Method and device for optimizing battery health state prediction model based on federated continuous learning

Abstract

The present invention discloses a method and device for optimizing a battery health state prediction model based on federated continuous learning, which relates to the technical field of battery health management. The method includes: S1. Locally train a battery health state (SOH) prediction model and a data monitoring model, upload local parameters to a central server, and after the central server integrates and distributes them, repeat the training until the loss function converges, and wait for local training and federated learning re-training instructions; S2. The SOH prediction model outputs the battery health degree and the data monitoring model outputs the data drift abnormality degree; S3. Judge the accuracy of the output result of the SOH prediction model according to the data drift abnormality degree, and continuously optimize the SOH prediction model through parameter fine-tuning to trigger local training and federated re-learning triggered by model re-training. The present invention realizes the joint optimization of battery prediction models on multiple in-vehicle computer terminals, continuously optimizes the model performance during use, enables the model to continuously adapt to new data, and realizes the lifelong learning of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of battery health management, in particular to a method and device for optimizing a battery health state prediction model based on federated continuous learning. Background Art

[0002] With the rapid development of electric vehicles and renewable energy systems, the performance of the Battery Management System (BMS) has been required to be higher and higher. The performance of the BMS is directly related to the cruising range, charging efficiency and safety of the vehicle. Therefore, accurately evaluating the State of Health (SOH) of the battery is of great significance for improving the performance of the BMS. At the same time, in the current era, data privacy has received increasing attention. In the process of processing data, it is urgent to protect data privacy and solve the data island problem.

[0003] Most traditional battery SOH evaluation methods rely on single measurement indicators, such as voltage, capacity and internal resistance. These methods often ignore the comprehensive influence of multiple factors on the battery health state, and have high requirements for the accuracy of measurement equipment, thus resulting in limited estimation accuracy. In addition, an equivalent circuit model constructed based on a partial differential equation with physical constraints is also applied to the evaluation of the lithium battery health state. Although this method is simple and practical and has a low computational complexity, its reliability and accuracy still need to be further improved due to interference from multiple uncertain factors.

[0004] With the continuous progress of artificial intelligence technology, data-driven algorithms have been widely applied in the field of battery health state evaluation. The emergence of federated continuous learning mainly stems from the consideration of data privacy protection and data island problems. Federated learning provides a distributed training solution to address these problems, and continuous learning endows the model with the ability to continuously learn in a continuous data stream. Federated continuous learning combines the advantages of both, and through knowledge sharing between in-vehicle computer terminals, it realizes the continuous optimization of the model and the adaptability to new tasks while protecting data privacy, and has important research and application value in multiple practical application scenarios. By combining a deep autoencoder with a Gaussian mixture model, a powerful unsupervised anomaly detection method is also provided. Its end-to-end training mode and the ability to retain key information make it perform excellently in multiple fields. Summary of the Invention

[0005] The purpose of the present invention is to propose a method and device for optimizing a battery health state prediction model based on federated continuous learning. Through a specific model structure, data processing method, parameter update and optimization strategy, the joint optimization of the battery health state SOH prediction model under data privacy protection is realized, the data requirements and communication volume are reduced, the universality and lifelong learning are enhanced.

[0006] To achieve the above object, the present invention proposes a method for optimizing a battery health state prediction model based on federated continuous learning, and the specific steps are as follows:

[0007] Step S1: Federated learning model training. An on-vehicle computer terminal and a central server build a battery health state (SOH) prediction model and a data monitoring model with the same architecture. Each on-vehicle computer terminal collects and preprocesses local battery charge and discharge data, trains the model locally, and then uploads the local model parameters to the central server. The central server integrates the parameters and distributes them back to each on-vehicle computer terminal, and repeats the training until the loss functions all converge;

[0008] Step S2: Use the battery health state (SOH) prediction model to output the battery health degree, and output the data drift abnormality degree through the data monitoring model to evaluate the performance of the battery health state (SOH) prediction model;

[0009] Step S3: Select the optimization method and continuous optimization. According to the data drift abnormality degree, select the optimization method for the battery health state (SOH) prediction model and perform continuous optimization according to the optimization method.

