A battery health estimation method considering adaptive representation of degradation information

Abstract

The present invention relates to a battery health estimation method considering adaptive characterization of degradation information, the method comprising the following steps: data preprocessing; data representation based on variational autoencoder VAE; taking the health factor of lithium-ion battery as input data, encoding it into hidden layer representation through the encoder of VAE, projecting it into a low-dimensional hidden variable space to extract the abstract features of the data, and then decoding the hidden layer representation into the original representation through the decoder; constructing an improved LSTM neural network, and optimizing the parameters in the neural network based on the training set by using an improved war strategy optimization algorithm; constructing an optimal multi-task model using optimal hyperparameters, performing SOH estimation in the training set, and performing SOH prediction on lithium-ion batteries in the test set. The present invention improves the global search capability of the optimization algorithm, realizes the adaptive representation of lithium-ion battery degradation information, and improves the model's ability to accurately estimate the SOH of lithium-ion batteries.

Description

Technical Field

[0001] The present invention relates to the technical field of power battery management, and in particular to a battery health estimation method that takes adaptive characterization of degradation information into consideration. Background Art

[0002] Electrochemical energy storage is an important technology to solve the energy problem of electric vehicles. Compared with lead-acid batteries, nickel-cadmium batteries, etc., lithium-ion batteries dominate the power battery and energy storage battery technology due to their excellent performance such as high energy density, low self-discharge rate, long service life, no memory and environmental friendliness. During the charge and discharge cycle of lithium-ion batteries, irreversible physical and chemical processes will occur inside them to form a solid electrolyte intermediate phase. The loss of active lithium and harsh external working conditions seriously affect the electrochemical reaction inside the battery, resulting in an increase in the internal resistance of the battery and a continuous decrease in capacity. This phenomenon is called battery degradation. Battery degradation will bring about a series of safety and reliability problems such as battery performance degradation and shortened battery life, and even lead to performance degradation of electrical equipment or system failure, causing fire and explosion accidents. In order to ensure the safe and reliable operation of electric vehicles and reduce their failure rate and operating costs, it is necessary to obtain accurate information on the health status of power batteries.

[0003] Domestic and foreign researchers have conducted a lot of research on battery SOH estimation methods. Estimation methods can be divided into model-based methods and data-driven methods. The parameter identification process of the model-based method usually depends on the convergence and convergence speed of the filtering algorithm. At the same time, it is affected by external conditions such as ambient temperature and battery type, and it is difficult to accurately identify all important parameters of the model. The data-driven SOH estimation technology measures the battery's current, voltage, and temperature parameters, extracts features that are highly correlated with SOH changes, and uses them as training data to build an estimation model, thereby realizing real-time estimation of lithium battery SOH. The data-driven SOH estimation method ignores the internal structure of the battery, only needs to extract health indicators (HIs) related to battery degradation, and uses the data-driven algorithm to establish a mapping relationship between the two, avoiding physical modeling and parameter identification processes. However, the accuracy of the data-driven method depends largely on the quantity and quality of the experimental data, and usually requires appropriate preprocessing of the original data. In addition, some data-driven methods also have inherent limitations, such as improper training modes that easily lead to overfitting and local optimization of the algorithm, and there are deficiencies such as high computational cost and high training complexity. Metaheuristic optimization algorithms are usually inspired by natural or social phenomena and can be used to solve complex optimization problems. Metaheuristic optimization algorithms can be used to search for the best neural network architecture, such as the number of neurons, number of layers, and connection methods, and can also be used to optimize the weights of neural networks as an alternative to traditional gradient descent methods. However, there are still many shortcomings that need further study, such as slow convergence speed and poor intelligent search capabilities. Therefore, how to optimize the parameters of the neural network in SOH estimation becomes another problem.

[0004] The existing technologies have the following main disadvantages: (1) Subjectivity in the selection of health factors. In actual engineering, the degradation process of lithium-ion batteries has complex nonlinear characteristics. Researchers usually select health factors (HIs) based on experience, which leads to a subjective selection process, and different researchers may select different health factors. This subjectivity increases the computational complexity and uncertainty of SOH estimation. (2) The problem of forgetting historical degradation information: When processing lithium-ion battery degradation information, traditional long short-term memory neural networks (LSTMs) have the problem of quickly forgetting historical information, which affects the accurate description of the battery degradation process and the precise estimation of SOH. (3) Difficulty in optimizing model hyperparameters: When constructing an SOH estimation model, selecting appropriate latent space representations and model hyperparameters is a complex optimization problem. Existing methods usually require a large number of experiments to determine the optimal parameter configuration, which is inefficient and time-consuming. Summary of the invention

[0005] In order to solve the shortcomings of the above-mentioned existing problems, the present invention proposes a lithium-ion battery SOH estimation method with adaptive representation of degradation information, which solves the problems of lack of adaptability in the process of using VAE to represent battery degradation information and the fast forgetting rate of traditional LSTM neural network for lithium-ion battery degradation information, thereby improving the estimation accuracy of the model for lithium-ion battery SOH.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A battery health estimation method considering adaptive characterization of degradation information, the method comprising the following steps:

[0008] Step 1: Data preprocessing;

[0009] Step 2: Data representation based on variational autoencoder VAE: The health factor of the lithium-ion battery is used as input data, encoded into a hidden layer representation through the encoder of VAE, projected into a low-dimensional hidden variable space to extract the abstract features of the data, and then decoded into the original representation through the decoder;

[0010] Step 3: Build an improved LSTM neural network and optimize the parameters in the neural network using an improved war strategy optimization algorithm based on the training set;

[0011] Step 4: Build the optimal multi-task model with the optimal hyperparameters to perform SOH estimation in the training set and SOH prediction for lithium-ion batteries in the test set.

[0012] As a further technical solution of the present invention, in step 1, the step of data preprocessing includes:

[0013] Carry out charge and discharge experiments on lithium-ion batteries, and record the current, voltage, temperature, time data of the lithium battery during the charge and discharge process, as well as the capacity of each full discharge. Analyze the power battery degradation process and aging mechanism based on the IC curve, and use the extreme values ​​on the curve and the difference between the extreme values ​​as health factors.

