Lithium ion battery health state prediction method and device and program product

Abstract

The invention discloses a lithium ion battery health state prediction method, a lithium ion battery health state prediction device and a program product, relates to the field of battery health detection, and aims to rapidly and accurately predict the health state of a new type of lithium ion batteries under the condition of few samples. The method comprises the steps of dividing a sample data set into a training sample set and a test sample set, performing subset division on the training sample set by adopting meta learning methods such as a model agnostic meta learning algorithm and the like, performing meta training on the training sample set by utilizing LSTM to complete network parameter initialization, and performing meta testing on an LSTM network on the test sample set to obtain a test result. The parameter adjustment of the LSTM network is completed, the health state of the future time step is predicted, and in the meta-learning process, the hyper-parameter of the LSTM network is optimized by using a particle swarm optimization method. According to the method, parameter adjustment of the prediction model can be achieved only through a small number of test samples, the method is suitable for prediction scenes with scarce data, and the method has a good convergence effect and high calculation efficiency.

Description

Technical Field

[0001] The present invention relates to the field of battery health detection, and in particular to a method, device and program product for predicting the health status of a lithium-ion battery. Background Art

[0002] Lithium-ion batteries are widely used in electric vehicles and aerospace due to their high energy density and durability. However, as lithium-ion batteries age, their capacity and internal resistance deteriorate, causing the battery's performance to gradually decline after repeated charge and discharge cycles. Once the battery performance declines beyond a certain threshold, internal short circuits and power drops may occur, leading to functional failures and even serious accidents.

[0003] State of Health (SOH) refers to the health status or performance level of the battery, reflecting the reduction of battery capacity and the increase of internal resistance, and characterizing the degree of battery aging. Predicting the health status of lithium-ion batteries is crucial to understanding battery deterioration and providing fault warnings, which can ensure the reliability of battery use.

[0004] Methods for estimating the health status of lithium-ion batteries can be roughly divided into two categories: traditional methods and deep learning-based methods.

[0005] Traditional methods rely on the physical and chemical properties of batteries, based on which knowledge-driven models for estimating the battery's state of health can be designed. Traditional methods conduct aging experiments on batteries, sample relevant operating parameters during the experiment, draw a chart that characterizes the battery aging process, and further fit an aging curve that characterizes the mapping relationship between the battery's state of health and the battery's operating parameters. With the help of this aging curve, the battery's state of health is matched by testing the operating parameters of the target battery. This method is severely affected by the aging experiment environment, operational norms, sampling accuracy, the number of samples tested, etc., the evaluation reliability is poor, and professionals are required to record experimental data for a long time.

[0006] The deep learning-based method can capture the complex nonlinear relationship in the operating parameters and can accurately predict the health status of the battery for the same type of battery. However, this method requires a large amount of data to train the inference model. Due to the differences in lithium-ion battery types, different types of lithium-ion batteries cannot share operating parameters to train the inference model, which is less practical in scenarios where only a small amount of data samples can be obtained.

[0007] In scenarios with limited data samples, some scholars have proposed transfer learning to solve the problem of few-sample training. This method pre-trains on data samples in the source domain and then transfers the pre-trained model to the target domain for further training, thus reducing the number of target domain data samples required. However, transfer learning relies heavily on the similarity between the source domain and the target domain. When the source domain and the target domain are very different, it may cause negative transfer, which in turn affects the learning effect in the target domain.

[0008] Therefore, it is of great significance to study how to improve the generalization ability of battery health status prediction under few sample conditions. Summary of the invention

[0009] The object of the present invention is to provide a method, device and program product for predicting the health status of lithium-ion batteries in order to solve all or part of the above problems, so as to achieve rapid and accurate prediction of the health status of new types of lithium-ion batteries with few samples.

