A method for lithium battery state of health estimation based on uncertainty quantification

Abstract

The application provides a lithium battery health state estimation method based on uncertainty quantification, and belongs to the field of lithium battery health state prediction, and comprises the following steps: in step S1, voltage, time and temperature data of a lithium battery are acquired in a lithium battery charging and discharging cycle; in step S2, multiple types of features are extracted according to the voltage, time and temperature data, and an optimal feature combination is selected to form a training data set; in step S3, a Bayesian neural network that fuses cognitive uncertainty and accidental uncertainty is built; and in step S4, the training data set obtained in step S2 is used to train the fused model. The cognitive uncertainty and the accidental uncertainty are comprehensively considered, and the accuracy and reliability of SOH estimation are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of lithium battery state of health prediction, and particularly relates to a lithium battery state of health estimation method based on uncertainty quantification. BACKGROUND

[0002] With the development of machine learning and deep learning, various data-driven machine learning techniques such as Gaussian process regression (GPR) and transfer learning (TL) have emerged to achieve accurate SOH (state of health of lithium battery) estimation. However, due to the limitations of data and technology, real-world prediction problems (either data-driven or model-driven) inevitably face two types of uncertainty: accidental uncertainty reflecting noise pollution in data collection and transmission and cognitive uncertainty reflecting the ignorance of the model itself. Current SOH estimation methods often ignore the uncertainty in model structure and data acquisition process, thereby affecting the accuracy and reliability of SOH estimation. SUMMARY

[0003] The purpose of the present application is to provide a lithium battery state of health estimation method based on uncertainty quantification, which comprehensively considers cognitive uncertainty and accidental uncertainty to improve the accuracy and reliability of SOH estimation.

[0004] The present application is achieved by the following technical solutions:

[0005] A lithium battery state of health estimation method based on uncertainty quantification, comprising the following steps:

[0006] Step S1, in the lithium battery charging and discharging cycle, obtaining voltage, time and temperature data of the lithium battery;

[0007] Step S2, extracting multiple types of features from the voltage, time and temperature data, and selecting the optimal feature combination to form a training data set;

[0008] Step S3, building a Bayesian neural network model that fuses two types of quantified uncertainty, specifically including:

[0009] Step S31, placing a prior distribution W~N(0,1) on the weights of the neural network, performing MC dropout training on the neural network, and obtaining a Bayesian neural network model;

[0010] Step S32, quantifying cognitive uncertainty: in MC dropout training, the parameter distribution of the neural network is sampled H times to obtain H different models, and the H models are used to predict the same sample respectively, and the cognitive uncertainty is quantified as where, the inherent noise of the input data, i.e. the variance of the input data, denotes the predicted mean value, denotes the model parameters obtained at the h-th sampling, x denotes the sample used for prediction, which is obtained from the training data set obtained in step S2, denotes the predicted value obtained at the h-th prediction, H is an integer, 1≤h≤H;

[0011] Step S33, quantifying the aleatoric uncertainty: using a loss function establishing an aleatoric uncertainty model, θ denotes the parameters of the aleatoric uncertainty model, N denotes the number of training samples used for the model, y i denotes the i-th true value, x i denotes the i-th training sample, which is obtained from the training data set obtained in step S2, denotes the variance of the input data;

[0012] Step S34, fusing the cognitive uncertainty and the quantified uncertainty into the Bayesian neural network model obtained in step S31, using a loss function training the fused model, wherein, , K denotes the number of training samples used for the model, which are obtained from the training data set obtained in step S2, denotes the variance of the output of the model, denotes the predicted value, denotes the true value;

[0013] Step S4, training the fused model using the training data set obtained in step S2.

[0014] Further, the step S2 specifically comprises:

[0015] Step S21, extracting multi-class features according to the voltage, time and temperature data obtained in step S1;

[0016] Step S22, calculating the correlation coefficient between the M-th class feature and the SOH estimation value using the formula wherein, is the l-th data point of the extracted M-th class feature, is the actual SOH value of the lithium battery corresponding to the M-th class feature, is the mean value of the L data points, is the mean value of the L actual SOH values;

[0017] Step S23, calculating the correlation coefficient between the M-th class feature and other class features using the formula wherein, ​​a decision coefficient of a multiple linear regression obtained when predicting the Mth class feature using other class features;

[0018] In step S24, features whose correlation coefficients with the SOH estimation value are greater than a set first threshold value and whose correlation coefficients with other class features are less than a set second threshold value are obtained, and at least two class features are selected from the features as the optimal feature combination.

