Battery remaining life prediction method capable of automatically identifying battery health characteristics

Abstract

The invention discloses a battery remaining life prediction method capable of automatically identifying battery health characteristics. Firstly, a lithium ion battery data set is obtained, an automatic health feature extractor is used for extracting health features related to the aging state of the battery, the extractor can efficiently recognize key features in the aging process of the battery, the accuracy and automation degree of feature extraction are improved, and errors caused by human intervention are reduced. Then, a health feature filter is used, contribution of each feature to prediction of the remaining life of the battery is analyzed and quantified through an SHAP model, and features with remarkable influence are screened out; and the screened features are divided into a training set and a test set, a BiLSTM model is trained, verification is performed through a forward and backward dependency relationship, and a feedback mechanism is added for optimization. And if the prediction precision of the model meets the requirement, the residual life of the battery data can be accurately predicted.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithium ion battery remaining service life prediction, and relates to a lithium ion battery health feature extraction and lithium ion battery remaining service life prediction method. Background Art

[0002] As the energy density and power density of lithium-ion batteries gradually increase, they have been widely used in portable electronic devices, electric vehicles, and large-scale energy storage systems. The safety performance and remaining service life prediction of lithium-ion batteries have become increasingly important. However, their performance will gradually decay with the increase of usage time. Accurate remaining service life (RUL) prediction helps to formulate reasonable battery charging and discharging plans, promptly remind users to replace or repair batteries, and reduce potential safety risks.

[0003] Currently, there are three main methods for predicting the remaining service life of lithium batteries: The first type of prediction of the remaining service life of lithium batteries relies on physical models, but due to the rapid development of battery technology and complex structure, it is difficult to establish a general mathematical model. The second type of method monitors the battery status through equivalent circuits and electrochemical models, which requires a lot of experimental tuning. The third type of data-driven method uses machine learning to analyze historical data and flexibly predict the life of lithium batteries. It does not rely on physical models and has been widely used in the prediction of the remaining service life of lithium batteries.

[0004] However, the current data-driven RUL prediction method still has some shortcomings. The data-driven model requires more computing resources and efficiency to process large amounts of data and complex model structures, and the model is more susceptible to the redundancy of health features. The selection of hyperparameters for most models is also a difficult problem. The patent (CN118261046A) discloses a lithium battery remaining life prediction method based on NGO-BiLSTM. The method extracts the degradation characteristic parameters of the remaining service life of the lithium-ion battery, uses the Northern Goshawk algorithm (NGO) to iteratively search for the optimal parameters of the bidirectional long short-term memory neural network (BiLSTM) and saves and constructs the NGO-BiLSTM model. It has the advantages of small error and high prediction accuracy, but the degradation health characteristics of the model need to be manually selected, and the generalization ability of the model is poor. The patent (CN118211474A) discloses a battery life prediction method based on the MPSO-LSSVM algorithm. First, the sample data set is generated by extracting the time series data of the voltage, current, remaining capacity and internal resistance parameters during the battery cycle discharge process, and the data set is divided into a test set and a training set; a lithium-ion battery remaining service life prediction model is established based on the training set, and the MPSO algorithm and the LSSVM algorithm are combined to iteratively optimize the lithium-ion battery remaining service life prediction model to obtain the optimal solution. This method uses the ampere-hour integration method for data processing to obtain energy indicators to evaluate the remaining service life of lithium-ion batteries, reduce battery life prediction errors, improve prediction accuracy, and optimize the calculation process, but requires complete battery data. The present invention uses self-supervised reconstruction and semi-supervised capacity prediction to solve the problem of insufficient data. Summary of the invention

[0005] The problem solved by the present invention is to provide a method for predicting the remaining life of a battery by automatically identifying the health characteristics of the battery. The method solves the following problems:

[0006] (1) Battery health features need to be manually extracted, resulting in low accuracy in predicting the remaining battery life;

[0007] (2) A large amount of complete battery cycle data is required to use data-driven methods to predict the remaining useful life of lithium batteries;

[0008] (3) The problem of poor generalization ability of the battery remaining service life prediction model.