[0010] Preferably, in step S1, the structure of the battery health state (SOH) prediction model is as follows: The preprocessed data passes through a feature extraction layer including multiple fully connected layers or convolutional layers to extract high-level features therefrom. Subsequently, a mapping layer composed of fully connected layers, with the help of a non-linear activation function, captures the complex relationships in the data, further transforms the feature data into specific feature representations and outputs the prediction results. Finally, a degradation simulation layer simulates the battery attenuation rate by adding physical constraints; the structure of the data monitoring model is as follows: Input the same data as the battery health state (SOH) prediction model and copy the parameters of the feature extraction layer to the fitting layer so that the two models have the same feature data output at the front end. The feature data outputs a feature vector through the mapping layer, and then the data is restored by the demapping layer to calculate the reconstruction error, and finally the model performance is calculated through the Gaussian mixture model.

[0011] Preferably, in step S1, the specific steps of the federated learning model training are as follows:

[0012] Step S11: The on-vehicle computer terminal downloads the current battery health state (SOH) prediction model and data monitoring model from the central server. The parameters include the number of hidden layers and the number of neurons H, the activation function, and decompose the local model parameters into globally shared parameters and specific parameters , parameter matrix , is the bias term;

[0013] Step S12: The on-vehicle computer terminal sorts out and prepares the charging data stored locally, performs data cleaning and preprocessing, and divides it according to the ratio of 6:2:2 to determine the training set, validation set, and test set;

[0014] Step S13: Use the training set data to train the downloaded battery state of health (SOH) prediction model and data monitoring model. Calculate the model output through forward propagation, calculate the loss function for backpropagation, calculate the gradient of the model parameters, and use the gradient descent method to update the global shared parameters of the model , specific parameters , bias term , and iterate multiple times until the loss functions of both models converge;

[0015] Step S14: The in-vehicle computer terminal sends the updated global shared parameters and bias term back to the central server through the homomorphic encryption algorithm, and wait for the central server to aggregate all the uploaded model parameters and issue new model parameters , and iterate for the next round of training until the loss functions of both models converge; where are the aggregated parameter matrix, global parameters, and bias term, respectively.

[0016] Preferably, in step S11, the federated learning model parameters include global shared parameters and specific parameters . After the local training of the in-vehicle computer terminal is completed, the global shared parameters are uploaded to the central server, and the specific parameters are left locally to adapt to the personalization of the in-vehicle computer terminal. The formula for the federated learning model parameters is as follows:

[0017] ;

[0018] where represents the element-wise multiplication of the corresponding elements of each network unit, is the federated parameter, is the global shared parameter, is the migration vector, is the specific parameter of the in-vehicle computer terminal.

[0019] Preferably, in step S12, the data acquisition steps are as follows: Collect the data during battery charging. Based on the in-vehicle terminal in the federated framework, record the monitoring data every 0.1 s. The data form is time-series streaming data, that is, the input feature of the th in-vehicle terminal is , where R is the resistance of the in-vehicle terminal, is the voltage of the in-vehicle terminal, is the current of the in-vehicle terminal, S is the mean square error of the in-vehicle terminal, Pk is the power of the vehicle terminal, D e is the slope of the vehicle terminal, is the charging time of the vehicle terminal, is the accumulated power of the vehicle terminal, is the slope of the curve of the vehicle terminal, is the temperature of the vehicle terminal;

[0020] The vehicle computer is based on The criterion cleans and preprocesses the local charging and discharging data, and normalizes the data at the same time. The formula is as follows:

[0021] ;

[0022] in, is the interquartile range, 0.7413 is the interquartile range A constant multiple of For the processed data, is the original data point, The median of the values in the data set.