[0014] As a further technical solution of the present invention, in step 2, based on the data representation of the variational autoencoder VAE; the health factor of the lithium-ion battery is used as input data, encoded into a hidden layer representation by the encoder of the VAE, and projected into a low-dimensional hidden variable space to extract abstract features of the data, and then the hidden layer representation is decoded into the original representation by the decoder, the steps include:

[0015] Step 2.1: The relationship between the input data and the hidden layer space of VAE is expressed as Equation 1;

[0016] Formula 1:

[0017] x and denote the true value and estimated value of the input data, z denotes the hidden layer space, θ denotes the parameters of VAE, and the optimization goal is to minimize x and The error between

[0018] Step 2.2: Assume that x is generated by z, and the conditional probability distribution of z to x is p(x|z), p(z) is the prior probability distribution of z, and the optimization goal is transformed into maximizing p(x). The marginal likelihood estimate is shown in Equation 2.

[0019] Formula 2: p θ (x) = ∫p θ (z)p θ (x|z)dz;

[0020] p in the formula θ (x) is expressed as the probability of x under the condition of θ. According to the Bayesian formula, Formula 2 is transformed into Formula 3:

[0021] Formula 3: p θ (z|x)=p θ (z)p θ (x|z) / p θ (x);

[0022] Step 2.3: Since the hidden layer space z and the parameters θ are unknown, this increases the complexity of marginal likelihood estimation, so q is introduced k (z|x) to approximate p θ (z|x), given the value of x, we can use qκ(z|x) with parameter κ to generate a distribution z=g(x;κ), given the value of z, we can estimate and get the possible value of x. Through this framework, and z=g(x,κ) can represent p θ (z|x) and q k (z|x);

[0023] According to the maximum likelihood estimation, the logarithm of the marginal likelihood estimate is expressed by Equation 4:

[0024] Formula 4:

[0025]

[0026] According to KL divergence, Equation 4 is expressed by Equation 5:

[0027] Equation 5: logp θ (x) = E z [logp θ (x|z)]-D KL (q k(z|x)||p θ (z))+D KL (q k (z|x)||p θ (z|x))≥-

[0028] D KL (q k (z|x)||p θ (z|x))+E qk [logp θ (x|z)] = L(θ, k, x);

[0029] Where L(θ, k, x) is the loss function of VAE based on variational inference theory, including q k (z|x) and p θ (z) and the expected reconstruction error, so the optimization objective of VAE is expressed by Equation 6:

[0030] Formula 6:

[0031] θ * and k * Obtained by stochastic gradient variational Bayes method;

[0032] From the above description of VAE, it can be concluded that the original input data is transformed into potential hidden information through the encoder layer, and then the potential hidden information is transformed into an approximate representation of the input data. This method can filter out noise and irrelevant data in the original data and extract the most representative degradation information for lithium-ion battery SOH estimation.

[0033] As a further technical solution of the present invention, in step 3, the steps of constructing an improved LSTM neural network and optimizing the parameters in the neural network using an improved war strategy optimization algorithm based on a training set include:

[0034] Step 3.1: Build an improved LSTM neural network model. The LSTM neural network consists of a forget gate, an input gate, and an output gate. t is the input of the neural network at time t, and the output of the hidden layer node in the current state is h t and the hidden layer node state c t The output h of the hidden layer node at the previous moment t-1 and the hidden layer node state c t-1 The calculation process of LSTM is defined as follows:

[0035] Formula 7: f t =σ(w f [h t-1 ,x t ]+b f);

[0036] Formula 8:i t =σ(w i [h t-1 ,x t ]+b i );

[0037] Formula 9:

[0038] Formula 10:

[0039] Formula 11: o t =σ(w o [h t-1 ,x t ]+b o );

[0040] Formula 12: h t =o t ⊙tanh(c t );

[0041] Among them, f t 、i t , o t They are the output of the forget gate, the output of the input gate, the output of the candidate state, and the output of the output gate, respectively. f 、w i 、w c 、w o Respectively represent the weights of the forget gate, input gate, candidate state and output gate, h t-1 and h t are the outputs of the hidden layer states at the previous time t-1 and the current time t, c t-1 and c t Respectively represent the cell states at time t-1 and time t, σ and tanh represent the sigmoid function and tanh function respectively;

[0042] Although LSTM can alleviate some limitations of RNN, such as gradient vanishing and explosion problems, it forgets historical information quickly, which affects the accurate description of the battery degradation process and the precise estimation of SOH.

[0043] The traditional LSTM is improved and a new process is constructed to learn the historical information of the previous two moments. The specific steps include: t-2 ,h t-1 , c t-2 and c t-1 Connect them separately, summarize the historical information through the fully connected layer, and get the degradation information h' at the current moment t and cell state c't , the forget gate f of the improved LSTM neural network t , input gate i t , output gate o t The relationship between the weight parameter W and the bias parameter b corresponding to each gate is shown in the following formula;

[0044] Formula 13: h' t =σ(W' t [h t-2 ,h t-1 ]+b' t );

[0045] Formula 14:

[0046] Formula 15: f t =σ(w f [h' t ,x t ]+b f );

[0047] Equation 16:i t =σ(w i [h' t ,x t ]+b i );

[0048] Formula 17:

[0049] Formula 18:

[0050] Formula 19: o t =σ(w o [h' t ,x t ]+b o );

[0051] Formula 20: h t =o t ⊙tanh(c t );

[0052] Among them, W' t , b' t and The degradation information of the new process and the weight matrix and bias matrix of the cell state are constructed. Finally, the Adam optimizer is used to update the weight matrix and bias matrix to improve the computational efficiency and reduce the computational consumption.