[0009] To solve the above problems, the present invention adopts the following technical solution, and the steps are as follows:

[0010] S1: Obtain battery voltage and battery current data during the aging cycle charge and discharge process in the lithium-ion battery data set;

[0011] S2: Perform linear interpolation on outliers or missing values ​​in the battery data set and use min-max for standardization;

[0012] S3: Automatically extract the health features in the health feature automatic extractor, the steps are as follows:

[0013] (1) Input the standardized charge and discharge voltage and current data, process the input voltage and current data through the battery feature encoding model, and extract the intermediate features of the data;

[0014] (2) Using all the intermediate features obtained, the reconstruction task module is sent to train to obtain reconstructed battery data, and the reconstructed battery data is sent to the reconstruction loss error model to calculate the reconstruction loss error. The reconstruction loss error and all the intermediate features are used as inputs, and the data feature 1, i.e., the compressed features of the standardized battery charge and discharge voltage and current data, is obtained through the feature selector 1;

[0015] (3) The intermediate features with capacity labels are sent to the capacity prediction module to obtain the predicted capacity, and the predicted capacity is sent to the capacity prediction loss error model to calculate the capacity prediction loss error. The capacity prediction loss error and the intermediate features with capacity labels are used as inputs, and after passing through feature selector 2, data feature 2, i.e., data features related to capacity change, is obtained.

[0016] (4) The reconstruction loss error model and the capacity prediction loss error model are weightedly combined to form a health feature fusion function, and the obtained data feature 1 and data feature 2 are input into the health feature fusion function to obtain the health feature error, and determine whether the health feature is less than the error index, and verify whether the health feature meets the extraction requirements. If the requirements are not met, repeat steps (1) to (3). If the requirements are met, output the health feature;

[0017] S4: Filter the health features in the health feature filter, calculate the SHAP value for each feature, analyze the importance of each feature in the prediction process, obtain important health features and unimportant health features, retain the health features with high importance, form an important health feature data set, discard the unimportant health feature data set composed of unimportant health features, and then divide the important health feature data set into training set data and test set data according to a certain ratio;

[0018] S5: Use the training set data to input into the BiLSTM model for training to obtain the lithium battery RUL training prediction value;

[0019] S6: Input the test set data into the BiLSTM network estimation model to verify whether the model accuracy meets the design requirements. If it does not meet the requirements, reconstruct the network model. If it meets the requirements, the battery data can be input into the prediction model to achieve accurate prediction of the RUL of lithium batteries.

[0020] In step S1, the historical cycle charge and discharge data of the lithium battery data set is obtained, the charge and discharge voltage and current values ​​are extracted at time intervals, and the lithium battery capacity is selected as the key indicator for measuring the health life of the battery.

[0021] In step S2, outliers and missing values ​​are first identified and processed. For each outlier or missing value, the linear relationship between the known values ​​before and after is calculated to determine the appropriate filling value. Then, the health feature dataset is standardized using min-max standardization. The calculation formula is: In the formula, x i is the health characteristic data, y i is the standardized data, max(x) and min(x) are the maximum and minimum values ​​of each row of the corresponding health feature data.

[0022] The step S3 of extracting the health feature reflecting the battery capacity degradation by using the health feature automatic extractor comprises the following steps:

[0023] (1) Record the charge and discharge voltage and current data of each cycle at time intervals from the battery aging dataset to construct a time series dataset [V 充 ,V 放 ] and [I 充 ,I 放 ], where V 充 、V 放 represents the charge and discharge voltage at time step t, I 充 ,I 放 represents the charge and discharge current at time step t;

[0024] (2) In the battery feature encoding model, the constructed time series dataset is positionally encoded to preserve the time series order. The query (Q), key (K), and value (V) vectors are generated for each time step. The attention weight is obtained by calculating the dot product of the query and the key and normalizing it using Softmax. The weight is then used to weight the value vector to generate the attention head output. The multi-head attention mechanism is used to concatenate the outputs of each head and linearly transform them to obtain the final output.

