A method for estimating the health of a lithium-ion battery

Abstract

The present application belongs to the technical field of new energy power system energy storage, and particularly relates to a lithium ion battery health degree estimation method. A feature vector reflecting battery aging is extracted from a charge and discharge cycle segment data, a feature matrix is constituted, and the feature matrix is preprocessed; important feature screening is performed based on the preprocessed feature matrix, and an important feature matrix is obtained; based on the important feature matrix, a Mogrifier-LSTM is used to construct a battery health degree estimation model, and the battery health degree estimation model is trained to obtain a trained battery health degree estimation model; and the trained battery health degree estimation model is used for battery health degree estimation. The present application improves key feature fusion efficiency and improves multi-dimensional information utilization rate through Mogrifier iterative reconstruction; and a triple LSTM+attention mechanism reduces complex decay mode fitting error.

Description

Technical Field

[0001] This invention belongs to the field of energy storage technology for new energy power systems, and specifically relates to a method for estimating the health of lithium-ion batteries. Background Technology

[0002] Lithium-ion batteries, due to their high energy density and long cycle life, have become a core component of energy storage devices in new energy power systems. Battery State of Health (SOH) is a key indicator characterizing the degree of battery aging and predicting Remaining Useful Life (RUL), which is crucial for the safe and stable operation, cascade utilization, and economic benefits of energy storage systems. Therefore, developing high-precision and robust battery SOH estimation methods is currently a research hotspot.

[0003] Currently, battery SOH estimation methods are mainly divided into three categories:

[0004] (1) Model-based methods: such as equivalent circuit models (ECM) or electrochemical models. These methods have clear physical meanings, but they usually rely on complex and difficult-to-identify parameters in real time, such as internal resistance and diffusion coefficient. They are sensitive to changes in the internal state of the battery, and the accuracy of the model is easily affected by fluctuations in operating conditions and individual differences.

[0005] (2) Feature-based methods: These methods estimate the state of battery (SOH) by analyzing the correlation between characteristic parameters during the battery charging and discharging process (such as constant current / constant voltage charging time, voltage plateau, temperature changes, and incremental capacity analysis (ICA) feature points). While relatively intuitive, feature engineering heavily relies on expert experience, and single or limited features often fail to fully capture complex nonlinear degradation processes, thus limiting universality and robustness. For example, inappropriate feature selection or changes in battery aging patterns can lead to estimation failure.

[0006] (3) Data-driven methods: especially deep learning-based methods, such as using recurrent neural networks (RNNs) and their variant, long short-term memory networks (LSTMs), to process battery degradation time series data. LSTMs, with their memory gate mechanism, have shown advantages in handling time series dependencies. However, existing LSTM-based SOH estimation methods still face challenges: on the one hand, the complex nonlinear interactions between the original time series features are not fully utilized, and the feature representation ability is limited; on the other hand, the choice of LSTM network structure (such as the number of layers and neurons) and training hyperparameters (such as the learning rate) has a significant impact on model performance, and traditional grid search or empirical trial-and-error methods are inefficient and difficult to obtain the optimal configuration. In addition, existing methods mostly focus on single sensor data streams or simple feature combinations, failing to effectively integrate multi-dimensional (voltage, current, temperature) information and extract the key features that best represent the aging state.

[0007] In summary, existing battery health estimation techniques have shortcomings in feature representation capabilities, depth of temporal feature interaction modeling, and efficient hyperparameter optimization, which limit the estimation accuracy and engineering applicability. Summary of the Invention

[0008] To overcome the problems in the prior art, this invention proposes a method for estimating the health of lithium-ion batteries.

[0009] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0010] This invention provides a method for estimating the health of lithium-ion batteries, comprising the following steps:

[0011] Step 100: Extract feature vectors reflecting battery aging from charge-discharge cycle data, construct a feature matrix, and preprocess the feature matrix; based on the preprocessed feature matrix, perform important feature screening to obtain an important feature matrix;

[0012] Step 200: Based on the important feature matrix, use Mogrifier-LSTM to construct a battery health estimation model, and train the battery health estimation model to obtain a trained battery health estimation model.

