Lithium ion battery health state estimation method based on deep learning and relaxation voltage

Abstract

The invention is suitable for the technical field of health management of lithium ion batteries, and provides a method comprising the following steps: extracting a relaxation voltage sequence after full charge; extracting deep-level features related to battery aging from the relaxation voltage sequence by adopting CNN (Convolutional Neural Network); processing the extracted features based on LSTM to obtain a preliminary battery SOH estimation result; and adopting a war strategy algorithm to optimize the deep learning model, obtaining an optimal hyper-parameter and an optimal SOH estimation model, and obtaining an optimal battery SOH estimation result. According to the method, deep aging features are extracted by using multilayer convolution, a complex mode and a nonlinear relation hidden in relaxation voltage reflecting battery aging are found, and time dependence is established through LSTM; wSO is used for optimizing related hyper-parameters, the solution space is searched more comprehensively and quickly, local optimum is not likely to happen, accurate hyper-parameter selection is achieved, the neural network prediction precision is effectively improved, and the lithium ion battery SOH estimation accuracy is improved.

Description

Technical Field

[0001] This invention belongs to the field of health management technology for lithium-ion batteries, and particularly relates to a method for estimating the health status of lithium-ion batteries based on deep learning and relaxation voltage. Background Technology

[0002] With the escalating energy crisis and environmental pollution, the transition to a sustainable energy system has become an urgent global need, and lithium-ion batteries (LIBs) play a crucial role in this transition. Due to their advantages such as long lifespan, ease of manufacturing, low cost, compact size, and high energy density, LIBs are finding applications ranging from portable electronic devices to electric vehicles and renewable energy systems.

[0003] Due to the inherent nature of electrochemical systems, battery batteries (LIBs) inevitably age over time during use or storage, leading to issues such as capacity degradation, increased internal resistance, and voltage drop. This complicates battery charging, usage management, and safe operation. Therefore, the Battery Management System (BMS) is crucial for ensuring the safe and reliable operation of batteries. The State of Health (SOH), the ratio of the battery's current capacity to its rated capacity, is a key indicator for assessing the degree of battery aging within the BMS. Thus, accurately estimating the LIB's SOH is an essential component of battery health management, allowing for replacement before the end of the battery's lifespan and ensuring its safe operation.

[0004] Currently, SOH estimation methods are mainly divided into model-based and data-driven methods. Model-based methods typically rely on electrochemical models, equivalent circuit models, and empirical models to describe the degradation trajectory of battery performance, treating model parameters directly related to battery SOH as state variables to be determined. However, model-based estimation methods are usually specific to certain battery types, and once the battery design or materials change, the original model may no longer be applicable and needs to be rebuilt. Compared with model-based methods, data-driven methods do not require in-depth research on the internal electrochemical reactions of the battery or analysis of its working mechanism; they only need to measure some parameters during battery operation. These advantages have led to the widespread application of data-driven methods in the field of LIB SOH prediction. Existing technologies have proposed machine learning methods that extract features from measurable cyclic data such as voltage, current, and temperature to learn the mapping relationship with SOH.

[0005] It is important to emphasize that the above research is based on the assessment of battery aging under standard charge-discharge modes. Various current charging strategies, such as multi-stage constant current (MSCC) with gradually decreasing current values ​​and pulse charging, are employed. In these strategies, the charging current can be reduced, interrupted, or replaced by short discharge pulses over a period of time, making it difficult to obtain a specific voltage range under constant current conditions. Compared to many studies that rely on charge / discharge information for battery state estimation, relaxation voltage has certain advantages due to its easier data acquisition. The battery's terminal voltage during the relaxation phase is less affected by the charging process, which is independent of the initial charging conditions and is applicable to various charging strategies. Furthermore, since electric vehicles are typically fully charged overnight, the relaxation voltage curve after full charge can be easily obtained from the BMS. This method does not require special hardware; it only requires measuring the battery voltage in the relaxation state, making it easy to implement in existing BMS systems. Moreover, research has demonstrated a correlation between the battery aging process and the voltage relaxation process. Therefore, capacity estimation using relaxed voltage distribution becomes possible. Current methods mainly extract statistical features, local geometric features, point value features, and circuit model parameter features from the relaxed voltage as input to the model. However, it is obvious that these methods usually rely on the quality of manually selected features, which is undoubtedly a complex and time-consuming process.