[0025] (3) Input the output of the multi-head attention into the feed-forward neural network (FFN) and generate the intermediate feature representation Z' through the ReLU activation function;

[0026] (4) All intermediate features Z' are input into the reconstruction task module to reconstruct the input data. These features are used to reconstruct the original charge and discharge voltage and current data. The MSE is used to calculate the error between the reconstructed battery data and the input battery charge and discharge voltage and current data, i.e., the reconstruction error. The formula is: Among them, x i For charge and discharge voltage and current data, The reconstructed data is output through feature selector 1, and data feature 1, that is, the compressed features of the standardized battery charging and discharging voltage and current data, is output;

[0027] (5) Input the intermediate feature Z'1 with capacity label into the capacity prediction module to obtain the predicted capacity. Send the predicted capacity into the capacity prediction loss error model to calculate the error. Use MSE to calculate the error between the predicted capacity and the actual capacity, that is, the capacity prediction error. The formula is: Among them, C i is the actual capacity value, is the predicted capacity value of the model, and through feature selector 2, outputs data feature 2, that is, the data feature related to capacity change;

[0028] (6) The reconstruction loss error model and the capacity prediction loss error model are weighted together to form a health feature fusion function, where the health feature fusion function is: Loss total =α·Loss reconstruction +β·Loss capacity In the formula, α and β are the hyperparameters for adjusting the weights of the reconstruction task and the capacity prediction module, respectively. The obtained data features 1 and 2 are input into the health feature fusion function together. After a lot of training, the health feature fusion function is continuously optimized to minimize the value of the health feature fusion function. Finally, the health factor Z related to the battery aging state is extracted.

[0029] In step S4, SHAP analysis is used to evaluate the influence of each feature on the battery remaining life prediction result. The specific steps are as follows: Calculate the SHAP value of each feature. For the input health feature Z and battery capacity C, the SHAP value φ i The calculation formula is: Among them, S is the i The feature subset of n is the total number of features, C(S∪{z i}) indicates adding feature z i The change in contribution to the model prediction results; the SHAP value graph is generated according to the formula. When the SHAP value is positive, it means that the feature has a positive contribution to the prediction value, otherwise it has a negative contribution; according to the SHAP analysis results, the important feature data that has a significant impact on the prediction results is retained and divided into training set data and test set data according to a certain ratio.

[0030] In step S5, BiLSTM consists of a forward LSTM and a backward LSTM. Each LSTM unit includes a forget gate, an input gate, and an output gate for selectively retaining and updating information. The specific process of BiLSTM is as follows: the first step starts with the input at the beginning of the sequence of the forward LSTM, and uses the forget gate f t =σ(W f ·[h t-1 ,x t ]+b f ) decides the old information that needs to be discarded and inputs gate i t =σ(W i ·[h t-1 ,x t ]+b i ) to selectively update information and Generate candidate states, combined with the previous state C t-1 Calculate the updated cell state C t Output gate ot = σ(W o ·[h t-1 ,x t ]+b o ) is used to generate the current hidden state h t =o t *tanh(C t ); The second step is that the backward LSTM starts from the end of the sequence and inputs the reverse direction, and the process is similar to the forward LSTM; the third step is to concatenate the outputs of the forward and backward LSTMs to obtain the final BiLSTM output.

[0031] In step S6, the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are selected as evaluation indicators to evaluate the RUL prediction accuracy of the battery to determine whether the expected results are met. The specific formula is: Where N is the number of experimental estimates, y n is the actual value of RUL, is the estimated value of RUL. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0033] In the attached picture:

[0034] Figure 1 is a flow chart of a method for predicting the remaining useful life of a lithium-ion battery according to the present invention;

[0035] Figure 2 It is a structural schematic diagram of the automatic extractor of health characteristics of the present invention;

[0036] Figure 3 It is a schematic diagram of the BiLSTM structure of the present invention; DETAILED DESCRIPTION