[0010] The technical solution adopted by the present invention is as follows: A method for predicting the health status of a lithium-ion battery, comprising: Dividing a sample data set into a training sample set and a test sample set; the sample data set includes health status data and corresponding unhealthy status data continuously sampled during the charge and discharge cycle of multiple categories of lithium-ion batteries; Based on the training sample set, a meta-training method is used to initialize the network parameters of a LSTM (Long Short Term Memory) network; the LSTM network independently extracts features of the health status data and the unhealthy status data, and combines the features to predict the health status of the lithium-ion battery; Based on the test sample set, the network parameters of the LSTM network are optimized based on the initialized network parameters using a meta-testing method, and the health status of the lithium-ion battery in the future time step is predicted on the test sample set; In the meta-learning of the LSTM network, the particle swarm optimization method is used to optimize the hyperparameters.

[0011] In view of the above problems, the present invention further provides a lithium-ion battery health status prediction device, which comprises: A storage module, storing a training sample set and a test sample set divided by a sample data set, respectively; the sample data set includes health status data and corresponding unhealthy status data continuously sampled during the charge and discharge cycle of multiple categories of lithium-ion batteries; The model training module is connected to the storage module and is configured to: based on the training sample set, initialize the network parameters of the LSTM network using a meta-training method; the LSTM network independently extracts the features of the healthy state data and the unhealthy state data, and combines the features to predict the healthy state of the lithium-ion battery; and, based on the test sample set, optimize the network parameters of the LSTM network based on the initialized network parameters using a meta-testing method, and predict the healthy state of the lithium-ion battery in the future time step on the test sample set; in the meta-learning of the LSTM network, optimize the hyperparameters using a particle swarm optimization method.

[0012] In response to the above problems, the present invention also provides another lithium-ion battery health status prediction device, including a storage medium and a processor, wherein the storage medium stores computer instructions, and when the processor runs the computer instructions, it executes the above lithium-ion battery health status prediction method.

[0013] In view of the above problems, the present invention further provides a computer program product, including a computer program, which can execute the above-mentioned lithium-ion battery health status prediction method when executed by a processor.

[0014] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: The lithium-ion battery health status prediction method proposed in this application, with the help of meta-learning ideas, performs meta-training on the training sample set to complete the initialization of the LSTM network, and further completes the parameter adjustment of the LSTM network with a small number of samples in the test sample set as the prediction task, reducing the dependence on the number of data samples, and greatly reducing the iterative steps of network parameter optimization. Compared with the idea of ​​transfer learning, this application does not need to pay attention to the similarity between the training sample set and the test sample set, and can achieve rapid forward optimization of the target task prediction model in scenarios with large category differences. In addition, this application uses the LSTM network to independently analyze the change rules of operating parameters from the two dimensions of healthy state data and unhealthy state data, thereby improving the context perception ability of timing parameters, and then can improve the ability to extract potential features of operating parameters, improve convergence speed and prediction performance, and further reduce the demand for the number of target task data samples. This application uses the particle swarm optimization method to optimize the hyperparameters (model parameters) of the LSTM network. With the help of its global information sharing characteristics, it avoids the phenomenon of the LSTM network falling into the local optimum, making the LSTM network more robust, and improving the convergence speed of hyperparameter optimization and prediction efficiency. This application adopts a dual-subnet design structure for the LSTM network, cooperates with the particle swarm optimization method to optimize the hyperparameters, and combines the meta-learning concept. The three work together to reduce the demand for the number of target task data samples and achieve model optimization under small sample conditions. Furthermore, in the network parameter adjustment stage, only a small amount of data samples (10%) are required to adjust the network parameters to accurately predict the future health status of lithium-ion batteries. Compared with the conventional meta-learning method that needs to use at least 50% of the test sample set for parameter adjustment, in addition to improving the parameter adjustment efficiency, more target data samples can be predicted under a limited test sample set, thereby improving the comprehensiveness and interpretability of the prediction of the health status of the target lithium-ion battery. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The present invention will now be described by way of example with reference to the accompanying drawings, in which: Figure 1 It is an implementation flow chart of the lithium-ion battery health status prediction method provided in the embodiment of the present application.

[0016] Figure 2 It is a data flow diagram of the operation of the lithium-ion battery health status prediction method provided in the embodiment of the present application.

[0017] Figure 3 It is a data flow diagram of the LSTM network operation in the embodiment of the present application. DETAILED DESCRIPTION

[0018] All features disclosed in this specification, or steps in all methods or processes disclosed, except mutually exclusive features and / or steps, can be combined in any manner.

