Satellite lithium ion battery SOH estimation method based on multi-working-condition distributed learning

Abstract

The invention provides a satellite lithium ion battery SOH estimation method based on multi-working-condition distributed learning, belongs to the technical field of battery life estimation, and solves the problems that in traditional transfer learning, feature alignment of sample data under a high stress condition and sample data under a satellite condition is difficult, and an existing centralized training model is poor in generalization ability and poor in reliability. The satellite lithium ion battery state-of-health estimation method comprises the following steps: S1, data acquisition; s2, building a distributed learning framework; s3, initializing a global LSTM model; s4, the client performs local training, calculates local model parameters and transmits the local model parameters to the server; s5, the server aggregates the global model parameters and transmits the global model parameters to the client; s6, verifying whether the loss function of the client side converges or not, and if not, repeating the steps S3-S4 until the loss function converges; and S7, testing the global model. The lithium ion battery SOH estimated value calculated by the method is basically matched with the real value at the initial stage of charging and discharging, and the error is smaller than 1% along with the increase of the period.

Description

Technical Field

[0001] The invention relates to a satellite lithium-ion battery SOH estimation method based on multi-operating-condition distributed learning, belonging to the technical field of battery life estimation. Background Art

[0002] Satellite lithium-ion batteries are widely used in various spacecraft due to their excellent energy density and long cycle life, providing long-lasting and reliable power support for aerospace equipment. However, during continuous charge and discharge cycles, irreversible electrochemical reactions occur within the battery, leading to performance degradation, mainly manifested as a loss of available capacity and power output. The state of health (SOH) of a lithium-ion battery is usually defined as the ratio of the current capacity to the initial capacity and is a key indicator for measuring the degree of battery degradation. The degradation of SOH not only affects the operating efficiency of the satellite battery system, but may also pose a safety hazard. Therefore, real-time monitoring and accurate estimation of the SOH of satellite lithium-ion batteries are crucial.

[0003] In recent years, data-driven approaches based on machine learning have become an effective means of estimating the state of health (SOH) of lithium-ion batteries (Li-ion batteries). However, the performance of these data-driven models is highly dependent on the availability of large amounts of high-quality training data. Only models trained with sufficient data can achieve satisfactory accuracy. Satellite Li-ion batteries typically operate at low charge / discharge rates (C-rates) and shallow depths of discharge (DOD), which results in slow SOH degradation and makes data acquisition difficult and costly. To accelerate battery degradation and quickly acquire sample data, accelerated life testing (ALT) or accelerated degradation testing (ADT) methods are often employed. These methods accelerate battery degradation by increasing the charge / discharge rate and depth of discharge, enhancing capacity decay and shortening the duration of each charge / discharge cycle. This allows for the rapid collection of high-stress sample data on the ground, which can then be used to develop models to estimate the SOH of satellite Li-ion batteries under low-stress conditions. Many studies have addressed this issue through transfer learning (TL). The basic idea is to freeze some hidden layers of the network and fine-tune the remaining fully connected layers to learn the characteristics of the target domain. However, existing methods still face the following specific challenges: (1) There is a significant distribution difference between sample data under high stress conditions and sample data under satellite conditions, which makes feature alignment difficult in traditional transfer learning; (2) The target domain data is scarce. The cost of obtaining SOH annotation data for satellite on-orbit batteries is high, and only about 200 sets of valid samples can be generated during the entire life cycle of a single satellite, which cannot meet the data requirements of traditional fine-tuning methods (usually requiring ≥1000 sets); (3) Model generalization limitations: Existing centralized training frameworks have difficulty integrating multiple sets of heterogeneous data and cannot fully utilize the data features under different conditions, resulting in limited generalization capabilities of the model in practical applications. Summary of the Invention

[0004] The present invention aims to solve the problem that in traditional transfer learning, it is difficult to align the features of sample data under high stress conditions with sample data under satellite conditions, and the existing centralized training model has poor generalization ability and is difficult to integrate multiple sets of heterogeneous data, resulting in inaccurate health status estimation of satellite lithium-ion batteries. A satellite lithium-ion battery SOH estimation method based on multi-working condition distributed learning is proposed.

[0005] The technical solution adopted by the present invention to solve the above problems is: the present invention comprises the following steps:

[0006] Step 1: Collect high stress and low stress working condition data of lithium-ion batteries and build a distributed learning architecture. The distributed learning architecture includes N clients C1, C2, ..., C N and an aggregation server node S;

[0007] Step 2: Use the LSTM model as the initial global model and send it to all clients through the server. Train the global model based on the client.

[0008] Step 3: The server integrates the training results of all clients and updates the global model parameters, and then sends the updated global model back to all clients.

