Lithium ion battery health state evaluation method based on LSTM (Long Short Term Memory) and Bayesian uncertainty quantification

Abstract

The invention discloses a lithium ion battery health state evaluation method based on LSTM (Long Short Term Memory) and Bayesian uncertainty quantification, and relates to a lithium ion battery health state evaluation method. The invention aims to solve the problem that a prediction result lacks statistical reliability due to the fact that the prior art focuses on single LSTM or CNN structure optimization and an uncertainty quantization mechanism is not systematically integrated. According to the method, the time sequence modeling capability of the LSTM is combined with the Bayesian uncertainty quantization framework, so that the cognitive uncertainty and the accidental uncertainty predicted by the model can be quantified while the time sequence dependence in the battery degradation process is captured, and the SOH estimation precision and reliability are improved. The invention belongs to the technical field of lithium ion battery health state prediction.

Description

Technical Field

[0001] The present invention relates to a method for evaluating the state of health of a lithium-ion battery, belonging to the technical field of predicting the state of health of lithium-ion batteries. Background Art

[0002] Lithium-ion batteries are widely used due to their excellent energy density and long cycle life. However, during continuous charge-discharge cycles, irreversible electrochemical reactions occur inside the battery, leading to performance degradation, mainly manifested as losses in available capacity and power output. The state of health of a lithium-ion battery is usually defined as the ratio of the current capacity to the initial capacity, which is a key indicator for measuring the degree of battery degradation. The degradation of SOH not only affects the operating efficiency of satellite battery systems but also may pose safety hazards. Therefore, it is crucial to monitor and accurately estimate the SOH of lithium-ion batteries in real time.

[0003] In recent years, data-driven methods based on machine learning have become effective means for estimating the SOH of lithium-ion batteries. For example, methods such as Gaussian process regression (GPR), convolutional neural network (CNN), and long short-term memory network (LSTM) have been widely used due to their good temporal modeling capabilities. However, due to data and technology limitations, real-world prediction problems inevitably face uncertainties. However, most existing SOH estimation methods ignore two important uncertainty factors: one is the aleatoric uncertainty caused by noise pollution during data collection and transmission; the other is the epistemic uncertainty caused by the structure of the model itself and the assumption bias during the training process. Traditional methods often assume that parameters and models are completely deterministic, but in practical applications, the model cannot perfectly capture the complex battery degradation process and external environmental factors, so its prediction results may be affected by uncertainties, reducing the accuracy and reliability of SOH estimation. Most existing patented technologies focus on optimizing a single LSTM or CNN structure and do not systematically integrate an uncertainty quantification mechanism, resulting in the lack of statistical reliability of prediction results. To address this problem, this paper proposes a method for estimating the state of health of lithium-ion batteries based on the combination of LSTM and Bayesian uncertainty quantification. By introducing the Bayesian uncertainty quantification framework, the prediction bias caused by data noise and model structure uncertainty can be effectively handled, thereby improving the accuracy and reliability of SOH estimation. Summary of the Invention

[0004] The present invention aims to solve the problem that most existing technologies focus on optimizing a single LSTM or CNN structure and do not systematically integrate an uncertainty quantification mechanism, resulting in the lack of statistical reliability of prediction results. Furthermore, a method for evaluating the state of health of a lithium-ion battery based on LSTM and Bayesian uncertainty quantification is proposed.

[0005] The technical solution adopted by the present invention to solve the above problems is as follows: The steps of the present invention include:

[0006] Step 1, data acquisition: Conduct lithium-ion battery charge and discharge experiments to collect battery aging data;

[0007] Step 2, feature extraction, generating a training set and a test set;

[0008] Step 3, building an LSTM Bayesian neural network model;

[0009] Step 4, uncertainty quantification and SOH prediction.

[0010] Furthermore, in Step 2, the terminal voltage and current data are used for feature extraction to obtain battery degradation features as the model input X, and the SOH data as the model output Y. The data set is divided into a training set and a test set according to 8:2.

[0011] Furthermore, the LSTM Bayesian neural network model built in Step 3 includes:

[0012] Input layer: Used to receive the battery degradation feature vector;

[0013] LSTM layer: Used to capture the long-term dependencies in the battery degradation process. Perform Bayesian parameterization, replacing the deterministic weight matrix of the traditional LSTM with random variables subject to a Gaussian distribution, that is, each weight Optimize the mean μ and variance σ through variational inference 2 ;

[0014] Monte Carlo Dropout layer: Used to add a Dropout layer after the LSTM layer. Neurons are randomly masked with a certain probability during both the training and inference phases, which is equivalent to sampling the posterior distribution of the parameters in the Bayesian network; through multiple forward propagations, a prediction distribution is generated to quantify the epistemic uncertainty;

[0015] Fully connected output layer: Output the estimated value μ of the current SOH and the output log variance logσ 2 , and obtain the aleatoric uncertainty σ through an exponential transformation 2 .