[0019] Any feature disclosed in this specification (including any additional claims and abstract), unless otherwise stated, may be replaced by other alternative features that are equivalent or have similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0020] In response to the problems that existing lithium-ion battery health status prediction methods are either unreliable, heavily dependent on the number of data samples, or have poor generalization ability, the present application proposes a lithium-ion battery health status prediction method, device and program product, which aims to quickly and accurately predict the health status of lithium-ion batteries in the target domain even when there are only a small number of samples in the target domain and the similarity with the source domain is not high.

[0021] like Figure 1 , Figure 2 As shown, the lithium-ion battery health status prediction method proposed in this application includes the following process: S1. Obtain a sample data set.

[0022] The sample data set includes healthy state data and corresponding unhealthy state data continuously sampled during the charge and discharge cycle of multiple categories of lithium-ion batteries.

[0023] The so-called health status data refers to data reflecting the health status of the lithium-ion battery. For example, the health status data includes the capacity calculated from the battery impedance.

[0024] Non-healthy state data refers to environmental parameters that are not directly related to the battery health state. For example, non-healthy state data includes temperature, voltage, and current.

[0025] For example, different single lithium-ion batteries or lithium-ion battery strings are used to distinguish the types of lithium-ion batteries. First, each type of lithium-ion battery is charged with a constant current (such as 1.5A) until the battery voltage reaches a predetermined value, and then charged with a constant voltage until the current decreases to a predetermined value; then each type of lithium-ion battery is discharged with a constant current (such as 2A) until the battery voltage decreases to a predetermined value. When the battery power drops by a predetermined proportion (such as 30%), an electrical signal with a frequency from 0.1Hz to 5KHz is used to measure impedance, and the internal impedance of each type of lithium-ion battery is obtained, and the capacity of the lithium-ion battery is calculated accordingly to obtain the health status data. Accordingly, while measuring the impedance, the ambient temperature and the voltage and current of each type of lithium-ion battery are recorded to obtain the unhealthy status data.

[0026] S2. Divide the sample data set into a training sample set and a test sample set.

[0027] The training sample set is used to train the inference model and belongs to the source domain, while the test sample set belongs to the target domain data sample for carrying out new prediction tasks.

[0028] S3. Based on the training sample set, the network parameters of the LSTM network are initialized using the meta-training method.

[0029] The meta-training method learns the common features of the LSTM network in multiple meta-task predictions by having the LSTM network perform predictions on multiple meta-tasks, thereby ignoring domain differences and model differences and retaining the common learning knowledge for predicting the health status of lithium-ion batteries.

[0030] As an optional implementation, the model agnostic meta learning algorithm (MAML, Model Agnostic Meta Learning) is used in the embodiments of the present application to carry out meta-training and meta-testing.

[0031] The above method of initializing the network parameters of the LSTM network using the meta-training method includes: S3-1: Divide the training sample set into different subsets according to different categories of lithium-ion batteries.

[0032] As mentioned above, each single lithium-ion battery can be regarded as a type (the same applies to lithium-ion battery strings). In this step, the healthy status data and the corresponding unhealthy status data sampled from each single lithium-ion battery can be regarded as a subset.

[0033] In addition, based on the meta-training concept, each subset is divided into a training set and a test set. The training set is used as a training sample when training the LSTM network with each subset, and the test set is used as a test sample when training the LSTM network with each subset, so as to adjust the network parameters of the LSTM network.

[0034] S3-2: With the goal of minimizing the loss of the LSTM network on each subset, the LSTM network is trained using each subset to obtain multiple sets of network parameters.

[0035] Assume that the training sample set is divided into K (K is a positive integer) subsets. In each batch, select N subsets to carry out the meta-task (N is a positive integer less than or equal to K). Assume that the dataset composed of N subsets in the jth batch is Each subset is trained once, and the LSTM network is trained N times to obtain N sets of network parameters. represents the training task carried out with the i-th subset, that is, the i-th training task, , in the i-th subset, the input sequence and the corresponding true value are represented as The network parameters (i.e., optimal network parameters) obtained by executing the i-th training task are , then in each training task, The update method is expressed as: , In the formula, Indicates the execution of training tasks The network parameters are randomly initialized when Indicates the network parameters The prediction results for the i-th subset are as follows: For training tasks Prediction results of loss, is the gradient of the loss, is a trainable parameter.