[0053] According to the battery aging principle, it is clear that the capacity decreases with the increase of the number of charge and discharge cycles, indicating that the battery health status degradation data is a long-term time series, but the traditional LSTM neural network forgets the battery degradation data quickly and it is difficult to fully utilize the historical degradation information of lithium-ion batteries. Unlike the traditional LSTM neural network, the improved LSTM neural network adds a new processing process to the original structure, which can learn from the previous two moments to obtain the output result of the current moment, so as to improve the learning ability of lithium-ion battery degradation information, and can solve the problem of limited accuracy and robustness of SOH estimation of traditional LSTM neural network;

[0054] As a further technical solution of the present invention, step 3.2 is also included: in order to further improve the global search capability of the model and the accuracy of SOH estimation, the present invention adopts an improved war strategy optimization algorithm (IWSO) to adaptively optimize the improved LSTM hyperparameters;

[0055] The war strategy optimization algorithm is a meta-heuristic optimization algorithm based on the strategic movement of the army in ancient wars. Its basic settings are as follows:

[0056] (1) Soldiers in the army are randomly distributed on the battlefield. The soldier with the strongest attacking power is the commander, and the king is the supreme leader of the army;

[0057] (2) The positions of soldiers are dynamically updated based on the positions of the king and commander;

[0058] (3) The king dynamically changes his search strategy through war drums;

[0059] (4) Soldiers change their positions according to the positions of nearby soldiers and the king;

[0060] (5) The injured or weakest soldiers are replaced by new soldiers or their positions are updated through a reset strategy;

[0061] (6) All soldiers have an equal opportunity to become commanders or kings;

[0062] A detailed description of the war strategy optimization algorithm is as follows;

[0063] (1) Offensive strategy: The positions of soldiers are updated according to the positions of the king and commander, which are defined as follows:

[0064] Formula 21: X i (t+1)=X i (t)+2r 1 (CK)+r 1 (W i KX i (t));

[0065] X in the formulai (t+1) and X i (t) represents the position of the soldier at the t+1th iteration and the tth iteration, C, K and W i Represent the commander's position, the king's position and the king's weight, r 1 is a random number between 0 and 1;

[0066] (2) Soldier level and weight update: The level of a soldier is an important factor affecting its position, and the level of each soldier depends on its fitness value in the war; if the fitness value of a soldier in a new position is smaller than that in the previous position, the previous position is retained and no position update is performed:

[0067] Formula 22: X i (t+1)=X i (t+1)*(F n ≥F p )+X i (t)*(F n ≤F p );

[0068] F in the formula n and F p Represents the fitness value of the soldier at the current position and the previous position respectively. When the soldier's position is updated, the soldier's level R i will increase, which is defined as follows:

[0069] Formula 23: R i =(R i +1)*(F n ≥F p )+R i *(F n ≤F p );

[0070] Based on the soldier's new level, the weights will be updated as follows;

[0071] Formula 24:

[0072] R in the formula i and α represent the rank and exponential change factor of the i-th soldier, respectively;

[0073] (3) Defensive strategy: Soldiers will change their positions according to the positions of nearby soldiers and the king to protect the king and try to win the battle. The mathematical expression is as follows;

[0074] Formula 25: X i (t+1)=X i (t)+2ρ(KX r (t))+r 1 Wi (cX i (t));

[0075] X in the formula i (t) represents the random position of the soldier at the tth iteration, ρ is a random number between 0 and 1;

[0076] (4) Replacement and update of soldiers: In each iteration, the soldier with the worst fitness value is replaced. There are two replacement strategies;

[0077] The random replacement strategy is defined as follows;

[0078] Formula 26: X w (t+1)=Lb+r*(Ub-Lb);

[0079] X in the formula w (t+1) represents the soldier replaced in the t+1 iteration, Ub and Lb are the upper and lower bounds of the search space;

[0080] Another strategy is defined as follows;

[0081] Formula 27: X w (t+1)=-(1-randn(X))*(X w (t)-median(X))+K;

[0082] X in the formula w (t) represents the soldier with low fitness value in the tth iteration, randn is the function used to generate random numbers that conform to the standard normal distribution, and median is the median function;

[0083] In order to overcome the problem that traditional war strategy optimization algorithm easily falls into local optimum and converges prematurely during the iteration process, an improved mathematical expression is proposed to enhance its global search capability.

[0084] As a further technical solution of the present invention, step 3.2 further includes:

[0085] First, considering that the initialization of the traditional war strategy optimization algorithm is a random process, the soldiers are unevenly distributed in space, which affects the global search performance. Therefore, the cubic chaos mapping is used to initialize the soldier positions to make the distribution of soldiers more uniform. The mathematical expression is as follows:

[0086] Formula 28: y(t+1)=4y(t) 3 -3y(t);

[0087] Assume that there are m individuals in total, the first individual is randomly generated, and then the mapping of the remaining m-1 individuals is iterated according to Formula 22. On this basis, these obtained cubic chaotic operators are mapped to the whole, which is defined as follows;

[0088] Formula 29: X t =X lb +(X ub -X lb )*y(t+1)*0.5;

[0089] X in the formula ub and X lb They represent the upper and lower bounds of the search space respectively.

[0090] As a further technical solution of the present invention, step 3.2 further includes:

[0091] Secondly, in order to improve the convergence performance of the war strategy optimization algorithm in the search process, an improved nonlinear convergence function is proposed, which is defined as follows;

[0092] Formula 30:

[0093] The r in the formula 2 is a random number between 0 and 1, d r represents the attenuation factor, and its value is set to 0.5. According to the proposed nonlinear convergence function, Equation 21 and Equation 25 can be redefined as Equation 31 and Equation 32;

[0094] Formula 31: X i (t+1)=X i (t)+2r 1 (CK)*f o +r 1 (W i KX i (t));

[0095] Formula 32: X i (t+1)=X i (t)+2ρ(KX r (t))*f o +r 1 W i (cX i (t));

[0096] The present invention adopts an improved war strategy optimization algorithm to simultaneously adjust the hyperparameters of VAE and improved LSTM to achieve collaborative optimization of SOH estimation under multi-tasks. The optimized hyperparameters include learning rate, the number of neurons of VAE encoder and decoder, and the number of neurons of improved LSTM neural network.