[0016] Furthermore, the uncertainty quantification and SOH prediction in Step 4 specifically include: Performing N forward propagations on the same test sample to obtain a prediction set Calculating the final prediction result;

[0017]

[0018] In formulas (1) and (2), T represents, and t represents;

[0019] Then according to Calculate the 95% confidence interval as a reliability measure for SOH prediction.

[0020] The beneficial effects of the present invention are as follows:

[0021] 1. The present invention realizes a probabilistic time-series modeling architecture. For the first time, the present invention deeply integrates the Bayesian probability framework with LSTM, and replaces the traditional deterministic parameters with a stochastic weight distribution (Gaussian distribution) to approximate the posterior distribution of the model parameters, fundamentally solving the defect that traditional LSTM cannot quantify cognitive uncertainty.

[0022] 2. The present invention realizes lightweight uncertainty propagation. Based on the multiple forward propagation mechanism of Monte Carlo Dropout, only by increasing the number of inferences can a prediction distribution be generated without introducing additional network parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a flowchart of the present invention;

[0024] Figure 2 is a schematic diagram of the Bayesian LSTM network architecture;

[0025] Figure 3 is a graph of battery degradation characteristics and SOH;

[0026] Figure 4 is a schematic diagram of the battery SOH prediction result. DETAILED DESCRIPTION OF THE INVENTION

[0027] DETAILED DESCRIPTION OF THE INVENTION I: As Figures 1 to 4 shown, a method for evaluating the state of health of a lithium-ion battery based on LSTM and Bayesian uncertainty quantification, the steps include:

[0028] Step 1. Data collection:

[0029] Conduct a charge and discharge experiment on a lithium-ion battery to collect battery aging data; the aging process of each lithium-ion battery will experience different battery life cycles. During each cycle, collect the terminal voltage, current, and actual capacity C of the battery 实际 , given the nominal capacity C of the battery 标称 , determine the state of health value of the lithium-ion battery based on the actual capacity and the nominal capacity of the lithium-ion battery;

[0030]

[0031] The lithium-ion battery includes, but is not limited to, a lithium iron phosphate (LFP) system battery, a lithium cobaltate (LCP) system battery, a lithium manganate (LMP) system battery, or a ternary system battery;

[0032] Step 2. Feature extraction

[0033] After data collection, the next step is to preprocess the system degradation data, which specifically includes steps such as degradation feature extraction, feature screening, feature dimensionality reduction, and data normalization; to facilitate subsequent modeling, effective features are first extracted from the collected data such as battery voltage and current; commonly used features include battery charge and discharge efficiency, internal resistance, capacity attenuation, etc.; feature screening and dimensionality reduction operations help to remove redundant information, reduce the computational burden, and improve the accuracy and generalization ability of the model;

[0034] Through data normalization processing, feature data with different dimensions can be mapped to the same scale to avoid biases during model training due to different ranges of feature values; finally, the features and SOH curves can be used as input-output pairs for subsequent modeling;

[0035] Step 3: Build an LSTM Bayesian neural network model;

[0036] Step 3-1: Input layer: Receive the battery degradation feature vector;

[0037] This layer receives the feature vector x t = [x 1t , x 2t , …, x nt during the battery degradation process, where n is the number of features and t is the time step;

[0038] Step 3-2: LSTM layer: Used to capture long-term dependencies during the battery degradation process. The number of hidden units in this layer is 128, and the activation function is tanh; perform Bayesian parameterization, replacing the deterministic weight matrix of the traditional LSTM with random variables subject to a Gaussian distribution, that is, each weight Optimize the mean μ and variance σ through variational inference 2 ;

[0039] h t = LSTM(h t-1 , x t ),

[0040] where h t is the hidden state at time step t, h t-1 is the hidden state of the previous time step, and x t is the input feature at the current time step;

[0041] Step 3-3: Add a Dropout layer after the LSTM layer. During both the training and inference phases, neurons are randomly masked with a probability p = 0.2, which is equivalent to sampling from the posterior distribution of the parameters in the Bayesian network; through multiple forward propagations, multiple prediction results are obtained, generating a prediction distribution to quantify the epistemic uncertainty;