[0036] The training tasks are carried out on N subsets respectively, and the obtained N groups of network parameters are expressed as .

[0037] S3-3: With the goal of minimizing the loss of the LSTM network on the above multiple subsets, the final network parameters are obtained by optimizing multiple sets of network parameters to complete the initialization of the network parameters of the LSTM network.

[0038] Step S3-2 belongs to the inner optimization step. In step S3-2, each training batch obtains N sets of network parameters. Under this batch, the final network parameters are obtained by iterative optimization using the following method: , In the formula, is a trainable parameter, represents the updated network parameters, According to the data set All training missions carried out , Indicated in the parameter The prediction results for the i-th subset are as follows: For training tasks Prediction results of loss, is the gradient of the loss.

[0039] Step S3-3 belongs to the outer optimization step, and the outer optimization is completed by the meta-learner, and each network parameter obtained by the inner optimization is passed to the meta-learner. The inner optimization step and the outer optimization step are nested and executed under each batch. In some feasible implementations, in the meta-training phase, each training task performs 3000 iterations of optimization, the learning rate in the inner optimization step is designed to be 0.001, and the learning rate in the outer optimization step is designed to be 0.0001, with 50% of the subset used for inner optimization and 50% for outer optimization.

[0040] After the above optimization, the final network parameters are obtained , the network parameters are initialized, and the initialized network parameters are used as the initial network parameters for the subsequent meta-prediction stage when training for the health status prediction task of a specific type of lithium-ion battery, so that the knowledge learned from the health status prediction method of lithium-ion batteries can be directly used in the specific recognition task. At the same time, through the type-independent prediction knowledge learned from multiple types of lithium-ion batteries, only a small number of target type lithium-ion battery data samples are needed to complete the parameter adjustment of the LSTM network, reducing the number of iterations of the training inference model when predicting the health status of a specific type of lithium-ion battery, improving the model fitting efficiency, and improving the prediction efficiency.

[0041] The LSTM network used in the embodiment of the present application adopts two branches to independently extract the features of the healthy status data and the unhealthy status data, and combines the features to predict the health status of the lithium-ion battery.

[0042] like Figure 3 As shown, in some optional embodiments, the LSTM network includes a first LSTM subnetwork, a second LSTM subnetwork, a feature fusion module and a feature mapping module. The first LSTM subnetwork extracts a first feature from the health status data; the second LSTM subnetwork extracts a second feature from the unhealthy status data; the feature fusion module uses a global average pooling layer to fuse the first feature and the second feature; the feature mapping module uses a fully connected layer to map the health status of the lithium-ion battery with the fused features. The first LSTM subnetwork and the second LSTM subnetwork each include multiple LSTM layers. Each LSTM layer includes three components: a forget gate, an input gate, and an output gate.

[0043] ‌Forget Gate‌: Responsible for deciding which information should be forgotten. It outputs a value between 0 and 1 through a sigmoid layer, indicating the degree of information retention or forgetting. 1 means "completely retained" and 0 means "completely forgotten".

[0044] ‌Input Gate‌: Responsible for processing new input information and adding it to the cell state. It passes through a sigmoid layer to decide which information should be updated and a tanh layer to generate new candidate values.

[0045] ‌Output gate‌: determines what information should be output. It determines the output content through a sigmoid layer and outputs it after processing through a tanh function.

[0046] The so-called ‌cell state, similar to the hidden state of a neural network, but with a richer internal structure, is responsible for storing and transmitting long-term dependency information. The cell state is selectively updated through a gating mechanism rather than directly undergoing linear or nonlinear transformations.

[0047] S4. Based on the test sample set, the meta-test method is used to optimize the network parameters of the LSTM network based on the initialized network parameters, and the health status of the lithium-ion battery in the future time step is predicted on the test sample set.