[0097] As a further technical solution of the present invention, step 3.2 includes:

[0098] Step 3.2.1: Use the cubic chaos mapping method to initialize the soldier population, including the soldier population U and the maximum number of iterations T m , and the upper bound U of the hyperparameters of VAE and improved LSTM neural network b and the lower bound L b , the hyperparameter information of the neural network is contained in the soldier population U, and its expression is as follows;

[0099] Formula 33: U = [lr i ,h zi ,h vaei ,h li ];

[0100] In the formula, lr i represents the learning rate of the i-th soldier, h zi 、h vaei and h li are the number of VAE encoder neurons, VAE decoder neurons, and improved LSTM neurons of the i-th soldier respectively;

[0101] Step 3.2.2: Calculate the fitness function, using the coefficient of determination R of the training set 2 (R-Square) is used as a fitness function to retain the information of the best soldiers;

[0102] Step 3.2.3: Update the soldier's position according to equations 22, 23, 26, 27, 31 and 32;

[0103] Step 3.2.4: Calculate the fitness value. If the previous fitness value is worse than the current one, update the parameters; otherwise, keep the previous information.

[0104] Step 3.2.5: If the maximum iteration is reached or the predefined early stopping criterion is met, exit the loop and output the optimal hyperparameter space.

[0105] The terms involved in the present invention are explained as follows:

[0106] SOH: State of health (SOH) is a quantitative indicator used to evaluate battery aging, usually expressed as the ratio of the actual capacity of the battery to the nominal capacity.

[0107] LSTM: Long short-term memory (LSTM) is a type of feedback deep learning neural network used to process time series data. The state at each moment is affected by the current input state and the previous input state, and has a memory effect.

[0108] VAE: Variational autoencoder (VAE) is a deep generative model based on the autoencoder structure. It is widely used in fields such as dimensionality reduction and feature extraction. The basic structure is to map samples to latent variables in a low-dimensional space through the encoding process, and then restore the latent variables to reconstructed samples through the decoding process.

[0109] WSO: War strategy optimization (WSO) is a new meta-heuristic optimization algorithm based on ancient war strategies. The algorithm is inspired by the attack and defense strategies in ancient wars. It solves the optimization problem by updating the positions of soldiers on the battlefield. It has the characteristics of strong optimization ability and fast convergence speed.

[0110] Compared with the prior art, the present invention has the following beneficial effects:

[0111] Firstly, a variational autoencoder (VAE) is used to filter out the noise and irrelevant information in the original data, extract the most representative information for lithium-ion battery SOH estimation, and realize the representation of lithium-ion battery degradation information based on the degradation health index.

[0112] Secondly, in order to solve the problem that the traditional LSTM neural network forgets the degradation information of lithium-ion batteries at a fast rate, an improved LSTM neural network model is constructed to improve the LSTM forget gate and input gate, which can make full use of the historical degradation information of lithium-ion batteries; in order to solve the problem that the war strategy optimization algorithm is prone to fall into the local optimum, an improved war strategy optimization algorithm is proposed to improve the global search ability of the optimization algorithm;

[0113] Finally, an improved war strategy optimization algorithm is used to adjust the hyperparameters of VAE and improved LSTM, and collaborative optimization of SOH estimation under multi-task conditions is performed, which realizes the adaptive representation of lithium-ion battery degradation information and improves the model's ability to accurately estimate lithium-ion battery SOH. BRIEF DESCRIPTION OF THE DRAWINGS

[0114] Figure 1 A structural diagram of a variational autoencoder provided in an embodiment of the present invention.

[0115] Figure 2 A structural diagram of an improved LSTM neural network provided in an embodiment of the present invention.

[0116] Figure 3 An optimization process diagram for the improved war strategy optimization provided by an embodiment of the present invention.

[0117] Figure 4A diagram of the SOH estimation process provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0118] The technical solution of this patent is further described in detail below in conjunction with specific implementation methods.

[0119] See also Figures 1 to 4 , a battery health estimation method considering adaptive characterization of degradation information, the method comprising the following steps:

[0120] Step 1: Data preprocessing;

[0121] Specifically, it includes: conducting charge and discharge experiments on lithium-ion batteries, and recording the current, voltage, temperature, time data of the lithium battery during the charge and discharge process, as well as the capacity of each full discharge, analyzing the power battery degradation process and aging mechanism based on the IC curve, and taking the extreme values ​​on the curve and the difference between the extreme values ​​as health factors.

[0122] Step 2: Data representation based on variational autoencoder VAE: The health factor of the lithium-ion battery is used as input data, encoded into a hidden layer representation through the encoder of VAE, projected into a low-dimensional hidden variable space to extract the abstract features of the data, and then decoded into the original representation through the decoder;

[0123] Specifically, it includes: Step 2.1: The relationship between the input data and the hidden layer space of VAE is expressed as formula 1;

[0124] Formula 1:

[0125] x and denote the true value and estimated value of the input data, z denotes the hidden layer space, θ denotes the parameters of VAE, and the optimization goal is to minimize x and The error between

[0126] Step 2.2: Assume that x is generated by z, and the conditional probability distribution of z to x is p(x|z), p(z) is the prior probability distribution of z, and the optimization goal is transformed into maximizing p(x). The marginal likelihood estimate is shown in Equation 2.

[0127] Formula 2: p θ (x) = ∫p θ (z)p θ (x|z)dz;

[0128] p in the formula θ (x) is expressed as the probability of x under the condition of θ. According to the Bayesian formula, Formula 2 is transformed into Formula 3:

[0129] Formula 3: p θ (z|x)=pθ (z)p θ (x|z) / p θ (x);

[0130] Step 2.3: Since the hidden layer space z and the parameters θ are unknown, this increases the complexity of marginal likelihood estimation, so q is introduced k (z|x) to approximate p θ (z|x), given the value of x, we can use qκ(z|x) with parameter κ to generate a distribution z=g(x;κ), given the value of z, we can estimate and get the possible value of x. Through this framework, and z=g(x,κ) can represent p θ (z|x) and q k (z|x);

[0131] According to the maximum likelihood estimation, the logarithm of the marginal likelihood estimate is expressed by Equation 4:

[0132] Formula 4:

[0133]

[0134] According to KL divergence, Equation 4 is expressed by Equation 5:

[0135] Equation 5: logp θ (x) = E z [logp θ (x|z)]-D KL (q k (z|x)||p θ (z))+D KL (q k (z|x)||p θ (z|x))≥-

[0136] D KL (q k (z|x)||p θ (z|x))+E qk [logp θ (x|z)] = L(θ, k, x);

[0137] Where L(θ, k, x) is the loss function of VAE based on variational inference theory, including q k (z|x) and p θ (z) and the expected reconstruction error, so the optimization objective of VAE is expressed by Equation 6:

[0138] Formula 6:

[0139] θ* and k * Obtained by stochastic gradient variational Bayes method;

[0140] From the above description of VAE, it can be concluded that the original input data is converted into potential hidden information through the encoder layer, and then the potential hidden information is converted into an approximate representation of the input data. This method can filter out noise and irrelevant data in the original data and extract the most representative degradation information for lithium-ion battery SOH estimation. The structure diagram of VAE is shown in Figure 1 shown.