[0037] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0038] Example: Figure 1 As shown, a method for predicting the remaining battery life by automatically identifying battery health characteristics is described, and the specific steps are as follows:

[0039] S1: Obtain the voltage and current data of the lithium-ion battery during the aging cycle. Taking the B0005 battery data set disclosed by NASA as an example, the battery charging and discharging experiment process is as follows: In the charging stage, the battery is first charged with a constant current of 1.5A (CC mode) until the voltage rises to 4.2V; then it is switched to constant voltage (CV mode) charging, and the voltage is maintained at 4.2V until the current drops to 20mA. In the discharge stage, the battery is discharged with a constant current of 2A, and the discharge ends when the voltage drops from 4.2V to the set cut-off voltage. The entire charging cycle lasts about 1100 seconds, and the data is recorded every 60 seconds, recording the corresponding charging voltage and current values. The discharge cycle lasts about 3600 seconds, and the data is recorded every 60 seconds, recording the corresponding discharge voltage and current values. Finally, these charge and discharge voltage and current data constitute a time series, providing a basis for subsequent analysis.

[0040] S2: In the process of battery data processing, linear interpolation and standardization methods are used to process outliers and missing values ​​in the data set. First, outliers and missing values ​​are identified and processed. When the duration of the charge and discharge cycle is less than 1100 seconds, if the proportion of missing data is too high, the cycle should be discarded; if the proportion of missing data is less than 10% of the normal cycle, for each missing value or outlier, interpolation is performed based on the linear relationship of the known data points before and after it to determine a reasonable filling value. After the missing value processing is completed, the health feature data set is subjected to min-max standardization processing, and the calculation formula is x i is the health characteristic data, y i is the standardized data, max(x) and min(x) are the maximum and minimum values ​​of each row of the corresponding health feature data. Through this method, the health feature data is scaled within the range of 0 to 1, reducing the impact of feature scale differences on model training.

[0041] S3: Use the health feature automatic extractor to automatically extract the health features, the steps are as follows:

[0042] (1) Extract the battery charge and discharge voltage and current data from the battery aging dataset. These data are recorded once every minute and a time series dataset is constructed as the input sequence of the model. The input of each time step is [V 充 ,V 放 ] and [I 充 ,I 放 ], where V 充 、V 放 represents the charge and discharge voltage at time step t, I 充 ,I 放 Represents the charge and discharge current at time step t. In order to maintain the order information of the time series, positional encoding is applied. The encoding formula is as follows:

[0043]

[0044] Where t represents the position of the time step (i.e., the index in the input sequence), i is the index of the current feature dimension, and d is the dimension of the input data;

[0045] (2) Use the battery encoding model to positionally encode the constructed time series dataset to preserve the time series order. For each input time step x t Generate the corresponding query (Q), key (K), and value (V) vectors, where Q t =x t W q ,K t =x t W k ,V t =x t W v , where W q , W k , W v is the learned weight matrix;

[0046] (3) By calculating the dot product of the query and all keys, the attention score is obtained. The formula is: d k is the dimension of the key. Next, the attention score is normalized using the Softmax function to obtain the attention weight;

[0047] (4) Use the calculated attention weights to perform weighted summation on the value vector to generate a single output of the attention head: t =∑ j Attention tj V j ;

[0048] (5) A multi-head attention mechanism is used to enhance the capture of battery aging feature data. The model can capture information from different subspaces. Each attention head calculates the output independently, and finally the outputs of all heads are spliced ​​to obtain the final output: t =Concat(O1,O2,...,O h )W o , where W o It is the linear transformation matrix after the final concatenation, and uses 8 attention heads (i.e. h = 8), the dimension of each head is 64, and the output dimension is 512.