[0048] Each test sample in the test sample set is the sample data of the target type of lithium-ion battery. In the meta-test phase, only a small amount of sample data (such as 10%) is needed for each target type of lithium-ion battery to adjust the LSTM network parameters of the initialized network parameters (the adjustment of the network parameters is completed by continuing to train the LSTM network), and the final network parameters are obtained. Then, the remaining sample data (such as 90%) is used to predict the health status of the lithium-ion battery in the future time step.

[0049] In addition, in this application, in the meta-learning of the LSTM network, the particle swarm optimization method is used to optimize the hyperparameters.

[0050] The hyperparameters of the LSTM network include the length of the input sequence, the number of hidden units, and the number of recurrent layers, which are model parameters rather than network parameters.

[0051] In some feasible implementations, the method of optimizing the above hyperparameters using the particle swarm optimization method includes: According to the set particle swarm size, initialize the velocity vector and position vector of each particle. Assuming the velocity vector is v and the position vector is m, the velocity vector of the i-th particle is expressed as , the position vector is expressed as , , M is the particle swarm size, that is, the total number of particles. The velocity vector and position vector of each particle include three dimensions: the length of the input sequence, the number of hidden units, and the number of recurrent layers, namely , The speed and position represent the length of the input sequence, the number of hidden units, and the number of recurrent layers respectively.

[0052] According to the set iteration rounds, the velocity vector and position vector of each particle are iteratively updated with the goal of minimizing the loss of the LSTM network.

[0053] In the particle swarm optimization process, the best position of each particle is recorded in each iteration. and the optimal position of the particle swarm , where the reciprocal of the loss value (absolute value) of the LSTM network is used as the fitness of the particle in each round of iteration. By comparing it with the fitness of the previous round of iteration, if the fitness is higher, the position vector of the current round is used as the optimal position of the particle. ; If the maximum fitness of all particles in the particle swarm is higher than the fitness of the best position of the particle swarm in the previous round of iteration, the best position of the particle swarm is updated with the position vector of the particle with the largest fitness in the current round . Then update the position vector of each particle using the following method: , In the formula, j represents the jth dimension of the particle, j={1,2,3}, j=1 represents the length dimension of the input sequence, j=2 represents the number of hidden units, and j=3 represents the number of recurrent layers. represents the j-th dimension of the velocity vector of the ith particle, represents the jth dimension of the position vector of the ith particle, t represents the current number of iterations, are acceleration constants, usually in the interval (0,2). are two independent random numbers in the range of [0,1], and w represents the inertia weight, which is used to adjust the search ability of the solution space.

[0054] With the goal of minimizing the loss of the LSTM network, the velocity vector and position vector are continuously optimized based on the set iteration rounds. Finally, the optimal position of the particle swarm obtained in the last round of iteration is used as the final position vector.

[0055] Parse the resulting position vector to get the length of the input sequence, the number of hidden units, and the number of recurrent layers.

[0056] As an optional implementation, the loss of the LSTM network mentioned in the above embodiment can adopt mean square error loss, which is represented by L and is calculated as follows: , In the formula, Respectively represent the input sequence and the corresponding true value of the LSTM network, D is the sample set, Represents all input sequences and corresponding true values ​​in the sample set D, Represents the LSTM network for the input sequence The predicted value of Express request The L2 norm of , N is the number of samples.

[0057] In the embodiment of the present application, the performance of the designed method is also verified. In the embodiment of the present application, NASAP CoE is used as the sample data set of lithium-ion batteries, which records the voltage, current, temperature (i.e., unhealthy state data) and capacity (i.e., healthy state data) collected by the lithium-ion battery during the charge and discharge cycle device. All data in the sample data set are standardized by minimum-maximum normalization.

[0058] As in the previous implementation, the sample data set used is divided into a training sample set and a test sample set. In the meta-training stage, sample data of lithium-ion batteries numbered #5 ("#" indicates a number), #6, #7, and #18 (i.e., healthy state data and corresponding unhealthy state data) are selected to perform 3000 rounds of iterative training on the LSTM network, and 50% of the sample data are still used for inner layer optimization, with the learning rate set to 0.001, and 50% of the sample data are used for outer layer optimization, with the learning rate set to 0.0001. During the meta-training, the particle swarm optimization method is used to optimize the hyperparameters of the three dimensions of the input sequence length, the number of hidden units, and the number of loop layers of the LSTM network, and the optimization results are 6, 3, and 3, respectively. The initialization network parameters of the LSTM network are obtained in the meta-training stage.