[0141] Step 3: Build an improved LSTM neural network and optimize the parameters in the neural network using an improved war strategy optimization algorithm based on the training set;

[0142] It specifically includes: Step 3.1: Build an improved LSTM neural network model. The LSTM neural network consists of a forget gate, an input gate, and an output gate. t is the input of the neural network at time t, and the output of the hidden layer node in the current state is h t and the hidden layer node state c t The output h of the hidden layer node at the previous moment t-1 and the hidden layer node state c t-1 The calculation process of LSTM is defined as follows:

[0143] Formula 7: f t =σ(w f [h t-1 ,x t ]+b f );

[0144] Formula 8: i t =σ(w i [h t-1 ,x t ]+b i );

[0145] Formula 9:

[0146] Formula 10:

[0147] Formula 11: o t =σ(w o [h t-1 ,x t ]+b o );

[0148] Formula 12: h t =o t ⊙tanh(c t );

[0149] Among them, f t 、i t , o t They are the output of the forget gate, the output of the input gate, the output of the candidate state, and the output of the output gate, respectively. f 、w i 、w c 、w o Respectively represent the weights of the forget gate, input gate, candidate state and output gate, h t-1 and h t are the outputs of the hidden layer states at the previous time t-1 and the current time t, c t-1 and c t Respectively represent the cell states at time t-1 and time t, σ and tanh represent the sigmoid function and tanh function respectively;

[0150] Although LSTM can alleviate some limitations of RNN, such as gradient vanishing and explosion problems, it forgets historical information quickly, which affects the accurate description of the battery degradation process and the precise estimation of SOH.

[0151] The traditional LSTM is improved and a new process is constructed to learn the historical information of the previous two moments. The specific steps include: t-2 ,h t-1 , c t-2 and c t-1 Connect them separately, summarize the historical information through the fully connected layer, and get the degradation information h' at the current moment t and cell state c' t , the forget gate f of the improved LSTM neural network t , input gate i t , output gate o t The relationship between the weight parameter W and the bias parameter b corresponding to each gate is shown in the following formula;

[0152] Formula 13: h' t =σ(W' t [h t-2 ,h t-1 ]+b' t );

[0153] Formula 14:

[0154] Formula 15: f t =σ(w f [h' t ,x t ]+b f );

[0155] Equation 16:it =σ(w i [h' t ,x t ]+b i );

[0156] Formula 17:

[0157] Formula 18:

[0158] Formula 19: o t =σ(w o [h' t ,x t ]+b o );

[0159] Formula 20: h t =o t ⊙tanh(c t );

[0160] Among them, W' t , b' t and The degradation information of the new process and the weight matrix and bias matrix of the cell state are constructed. Finally, the Adam optimizer is used to update the weight matrix and bias matrix to improve the computational efficiency and reduce the computational consumption.

[0161] According to the battery aging principle, it is clear that the capacity decreases with the increase of the number of charge and discharge cycles, indicating that the battery health degradation data is a long-term time series. However, the traditional LSTM neural network forgets the battery degradation data quickly and it is difficult to fully utilize the historical degradation information of lithium-ion batteries. Unlike the traditional LSTM neural network, the improved LSTM neural network adds a new processing process to the original structure, which can learn from the previous two moments to obtain the output result at the current moment, so as to improve the learning ability of lithium-ion battery degradation information and solve the problem of limited accuracy and robustness of SOH estimation of traditional LSTM neural network. The structural diagram of the proposed improved LSTM neural network is shown in the figure below: Figure 2 shown.

[0162] Step 3.2: In order to further improve the global search capability of the model and the accuracy of SOH estimation, the present invention adopts an improved war strategy optimization algorithm (IWSO) to adaptively optimize the improved LSTM hyperparameters;

[0163] The war strategy optimization algorithm is a meta-heuristic optimization algorithm based on the strategic movement of the army in ancient wars. Its basic settings are as follows:

[0164] (1) Soldiers in the army are randomly distributed on the battlefield. The soldier with the strongest attacking power is the commander, and the king is the supreme leader of the army;

[0165] (2) The positions of soldiers are dynamically updated based on the positions of the king and commander;

[0166] (3) The king dynamically changes his search strategy through war drums;

[0167] (4) Soldiers change their positions according to the positions of nearby soldiers and the king;

[0168] (5) The injured or weakest soldiers are replaced by new soldiers or their positions are updated through a reset strategy;

[0169] (6) All soldiers have an equal opportunity to become commanders or kings;

[0170] A detailed description of the war strategy optimization algorithm is as follows;

[0171] (1) Offensive strategy: The positions of soldiers are updated according to the positions of the king and commander, which are defined as follows:

[0172] Formula 21: X i (t+1)=X i (t)+2r 1 (CK)+r 1 (W i KX i (t));

[0173] X in the formula i (t+1) and X i (t) represents the position of the soldier at the t+1th iteration and the tth iteration, C, K and W i Represent the commander's position, the king's position and the king's weight, r 1 is a random number between 0 and 1;

[0174] (2) Soldier level and weight update: The level of a soldier is an important factor affecting its position, and the level of each soldier depends on its fitness value in the war; if the fitness value of a soldier in a new position is smaller than that in the previous position, the previous position is retained and no position update is performed:

[0175] Formula 22: X i (t+1)=X i (t+1)*(F n ≥F p )+X i (t)*(F n ≤F p );