[0023] Preferably, in step S14, the parameter update of the on-board computer is achieved by stochastic gradient descent, and the central server aggregates all uploaded model parameters in two parts, including training the local data in batches on the on-board computer, calculating the gradient of each batch, accumulating multiple rounds of gradients, and sending them to the server for aggregation at one time and adding random noise to protect privacy; during retraining, the on-board computer determines the weight of the updated parameter according to the abnormality degree of the outlier value, and increases the weight when the abnormality degree is large, and decreases the weight when the abnormality degree is small; wherein, the calculation formula of the central server parameter aggregation is as follows:

[0024] ;

[0025] in, is a global parameter, is the learning rate, N is the number of onboard computers, For the The gradient of the onboard computer, For vehicle computer dataset, is the weight of the onboard computer, is random noise.

[0026] Preferably, in step S3, for optimization method selection and continuous optimization, the data monitoring model outputs the data drift anomaly degree as a performance indicator for judging the performance of the battery state of health (SOH) prediction model. According to the data drift anomaly degree, the update method of the battery state of health (SOH) prediction model is selected. The central server records the number of times with a high anomaly degree and the number of normal times. When the output result has a low anomaly degree, the model fine-tunes the parameters of the local in-vehicle computer terminal according to the input data; when the output result has a high anomaly degree, the central server records an anomaly count once and calculates the F1 score. If the F1 score is lower than the threshold, all models are retrained using a new data set and the global parameters are updated.

[0027] Preferably, the loss function formula of the data monitoring model is as follows:

[0028] ;

[0029] where is the loss function of the data monitoring model, is the reconstruction loss, is the true value of the th in-vehicle computer terminal, is the predicted value of the model for the th in-vehicle computer terminal, is the weight hyperparameter, is the weight hyperparameter, is the feature combiner loss, is the th in-vehicle computer terminal's feature representation obtained through the feature extractor , is the regularization term;

[0030] The loss function formula of the battery state of health (SOH) prediction model is as follows:

[0031] ;

[0032] where is the total loss function, , , are the weights of each loss function, is the predicted value of the model for the charging time , is the predicted value of the model for the spatial position, is the true value of the battery health, is the predicted value of the battery health, , is the predicted value of the adjacent time step, H (·) is the solution function of the partial differential equation for the prediction of the output and input features of the model, M is the number of time steps, is the number of battery cycles;

[0033] The calculation formula for feature extraction is as follows:

[0034] ;

[0035] where, T is the output feature vector, is the start time for intercepting the charging curve, is the end time for intercepting the charging curve, is the current value, is the starting voltage value, is the ending voltage value, is the peak value of the IC curve, F (·) is the feature solving function.

[0036] The present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method for optimizing the battery health state prediction model based on federated continuous learning are implemented.

[0037] Therefore, the present invention proposes a method and device for optimizing the battery health state prediction model based on federated continuous learning, and its beneficial effects are as follows:

[0038] (1) For the method for optimizing the battery health state prediction model based on federated continuous learning proposed by the present invention, by using the federated learning architecture, only the global parameters are uploaded at the in-vehicle computer side, and the local data does not leave the side, effectively protecting data privacy. At the same time, it reduces the data collection cost and communication volume, realizes the joint optimization of the model with multi-terminal data, and improves the data utilization efficiency.

[0039] (2) For the method for optimizing the battery health state prediction model based on federated continuous learning proposed by the present invention, through the unique model structure design and training method, combined with multi-dimensional data acquisition and cleaning, it enhances the prediction ability of the model for the battery health degree; the continuous learning mechanism enables the model to continuously optimize according to new data, improving the accuracy and stability.

[0040] (3) For the method and device for optimizing the battery health state prediction model based on federated continuous learning proposed by the present invention, it can dynamically adjust the model according to the performance prediction result, adapt to different battery state changes, and avoid catastrophic forgetting; the central server and the in-vehicle computer side work together, can be flexibly expanded, adapt to vehicle data of different scales, and improve the overall adaptability and scalability of the system.