[0013] Step 300: Use the trained battery health estimation model to estimate battery health.

[0014] Furthermore, step 100 also includes: calculating the battery health label for each charge-discharge cycle based on charge-discharge cycle fragment data, and preprocessing the battery health label.

[0015] Further, in step 100, important feature filtering is performed based on the preprocessed feature matrix to obtain an important feature matrix, including: calculating feature weights using a random forest method based on the preprocessed feature matrix; sorting the feature weights according to their size; and selecting the top-ranked features. N Each feature is taken as an important feature, and the important feature matrix is ​​obtained.

[0016] Furthermore, the formula for calculating the feature weights is as follows:

[0017] ;

[0018] In the above formula, Indicates the first j Feature weights of each feature This indicates the number of iterations for calculating the weights in the random forest. Indicates the first k During the nth iteration j The reduction in Gini impurity of a normalized feature.

[0019] Further, in step 200, based on the important feature matrix, a battery health estimation model is constructed using Mogrifier-LSTM, including:

[0020] Step 210: Based on the important feature matrix and combined with the feature-state interaction and enhancement of the initial hidden state matrix under the Mogrifier mechanism, the final enhanced features and optimized state are obtained;

[0021] Step 220: Based on the final enhanced features and optimized state, extract temporal features layer by layer through three cascaded LSTM neural network layers to aggregate the global degradation representation of the battery;

[0022] Step 230: Based on the global degradation characterization of the battery, use an attention mechanism to calculate the essential feature vector of the battery's full life cycle degradation process;

[0023] Step 240: Using a fully connected layer, the calculated feature vector is mapped to a battery health estimate through a linear transformation.

[0024] Further, step 210 specifically includes:

[0025] Through iterative bidirectional gating interaction with a preset number of rounds R, the important feature matrix is... F With the hidden state matrix H Perform nonlinear cross-reconstruction:

[0026] For each round r , Perform the following operations:

[0027] Even-numbered wheels:

[0028] ;

[0029] Odd-numbered wheels:

[0030] ;

[0031] In the above formula, Represents the characteristic gate matrix, Represents the state gating matrix; Indicates random initialization r When it is an even number, the first r The weight matrix of the round, express r When it is an even number, the first r Wheel offset, express r When it is an odd number, the first r The weight matrix of the round, express rWhen it is an odd number, the first r The wheel's bias is updated during training;

[0032] This represents element-wise multiplication; This represents the Sigmoid activation function; Indicates enhanced features; Indicates an optimized state;

[0033] After each iteration, the enhanced features and optimization state are updated, and so on. R The process iterates through rounds, outputting the final enhanced features. and optimization state .

[0034] Furthermore, step 220 specifically includes: constructing three cascaded LSTM neural network layers, the first LSTM neural network layer and the second LSTM neural network layer are used to calculate the hidden state at each time step and pass it to the next layer, and the third LSTM neural network layer outputs the hidden state matrix of the third LSTM neural network layer by splicing the outputs of the first two layers, thus aggregating the global degradation representation of the battery.

[0035] Furthermore, step 230 specifically includes:

[0036] Based on the hidden state matrix output by the third LSTM neural network layer, an attention mechanism is used to calculate the essential feature vector c of the degradation process of the condensed battery throughout its entire life cycle:

[0037] ;

[0038] In the above formula, This represents the hidden state matrix output by the third LSTM neural network layer. At time step i The hidden state; It is a linearly changing weight matrix, obtained by random initialization and updated during training; The attention weight vector is also randomly initialized and updated during training.

