent, the direction of each gradient descent in batch training is determined by a group of data rather than a single data, thus reducing randomness and improving convergence performance.

[0076] Calculate the MMD between all tasks in a batch and the target test task, and sort the tasks in the batch according to the size of the MMD. The smaller the MMD, the higher the ranking. Then, input the support set in the sorted tasks into the network in turn for forward calculation to obtain the network output, calculate the error loss function between the predicted value and the true value, and calculate the parameters of the copy model based on the gradient of the loss function; then input the test set into the copy model and calculate the loss function value for the second time. After all the tasks in the batch have gone through the above two operations, sum the meta-test data set loss values ​​of all tasks in the batch, and calculate the gradient. According to this gradient value, the training model parameters are actually updated to complete the gradient descent of the outer layer of the task.

[0077] Repeat the above two steps for several epochs. In common neural network methods, epoch refers to the training process in which all data in the training data set are used once (and only once). In an epoch, the training algorithm inputs all training data into the model in a set order for forward propagation, loss calculation, backpropagation, and parameter update. In meta-learning, epoch refers to the process of using all tasks in a batch once. The value range of epoch is usually between hundreds and thousands, and the meta-model training is completed.

[0078] By inputting battery voltage data into the trained meta-model, the estimated value of battery SOH can be output, realizing accurate estimation of the health status of a small sample of lithium batteries.

[0079] Finally, based on the NASA dataset and the MIT dataset, the common Gaussian process regression model (GPR), the transfer learning method (TL) of the same network and the meta-learning without task sorting (MAML) method are compared to show the comparative experimental results of the methods and verify the effectiveness of the meta-learning and task sorting strategies.

[0080] like Figure 2 As shown, according to the results of three experiments verified by NASA dataset, the results of each indicator are listed in Table 1;

[0081] Table 1. NASA dataset experimental SOH estimation performance test results

[0082]

[0083] Obviously, for the three experiments, the proposed domain similarity meta-learning method DS-MAML has the smallest error, R 2 As calculated in Table 1, the average MSE of the three experiments decreased by 69%, 77% and 88% compared with the MAML, TL and GPR methods respectively. 2 Compared with MAML, TL, and GPR methods, the results are improved by 6.5%, 92%, and 280%, respectively, indicating that the DS-MAML method has the best lithium battery SOH estimation effect in the case of small samples. The MAML method is second only to DS-MAML in accuracy, and the estimation result of MAML is better than that of transfer learning.

[0084] This shows the advantage of the meta-learning framework in adapting to new tasks. Although the transfer learning method can learn certain general knowledge through pre-training, it mainly focuses on adapting the model to specific target tasks rather than a general learning ability. Since the first 30% of the test set data is used during fine-tuning, it is easy to overfit the first half of the data, resulting in poor estimation results for the second half. Since the meta-learning method involves gradient calculations of both the inner and outer layers, the model training speed is slower than other methods. However, since the improved DS-MAML method sorts the tasks of the training set in advance, it has a faster convergence speed than MAML, and the running speed is increased by about 24%. Since the GPR model has a small number of parameters and is an analytical solution derived directly from the data, it does not require iterative optimization through optimization methods such as gradient descent, so the training time is the shortest, but its ability to adapt to new tasks is poor.

[0085] like Figure 3 As shown, three experimental results verified using the MIT dataset are listed in Table 2 for each indicator.

[0086] Table 2 Experimental SOH estimation performance test results of MIT dataset

[0087]

[0088]

[0089] It can be seen that since the SOH curve differences between batteries in the MIT dataset are smaller than those in the NASA dataset, all methods have good estimation effects. However, compared with the other three methods, the proposed DS-MAML method still has the smallest error and the highest SOH estimation accuracy for batteries. It can be calculated from the table that the average MSE of the three verification experiments of DS-MAML on the MIT dataset is 56%, 85% and 65% lower than that of MAML, TL and GPR methods respectively. The average R 2 Compared with MAML, TL, and GPR methods, the performance is improved by 0.63%, 3.9%, and 1.5% respectively. Compared with ordinary MAML, the addition of task sorting strategy makes the DS-MAML method have a faster model training speed and shortens the running time by 16%.