[0041] The technical solution of the present invention will be further described in detail below through the drawings and embodiments. Description of the Drawings

[0042] Figure 1This is the overall implementation flowchart of the method for optimizing the battery health state prediction model based on federated continuous learning in the present invention;

[0043] Figure 2 This is the flowchart of the federated continuous learning process;

[0044] Figure 3 This is the flowchart of the model optimization method in the present invention. Detailed implementation manners

[0045] To make the technical solutions, advantages and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be described clearly and completely below. The described embodiments are part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the described embodiments of the present invention without creative efforts belong to the protection scope of this application.

[0046] Unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meaning understood by those with ordinary skills in the field to which the present invention belongs.

[0047] Embodiment 1

[0048] As Figure 1 shown, the present invention provides a method for optimizing the battery health state prediction model based on federated continuous learning, and the specific steps are as follows:

[0049] Step S1, Federated learning model training. The on-vehicle computer terminal and the central server build the battery health state SOH prediction model and data monitoring model with the same architecture. Each on-vehicle computer terminal collects and preprocesses the local battery charge and discharge data, trains the model locally and then uploads the local model parameters to the central server, and the central server integrates the parameters and distributes them back to each on-vehicle computer terminal, repeating the training until the loss functions all converge;

[0050] The structure of the battery health state SOH prediction model is as follows: The preprocessed data passes through a feature extraction layer containing multiple fully connected layers or convolutional layers to extract high-level features from it. Subsequently, a mapping layer composed of fully connected layers, with the help of a non-linear activation function, captures the complex relationships in the data, further transforms the feature data into specific feature representations and outputs the prediction results. Finally, the degradation simulation layer simulates the battery attenuation rate by adding physical constraints; The structure of the data monitoring model is as follows: The same data as the battery health state SOH prediction model is input and the parameters of the feature extraction layer are copied to the fitting layer so that the two models have the same feature data output at the front end. The feature data outputs the feature vector through the mapping layer, and then the data is restored by the demapping layer to calculate the reconstruction error, and finally the model performance is calculated through the Gaussian mixture model.

[0051] As Figure 2As shown below, the specific steps for training the federated learning model are as follows:

[0052] Step S11: The in-vehicle computer downloads the current State of Health (SOH) prediction model and data monitoring model from the central server. The parameters include the number of hidden layers and the number of neurons H, the activation function, and the local model parameters are decomposed into globally shared parameters and specific parameters , parameter matrix , is the bias term;

[0053] The federated learning model parameters include globally shared parameters and specific parameters . After local training on the in-vehicle computer is completed, the globally shared parameters are uploaded to the central server, and the specific parameters remain local to adapt to the personalization of the in-vehicle computer. The formula for the federated learning model parameters is as follows:

[0054] ;

[0055] where, represents the element-wise multiplication of the corresponding elements of each network unit, is the federated parameter, is the globally shared parameter, is the transfer vector, is the specific parameter of the in-vehicle computer.

[0056] Step S12: The in-vehicle computer organizes and prepares the charging data stored locally, performs data cleaning and preprocessing, and divides it according to the ratio of 6:2:2 to determine the training set, validation set, and test set;

[0057] The data collection steps are as follows: Collect data during battery charging. Based on the in-vehicle terminal in the federated framework , monitor data is recorded every 0.1 s. The data form is time-series streaming data, that is, the input feature of the th in-vehicle terminal is , where, R is the resistance of the in-vehicle terminal, is the voltage of the in-vehicle terminal, is the current of the in-vehicle terminal, S is the mean square error of the in-vehicle terminal, P k is the power of the in-vehicle terminal, D e is the slope of the in-vehicle terminal, is the charging time of the in-vehicle terminal, is the cumulative power of the in-vehicle terminal, is the curve slope of the vehicle-mounted terminal, is the temperature of the vehicle-mounted terminal;

[0058] Based on criterion, the vehicle-mounted computer terminal performs data cleaning and preprocessing on the local charge and discharge data, and at the same time performs data normalization. The formula is as follows:

[0059] ;

[0060] Among them, is the interquartile range, and 0.7413 is a constant multiple of the interquartile range of, is the processed data, is the original data point, is the median of the dataset median.