[0176] F in the formulan and F p Represents the fitness value of the soldier at the current position and the previous position respectively. When the soldier's position is updated, the soldier's level R i will increase, which is defined as follows:

[0177] Formula 23: R i =(R i +1)*(F n ≥F p )+R i *(F n ≤F p );

[0178] Based on the soldier's new level, the weights will be updated as follows;

[0179] Formula 24:

[0180] R in the formula i and α represent the rank and exponential change factor of the i-th soldier, respectively;

[0181] (3) Defensive strategy: Soldiers will change their positions according to the positions of nearby soldiers and the king to protect the king and try to win the battle. The mathematical expression is as follows;

[0182] Formula 25: X i (t+1)=X i (t)+2ρ(KX r (t))+r 1 W i (cX i (t));

[0183] X in the formula i (t) represents the random position of the soldier at the tth iteration, ρ is a random number between 0 and 1;

[0184] (4) Replacement and update of soldiers: In each iteration, the soldier with the worst fitness value is replaced. There are two replacement strategies;

[0185] The random replacement strategy is defined as follows;

[0186] Formula 26: X w (t+1)=Lb+r*(Ub-Lb);

[0187] X in the formula w (t+1) represents the soldier replaced in the t+1 iteration, Ub and Lb are the upper and lower bounds of the search space;

[0188] Another strategy is defined as follows;

[0189] Formula 27: Xw (t+1)=-(1-randn(X))*(X w (t)-median(X))+K;

[0190] X in the formula w (t) represents the soldier with low fitness value in the tth iteration, randn is the function used to generate random numbers that conform to the standard normal distribution, and median is the median function;

[0191] In order to overcome the problem that traditional war strategy optimization algorithm easily falls into local optimum and converges prematurely during the iteration process, an improved mathematical expression is proposed to enhance its global search capability.

[0192] First, considering that the initialization of the traditional war strategy optimization algorithm is a random process, the soldiers are unevenly distributed in space, which affects the global search performance. Therefore, the cubic chaos mapping is used to initialize the soldier positions to make the distribution of soldiers more uniform. The mathematical expression is as follows:

[0193] Formula 28: y(t+1)=4y(t) 3 -3y(t);

[0194] Assume that there are m individuals in total, the first individual is randomly generated, and then the mapping of the remaining m-1 individuals is iterated according to Formula 22. On this basis, these obtained cubic chaotic operators are mapped to the whole, which is defined as follows;

[0195] Formula 29: X t =X lb +(X ub -X lb )*y(t+1)*0.5;

[0196] X in the formula ub and X lb They represent the upper and lower bounds of the search space respectively.

[0197] Secondly, in order to improve the convergence performance of the war strategy optimization algorithm in the search process, an improved nonlinear convergence function is proposed, which is defined as follows;

[0198] Formula 30:

[0199] The r in the formula 2 is a random number between 0 and 1, d r represents the attenuation factor, and its value is set to 0.5. According to the proposed nonlinear convergence function, Equation 21 and Equation 25 can be redefined as Equation 31 and Equation 32;

[0200] Formula 31: X i (t+1)=Xi (t)+2r 1 (CK)*f o +r 1 (W i KX i (t));

[0201] Formula 32: X i (t+1)=X i (t)+2ρ(KX r (t))*f o +r 1 W i (cX i (t));

[0202] The present invention uses an improved war strategy optimization algorithm to adjust the hyperparameters of VAE and improved LSTM at the same time to achieve collaborative optimization of SOH estimation under multi-tasks. The optimized hyperparameters include learning rate, the number of neurons of VAE encoder and decoder, and the number of neurons of improved LSTM neural network. The whole process of optimization using the improved war optimization algorithm is as follows: Figure 3 shown.

[0203] Step 3.2.1: Use the cubic chaos mapping method to initialize the soldier population, including the soldier population U and the maximum number of iterations T m , and the upper bound U of the hyperparameters of VAE and improved LSTM neural network b and the lower bound L b , the hyperparameter information of the neural network is contained in the soldier population U, and its expression is as follows;

[0204] Formula 33: U = [lr i ,h zi ,h vaei ,h li ];

[0205] In the formula, lr i represents the learning rate of the i-th soldier, h zi 、h vaei and h li are the number of VAE encoder neurons, VAE decoder neurons, and improved LSTM neurons of the i-th soldier respectively;

[0206] Step 3.2.2: Calculate the fitness function, using the coefficient of determination R of the training set 2 (R-Square) is used as a fitness function to retain the information of the best soldiers;

[0207] Step 3.2.3: Update the soldier's position according to equations 22, 23, 26, 27, 31 and 32;

[0208] Step 3.2.4: Calculate the fitness value. If the previous fitness value is worse than the current one, update the parameters; otherwise, keep the previous information.

[0209] Step 3.2.5: If the maximum iteration is reached or the predefined early stopping criterion is met, exit the loop and output the optimal hyperparameter space.

[0210] Step 4: Use the optimal hyperparameters to build the optimal multi-task model, perform SOH estimation in the training set, and perform SOH prediction for lithium-ion batteries in the test set. The process of the proposed lithium-ion battery SOH estimation method based on the improved LSTM neural network based on the war strategy optimization algorithm is as follows Figure 4 shown.

[0211] At present, the data-driven SOH estimation method for lithium-ion batteries mainly includes three aspects: constructing a battery SOH dataset, extracting SOH characteristic parameters, and establishing a SOH estimation model. SOH, which is usually defined by capacity and internal resistance, is difficult to estimate directly. Currently, characteristic parameters such as voltage, current, and temperature that can reflect changes in the battery's health status are generally used as health factors. The strength of the correlation between the health factor and SOH directly affects the accuracy of the estimation model. The basic idea is to build an optimal mapping model between input and output, that is, first build a battery dataset, then extract the health factor from it as the input of the established model, and then continuously train, optimize and adjust the model to achieve accurate estimation of the battery SOH.

[0212] The selection of battery health factors is crucial to the accuracy of battery SOH model estimation. Inappropriate health factors have little effect on the accurate estimation of SOH and will also increase the computational complexity and estimation uncertainty.