[0019] Further, the first threshold value is set to 0.8, and the second threshold value is set to 10.

[0020] Further, in step S31, when performing MC dropout training, the minimization target is wherein, , is a parameter distribution to be optimized, p is a random inactivation rate in MC dropout, is the number of training samples, is the parameter obtained by the Mth sampling, is the (i)th training sample, the training sample is obtained from the training data set obtained in step S2, is the (i)th true value, represents the variance of the input data.

[0021] Further, the optimal feature combination is the peak data of the capacity increment curve and the left peak data of the temperature change rate curve.

[0022] Further, in step S21, the battery capacity increment curve and the temperature change rate curve of the lithium battery are obtained according to the voltage, time and temperature data, the extreme points, slopes and mean values of each battery capacity increment curve and each temperature change rate curve are extracted, and the multiple class features are obtained.

[0023] Further, the SOH is defined by the formula wherein, is the current battery capacity data, is the maximum capacity data in this charge and discharge cycle.

[0024] Further, the charge and discharge cycle is a charge and discharge cycle of the first 60% of the lithium battery capacity.

[0025] The present application has the following beneficial effects:

[0026] ​​​1. This invention first obtains the voltage, time, and temperature data of lithium batteries, then extracts multiple features based on these data and selects the optimal feature combination to form a training dataset. Next, it integrates quantified cognitive uncertainty and random uncertainty into a Bayesian network model to obtain a fused model. Finally, it uses the training dataset to train the fused model to obtain a prediction model, thereby comprehensively considering cognitive uncertainty and random uncertainty to effectively improve the accuracy and reliability of SOH estimation. Attached Figure Description

[0027] The present invention will now be described in further detail with reference to the accompanying drawings.

[0028] Figure 1 This is a flowchart of the present invention.

[0029] Figure 2 This is a comparison chart showing the effectiveness of the present invention with existing prediction methods.

[0030] Figure 3 This is a comparison chart of MAE between the present invention and existing prediction methods.

[0031] Figure 4 This is a comparison chart of the RMSE of the present invention and existing prediction methods. Detailed Implementation

[0032] like Figure 1 As shown, the method for estimating the state of health of lithium batteries based on uncertainty quantification includes the following steps:

[0033] Step S1: During the lithium battery charge-discharge cycle, acquire the voltage, time, and temperature data of the lithium battery;

[0034] Specifically, the above data is obtained through the battery management system during the first 60% charge-discharge cycle of the lithium battery, and this data is tabulated to observe the changing trends.

[0035] Step S2: Extract multiple features based on voltage, time, and temperature data, and select the optimal feature combination to form a training dataset;

[0036] Specifically, it includes:

[0037] Step S21: Extract multiple features based on the voltage, time, and temperature data obtained in step S1;

[0038] Specifically, based on voltage, time, and temperature data, the battery capacity increment curve (horizontal axis: voltage, vertical axis: battery capacity IC) and temperature change rate curve (horizontal axis: voltage, vertical axis: temperature change DT) of the lithium battery are obtained. The extreme points, slopes, and mean values ​​of each battery capacity increment curve and temperature change rate curve are extracted to obtain multiple features. The formula for curve IC is as follows: , Q is the battery charge capacity, unit mAh, V is the battery voltage, unit V, AQ is the battery charge capacity change in a small interval; AV is the battery voltage change in the interval; the curve DT formula is , T is the temperature; L is the sampling interval 40; the specific process of this part is prior art;

[0039] Step S22, the correlation coefficient between each type of feature and the SOH estimation value is calculated by using Pearson correlation analysis method, that is, according to the formula The correlation coefficient between the Mth feature and the SOH estimation value is calculated , wherein, is the lth data point of the Mth feature extracted, is the actual SOH value of the lithium battery corresponding to the Mth feature, is the mean value of the L data points, is the mean value of the L actual SOH values, and M is an integer;

[0040] Step S23, the correlation coefficient between each type of feature is calculated by using the variance inflation factor, that is, according to the formula The correlation coefficient between the Mth feature and other features is calculated , wherein, is the determination coefficient of the multiple linear regression obtained when the Mth feature is predicted by using other features;