[0049] (6) The output of the multi-head attention is input into the feedforward neural network (FFN) and the dimension of the front hidden layer is set to 2048. It is linearly transformed through a RELU activation function, where the activation function is FFN(x) = max(0, xW1+b1)W2+b2, which helps the model to further transform the output of the multi-head attention and generate a more informative intermediate feature representation Z';

[0050] (7) All intermediate features Z' are input into the reconstruction task module to reconstruct the input data. These features are used to reconstruct the original charge and discharge voltage and current data. The MSE is used to calculate the error between the reconstructed battery data and the input battery charge and discharge voltage and current data, i.e., the reconstruction error. The formula is: Among them, x i For charge and discharge voltage and current data, The reconstructed data is selected through feature selector 1, and the features whose MSE changes rank in the top 30% among all features are output, and data feature 1, that is, the compressed features of the standardized battery charge and discharge voltage and current data, are output;

[0051] (8) 20% of the battery data with capacity labels are extracted from the data set, and the capacity labels are removed from the remaining data. The intermediate feature Z'1 with capacity labels is input into the capacity prediction module to obtain the predicted capacity. The predicted capacity is sent to the capacity prediction loss error model to calculate the error. The MSE is used to calculate the error between the predicted capacity and the actual capacity, that is, the capacity prediction error. The formula is: Among them, C i is the actual capacity value, is the predicted capacity value of the model, and through feature selector 2, the features whose MSE changes rank in the top 30% among all features are selected, and data feature 2, i.e., the data features related to capacity changes, are output;

[0052] (9) The reconstructed voltage and current errors and the capacity prediction errors are weighted and combined to form a health feature fusion function. The health feature fusion function is: Losstotal =α·Loss reconstruction +β·Loss capacity , where α and β are hyperparameters for adjusting the weights of the two tasks. The initial default values ​​are set to α as 0.75 and β as 0.5. The obtained data features 1 and 2 are input into the health feature fusion function together. , The model gradually optimizes its parameters through a large number of iterative training, so that the health feature fusion function converges to the minimum value, and finally extracts the final health factor Z.

[0053] S4: In step S4, SHAP analysis is used to evaluate the influence of each feature on the battery remaining life prediction result. The specific steps are: calculate the SHAP value of each feature. For the input health feature Z and battery capacity C, the SHAP value Φ i The calculation formula is: Among them, S is the i The feature subset of n is the total number of features, C(S∪{z i}) indicates adding feature z i The change in contribution to the model prediction results; the SHAP value graph is generated according to the formula. When the SHAP value is positive, it means that the feature has a positive contribution to the prediction value, otherwise it has a negative contribution; according to the SHAP analysis results, the data with a contribution less than 30% of the average SHAP value are eliminated, and only the important feature data that have a significant impact on the prediction results are retained, and divided into training set data and test set data in a ratio of 7:3.

[0054] S5: Use the training set data to input into the bidirectional long short-term memory network (BiLSTM) model for training; BiLSTM can better capture the bidirectional dependencies in the battery charging and discharging process by combining the capabilities of forward and backward LSTM. At each time step, the forward LSTM processes the forward information from the beginning of the sequence to the current time, and the backward LSTM processes the reverse information from the end of the sequence to the current time, and the outputs of the two are concatenated as the final output of BiLSTM. Among them, the hidden layer dimension is set to 128, the batch size is set to 32, and the learning rate is set to 0.001. After two layers of BiLSTM, the RUL of the lithium battery is finally obtained through the forget gate, input gate, and output gate.

[0055] S6: Input the test set data into the BiLSTM network estimation model to verify whether the model accuracy meets the design requirements. The mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are selected as evaluation indicators to evaluate the prediction accuracy of the RUL prediction model to determine whether it meets the expected results. The specific formulas are: Where N is the number of experimental estimates, y n is the actual value of RUL, is the RUL estimate. If it does not meet the requirements, the model is rebuilt.