[0213] Existing technologies directly extract features related to battery degradation from the battery charging voltage curve based on the collected parameters such as temperature, voltage, and current. The present invention analyzes the power battery degradation process and aging mechanism according to the IC curve, and uses the height of the peak on the curve and the distance between different peaks as health factors; on this basis, these health factors are used as the input of the variational autoencoder VAE to complete the reconstruction of the lithium-ion battery degradation information in the latent space, retaining the health factors that are more correlated with SOH, and eliminating irrelevant or redundant health factors. Then, the degradation information in the latent space reconstructed by VAE is used as the input of the improved LSTM neural network for training the SOH estimation model.

[0214] The optimization algorithm of the prior art is only used to optimize the structure of the neural network. The present invention uses an improved war strategy optimization algorithm to collaboratively optimize the parameters of the VAE and the improved LSTM neural network, which can achieve adaptive representation of lithium-ion battery degradation information and improve the SOH estimation accuracy.

[0215] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0216] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A battery health estimation method considering adaptive characterization of degradation information, characterized in that: The method comprises the following steps: Step 1: Data preprocessing; Step 2: Data representation based on variational autoencoder VAE: The health factor of the lithium-ion battery is used as input data, encoded into a hidden layer representation through the encoder of VAE, projected into a low-dimensional hidden variable space to extract the abstract features of the data, and then decoded into the original representation through the decoder; Step 3: Build an improved LSTM neural network and optimize the parameters in the neural network using an improved war strategy optimization algorithm based on the training set; Step 4: Build the optimal multi-task model with the optimal hyperparameters to perform SOH estimation in the training set and S0H prediction for lithium-ion batteries in the test set; In step 3, the steps of constructing an improved LSTM neural network and optimizing the parameters in the neural network using an improved war strategy optimization algorithm based on the training set include: Step 3.1: Build an improved LSTM neural network model. The LSTM neural network consists of a forget gate, an input gate, and an output gate. t is the input of the neural network at time t, and the output of the hidden layer node in the current state is h t and the hidden layer node state c t The output h of the hidden layer node at the previous moment t-1 and the hidden layer node state c t-1 The calculation process of LSTM is defined as follows: Equation 7: f t = σ(w f [h t-1 , x t + b f ); Equation 8: i t = σ(w i [h t-1 , x t + b i ); Formula 9: Formula 10: Equation 11: o t = σ(w o [h t-1 , x t + b o ); Formula 12: h t =o t ⊙tanh(c t ); Among them, f t 、i t , o t They are the output of the forget gate, the output of the input gate, the output of the candidate state, and the output of the output gate, respectively. f 、w i 、w c 、w o Respectively represent the weights of the forget gate, input gate, candidate state and output gate, h t-1 and h t are the outputs of the hidden layer states at the previous time t-1 and the current time t, c t-1 and c t Respectively represent the cell states at time t-1 and time t, σ and tanh represent the sigmoid function and tanh function respectively; The traditional LSTM is improved and a new process is constructed to learn the historical information of the previous two moments; the specific steps include: t-2 ,h t-1 , c t-2 and c t-1 Connect them separately to get the degradation information h′ at the current moment t and cell state c′ t , the forget gate f of the improved LSTM neural network t , input gate i t , output gate o t The relationship between the weight parameter W and the bias parameter b corresponding to each gate is shown in the following formula; Formula 13: h′ t =σ(W′ t [h t-2 ,h t-1 ]+b′ t ); Formula 14: Equation 15: f t = σ(w f [h′ t , x t + b f ); Equation 16: i t = σ(w i [h′ t , x t + b i ); Formula 17: Formula 18: Equation 19: o t = σ(w o [h′ t , x t + b o ); Formula 20: h t =o t ⊙tanh(c t ); Among them, W′ t , b′ t and It is the degradation information of the new process and the weight matrix and bias matrix of the cell state. Finally, the Adam optimizer is used to update the weight matrix and bias matrix.

2. The method for battery health estimation considering adaptive characterization of degradation information according to claim 1, characterized in that: In step 1, the data preprocessing steps include: Carry out charge and discharge experiments on lithium-ion batteries, and record the current, voltage, temperature, time data of the lithium battery during the charge and discharge process, as well as the capacity of each full discharge. Analyze the power battery degradation process and aging mechanism based on the IC curve, and use the extreme values ​​on the curve and the difference between the extreme values ​​as health factors.

3. The method for battery health estimation considering adaptive characterization of degradation information according to claim 2, characterized in that: In step 2, based on the data representation of variational autoencoder VAE; the health factor of the lithium-ion battery is used as input data, encoded into a hidden layer representation through the encoder of VAE, projected into a low-dimensional hidden variable space to extract the abstract features of the data, and then, the steps of decoding the hidden layer representation into the original representation through the decoder include: Step 2.1: The relationship between the input data and the hidden layer space of VAE is expressed as Equation 1; Formula 1: x and denote the true value and estimated value of the input data, z denotes the hidden layer space, θ denotes the parameters of VAE, and the optimization goal is to minimize x and The error between Step 2.2: Assume that x is generated by z, and the conditional probability distribution of z to x is p(x|z), p(z) is the prior probability distribution of z, and the optimization goal is transformed into maximizing p(x). The marginal likelihood estimate is shown in Equation 2. Equation 2: p θ (x) = ∫ p θ (z) p θ (x|z) dz; p in the formula θ( x) is expressed as the probability of x under the condition of θ. According to the Bayesian formula, Formula 2 is transformed into Formula 3: Formula 3: p θ (z|x) = p θ (z)p θ (x|z) / p θ (x); Step 2.3: Introduce q k (z|x) to approximate p θ (z|x), given the value of x, use qκ(z|x) with parameter κ to generate a distribution z=g(x;κ), given the value of z, estimate and get the possible value of x, and z=g(x,κ) respectively represent p θ (z|x) and q k (z|x); According to the maximum likelihood estimation, the logarithm of the marginal likelihood estimate is expressed by Equation 4: Formula 4: According to KL divergence, Equation 4 is expressed by Equation 5: Equation 5: log p θ (x) = E z [log p θ (x|z)] - D KL (q k (z|x) || p θ (z)) + D KL (q k (z|x) || p θ (z1x)) ≥ -D KL (q k (z|x) || p θ (z|x)) + E qk [log p θ (x|z)] = L(θ, k, x); Where L(θ, k, x) is the loss function of VAE based on variational inference theory, including q k (z|x) and p θ (z) and the expected reconstruction error, so the optimization objective of VAE is expressed by Equation 6: Formula 6: θ * and k * Obtained by stochastic gradient variational Bayes method.