[0009] Step 4: Repeat steps 2-3 to iteratively update the global model until the loss function convergence;

[0010] Step 5: Use the collected low stress test data set D test The global model after iterative training is tested. If the difference between the predicted SOH value of the lithium-ion battery obtained by the test and the true value is less than the preset range, the global model is adjusted and steps 2-3 are repeated until the difference between the predicted SOH value of the lithium-ion battery and the true value meets the preset range.

[0011] Furthermore, the collection of high-stress operating condition data and low-stress operating condition data of the lithium-ion battery in step 1 specifically includes:

[0012] Step 1.1: Conduct charge and discharge tests on the lithium-ion battery by setting different high-stress conditions and low-stress conditions simulating space conditions, wherein the different high-stress conditions include at least different charge and discharge rates and different charge and discharge depths;

[0013] Step 1.2: During each high-stress and low-stress charge-discharge test cycle, collect the terminal voltage, current, and actual capacity of the lithium-ion battery. Combined with the nominal capacity of the lithium-ion battery, obtain the state of health (SOH) value of the lithium-ion battery. The terminal voltage and current of the collected lithium-ion battery are preprocessed.

[0014] Furthermore, the pre-processing steps in step 1.2 include:

[0015] Step 1.2.1: Perform degradation feature extraction, feature screening, feature dimensionality reduction, and data normalization on the terminal voltage and current of the lithium-ion battery collected during the high-stress charge-discharge test and the low-stress charge-discharge test, respectively. Map the terminal voltage and current of the lithium-ion battery to the same scale, and use the normalized battery degradation features as the input of the global model.

[0016] Step 1.2.2: Construct client local data {X based on the normalized battery degradation characteristics and the corresponding lithium-ion battery SOH value under each high stress condition i ,Y i}, where X iis the normalized battery degradation characteristic under the current high stress condition input to the global model, Y i The SOH data of lithium-ion batteries under current high stress conditions output by the global model;

[0017] Step 1.2.3: Take the normalized battery degradation characteristics under low stress conditions and the corresponding lithium-ion battery SOH value under low stress conditions as the low stress test data set

[0018] Furthermore, step 2 specifically includes:

[0019] Step 2.1: Initialize the global model parameter ω0 and send the initialized global model to each client;

[0020] Step 2.2: Each client C i After receiving the initialized global model, according to the local data {X i ,Y i} Train it and use Adam optimizer to calculate the local model parameters ω i ;

[0021] Step 2.3: Update the local model parameters ω of all clients i Upload to the server via an encrypted channel.

[0022] Furthermore, the local model parameter ω in step 2.2 i The calculation formula is:

[0023]

[0024] In formula (1), η is the learning rate, is the loss function Gradient of local data, k is the training round;

[0025] Based on client C i Local data {X i ,Y i}Calculate the loss function Combined loss function Gradient of local data to calculate local model parameters ω i .

[0026] Furthermore, step 3 specifically includes:

[0027] The server uses the weighted average method to receive the local model parameters uploaded by all clients. Perform integration and update the global model parameters ω k+1 , and the updated global model parameters ω k+1 Sent to all clients for the next round of training;

[0028] Global model parameter ω k+1 The update formula is:

[0029]

[0030] The beneficial effects of the present invention are:

[0031] 1. This invention builds a distributed learning architecture, enabling each client to share only local model updates rather than directly share data. This ensures data privacy for each client while guaranteeing model adaptability for a single operating condition. This is particularly applicable to satellite applications where ground testing and on-orbit operation differ. By leveraging distributed computing resources, models can be trained across multiple operating conditions, improving computational efficiency when data volumes are large or clients are widely distributed.

[0032] 2. The present invention conducts charge and discharge tests on lithium-ion batteries under different working conditions (high stress and low stress), and uses these data as local training data and global test data of the client respectively. This method enables the model to better adapt to diverse actual working environments to accurately estimate the SOH of spatial lithium-ion batteries.

[0033] 3. The present invention provides feature dimensionality reduction and data normalization operations on the data collected from high-stress charge-discharge tests and low-stress charge-discharge tests, so that the features of the sample data under high-stress conditions are aligned with the features of the sample data under satellite conditions, which is beneficial to subsequent calculations and improves the accuracy of the health status estimation of lithium-ion batteries. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A schematic flow chart of a satellite lithium-ion battery SOH estimation method based on distributed learning in multiple operating conditions provided by the present invention;

[0035] Figure 2 A schematic diagram of the distributed learning architecture provided by the present invention;

[0036] Figure 3 The battery degradation characteristics and SOH curve after normalization provided by the present invention;

[0037] Figure 4 This is a schematic diagram of the health status estimation results of the lithium-ion battery provided by the present invention. DETAILED DESCRIPTION

[0038] Combine Figure 1-4 This embodiment is described as follows. Figure 1 As shown, the steps of the satellite lithium-ion battery SOH estimation method of multi-operating-condition distributed learning described in this embodiment include:

[0039] S1: Data collection;

[0040] S101: Perform charge and discharge tests on lithium-ion batteries under different high-stress conditions (such as different temperatures, different charge and discharge rates, etc.) and low-stress conditions simulating space working conditions, and collect battery aging data. The lithium-ion batteries used in the test of this embodiment include lithium iron phosphate (LFP) system batteries, lithium cobalt oxide (LCP) system batteries, lithium manganese oxide (LMP) system batteries or ternary system batteries. The aging process of each lithium-ion battery will go through different battery life cycles. In each cycle, the terminal voltage, current and actual capacity C of the battery are collected. 实际 , the battery nominal capacity C is known 标称 , the health status value of the lithium-ion battery is determined based on the actual capacity and nominal capacity of the lithium-ion battery, as shown in formula (1):

[0041]

[0042] S102: Data preprocessing. The preprocessing adopted in this embodiment includes: degradation feature extraction, feature screening, feature dimensionality reduction and data normalization. In order to facilitate subsequent modeling, this embodiment first extracts effective features from the collected battery voltage, current and other data. Common features include battery charge and discharge efficiency, internal resistance, capacity attenuation, etc. Feature screening and dimensionality reduction operations help to remove redundant information, reduce computational burden, and improve the accuracy and generalization ability of the model. Through data normalization processing, feature data of different dimensions can be mapped to the same scale to avoid deviations during model training due to different feature value ranges. Construct client local data {X based on the normalized battery degradation characteristics under each high stress condition and the corresponding lithium-ion battery SOH value i ,Y i}, where X i is the normalized battery degradation characteristic under the current high stress condition, Y i is the SOH value of the lithium-ion battery under the current high stress condition; the normalized battery degradation characteristics under low stress condition and the SOH value of the lithium-ion battery under low stress condition are used as the low stress test data set D test ,Finally, the features and SOH curves can be used as input-output pairs, such as Figure 3 for subsequent modeling.

[0043] This step aligns the features of the sample data under high stress conditions with those under satellite conditions, which is beneficial for subsequent calculations and improves the accuracy of the health status estimation of lithium-ion batteries.

[0044] S2: Build a distributed learning framework;

[0045] The distributed learning framework built in this implementation is as follows Figure 2 As shown in Figure 1, this framework consists of N clients and a trusted aggregation server node. The server is responsible for managing the global model and verifying the availability of the clients, ensuring bidirectional communication between the client and the server. After passing verification, the client participates in the federated learning training process, sending updated model parameters without sharing local data. Each client is responsible for maintaining its own local data and performing local training, using local high-stress data to train the SOH estimation model. The server is responsible for receiving the model parameters uploaded by the clients, aggregating the global model, and sending the updated global model back to each client.

[0046] S3: global LSTM model initialization;

[0047] In this embodiment, the LSTM model is used as the initial global model and the model parameter ω0 is initialized. The purpose of the LSTM network is to capture the time series characteristics of the battery degradation process to predict the battery SOH.

[0048] The structure of the LSTM model includes an input layer, an LSTM layer, and a fully connected output layer. Its mathematical model is:

[0049] h t =LSTM(h t-1 ,x t )(2);

[0050] In formula (2), h t is the hidden state at time step t, h t-1 is the hidden state of the previous time step, x t The input feature of the current time step is used, and LSTM controls the retention and forgetting of information through memory units to capture the long-term dependencies of the battery.

[0051] S4: Client local training, calculation of local model parameters and transmission to the server;

[0052] The server sends the initialized global model to each client. Each client C i After receiving the global LSTM model, then in the local data {X i ,Y i During the training process, the client uses the Adam optimizer to update the local model parameters ω i :

[0053]

[0054] In formula (3), η is the learning rate, is the loss function of client i, k is the iterative training round, The calculation formula is:

[0055]

[0056] The updated local model parameters are uploaded to the server via an encrypted channel:

[0057]

[0058] S5: The server aggregates the global model parameters and passes them to the client;

[0059] The server receives all local model parameters uploaded by the client Update the global model parameters ω using the weighted average method k+1 :

[0060]

[0061] Updated global parameter ω k+1 Will be sent back to each client to enter the next round of training:

[0062] ω k+1 →C1,C2,...,C N (7);

[0063] S6: Verify whether the client's loss function converges. If so, repeat S3-S4 until convergence.

[0064] Repeat S4-S5 for multiple rounds of training until the loss function Convergence. In each iteration, the server updates the global model based on the model parameters uploaded by each client, gradually optimizing it to adapt to diverse battery degradation data.

[0065] S7: Global model testing.

[0066] After completing global model training, use the low stress dataset The global model was tested to evaluate its performance under real working conditions. Figure 4 As shown, if the difference between the predicted SOH value of the lithium-ion battery obtained by the test and the true value is less than the preset range, the LSTM model is adjusted and S3-S6 are repeated until the difference between the predicted SOH value of the lithium-ion battery and the true value meets the preset range.