[0006] Therefore, developing an end-to-end lithium-ion battery health prediction method based on relaxation voltage for electric vehicles is crucial for convenient and accurate predictive maintenance of lithium-ion batteries. Summary of the Invention

[0007] The purpose of this invention is to provide a method for estimating the state of health of lithium-ion batteries based on deep learning and relaxation voltage, aiming to solve the problems mentioned in the background art.

[0008] The present invention is implemented as follows: a lithium-ion battery health state estimation method based on deep learning and relaxation voltage includes the following steps:

[0009] Extract the relaxation voltage sequence after full charge;

[0010] CNN is used to extract deep features related to battery aging from the relaxation voltage sequence;

[0011] Based on the features extracted by LSTM processing, preliminary SOH estimation results for the battery are obtained;

[0012] The deep learning model is optimized using a war strategy algorithm to obtain the best hyperparameters and the optimal SOH estimation model, thus obtaining the best battery SOH estimation result.

[0013] The lithium-ion battery state of health estimation method based on deep learning and relaxation voltage provided in this invention avoids the use of charging / discharging and other additional information, bringing more convenience to the data acquisition of existing battery management systems. It is a novel feature extraction method based on CNN, which uses multi-layer convolution for deep aging features to discover complex patterns and nonlinear relationships hidden in the relaxation voltage reflecting battery aging, and establishes time dependence through LSTM. Considering the performance dependence of neural networks on parameter selection, WSO is used to optimize relevant hyperparameters. Compared with traditional hyperparameter optimization algorithms, which have problems such as insufficient global search capability and slow convergence speed, WSO can achieve a more comprehensive and faster search of the solution space and is not prone to getting trapped in local optima. It achieves accurate hyperparameter selection, effectively improves the prediction accuracy of neural networks, and improves the accuracy of lithium-ion battery SOH estimation. Attached Figure Description

[0014] Figure 1 A flowchart illustrating a lithium-ion battery health state estimation method based on deep learning and relaxation voltage, provided in an embodiment of the present invention.

[0015] Figure 2 The charge / discharge curves and relaxation voltage curves provided for embodiments of the present invention;

[0016] Figure 3 This is a schematic diagram of feature extraction using a convolutional neural network provided in an embodiment of the present invention;

[0017] Figure 4 This is a schematic diagram illustrating the working principle of a long short-term memory neural network provided in an embodiment of the present invention.

[0018] Figure 5 This is a flowchart of WSO optimization hyperparameters provided in an embodiment of the present invention;

[0019] Figure 6 The SOH estimation results provided in the embodiments of the present invention;

[0020] Figure 7 The error index results for CY25-05-1#5 provided in the embodiments of the present invention;

[0021] Figure 8 The error index results of CY25-05-1#10 provided in the embodiments of the present invention;

[0022] Figure 9 The error index results of CY45-05-1#11 provided in the embodiments of the present invention;

[0023] Figure 10 The error index results of CY45-05-1#27 provided for the embodiments of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0025] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0026] like Figure 1 The diagram shows a flowchart of a lithium-ion battery health state estimation method based on deep learning and relaxation voltage provided in an embodiment of the present invention, which specifically includes the following steps:

[0027] S1. Extract the relaxation voltage sequence after full charging:

[0028] The dataset used in this embodiment of the invention is a publicly available dataset developed by Tongji University, generated from several commercial 18650 lithium-ion batteries under various cycling conditions. First, the battery is charged at a constant current of 4.2V, with a current range of 0.25-0.5C; then, it is charged at a constant voltage of 4.2V until the current reaches 0.05C; finally, it is discharged at a constant current to 2.65V. Detailed battery parameters are shown in Table 1.