[0042] Step 3-4: The fully connected output layer outputs the estimated value μ of the current SOH and the output log variance logσ 2 (To avoid the negative value problem), obtain the aleatoric uncertainty σ through exponential transformation 2 ;

[0043] Step 4: Uncertainty quantification and SOH prediction;

[0044] Perform forward propagation N times on the same test sample to obtain a prediction set Calculate the final prediction result:

[0045]

[0046] Then according to Calculate the 95% confidence interval as a reliability measure for SOH prediction;

[0047] Thus, the predicted value of the lithium-ion battery and its 95% confidence interval are as Figure 4 shown

[0048] Finally, through the Bayesian uncertainty quantification method, we can not only obtain the estimated value of the battery SOH, but also provide a reliability measure for this estimated value, that is, the confidence interval of the prediction result; in this way, not only the accuracy of SOH prediction is improved, but also more reference basis can be provided for subsequent battery management decisions.

[0049] Working principle

[0050] In the present invention, first, the LSTM network is used to process the time-series data of battery degradation, and the uncertainty of the model parameters of the LSTM is quantified through the Bayesian framework; the traditional LSTM model usually uses a deterministic weight matrix, while in the present invention, the weight matrix is transformed into a random variable obeying a Gaussian distribution, and its mean and variance are optimized through variational inference to approximate the posterior distribution of the model parameters; the present invention can effectively handle the epistemic uncertainty in the model, making the SOH prediction result more reliable.

[0051] In addition, in order to further improve the effect of uncertainty quantification, the present invention also adopts the Monte Carlo Dropout method; after the LSTM layer, a Dropout layer is added, and multiple prediction results are generated through multiple forward propagations to quantify the epistemic uncertainty; by calculating the distribution of the prediction results and the 95% confidence interval, a reliable confidence measure is provided for SOH estimation to help better evaluate the health state of the battery in practical applications.

[0052] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Although the present invention has been disclosed above with the preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art, within the scope of the technical solution of the present invention, can make some changes or modifications to the above-disclosed technical content to obtain equivalent embodiments with equivalent changes. However, as long as it does not depart from the technical solution content of the present invention, and based on the technical essence of the present invention, any simple modification, equivalent replacement, and improvement made to the above embodiments within the spirit and principle of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A lithium-ion battery health status assessment method based on LSTM and Bayesian uncertainty quantification, characterized in that: The specific steps include: Step 1, data collection: conduct lithium-ion battery charge and discharge experiments and collect battery aging data; Step 2: Feature extraction, generating training set and test set; Step 3: Build an LSTM Bayesian neural network model; Step 4: Uncertainty quantification and SOH prediction.

2. The lithium-ion battery health status assessment method based on LSTM and Bayesian uncertainty quantification according to claim 1 is characterized in that: In step 2, the terminal voltage and current data are used for feature extraction to obtain the battery degradation features as the model input X, and the SOH data is used as the model output Y. The data set is divided into a training set and a test set according to an 8:2 ratio.

3. The lithium-ion battery health status assessment method based on LSTM and Bayesian uncertainty quantification according to claim 1, characterized in that: The LSTM Bayesian neural network model built in step 3 includes: Input layer: used to receive the battery degradation feature vector; LSTM layer: used to capture the long-term dependencies in the battery degradation process. Bayesian parameterization is performed to replace the deterministic weight matrix of the traditional LSTM with a random variable that obeys a Gaussian distribution, that is, each weight Optimize mean μ and variance σ through variational inference 2 ; Monte Carlo Dropout layer: used to add a Dropout layer after the LSTM layer. In both the training and inference stages, neurons are randomly masked with a certain probability, which is equivalent to sampling the posterior distribution of parameters in the Bayesian network. Through multiple forward propagations, a predictive distribution is generated to quantify epistemic uncertainty. Fully connected output layer: outputs the estimated value μ of the current SOH and the output logarithmic variance logσ 2 , the random uncertainty σ is obtained by exponential transformation 2 .

4. The lithium-ion battery health status assessment method based on LSTM and Bayesian uncertainty quantification according to claim 1, characterized in that: The uncertainty quantification and SOH prediction in step 4 specifically include: performing N forward propagations on the same test sample to obtain a prediction set Calculate the final prediction result; In formulas (1) and (2), T represents, t represents; According to The 95% confidence interval was calculated as a measure of the reliability of the SOH prediction.

Images