[0061] Step S13: Use the training set data to train the downloaded model, calculate the model output through forward propagation, calculate the loss function for backpropagation, calculate the gradient of the model parameters, and use the gradient descent method to update the global shared parameters of the model , specific parameter , bias term , iterate multiple times until the loss functions of both models converge;

[0062] Step S14: The vehicle-mounted computer terminal sends the updated global shared parameters and the bias term back to the central server through the homomorphic encryption algorithm, waits for the central server to aggregate all the uploaded model parameters, and issues new model parameters , and iteratively performs the next round of training until the loss functions of both models converge; among them, are the aggregated parameter matrix, global parameters, and bias term respectively.

[0063] The parameter update of the vehicle-mounted computer terminal is achieved through stochastic gradient descent. The central server aggregates all the uploaded model parameters in two parts, including training the local data in batches on the vehicle-mounted computer terminal, calculating the gradient of each batch, accumulating the gradients of multiple rounds, and sending them to the server for aggregation and adding random noise to protect privacy at one time; when retraining, the vehicle-mounted computer terminal determines the weight of the updated parameters according to the degree of abnormality of the outliers, increasing the weight when the degree of abnormality is large and decreasing the weight when the degree of abnormality is small; the calculation formula for the central server parameter aggregation is as follows:

[0064] ;

[0065] Among them, are the global parameters, is the learning rate, Nis the quantity of in - vehicle computer terminals, is the gradient of the th in - vehicle computer terminal, is the dataset of the in - vehicle computer terminal is the weight of the in - vehicle computer terminal, is the random noise.

[0066] Step S2: Use the battery state of health (SOH) prediction model to output the battery health degree, and use the data monitoring model to output the data drift anomaly degree to evaluate the performance of the battery state of health (SOH) prediction model;

[0067] Step S3: Select the optimization method and continuously optimize. According to the data drift anomaly degree, select the optimization method of the battery state of health (SOH) prediction model and continuously optimize according to the optimization method.

[0068] As Figure 3 shown, select the optimization method and continuously optimize. The data monitoring model outputs the data drift anomaly degree as the performance index for judging the battery state of health (SOH) prediction model. According to the data drift anomaly degree, select the update method of the battery state of health (SOH) prediction model. The central server records the number of times with a high anomaly degree and the normal times. When the output result has a low anomaly degree, the model fine - tunes the local in - vehicle computer terminal parameters according to the input data; when the output result has a high anomaly degree, the central server records an abnormal number of times and calculates the F1 score. If the F1 score is lower than the threshold, retrain all models with a new dataset and update the global parameters.

[0069] The loss function formula of the data monitoring model is as follows:

[0070] ;

[0071] Among them, is the loss function of the data monitoring model, is the reconstruction loss, is the true value of the th in - vehicle computer terminal, is the predicted value of the model for the th in - vehicle computer terminal, is the weight hyperparameter, is the weight hyperparameter, is the feature combiner loss, is the th in - vehicle computer terminal through the feature extractor to obtain the feature representation, is the regularization term;

[0072] The loss function formula of the battery state of health (SOH) prediction model is as follows:

[0073] ;

[0074] Among them, is the total loss function, , , are the weights of each loss function, is the predicted value of the charging time by the model, is the predicted value of the model's spatial position, is the true value of the battery health, is the predicted value of the battery health, , is the predicted value at adjacent time steps, H (·) is the solution function of the partial differential equation for predicting the output and input features of the model, M is the number of time steps, is the number of battery cycles;

[0075] The calculation formula for feature extraction is as follows:

[0076] ;

[0077] Among them, T is the output feature vector, is the start time for intercepting the charging curve, is the end time for intercepting the charging curve, is the current value, is the starting voltage value, is the ending voltage value, is the peak value of the IC curve, F (·) is the feature solution function.

[0078] Embodiment 2

[0079] The present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the method for optimizing the battery health state prediction model based on federated continuous learning.