[0039] Further, in step 200, the battery health estimation model is trained to obtain a trained battery health estimation model, including:

[0040] The number of LSTM neurons, learning rate, and number of Mogrifier interaction rounds are used as hyperparameters to establish a three-dimensional hyperparameter search space;

[0041] Construct an objective function, which is to minimize the error between the battery health estimate and the battery health label;

[0042] Based on the three-dimensional hyperparameter search space, a tree-structured Parzen estimator is used to perform multiple rounds of hyperparameter combination evaluation, retaining the optimal hyperparameter combination in each round; after the evaluation is completed, the hyperparameter combination with the smallest objective function is selected as the optimal hyperparameter combination.

[0043] The battery health estimation model is initialized using the optimal hyperparameter combination, and then trained in multiple rounds to output the trained battery health estimation model.

[0044] Furthermore, the battery health estimation model is initialized using the optimal hyperparameter combination, and multiple rounds of training are performed to output the trained battery health estimation model. Specifically, this includes: using the important feature matrix as input, initializing the constructed battery health model using the optimal hyperparameter combination, setting the objective function as the loss function, updating the model parameters using the Adam optimizer, training the battery health estimation model, and obtaining the trained battery health model.

[0045] Compared with the prior art, the present invention has the following technical effects:

[0046] (1) This invention achieves nonlinear cross-reconstruction of important feature matrices and hidden state matrices through iterative bidirectional gating interaction under the Mogrifier mechanism. Compared with traditional LSTM which only relies on unidirectional state transmission, the Mogrifier mechanism adjusts the weight distribution of features and states through multiple rounds of dynamic gating, enabling the model to adaptively capture key feature interaction patterns at different aging stages. The Mogrifier mechanism gradually strengthens these complex nonlinear relationships through rounds of iteration, generating enhanced features and optimized states, significantly improving the feature matrix's ability to represent the battery degradation process, thereby enhancing the model's generalization ability to different operating conditions and individual differences.

[0047] (2) This invention employs a cascaded three-layer LSTM network, combined with an attention mechanism to construct a global degradation representation. This hierarchical design enables the model to simultaneously capture the short-term dynamics and long-term evolution of battery degradation. Furthermore, the attention mechanism, through dynamic weight allocation, highlights the time steps that have the most significant impact on battery health throughout the entire life cycle, thereby condensing a more discriminative essential feature vector. Attached Figure Description

[0048] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram of the process of the present invention;

[0050] Figure 2 This is charge / discharge cycle segment data for the present invention;

[0051] Figure 3 The top 10 most important features of random forests;

[0052] Figure 4 This is a comparison chart of the estimated battery health and the actual battery health. Detailed Implementation

[0053] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the specific implementation methods, structures, features, and effects of the technical solutions proposed according to the present invention are described in detail below with reference to the accompanying drawings and preferred embodiments. Specific features, structures, or characteristics in one or more embodiments may be combined in any suitable form. Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0054] In one embodiment of the present invention, reference is made to... Figures 1-4 A method for estimating the health of lithium-ion batteries is provided, comprising the following steps:

[0055] Step 100: Extract feature vectors reflecting battery aging from charge-discharge cycle data, construct a feature matrix, and preprocess the feature matrix; based on the preprocessed feature matrix, perform important feature screening to obtain an important feature matrix;

[0056] Step 200: Based on the important feature matrix, use Mogrifier-LSTM to construct a battery health estimation model, and train the battery health estimation model to obtain a trained battery health estimation model.

[0057] Step 300: Use the trained battery health estimation model to estimate battery health.

[0058] The following is a detailed explanation of each of the above steps:

[0059] Step 100: Extract feature vectors reflecting battery aging from charge-discharge cycle data, construct a feature matrix, and preprocess the feature matrix; based on the preprocessed feature matrix, perform important feature screening to obtain an important feature matrix.

[0060] As an example, step 100 specifically includes the following sub-steps:

[0061] Step 110: Extract charge / discharge cycle segment data from the battery cycle data and calculate the battery health label for each charge / discharge cycle.

[0062] Charge-discharge cycle segment data refers to: a charging process from the start to the end of charging is called a charging segment; a discharging process from the start to the end of discharging is called a discharging segment. Charge-discharge segment data is the data during the charging and discharging process.