[0090] In summary, multiple experiments on the two data sets have shown that the proposed algorithm has good lithium battery SOH estimation performance in small sample conditions and can quickly adapt to battery data under different working conditions. It has become a specific example of accurate battery state estimation and health management for effective small sample spacecraft lithium-ion battery data.

[0091] Therefore, the present invention adopts the above-mentioned small sample lithium battery SOH estimation method based on domain similarity meta-learning, and improves the effectiveness and adaptability of the meta-learning algorithm through mathematical tools.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A small sample lithium battery SOH estimation method based on domain similarity meta-learning, characterized in that: The following steps are involved: S1. Design a dynamic-static feature extraction module, use the dynamic-static feature extraction module to extract dynamic features and static features from the battery data set, and connect the obtained features through the grey correlation coefficient to collect sufficient information; S2. Design a task sorting module based on domain similarity. The task sorting module uses the maximum mean difference (MMD) algorithm to calculate the domain similarity between training tasks, sort the training tasks, and reduce the retraining time. S3. Add an LSTM meta-learning module to the meta-learning framework. The LSTM meta-learning module uses the LSTM model as a learner. The LSTM meta-learning module captures the long-term dependency between SOH and voltage time series and embeds the post-sorting training task into the meta-learning process.

2. According to claim 1, a small sample lithium battery SOH estimation method based on domain similarity meta-learning is characterized by: In S1, the dynamic-static feature extraction module includes a direct static feature extraction part, a direct dynamic feature extraction part, and an internal dynamic feature extraction part. In the process of extracting features in the direct static feature extraction part, a fully connected neural network FCNN with three hidden layers is used. DS Fit the voltage curve, then take 20 points at equal intervals from the voltage range of the battery dataset and feed the obtained points into FCNN DS The corresponding time value obtained is the equal voltage difference discharge time, which is recorded as characteristic H TUd ; Among them, U i is the input voltage sequence, variable i is the discrete time, m is the number of cycles, and the fully connected neural network FCNN DS For the network used to directly extract static features, the input voltage sequence U i Fitted to a continuous voltage curve U(t); The direct dynamic feature extraction part extracts the features of the incremental capacity IC curve in the battery data set as the mechanism parameter features that reflect the battery health status, and establishes the following feature set: The IC curve reflects the battery status information, where H peak1 is the horizontal coordinate of the peak value of the IC curve at the lower right corner, H peak2 is the horizontal coordinate of the peak value of the IC curve at the upper left corner, H peak1 and H peak2 Both change with the number of cycles; H A and H V Respectively represent the mean and variance of the IC curve; The internal dynamic feature extraction part uses the physical information neural network PINN to perform feature identification using the linear relationship between the ohmic resistance and the battery identified by the discharge curve in the battery data set to obtain the internal dynamic feature ohmic resistance value H R ; Among them, PINN is a physical information neural network, MLP is a multi-layer perceptron, and RNN is a recurrent neural network. After connecting PINN, MLP and RNN in series, the input voltage sequence U i and battery capacity Q i Identify and process to identify the internal ohmic resistance characteristics of the battery; The extracted dynamic features and static features are processed and fused using the grey correlation coefficient to obtain the feature set Feature Set Used to input into subsequent networks.

3. The method for estimating SOH of a small sample lithium battery based on domain similarity meta-learning according to claim 2, characterized in that: In S2, the calculation of the maximum mean difference MMD algorithm is based on the reproducing kernel Hilbert space. The MMD algorithm is to calculate the difference between two distribution sample points in the feature set The average distance in space is used to quantify the gap between two sample points. The calculation method is as follows: Among them, φ(·) is a nonlinear mapping function, H represents the mapping of the reproducing kernel Hilbert space, represents a randomly selected task, n is The total number of data in, m is The total number of data in, i and j are used for summing up the count, representing the i-th and j-th data respectively. Represents the training data during the meta-learning process.