[0080] Therefore, the present invention provides a method and device for optimizing the battery health state prediction model based on federated continuous learning, which realizes the joint optimization of multiple types of battery prediction models on multiple in-vehicle computer terminals, requires a smaller dataset and lower data communication volume, effectively improves the universality of the model; at the same time, in the actual application process, the model can continuously self-optimize, continuously improve its performance, avoid catastrophic forgetting, further enhance the universality, achieve the lifelong learning of the model, and ensure its long-term stable and efficient operation.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for optimizing a battery health state prediction model based on federated continuous learning, characterized in that, The specific steps are as follows: Step S1: Federated learning model training. The vehicle-mounted computer terminal and the central server build a battery state of health (SOH) prediction model and a data monitoring model with the same architecture. Each vehicle-mounted computer terminal collects and preprocesses local battery charge and discharge data, trains the model locally, uploads the local model parameters to the central server, and the central server integrates the parameters and distributes them back to each vehicle-mounted computer terminal. Repeat the training until the loss functions converge; Step S2: Use the battery state of health (SOH) prediction model to output the battery health degree, and output the data drift abnormality degree through the data monitoring model to evaluate the performance of the battery state of health (SOH) prediction model; Step S3: Select the optimization method and continuously optimize. According to the data drift abnormality degree, select the optimization method for the battery state of health (SOH) prediction model and perform continuous optimization according to the optimization method; In step S1, the structure of the battery state of health (SOH) prediction model is as follows: The preprocessed data passes through a feature extraction layer containing multiple fully connected layers or convolutional layers to extract high-level features. Subsequently, a mapping layer composed of fully connected layers, with the help of a non-linear activation function, captures the complex relationships in the data, further transforms the feature data into specific feature representations and outputs the prediction results. Finally, the degradation simulation layer simulates the battery attenuation rate by adding physical constraints; The structure of the data monitoring model is as follows: Input the same data as the battery state of health (SOH) prediction model and copy the parameters of the feature extraction layer to the fitting layer so that the two models have the same feature data output at the front end. The feature data outputs the feature vector through the mapping layer, and then the data is restored by the demapping layer to calculate the reconstruction error. Finally, the model performance is calculated through the Gaussian mixture model; The specific steps of federated learning model training are as follows: Step S11: The in-vehicle computer downloads the current battery health state of health (SOH) prediction model and data monitoring model from the central server. The parameters include the number of hidden layers and the number of neurons H, the activation function, and decompose the local model parameters into globally shared parameters and specific parameters , parameter matrix , is the bias term; Step S12: The vehicle-mounted computer terminal sorts out and prepares the charging data stored locally, performs data cleaning and preprocessing, and divides it according to the ratio of 6:2:2 to determine the training set, validation set, and test set; Step S13: Use the training set data to train the downloaded battery state of health (SOH) prediction model and data monitoring model. Calculate the model output through forward propagation, perform backpropagation by calculating the loss function, calculate the gradient of the model parameters, and use the gradient descent method to update the global shared parameters of the model , specific parameters and bias terms , and iterate multiple times until the loss functions of both models converge; Step S14: The in-vehicle computer terminal sends the updated global shared parameters and the bias term back to the central server through the homomorphic encryption algorithm, waiting for the central server to aggregate all the uploaded model parameters and send down the new model parameters , and iteratively perform the next round of training until the loss functions of both models converge; where are the aggregated parameter matrix, global parameters, and bias term respectively.

2. The method for optimizing the battery health state prediction model based on federated continuous learning according to claim 1, wherein In step S11, the federated learning model parameters include globally shared parameters and specific parameters . After the local training on the in-vehicle computer side is completed, the globally shared parameters are uploaded to the central server, and the specific parameters are left locally to adapt to the personalization of the in-vehicle computer side. The formula for the federated learning model parameters is as follows: ； Among them, represents the element-wise multiplication of each network unit, is the federal parameter, is the globally shared parameter, is the migration vector, is the specific parameter on the in-vehicle computer side.