[0063] The charge / discharge cycle segment data includes voltage, current, temperature, and charge / discharge quantity, which are acquired by power electronic sensors. In a specific embodiment, the voltage, current, and temperature are acquired by an IVT 300-2U1 isolated voltage sensor, an ITB 900-S7 closed-loop Hall effect sensor, and a DS18B20 digital temperature sensor, respectively. The charge / discharge quantity can be calculated by the BMS (Battery Management System) based on the acquired current.

[0064] Calculate the battery health label for each charge-discharge cycle. :

[0065] ;

[0066] in, The initial discharge capacity in one charge-discharge cycle. This is the discharge cutoff capacity. This refers to the nominal capacity.

[0067] Step 120: Extract feature vectors reflecting battery aging from the data segments of each charge-discharge cycle to form a feature matrix.

[0068] Reference Figure 2 The battery aging feature vectors of each charge / discharge front end are extracted from the obtained charge / discharge cycle fragment data to form a feature matrix. ,in, Indicates the first i The feature vector of a charge-discharge cycle , Indicates the first i The first charge-discharge cycle j One characteristic, m This represents the total number of battery cycles. n This represents the number of features extracted for each charge-discharge cycle segment.

[0069] In this implementation, 20 features are extracted from each charge-discharge cycle segment of data. The feature names and physical meanings are shown in Table 1 below:

[0070] Table 1. Feature Names and Physical Significance

[0071]

[0072] Step 130: Normalize the battery health label and feature matrix.

[0073] ;

[0074] ;

[0075] In the above formula, This represents the normalized eigenvalue matrix. Indicates the first j One feature; This represents the normalized battery health label. For the first i The battery health estimate for each record.

[0076] Step 140: Based on the normalized feature matrix, use random forest to filter features and obtain the important feature matrix.

[0077] Random forests assess feature importance by calculating the Gini impurity decrease during decision tree splits. A higher feature weight indicates a greater contribution of that feature to the predicted objective (SOH).

[0078] The feature weight formula is as follows:

[0079] ;

[0080] In the above formula, Indicates the first j Feature weights of each feature This indicates the number of iterations for calculating the weights in the random forest. Indicates the first k During the nth iteration j The reduction in Gini impurity of a normalized feature.

[0081] according to Select the high-weighted top N These features are used as selection features. In this embodiment, N Choosing 7 means selecting the top 7 features with the highest weights as the filtering features, resulting in the filtered important feature matrix F, which is then referred to... Figure 3 .

[0082] Step 200: Based on the important feature matrix, construct a battery health estimation model and train the battery health estimation model to obtain a trained battery health estimation model.

[0083] As an example, this step 200 specifically includes the following sub-steps:

[0084] Step 210: Based on the filtered important feature matrix, and combined with the feature-state interaction and enhancement of the initial hidden state matrix under the Mogrifier mechanism, the final enhanced features and optimized state are obtained.

[0085] Initialize the initial hidden state matrix H It is a matrix of all zeros, and based on the selected important feature matrix F. ， The Mogrifier mechanism is used to iteratively adjust the feature matrix and hidden state through multiple rounds to enhance the feature representation capability, resulting in the final enhanced features and optimized state. The specific steps are as follows:

[0086] Iteration rounds The preset number of iterations R is used in this embodiment. R =5;

[0087] Random initialization r When it is an even number, the first r Weight matrix of the wheel , r When it is an even number, the first r wheel offset , r When it is an odd number, the first r Weight matrix of the wheel , r When it is an odd number, the first r wheel offset And updated during training; feature gating matrix and state gating matrix Dynamically generated using the Sigmoid function;

[0088] For each round r Perform the following operations:

[0089] Even-numbered wheels:

[0090] ;

[0091] Odd-numbered wheels:

[0092] ;

[0093] In the above formula, This represents element-wise multiplication; This represents the Sigmoid activation function; Indicates enhanced features; This indicates an optimized state.