4. The method for battery health estimation considering adaptive characterization of degradation information according to claim 1, characterized in that: It also includes step 3.2: using the improved war strategy optimization algorithm to adaptively optimize the improved LSTM hyperparameters; The war strategy optimization algorithm is a meta-heuristic optimization algorithm based on the strategic movement of the army in ancient wars. Its basic settings are as follows: (1) Soldiers in the army are randomly distributed on the battlefield. The soldier with the strongest attacking power is the commander, and the king is the supreme leader of the army; (2) The positions of soldiers are dynamically updated based on the positions of the king and commander; (3) The king dynamically changes his search strategy through war drums; (4) Soldiers change their positions according to the positions of nearby soldiers and the king; (5) The injured or weakest soldiers are replaced by new soldiers or their positions are updated through a reset strategy; (6) All soldiers have an equal opportunity to become commanders or kings; A detailed description of the war strategy optimization algorithm is as follows; (1) Offensive strategy: The positions of soldiers are updated according to the positions of the king and commander, which are defined as follows: Formula 21: X i (t+1)=X i (t)+2r1(CK)+r1(W i KX i (t)); X in the formula i (t+1) and X i (t) represents the position of the soldier at the t+1th iteration and the tth iteration, C, K and W i They represent the commander's position, the king's position and the king's weight respectively, and r1 is a random number between 0 and 1; (2) Soldier level and weight update: The level of a soldier is an important factor affecting its position, and the level of each soldier depends on its fitness value in the war; if the fitness value of a soldier in a new position is smaller than that in the previous position, the previous position is retained and no position update is performed: Formula 22: X i (t + 1) = X i (t + 1)*(F n ≥F p ) + X i (t)*(F n ≤F p ) F in the formula n and F p Represents the fitness value of the soldier at the current position and the previous position respectively. When the soldier's position is updated, the soldier's level R i will increase, which is defined as follows: Formula 23: R i =(R i +1)*(F n ≥F p )+R i *(F n ≤F p ); Based on the soldier's new level, the weights will be updated as follows; Formula 24: R in the formula i and α represent the rank and exponential change factor of the i-th soldier, respectively; (3) Defensive strategy: Soldiers will change their positions according to the positions of nearby soldiers and the king to protect the king and try to win the battle. The mathematical expression is as follows; Equation 25: X i (t + 1) = X i (t) + 2ρ(K - X r (t)) + r1W i (c - X i (t)); X in the formula i (t) represents the random position of the soldier at the tth iteration, ρ is a random number between 0 and 1; (4) Replacement and update of soldiers: In each iteration, the soldier with the worst fitness value is replaced. There are two replacement strategies; The random replacement strategy is defined as follows; Formula 26: X w (t + 1) = Lb + r * (Ub - Lb); X in the formula w (t+1) represents the soldier replaced in the t+1 iteration, Ub and Lb are the upper and lower bounds of the search space; Another strategy is defined as follows; Equation 27: X w (t + 1) = -(1 - randn(X)) * (X w (t) - median(X)) + K; X in the formula w (t) represents the soldier with low fitness value in the tth iteration, randn is the function used to generate random numbers that conform to the standard normal distribution, and median is the median function.

5. The method for battery health estimation considering adaptive characterization of degradation information according to claim 4, characterized in that: Step 3.2 also includes: The cubic chaotic mapping is used to initialize the soldier positions, making the distribution of soldiers more uniform. Its mathematical expression is as follows: Equation 28: y(t + 1) = 4y(t) 3 - 3y(t); Assume that there are m individuals in total, the first individual is randomly generated, and then the mapping of the remaining m-1 individuals is iterated according to Formula 22. On this basis, the obtained cubic chaotic operator is mapped to the whole, which is defined as follows; Equation 29: X t = X lb + (X ub - X lb ) * y(t + 1) * 0.5; X in the formula ub and X lb They represent the upper and lower bounds of the search space respectively.

6. A battery health estimation method considering adaptive characterization of degradation information according to claim 5, characterized in that: Step 3.2 also includes: An improved nonlinear convergence function is proposed, which is defined as follows; Formula 30: In the formula, r2 is a random number between 0 and 1, and d r represents the attenuation factor. According to the proposed nonlinear convergence function, Equation 21 and Equation 25 are redefined as Equation 31 and Equation 32; Formula 31: X i (t+1)=X i (t)+2r1(CK)*f o +r1(W i KX i (t)); Equation 32: X i (t + 1) = X i (t) + 2ρ(K - X r (t)) * f o + r1W i (c - X i (t)).

7. A battery health estimation method considering adaptive characterization of degradation information according to claim 6, characterized in that: Step 3.2 includes: Step 3.2.1: Use the cubic chaos mapping method to initialize the soldier population, including the soldier population U and the maximum number of iterations T m , and the upper bound U of the hyperparameters of VAE and improved LSTM neural network b and the lower bound L b , the hyperparameter information of the neural network is contained in the soldier population U, and its expression is as follows; Equation 33: U = [lr i , h zi , h vaei , h li ; In the formula, lr i represents the learning rate of the i-th soldier, h zi 、h vaei and h li are the number of VAE encoder neurons, VAE decoder neurons, and improved LSTM neurons of the i-th soldier respectively; Step 3.2.2: Calculate the fitness function, using the coefficient of determination R of the training set 2 As a fitness function, the information of the best soldiers is retained; Step 3.2.3: Update the soldier's position according to equations 22, 23, 26, 27, 31 and 32; Step 3.2.4: Calculate the fitness value. If the previous fitness value is worse than the current one, update the parameters; otherwise, keep the previous information. Step 3.2.5: If the maximum iteration is reached or the predefined early stopping criterion is met, exit the loop and output the optimal hyperparameter space.

Images