[0059] In the meta-test phase, lithium-ion batteries numbered #45, #45, #47, #54, and #56 are respectively used as target types of lithium-ion batteries (referred to as test batteries) for health status prediction. For each test battery, the first 10% of the sample data are selected for initializing the network parameters of the LSTM network, the iteration round is set to 1000 times, the learning rate is set to 0.001, and the remaining 90% of the sample data are used to test the performance of the LSTM network. In the embodiment of the present application, the network parameter optimization process is optimized using the Adam optimizer.

[0060] In order to verify the performance of the solution of this application, the S-LSTM network, RNN network and P-LSTM network were selected as baseline models to compare their performance with the method proposed in this application (called Ours).

[0061] The performance evaluation indicators include mean absolute error (MAE), root mean square error (RMSE), mean square error (MSE) and mean absolute percentage error (MAPE).

[0062]

[0063] In the formula, represents the sample data of the lithium-ion battery tested for the kth time, N is the number of samples in the sample data, express All input sequences in and the corresponding truth value , Represents the input sequence The predicted value of .

[0064] The RMSE measurement results in the experimental results are shown in Table 1.

[0065] Table 1 Lithium-ion battery prediction performance index (RMSE)

[0066] The lower the performance index, the better the performance. It can be clearly seen from the experimental results in Table 1 that the prediction performance of the method designed in this application on each test battery is almost entirely due to the baseline model, and the performance is greatly improved. For lithium-ion batteries numbered #45, #46, #47, #54 and #55, the performance is improved by 29%, 26%, 40%, 46% and 20% respectively compared with the best network in the baseline model.

[0067] In addition, some studies have demonstrated the predictive performance indicators for #18 lithium-ion batteries. In the examples of this application, performance tests were also conducted on #18 batteries, and the experimental results are shown in Table 2.

[0068] Table 2 Lithium-ion battery prediction performance indicators (RMSE, MAE)

[0069] As can be seen from Table 2, compared with the errors of recurrent neural networks such as RNN, GRU and DDAN, the method provided in this application has excellent prediction performance.

[0070] Based on the concept of the present application, an embodiment of the present application further provides a lithium-ion battery health status prediction device, which includes: The storage module stores the training sample set and the test sample set divided by the sample data set respectively; the sample data set includes the health status data and the corresponding unhealthy status data continuously sampled during the charge and discharge cycle of multiple categories of lithium-ion batteries.

[0071] The model training module is connected to the storage module and is configured as follows: based on the training sample set, the network parameters of the LSTM network are initialized using a meta-training method; the LSTM network independently extracts the features of the healthy state data and the unhealthy state data, and combines the features to predict the healthy state of the lithium-ion battery; and, based on the test sample set, the network parameters of the LSTM network are optimized based on the initialized network parameters using a meta-testing method, and the healthy state of the lithium-ion battery in the future time step is predicted on the test sample set; in the meta-learning of the LSTM network, the particle swarm optimization method is used to optimize the hyperparameters.

[0072] When the present application is implemented, the data specifically configured for the storage module and the model training module in the lithium-ion battery health status prediction device can refer to the relevant features in the above method embodiments.

[0073] On the other hand, an embodiment of the present application further provides a lithium-ion battery health status prediction device, including a storage medium and a processor, wherein the storage medium stores computer instructions, and when the processor runs the computer instructions, the lithium-ion battery health status prediction method of the above embodiment is executed.

[0074] In addition, a computer program product is provided in an embodiment of the present application, including a computer program. When the computer program is executed by a processor, the lithium-ion battery health status prediction method of the above embodiment can be executed.

[0075] The present invention is not limited to the above-mentioned specific embodiments, but extends to any new features or any new combination disclosed in this specification, as well as any new method or process steps or any new combination disclosed.