[0094] After each iteration, the enhanced features and optimization state are updated, and so on. R The process iterates through rounds, outputting the final enhanced features. and optimization state .

[0095] Step 220: Based on the final enhanced features and optimized state, extract temporal features layer by layer through three cascaded LSTM neural network layers to aggregate the global degradation representation of the battery.

[0096] A cascaded LSTM neural network is constructed. The first LSTM layer extracts dynamic voltage / current features, the second LSTM layer fuses temperature and capacity decay features, and the third LSTM layer learns long-term degradation trends and aggregates global representations by concatenating the outputs of the first two layers. The output of each LSTM layer serves as the input to the next layer, refining degradation features layer by layer. Specifically, the first LSTM layer receives the input as the final enhancement feature. and optimization state The first LSTM layer outputs primary temporal features; the second LSTM layer receives the primary temporal features from the first layer and extracts higher-order temporal dependencies; the third LSTM layer aggregates the features from the first two layers to generate a global degenerate representation, as shown in the following formula:

[0097] ;

[0098] In the above formula, Indicates the LSTM layer. , , Used to initialize the l The hidden state matrix of the LSTM layer, Indicates the first l The hidden state matrix of the LSTM layer; Indicates the first l Output features of layer −1; Indicates the first l LSTM layer.

[0099] Step 230: Based on the global degradation characterization of the battery, use an attention mechanism to calculate the essential feature vector of the condensed battery life cycle degradation process.

[0100] The hidden state matrix output from the last LSTM layer in step 220 is input using an attention mechanism. The essential feature vector c, which encapsulates the entire battery lifecycle degradation process, is calculated using the following formula:

[0101] ;

[0102] In the above formula, This represents the hidden state matrix output by the last LSTM layer in step 220. At time step i The hidden state; It is a linearly changing weight matrix, obtained by random initialization and updated during training; The attention weight vector is also randomly initialized and updated during training; m This represents the total number of battery cycles.

[0103] Step 240: Using a fully connected layer, the calculated feature vector c is mapped to a battery health estimate through a linear transformation.

[0104] ;

[0105] In the above formula, This is an estimate of battery health. This represents the weight matrix of the fully connected layer. This represents the bias of the fully connected layer, which is obtained through random initialization and updated during training.

[0106] Step 250: Train the constructed battery health estimation model based on Mogrifier-LSTM to obtain the trained battery health estimation model.

[0107] Through hyperparameter optimization and model training, a high-performance battery health estimation model is obtained. Specifically, this includes: hyperparameter search, which optimizes the number of LSTM neurons, learning rate, and Mogrifier interaction rounds; and model training, which initializes the model with the optimal hyperparameters and minimizes the estimation error using the Adam optimizer.

[0108] As an example, step 250 includes the following sub-steps:

[0109] Step 2501: Use the number of LSTM neurons, learning rate, and number of Mogrifier interaction rounds as hyperparameters to establish a three-dimensional hyperparameter search space.

[0110] The number of LSTM neurons, learning rate, and number of Mogrifier interaction rounds are selected as hyperparameters for model training. The range of hyperparameter values ​​is set, and different sampling methods are used to sample the hyperparameters to establish a three-dimensional hyperparameter search space.

[0111] In this embodiment, the number of LSTM neurons is set to a discrete integer space ranging from 20 to 100, and the sampling step size is 10 (i.e., candidate values ​​include 20, 30, ..., 100); the learning rate is set to a logarithmic space ranging from 10 to 100. The optimal learning rate is found by sampling with an exponential distribution; the interaction rounds of the Mogrifier are set to an integer space of 1-5 to control the interaction depth between the input and the hidden state.

[0112] Step 2502: Construct an objective function, which is to minimize the error between the battery health estimate and the battery health label.

[0113] The important feature matrix F and the corresponding battery health labels are combined to form the dataset required for training the model. The dataset is then randomly split into a training set and a validation set. In this embodiment, the ratio of the training set to the validation set is 8:2.