[0029] Table 1 Battery Parameters

[0030]

[0031] By using test data, the true value of SOH (State of Health) of a lithium-ion battery can be obtained through the definition of SOH in each cycle. SOH is defined as an expression of performance characteristic parameters, thereby quantifying the degree of battery aging. Battery capacity is usually used as the evaluation index for battery SOH. The expression for SOH is as follows:

[0032]

[0033] Among them, C now and C rate These represent the current capacity and the rated capacity, respectively.

[0034] It is generally believed that when the maximum discharge capacity of a lithium-ion battery in the current cycle decays to 70%-80% of its initial maximum discharge capacity, the power battery reaches the end of its life.

[0035] The relaxation voltage of a lithium-ion battery refers to the process by which the battery voltage gradually stabilizes over time after charging and discharging have stopped. Even after charging and discharging are complete, electrochemical reactions and ion diffusion continue within the battery, causing the voltage to fluctuate. The relaxation voltage collected at a specific time is related to the battery's aging state. However, the relaxation process after a full charge is relatively unaffected by the charging phase. Therefore, this invention estimates the battery's health status based on the relaxation voltage. Figure 2 Complete charge-discharge curves and relaxation voltage curves are presented. It can be seen that the relaxation voltage decreases exponentially over time, which is the same as the aging trend of the battery. Therefore, it is promising to extract features from the relaxation voltage to estimate the SOH of the battery.

[0036] S2. Use CNN to extract deep features related to battery aging from the relaxation voltage sequence:

[0037] Deep convolutional neural networks (CNNs) are deep learning models capable of processing spatially structured data. They have wide applications not only in image, speech, and audio signal processing but also demonstrate excellent capabilities in time series prediction. Generally, CNNs used for time series modeling applications primarily employ one-dimensional and two-dimensional convolutions. Compared to one-dimensional CNNs, two-dimensional CNNs increase the channel dimension, thus capturing features of the input information at a deeper level. The convolution process of a 2D CNN is as follows... Figure 3 As shown; CNNs primarily use convolution operations to identify local and global features in data. During convolution, the weights of the convolution kernels are shared, which reduces network parameters and speeds up network computation. The non-linear function ReLU is often used as the activation function of convolutional layers, enabling the learning of complex patterns and improving training speed without significantly affecting model generalization. The feature extraction process of CNNs is as follows: Figure 1 As shown, in this embodiment of the invention, a convolution kernel size of 3 and a stride of 1 are used, and three convolutions are performed, with the number of channels being 16, 32, and 64 respectively. The convolution formula of the CNN is as follows:

[0038]

[0039] in, and These represent the outputs of the l-th layer and the (l-1)-th layer, respectively. M represents the i-th convolutional kernel in the o-th feature map of the l-th layer; o Represents the input feature map; ReLU(·) represents the activation function;

[0040] To accelerate the training and convergence process of neural networks, a Batch Normalization (BN) layer is added when constructing a convolutional neural network. This layer normalizes the input data, ensuring that the input values ​​of the activation function fall within the sensitive region of the function. This reduces training time, speeds up network convergence, and minimizes gradient vanishing and exploding. The calculation formula for the BN layer is shown below:

[0041]

[0042] Among them, x i Represents input features; μ B and σ B These represent the mean and variance, respectively; m represents the number of input samples. ε is the standardized value of the input features; ε is a very small number to prevent the denominator from being 0; γ and β are learnable parameters; y i It is the output of the BN layer.

[0043] S3. Based on the features extracted using LSTM processing, preliminary SOH estimation results for the battery are obtained:

[0044] Long Short-Term Memory Network (LSTM) is a unique type of Recurrent Neural Network (RNN) that addresses the long-term dependency issue in traditional RNNs by introducing a gating mechanism to control the information flow between different layers. Compared to other neural networks, LSTM is better suited for simulating and predicting different time series. The basic structure of an LSTM model includes an input gate, a forget gate, an output gate, and a unit state. The input gate is responsible for information transmission and updating the unit state. The output gate calculates the output information based on the previous output and the unit state. The forget gate determines which information to discard and which to retain based on the result of the input gate. A diagram illustrating the working principle of LSTM is shown below. Figure 4 As shown, the computation process of LSTM can be expressed mathematically as follows:

[0045] i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i );

[0046] f t =σ(W xf x t +W ht h t-1 +W cf ct-1 +b f );

[0047] o t =σ(W xo x t +W ho h t-1 +W co c t-1 +b o );

[0048]

[0049] h t =o t ⊙tanh(c t );

[0050] Among them, x t It is the input at time t; h t It is the output of the hidden layer; i t f t and o t These are the outputs of the input gate, forget gate, and output gate at time t, respectively; c t and Let represent the cell state and candidate memory unit at time t, respectively; ⊙ represents the Hadamard product; σ(·) and tanh(·) represent the sigmoid and tanh activation functions, respectively; W(·) and b(·) represent the weight matrix and bias vector, respectively.

[0051] This invention utilizes the excellent time series modeling capabilities of LSTM, inputting the deep aging features extracted by CNN into LSTM for battery SOH estimation. The optimizer is selected as the Adam optimizer, and the loss function is the mean squared error (MSE), expressed as follows:

[0052]

[0053] Where n represents the length of the test set; y i and These represent reference values ​​and estimated values, respectively.

[0054] To verify the effectiveness of the proposed method, three typical estimation error metrics were used to evaluate the performance of the designed model: root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 Their calculation formulas are as follows:

[0055]

[0056]

[0057] Among them, y i , and These represent the reference value, the estimated value, and the average of the reference values, respectively.

[0058] S4. Optimize the deep learning model using a war strategy algorithm to obtain the optimal hyperparameters and the optimal SOH estimation model, thus obtaining the best battery SOH estimation result:

[0059] Since the accuracy of neural network predictions is highly correlated with the choice of hyperparameters, this embodiment of the invention chooses to use the War Strategy Optimization Algorithm (WSO) for hyperparameter search. In WSO, the war strategy is modeled as an optimization process in which each soldier dynamically moves towards the optimal value. The proposed algorithm simulates two popular war strategies, namely attack and defense strategies. The soldier's position on the battlefield is updated according to the implemented strategy. To improve the convergence and robustness of the algorithm, a new weight update mechanism and a weak soldier relocation strategy are introduced. The war strategy algorithm achieves a good balance between the exploration and development phases. The hyperparameters optimized in the optimization algorithm include the learning rate lr, the batch size, and the normalization parameter L2. The search strategy in WSO involves the following steps:

[0060] Attack strategy:

[0061] X i (t+1)=X i (t)+2×ρ×(CK)×rand×(W i ×KX i (t));

[0062] Among them, X i (t+1) represents the updated position; X i (t) represents the old position; C and K are the positions of the commander and the king, respectively; W i The weight is represented by ρ; rand is a random number between 0 and 1, and ρ is the control step size, which is 1 in this embodiment of the invention.

[0063] Defense strategy:

[0064] X i (t+1)=X i (t)+2×ρ×(KX rand (t))+rand×W i ×(CX i (t));

[0065] Among them, X rand Indicates the location of a random soldier;

[0066] Sorting and weighting:

[0067] X i (t+1)=X i (t+1)×(F n ≥F p )+X i (t)×(F n ×F p );

[0068] R i =(R i +1)×(F n ×F p )+R i ×(F n <F p );

[0069]

[0070] Among them, F n and F p R represents the soldier's attack power (i.e., fitness value) for the new and old positions, respectively; i The rank represents the soldier's rank; α is a factor that controls the intensity of weight changes; T represents the current iteration number.

[0071] Weak soldier replacement strategy:

[0072] X w (t+1)=rand×(ub-lb)+lb;

[0073] X w (t+1)=-(1-N(0,1))×(X w (t)-median(X))+K;

[0074] Among them, X w (·) represents the position of the weak soldier; ub and lb represent the upper and lower bounds of the search, respectively; N(0,1) is a random number from a standard normal distribution; median(·) is a median function;

[0075] The process of WSO optimizing the hyperparameters of deep learning models is as follows: Figure 5 As shown, R is selected during the optimization process. 2 As the fitness value for optimization, the SOH estimate is initially obtained using the LSTM model to calculate R. 2The data is then input into WSO and iterated continuously until the stopping condition is met to obtain the optimal combination of hyperparameters, thereby obtaining the optimal estimation model and the best SOH estimation result.