Claims

1. A method for predicting the health status of a lithium-ion battery, characterized in that: include: Divide the sample data set into a training sample set and a test sample set; The sample data set includes health status data and corresponding unhealthy status data continuously sampled during the charge and discharge cycle of multiple categories of lithium-ion batteries; Based on the training sample set, the network parameters of the LSTM network are initialized using a meta-training method; the LSTM network independently extracts features of the health status data and the unhealthy status data, and combines the features to predict the health status of the lithium-ion battery; Based on the test sample set, the network parameters of the LSTM network are optimized based on the initialized network parameters using a meta-testing method, and the health status of the lithium-ion battery in the future time step is predicted on the test sample set; In the meta-learning of the LSTM network, the particle swarm optimization method is used to optimize the hyperparameters.

2. The method for predicting the health status of a lithium-ion battery according to claim 1, wherein: The LSTM network includes a first LSTM subnetwork, a second LSTM subnetwork, a feature fusion module and a feature mapping module; the first LSTM subnetwork extracts a first feature from the health status data; the second LSTM subnetwork extracts a second feature from the unhealthy status data; the feature fusion module uses a global average pooling layer to fuse the first feature and the second feature; the feature mapping module uses a fully connected layer to map the health status of the lithium-ion battery with the fused features.

3. The method for predicting the health status of a lithium-ion battery according to claim 1, wherein: Based on the training sample set, the network parameters of the LSTM network are initialized using a meta-learning method, including: Dividing the training sample set into different subsets according to different categories of lithium-ion batteries; With the goal of minimizing the loss of the LSTM network on each subset, the LSTM network is trained using multiple subsets to obtain multiple sets of network parameters; With the goal of minimizing the loss of the LSTM network on the multiple subsets, the final network parameters are obtained by optimizing the multiple groups of network parameters, thereby completing the initialization of the network parameters of the LSTM network.

4. The method for predicting the health status of a lithium-ion battery according to claim 1, wherein: The hyperparameters include the length of the input sequence of the LSTM network, the number of hidden units, and the number of recurrent layers.

5. The method for predicting the health status of a lithium-ion battery according to claim 4, wherein: The method of optimizing hyper parameters by using particle swarm optimization includes: Initialize the velocity vector and position vector of each particle according to the set particle swarm size, where the velocity vector and position vector of each particle include three dimensions: the length of the input sequence, the number of hidden units, and the number of recurrent layers; According to the set iteration rounds, the velocity vector and position vector of each particle are iteratively updated with the goal of minimizing the loss of the LSTM network; Parse the resulting position vector to get the length of the input sequence, the number of hidden units, and the number of recurrent layers.

6. The method for predicting the health status of a lithium-ion battery according to claim 3 or 5, characterized in that: The loss is the mean square error loss.

7. The method for predicting the health status of a lithium-ion battery according to claim 1, wherein: In the sample data set, each single lithium-ion battery or each lithium-ion battery string is taken as a category.

8. A lithium-ion battery health status prediction device, characterized in that: include: A storage module, storing respectively a training sample set and a test sample set divided by the sample data set; The sample data set includes health status data and corresponding unhealthy status data continuously sampled during the charge and discharge cycle of multiple categories of lithium-ion batteries; The model training module is connected to the storage module and is configured to: based on the training sample set, initialize the network parameters of the LSTM network using a meta-training method; the LSTM network independently extracts the features of the healthy state data and the unhealthy state data, and combines the features to predict the healthy state of the lithium-ion battery; and, based on the test sample set, optimize the network parameters of the LSTM network based on the initialized network parameters using a meta-testing method, and predict the healthy state of the lithium-ion battery in the future time step on the test sample set; in the meta-learning of the LSTM network, optimize the hyperparameters using a particle swarm optimization method.

9. A lithium-ion battery health status prediction device, comprising a storage medium and a processor, wherein the storage medium stores computer instructions, characterized in that: When the processor runs the computer instructions, it executes the lithium-ion battery health status prediction method as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for predicting the health status of a lithium-ion battery as claimed in any one of claims 1 to 7 can be executed.

Images