[0114] During model training, for each set of hyperparameters, the constructed battery health estimation model is initialized. Multiple rounds (e.g., 10 rounds) of training are performed using the key feature matrices from the training set. The error between the battery health estimation results and the actual battery health estimation results is calculated on the validation set. Figure 4 Specifically, the error between the estimated result and the actual result is calculated on the validation set. error :

[0115] ;

[0116] In the above formula, To verify the number of data entries in the validation set, For the first i Battery health estimate for each record. For the first i The actual value of each record, i.e., the battery health label.

[0117] Step 2503: Based on the three-dimensional hyperparameter search space, a tree-structured Parzen estimator is used to perform multiple rounds of hyperparameter combination evaluation, retaining the optimal hyperparameter combination in each round; after the evaluation is completed, the hyperparameter combination with the smallest objective function is selected as the optimal hyperparameter combination.

[0118] The sampling strategy is automatically assigned by the Parzen estimator algorithm. After evaluation, the parameter combination with the smallest validation set error is selected as the model initialization parameters. In this embodiment, the number of parameter combination evaluation rounds is set to 20.

[0119] Step 2504: Initialize the battery health estimation model using the optimal hyperparameter combination, perform multiple rounds of training, and output the trained battery health estimation model.

[0120] Using the obtained important feature matrix F as input, the battery health model is initialized with the obtained optimal parameters. The objective function is set as the loss function, the Adam optimizer is used to update the model parameters, and the model is trained to obtain the trained model. .

[0121] In this embodiment, the model training rounds are selected as 10 rounds, and the batch size of the input data is 16.

[0122] Step 300: Use the trained battery health estimation model to estimate battery health.

[0123] Feature vectors reflecting battery aging are extracted from the battery operation data to be estimated, and a feature matrix is ​​constructed. The feature matrix is ​​then normalized. Based on the normalized feature matrix, a random forest is used to select features to obtain the important feature matrix. The important feature matrix is ​​then input into the trained battery health estimation model to obtain the SOH estimation result.

[0124] ;

[0125] In the above formula, For the screening of batteries to be estimated, The result is the SOH estimate.

[0126] Table 2 Test Results of Lithium Battery Dataset

[0127]

[0128] Referring to Table 2, this lithium battery dataset test focused on lithium batteries of different models and chemical systems (lithium iron phosphate and ternary lithium). The experimental results show that:

[0129] (1) The overall performance of SOH estimation error is good: In the tests of lithium batteries of different models and different chemical systems, the SOH estimation error MAE (mean absolute error) does not exceed 0.7%, indicating that the estimation method used has high accuracy.

[0130] (2) The error fluctuation is small under different conditions: Under different battery models and different charge and discharge rates, the error fluctuation is controlled within ±0.7%, which shows that the method has strong stability and adaptability under different working conditions.