[0067] Depend on Figure 4 It can be seen that through the distributed framework proposed in the present invention, the SOH of space lithium-ion batteries can be estimated more accurately, and the estimated SOH value of the lithium-ion battery is basically consistent with the true value in the initial stage of charging and discharging. Even as the cycle increases, the error is less than 1%. Under complex working conditions, the generalization ability and reliability of the model can be improved through different training conditions, providing an effective solution for the health management of satellite batteries.

[0068] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the present profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical content disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement of the above embodiments made according to the technical essence of the present invention, within the spirit and principles of the present invention, without departing from the content of the technical solution of the present invention, shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A satellite lithium-ion battery SOH estimation method based on multi-operating-condition distributed learning, characterized in that: include: Step 1: Collect high stress working condition data and low stress working condition data of lithium ion batteries and build a distributed learning architecture. The distributed learning architecture includes N clients C1, C2, ..., C N and an aggregation server node S; Step 2: Use the LSTM model as the initial global model and send it to all clients through the server. Train the global model based on the client. LSTM stands for Long Short-Term Memory. Step 3: The server integrates the training results of all clients and updates the global model parameters, and then sends the updated global model back to all clients. Step 4: Repeat steps 2-3 to iteratively update the global model until the loss function convergence; Step 5: Utilize a low-stress test dataset The global model after iterative training is tested. If the difference between the predicted SOH value of the lithium-ion battery obtained by the test and the true value is less than the preset range, the global model is adjusted and steps 2-3 are repeated until the difference between the predicted SOH value of the lithium-ion battery and the true value meets the preset range.

2. The satellite lithium-ion battery SOH estimation method based on multi-operating-condition distributed learning according to claim 1 is characterized in that: The high stress working condition data and low stress working condition data of the lithium-ion battery collected in step 1 specifically include: Step 1.1: Conduct charge and discharge tests on the lithium-ion battery by setting different high-stress conditions and low-stress conditions simulating space conditions, wherein the different high-stress conditions include at least different charge and discharge rates and different charge and discharge depths; Step 1.2: During each high-stress and low-stress charge-discharge test cycle, collect the terminal voltage, current, and actual capacity of the lithium-ion battery. Combined with the nominal capacity of the lithium-ion battery, obtain the state of health (SOH) value of the lithium-ion battery. The terminal voltage and current of the collected lithium-ion battery are preprocessed.

3. The satellite lithium-ion battery SOH estimation method based on multi-operating-condition distributed learning according to claim 2 is characterized in that: The preprocessing steps in step 1.2 include: Step 1.2.1: Perform degradation feature extraction, feature screening, feature dimensionality reduction, and data normalization on the terminal voltage and current of the lithium-ion battery collected during the high-stress charge-discharge test and the low-stress charge-discharge test, respectively. Map the terminal voltage and current of the lithium-ion battery to the same scale, and use the normalized battery degradation features as the input of the global model. Step 1.2.2: Construct client local data {X based on the normalized battery degradation characteristics and the corresponding lithium-ion battery SOH value under each high stress condition i ,Y i }, where X i is the normalized battery degradation characteristic under the current high stress condition input to the global model, Y i The SOH data of lithium-ion batteries under current high stress conditions output by the global model; Step 1.2.3: Take the normalized battery degradation characteristics under low stress conditions and the corresponding lithium-ion battery SOH value under low stress conditions as the low stress test dataset D test .

4. The satellite lithium-ion battery SOH estimation method based on multi-operating-condition distributed learning according to claim 1 is characterized in that: Step 2 specifically includes: Step 2.1: Initialize the global model parameter ω0 and send the initialized global model to each client; Step 2.2: Each client C i After receiving the initialized global model, according to the local data {X i ,Y i } Train it and use Adam optimizer to calculate the local model parameters ω i ; Step 2.3: Update the local model parameters ω of all clients i Upload to the server via an encrypted channel.

5. The satellite lithium-ion battery SOH estimation method based on multi-operating-condition distributed learning according to claim 4 is characterized in that: The local model parameter ω in step 2.2 i The calculation formula is: In formula (1), η is the learning rate, is the loss function Gradient of local data, k is the training round; Based on client C i Local data {X i ,Y i }Calculate the loss function Combined loss function Gradient of local data to calculate local model parameters ω i .

6. The satellite lithium-ion battery SOH estimation method based on multi-operating-condition distributed learning according to claim 1 is characterized in that: Step 3 specifically includes: The server uses the weighted average method to receive the local model parameters uploaded by all clients. Perform integration and update the global model parameters ω k+1 , and the updated global model parameters ω k+1 Sent to all clients for the next round of training; Global model parameter ω k+1 The update formula is:

Images