4. The method for estimating SOH of a small sample lithium battery based on domain similarity meta-learning according to claim 3 is characterized by: In S3, meta-learning includes two steps: meta-training process and meta-testing process. Tasks are introduced into the meta-learning framework and used as training units. First, the task data to be studied is used as the meta-test data set, and the task data of other categories are used as the meta-training data set; The samples in each task in the meta-training dataset and the meta-testing dataset are divided into training data and test data, which are called support set and query set in meta-learning. In the meta-training phase, each task includes the same number of samples in the support set and the same number of samples in the query set; When executing the gradient descent in the meta-learning algorithm, several tasks are randomly selected to form a batch. The meta-training stage includes a two-layer learning process, namely the task-level inner layer and the task-level outer layer. Each layer has a gradient update. The purpose of the task-level inner layer is to use the basic learner to find the optimal parameters for each task in the support set. Assume that the training model is f(θ), where θ represents the task parameters in the support set. In the inner loop, the training model is copied into multiple layers and used The tasks in the task layer are used to calculate the copy model loss respectively, and the first gradient descent is performed to calculate the parameter meta-level of the copy model; the outer layer of the task layer learns a common parameter representation across multiple tasks based on the query set, with the aim of quickly adapting to new tasks.

5. The method for estimating SOH of a small sample lithium battery based on domain similarity meta-learning according to claim 4, characterized in that: The specific steps of the meta-training process after task sorting are as follows: Input the meta-training dataset and the meta-test support set to initialize the meta-learning learner; Sample a batch of tasks from the meta-training dataset, calculate the domain similarity MMD between the batch of tasks and the meta-test support set, and re-rank the tasks according to the calculation results; The reordered tasks use the gradient descent method to calculate the learner parameters in turn and update the system model.

6. The method for estimating SOH of a small sample lithium battery based on domain similarity meta-learning according to claim 5, characterized in that: An LSTM model is built using Python language, Tensorflow framework and keras library. The LSTM meta-learning module uses the LSTM model as the basic learner for meta-learning. The parameter training of the LSTM network is different from the error back propagation and gradient descent method of ordinary neural networks. Instead, it adopts the task sorting meta-learning method.

7. The method for estimating SOH of a small sample lithium battery based on domain similarity meta-learning according to claim 6, characterized in that: The specific steps of the task ranking meta-learning method are as follows: First, randomly sample a batch of tasks from the meta-training dataset, and update the parameters based on the data of a batch. When using a batch to train the LSTM model, the calculation of each time step is performed in parallel, that is, after processing the first time step, the LSTM model will simultaneously process the second time step of all tasks in the batch, and so on; Calculate the MMD between all tasks in a batch and the target test task, and sort the tasks in the batch according to the size of the MMD. The smaller the MMD, the higher the ranking. Then, input the support sets in the sorted tasks into the LSTM network in turn for forward calculation to obtain the output of the LSTM network, calculate the error loss function between the predicted value and the true value, and calculate the parameters of the replication model based on the gradient of the loss function; then input the meta-test data set into the replication model and calculate the loss function value for the second time; after all the tasks in the batch have gone through the above two operations, sum the meta-test data set loss values ​​of all the tasks in the batch, and calculate the gradient value. According to this gradient value, the training model parameters are actually updated to complete the gradient descent of the outer layer of the task; After repeating the above process for several epochs, the training model f(θ) is trained. The trained model is called a meta-model. In ordinary neural networks, an epoch refers to a training process in which all data in the training data set are used only once. In an epoch, the training algorithm inputs all training data into the training model f(θ) in a set order for forward propagation, loss calculation, back propagation, and parameter update. In meta-learning, an epoch refers to a process in which all tasks in a batch are used once. By inputting battery voltage data into the meta-model, the estimated value of battery SOH can be output, realizing accurate estimation of the health status of a small sample of lithium batteries.

Images