[0131] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for estimating the health of a lithium-ion battery, characterized in that, Includes the following steps: Step 100: Extract feature vectors reflecting battery aging from charge-discharge cycle data, construct a feature matrix, and preprocess the feature matrix; Based on the preprocessed feature matrix, important features are filtered to obtain the important feature matrix; Step 200: Based on the important feature matrix, use Mogrifier-LSTM to construct a battery health estimation model, and train the battery health estimation model to obtain a trained battery health estimation model. Based on the key feature matrix, a battery health estimation model is constructed, including: Step 210: Based on the important feature matrix and combined with the feature-state interaction and enhancement of the initial hidden state matrix under the Mogrifier mechanism, the final enhanced features and optimized state are obtained; Through iterative bidirectional gating interaction with a preset number of rounds R, the important feature matrix is... F With the hidden state matrix H Perform nonlinear cross-reconstruction: For each round r , Perform the following operations: Even-numbered wheels: ； Odd-numbered wheels: ； In the above formula, Represents the characteristic gating matrix, Represents the state gating matrix; Indicates random initialization r When it is an even number, the first r The weight matrix of the round, express r When it is an even number, the first r Wheel offset, express r When it is an odd number, the first r The weight matrix of the round, express r When it is an odd number, the first r The wheel's bias is updated during training; This represents element-wise multiplication; This represents the Sigmoid activation function; Indicates enhanced features; Indicates an optimized state; After each iteration, the enhanced features and optimization state are updated, and so on. R The process iterates through rounds, outputting the final enhanced features. and optimization state ; Step 220: Based on the final enhanced features and optimized state, extract temporal features layer by layer through three cascaded LSTM neural network layers to aggregate the global degradation representation of the battery; Three cascaded LSTM neural network layers are constructed. The first and second LSTM neural network layers are used to calculate the hidden state at each time step and pass it to the next layer. The third LSTM neural network layer outputs the hidden state matrix of the third LSTM neural network layer by splicing the outputs of the first two layers, thus aggregating the global degradation representation of the battery. Step 230: Based on the global degradation characterization of the battery, use an attention mechanism to calculate the essential feature vector of the battery's full life cycle degradation process; Based on the hidden state matrix output by the third LSTM neural network layer, an attention mechanism is used to calculate the essential feature vector c of the degradation process of the condensed battery throughout its entire life cycle: ； In the above formula, This represents the hidden state matrix output by the third LSTM neural network layer. At time step i The hidden state; It is a linearly changing weight matrix, obtained by random initialization and updated during training; The attention weight vector is also randomly initialized and updated during training; Step 240: Using a fully connected layer, the calculated feature vector is mapped to a battery health estimate through a linear transformation; Step 300: Use the trained battery health estimation model to estimate battery health.

2. The method for estimating the health of a lithium-ion battery according to claim 1, characterized in that, Step 100 further includes: calculating the battery health label for each charge-discharge cycle based on charge-discharge cycle fragment data, and preprocessing the battery health label.

3. The method for estimating the health of a lithium-ion battery according to claim 1, characterized in that, In step 100, important features are screened based on the preprocessed feature matrix to obtain an important feature matrix, including: calculating feature weights using a random forest method based on the preprocessed feature matrix; sorting the feature weights according to their size; and selecting the top-ranked features. N Each feature is taken as an important feature, and the important feature matrix is ​​obtained.

4. The method for estimating the health of a lithium-ion battery according to claim 3, characterized in that, The formula for calculating the feature weights is as follows: ； In the above formula, Indicates the first j Feature weights of each feature This indicates the number of iterations for calculating the weights in the random forest. Indicates the first k During the nth iteration j The reduction in Gini impurity of a normalized feature.

5. The method for estimating the health of a lithium-ion battery according to claim 1, characterized in that, In step 200, the battery health estimation model is trained to obtain a trained battery health estimation model, including: The number of LSTM neurons, learning rate, and number of Mogrifier interaction rounds are used as hyperparameters to establish a three-dimensional hyperparameter search space; Construct an objective function, which is to minimize the error between the battery health estimate and the battery health label; Based on the three-dimensional hyperparameter search space, a tree-structured Parzen estimator is used to perform multiple rounds of hyperparameter combination evaluation, retaining the optimal hyperparameter combination in each round; after the evaluation is completed, the hyperparameter combination with the smallest objective function is selected as the optimal hyperparameter combination. The battery health estimation model is initialized using the optimal hyperparameter combination, and then trained in multiple rounds to output the trained battery health estimation model.

6. The method for estimating the health of a lithium-ion battery according to claim 5, characterized in that, The battery health estimation model is initialized using the optimal hyperparameter combination, trained in multiple rounds, and output as a trained battery health estimation model. Specifically, this involves: using an important feature matrix as input, initializing the constructed battery health model using the optimal hyperparameter combination, setting the objective function as the loss function, updating the model parameters using the Adam optimizer, training the battery health estimation model, and obtaining the trained battery health model.

Images