3. The method for optimizing a battery health state prediction model based on federated continuous learning according to claim 1, wherein In step S12, the data acquisition step is: collecting data during battery charging, based on the on-vehicle terminal in the federated framework , recording the monitoring data every 0.1 s. The data form is time-series based streaming data, that is, the input feature of the th on-vehicle terminal is , where R is the resistance of the on-vehicle terminal, is the voltage of the on-vehicle terminal, is the current of the on-vehicle terminal, is the mean square error of the on-vehicle terminal, is the power of the on-vehicle terminal, is the slope of the on-vehicle terminal, is the charging time of the on-vehicle terminal, is the cumulative power of the on-vehicle terminal, is the curve slope of the on-vehicle terminal, is the temperature of the on-vehicle terminal; The in-vehicle computer terminal performs data cleaning and preprocessing on the local charge and discharge data based on the criterion, and at the same time performs data normalization. The formula is as follows: ； Among them, is the interquartile range, and 0.7413 is a constant multiple of the interquartile range ; is the processed data, is the original data point, and is the median of the dataset median value.

4. The method for optimizing the battery health state prediction model based on federated continuous learning according to claim 1, wherein In step S14, the parameter update of the vehicle-mounted computer terminal is realized through stochastic gradient descent. The central server aggregates all the uploaded model parameters in two parts, including training the local data in batches on the vehicle-mounted computer terminal, calculating the gradient of each batch, accumulating the gradients for multiple rounds, and sending them to the server for aggregation at one time and adding random noise to protect privacy; When retraining, the vehicle-mounted computer terminal determines the weight of the updated parameters according to the abnormality degree of the outliers. When the abnormality degree is large, the weight is increased, and when the abnormality degree is small, the weight is decreased; Among them, the calculation formula for the central server parameter aggregation is as follows: ； Among them, is a global parameter, is the learning rate, N is the number of in-vehicle computer terminals, is the gradient of the th in-vehicle computer terminal, is the dataset of the in-vehicle computer terminal is the weight of the in-vehicle computer terminal, is random noise.

5. The method for optimizing a battery health state prediction model based on federated continuous learning according to claim 1, wherein In step S3, the optimization method is selected and continuously optimized. The data monitoring model outputs the data drift abnormality degree as a performance index for judging the performance of the battery state of health (SOH) prediction model. According to the data drift abnormality degree, the update method of the battery state of health (SOH) prediction model is selected. The central server records the number of times with a high abnormality degree and the number of normal times. When the output result has a low abnormality degree, the model fine-tunes the parameters of the local in-vehicle computer terminal according to the input data; when the output result has a high abnormality degree, the central server records an abnormality count once and calculates the F1 score. If the F1 score is lower than the threshold, all models are retrained using a new data set and the global parameters are updated.

6. The method for optimizing a battery health state prediction model based on federated continuous learning according to claim 1, wherein The loss function formula of the data monitoring model is as follows: ； Among them, is the loss function of the data monitoring model, is the reconstruction loss, is the true value of the th in-vehicle computer terminal, is the predicted value of the model for the th in-vehicle computer terminal, is the weight hyperparameter, is the weight hyperparameter, is the feature combiner loss, is the th feature representation obtained by the in-vehicle computer terminal through the feature extractor ; is the regularization term; The loss function formula of the battery state of health (SOH) prediction model is as follows: ； Among them, is the total loss function, , , are the weights of each loss function, is the predicted value of the charging time by the model, is the predicted value of the model's spatial position, is the true value of the battery health, is the predicted value of the battery health, , is the predicted value of adjacent time steps, is the solution function of the partial differential equation for the prediction of the output and input features of the model, M is the number of time steps, is the number of battery cycles; The calculation formula for feature extraction is as follows: ； Among them, T is the output feature vector, is the start time for intercepting the charging curve, is the end time for intercepting the charging curve, is the current value, is the starting voltage value, is the ending voltage value, is the peak value of the IC curve, is the feature solving function.

7. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for optimizing the battery state of health prediction model based on federated continuous learning according to any one of claims 1-6.

Images