[0056] Finally, it should be noted that the above description is only an example of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting remaining battery life by automatically identifying battery health characteristics, the method comprising the following steps: S1: Obtaining battery voltage and battery current data during the aging cycle charge and discharge process of the lithium-ion battery; S2: Perform linear interpolation on outliers or missing values ​​in the battery data set and use min-max for standardization; S3: Using the standardized battery voltage and battery current data as input, the health feature automatic extractor automatically extracts the health feature. The specific operation steps are as follows: (1) Input the standardized charge and discharge voltage and current data, process the input voltage and current data through the battery feature encoding model, and extract the intermediate features of the data; (2) Sending all the intermediate features obtained to the reconstruction task module for training to obtain reconstructed battery data, sending the reconstructed battery data to the reconstruction loss error model, calculating the reconstruction loss error, and taking the reconstruction loss error and all the intermediate features as input, through the feature selector 1, obtaining data feature 1, that is, the compressed features of the standardized battery charge and discharge voltage and current data; (3) The intermediate features with capacity labels are sent to the capacity prediction module to obtain the predicted capacity, and the predicted capacity is sent to the capacity prediction loss error model to calculate the capacity prediction loss error. The capacity prediction loss error and the intermediate features with capacity labels are used as inputs, and the data features 2, i.e., the data features related to the capacity change, are obtained through feature selector 2. (4) The reconstruction loss error model and the capacity prediction loss error model are weightedly combined to form a health feature fusion function. The obtained data feature 1 and data feature 2 are used as inputs. After the health feature fusion function is used, the health feature error is obtained. It is determined whether the health feature is less than the error index. If the requirement is not met, steps (1) to (3) are repeated. If the requirement is met, the health feature is output. S4: The output health features are used as input, and passed through a health feature filter to obtain an important health feature data set. The health feature filter is designed as follows: first, the SHAP value of each feature is calculated to calculate the importance score of each feature, and the importance of each feature in the prediction process is analyzed according to the importance score, and important health features and unimportant health features are obtained respectively. Health features with high importance are retained to form an important health feature data set, and an unimportant health feature data set composed of unimportant health features is discarded. Then, the important health feature data set is divided into training set data and test set data according to a certain ratio; S5: input the training set data into a bidirectional long short-term memory network (BiLSTM) model for training to obtain a training prediction value of the remaining useful life (RUL) of the lithium battery; S6: Input the test set data into the BiLSTM prediction model to verify whether the RUL prediction error accuracy meets the design requirements. If it does not meet the requirements, return to step S2 and automatically extract the health features again. If it meets the requirements, output the lithium battery RUL prediction value.

2. A method for predicting remaining battery life by automatically identifying battery health characteristics according to claim 1, characterized in that: In step S1, the battery data set content is obtained as follows: historical cycle charge and discharge data, charge and discharge voltage and current values ​​are extracted at time intervals, and lithium battery capacity is selected as a key indicator to measure the health life of the battery.

3. The method for predicting remaining battery life by automatically identifying battery health characteristics according to claim 1, characterized in that: The specific data processing operations in step S2 are as follows: first, identify and process outliers and missing values. For each outlier or missing value, calculate the linear relationship between the known values ​​before and after it and determine the appropriate filling value; then use min-max standardization to standardize the health feature data set. The calculation formula is: In the formula, x i is the health characteristic data, y i is the standardized data, max(x) and min(x) are the maximum and minimum values ​​of each row of the corresponding health feature data.

4. The method for predicting remaining battery life by automatically identifying battery health characteristics according to claim 1, characterized in that: The battery feature encoding model process described in step S3 is as follows: (1) First, the charge and discharge voltage and current data of each cycle are recorded at time intervals from the battery aging dataset to construct a time series dataset [V 充 ,V 放 ] and [I 充 ,I 放 ], where V 充 、V 放 represents the charge and discharge voltage at time step t, I 充 ,I 放 represents the charge and discharge current at time step t; (2) Position encoding is performed on the constructed time series dataset to preserve the time series order. Query (Q), key (K), and value (V) vectors are generated for each time step. The attention weight is obtained by calculating the dot product of the query and the key and normalizing it using Softmax. The weight is then used to weight the value vector and generate the attention head output. The multi-head attention mechanism is used to concatenate the outputs of each head and linearly transform them to obtain the final output. (3) The output of the multi-head attention is input into the feed-forward neural network (FFN) and the intermediate feature representation Z' is generated through the activation function ReLU.