[0076] To verify the effectiveness of the method proposed in this embodiment of the invention, the Tongji University lithium-ion battery dataset was used for verification. The naming rule for the batteries is as follows: the naming rule for the batteries is CYX-Y / Z, where X represents the temperature and Y / Z represents the charge / discharge current rate ratio.

[0077] The SOH was estimated using the method provided in the embodiments of the present invention, and the results are as follows: Figure 6 As shown, the method provided in this embodiment of the invention achieves accurate estimation;

[0078] Two machine learning methods, Random Forest Regression (RFR) and Support Vector Regression (SVR), and two deep learning methods, Transformer and Temporal Convolutional Network (TCN), were selected to test the state of harmonic equilibrium (SOH) of the battery under different conditions. Detailed results are as follows: Figure 7-10 As shown, the method provided in this embodiment of the invention achieves minimal error under different working conditions.

[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for estimating the state of health of lithium-ion batteries based on deep learning and relaxation voltage, characterized in that, Includes the following steps: Extract the relaxation voltage sequence after full charge; CNN is used to extract deep features related to battery aging from the relaxation voltage sequence; Based on the features extracted by LSTM processing, preliminary SOH estimation results for the battery are obtained; The deep learning model is optimized using a war strategy algorithm to obtain the best hyperparameters and the optimal SOH estimation model, thus obtaining the best battery SOH estimation result.

2. The lithium-ion battery health state estimation method based on deep learning and relaxation voltage according to claim 1, characterized in that, The step of extracting deep-level features related to battery aging from the relaxed voltage sequence using CNN specifically includes: Feature extraction from the relaxed voltage sequence is performed using a multi-layer CNN. The convolution formula for the CNN is as follows: in, and These represent the outputs of the l-th layer and the (l-1)-th layer, respectively. M represents the i-th convolutional kernel in the o-th feature map of the l-th layer; o Represents the input feature map; ReLu(·) represents the activation function.

3. The lithium-ion battery health state estimation method based on deep learning and relaxation voltage according to claim 2, characterized in that, The CNN includes a batch normalization layer for normalizing the input data.

4. The lithium-ion battery health state estimation method based on deep learning and relaxation voltage according to claim 1, characterized in that, The step of obtaining preliminary battery SOH estimation results based on features extracted using LSTM processing specifically includes: The LSTM processing procedure is as follows: i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i ); f t =σ(W xf x t +W ht h t-1 +W cf c t-1 +b f ); o t =σ(W xo x t +W ho h t-1 +W co c t-1 +b o ); h t =o t ⊙tanh(c t ); Among them, x t It is the input at time t; h t It is the output of the hidden layer; i t f t and o t These are the outputs of the input gate, forget gate, and output gate at time t, respectively; c t and denoted as cell state and candidate memory unit at time t, respectively; ⊙ represents Hadamard product; σ(·) and tanh(·) represent sigmoid and tanh activation functions, respectively; W(·) and b(·) represent weight matrix and bias vector, respectively.

5. The lithium-ion battery health state estimation method based on deep learning and relaxation voltage according to claim 4, characterized in that, The LSTM uses the Adam optimizer, and the loss function is the mean squared error, expressed as follows: Where n represents the length of the test set; y i and These represent reference values ​​and estimated values, respectively.

6. The lithium-ion battery health state estimation method based on deep learning and relaxation voltage according to claim 1, characterized in that, The steps described above involve using a war strategy algorithm to optimize the deep learning model, obtain the optimal hyperparameters and the optimal SOH estimation model, and obtain the best battery SOH estimation result. The hyperparameters include the learning rate lr, the batch number, and the normalization parameter L2.

Images