5. The method for predicting remaining battery life by automatically identifying battery health characteristics according to claim 1, characterized in that: The reconstruction loss error model and capacity prediction loss error model described in step S3 are specifically designed as follows: (1) The mean square error (MSE) is used as the evaluation index of the reconstruction loss error model to calculate the error between the reconstructed battery data and the input battery charge and discharge voltage and current data, that is, the reconstruction error. The formula is: Among them, x i For charge and discharge voltage and current data, is the reconstructed data; (2) Using MSE as the evaluation index of the capacity prediction loss error model, the error between the predicted capacity and the actual capacity, i.e., the capacity prediction error, is calculated. The formula is: Among them, C i is the actual capacity value, is the predicted capacity value of the model.

6. The method for predicting remaining battery life by automatically identifying battery health characteristics according to claim 1, characterized in that: The feature selector described in step S3 is designed as follows: (1) First, cover each feature one by one and observe whether the error MSE of the reconstruction loss error model and the capacity prediction loss error model increases when the feature is not present; (2) Based on the MSE change of each feature obtained in the above steps, the feature selector calculates the contribution of each feature. That is, when a feature is masked, if the MSE increases significantly, it means that the feature is more important to the model performance; otherwise, it is less important. (3) Set the threshold of MSE change, filter out the features that have a significant impact on the task, retain those features that can significantly reduce MSE, and form the final data features 1 and 2.

7. A method for predicting remaining battery life by automatically identifying battery health characteristics according to claim 1, characterized in that: The health feature fusion function described in step S3 is: Loss total =α·Loss reconstruction +β·Loss capacity , where α and β are the hyperparameters for adjusting the weights of the reconstruction task and the capacity prediction module, respectively. Through a large number of iterative training, the health feature fusion function reaches the minimum value, and finally the health factor Z related to the battery aging state is extracted.

8. The method for predicting remaining battery life by automatically identifying battery health characteristics according to claim 1, characterized in that: In step S4, SHAP analysis is used to evaluate the influence of each feature on the battery remaining life prediction result. The specific steps are as follows: Calculate the SHAP value of each feature. For the input health feature Z and battery capacity C, the SHAP value Φ i The calculation formula is: Among them, S is the i The feature subset of n is the total number of features, C(S∪{z i }) indicates adding feature z i The change in contribution to the model prediction results; the SHAP value graph is generated according to the formula. When the SHAP value is positive, it means that the feature has a positive contribution to the prediction value, otherwise it has a negative contribution; according to the SHAP analysis results, the important feature data that has a significant impact on the prediction results is retained and divided into training set data and test set data according to a certain ratio.

9. The method for predicting remaining battery life by automatically identifying battery health characteristics according to claim 1, characterized in that: In step S5, BiLSTM is composed of a forward LSTM and a backward LSTM. Each LSTM unit includes a forget gate, an input gate, and an output gate for selectively retaining and updating information. The specific process of BiLSTM is as follows: the first step is to start inputting at the beginning position of the sequence of the forward LSTM, and use the forget gate f t =σ(W f ·[h t-1 ,x t ]+b f ) decides the old information that needs to be discarded and inputs gate i t =σ(W i ·[h t-1 ,x t ]+b i ) to selectively update information and Generate candidate states, combined with the previous state C t-1 Calculate the updated cell state C t ; Output gate o t =σ(W o ·[h t-1 ,x t ]+b o ) is used to generate the current hidden state h t =o t *tanh(C t ); The second step is that the backward LSTM starts from the end of the sequence and inputs the input in reverse, and the process is similar to the forward LSTM; The third step is to concatenate the outputs of the forward and backward LSTMs to obtain the final BiLSTM output.

10. The method for predicting remaining battery life by automatically identifying battery health characteristics according to claim 1, characterized in that: In step S6, the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are selected as evaluation indicators to evaluate the RUL prediction accuracy of the battery to determine whether the expected results are met. The specific formula is: Where N is the number of experimental estimates, y n is the actual value of RUL, is the estimated value of RUL.

Images