[0041] In this example, the SOH is defined by the formula , wherein, is the current battery capacity data, is the maximum capacity data in this charge and discharge cycle;

[0042] Step S24, the features whose correlation coefficient with the SOH estimation value is greater than a set first threshold value and whose correlation coefficient with other features is less than a set second threshold value are obtained, and at least two types of features are selected from these features as the optimal feature combination;

[0043] In this embodiment, the first threshold value is set to 0.8, and the second threshold value is set to 10. The peak value data of the capacity increment curve and the left peak value data of the temperature change rate curve are selected as the optimal feature combination, and the correlation between the two types of features and the SOH is high and the correlation between them is low.

[0044] Step S3, a Bayesian neural network model integrating cognitive uncertainty and accidental uncertainty is built, specifically including:

[0045] Step S31, a prior distribution W~N(0, 1) is placed on the weight of the neural network, the neural network is trained by MC dropout, and a Bayesian neural network model is obtained.

[0046] MCdropout is a variational Bayesian approximation where the approximate distribution is a mixture of two Gaussian distributions with small variances, one of which has a fixed mean of zero.

[0047] When training with MCdropout, the minimization objective is ,in, The negative log-likelihood. For regularization parameters, Let p be the parameter distribution to be optimized (i.e., the weight matrix in the case of school dropout), and let p be the random inactivation rate in MCdropout. The number of training samples. For the first The parameters obtained from the second sampling. For the first The training samples are obtained from the training dataset obtained in step S2. For the first Given a set of true values, the negative likelihood in regression can be further simplified to: , The variance of the input data (training samples) represents the inherent noise, i.e., the variance of the input data, and ∝ indicates a direct proportion.

[0048] Step S32, Quantifying Cognitive Uncertainty: In MC dropout training, the parameter distribution of the neural network is sampled H times to obtain H different models. These H models are then used to predict the same sample, and the cognitive uncertainty is quantified as follows: ,in, This represents the variance of the input data. This represents the mean of the predictions; this part measures the model's uncertainty about its predictions. The mean is zero when the parameter uncertainty is zero (i.e., when all models are plotted). When the same constant value is used, this term will disappear. Let represent the model parameters obtained from the h-th sampling, and x represent the sample used for prediction, which is obtained from the training dataset obtained in step S2. H represents the predicted value obtained from the h-th prediction, where H is an integer, 1 ≤ h ≤ H;

[0049] Step S33, Quantify random uncertainty: Use a loss function Establish a model of random uncertainty, the output of which is: , Let θ represent the predicted value, θ represent the parameters of the random uncertainty model, N represent the number of training samples used for the model, and y represent the predicted value. i Let x represent the i-th true value. idenotes the i-th training sample, which is obtained from the training data set obtained in step S2, denotes the square of the norm, denotes the variance of the input data of the model;

[0050] Step S34, fusing the cognitive uncertainty and the quantitative uncertainty into the Bayesian neural network model trained in step S31, two groups of outputs are needed, one is the final prediction mean , and the other is the prediction variance , that is , the fused model is trained by minimizing the loss function , wherein K denotes the number of training samples for the model, which is obtained from the training data set obtained in step S2, denotes the variance of the output of the model, denotes the k-th prediction value, denotes the k-th true value, this loss includes two parts: one part is the residual error of the regression model, which is used to capture the cognitive uncertainty of the model parameters; the other part is the accidental uncertainty of the data, which acts as a regularization term to prevent the network from predicting infinite uncertainty for all data points, resulting in a zero loss value, in the embodiment, the , (:= denotes definition) because it is more stable than the regression variance numerically, the loss avoids potential division by zero, then the loss function is transformed into ;

[0051] In this fused model, the prediction uncertainty of y can be approximated as , using a group of T sampling outputs , is a random mask weight , 1≤t≤T.

[0052] Step S4, training the fused model using the training data set obtained in step S2.

[0053] Figure 2 In the experiment, the present application (BNN+e+a) is compared with several existing SOH prediction models (LSTM, RNN, CNN) on the experimental data set, the abscissa is the cycle number, and the ordinate is the SOH, from the figure, it can be seen that the performance of the present application is more stable than other models both in the early and late training, because LSTM, RNN and CNN may have limitations in capturing the nonlinear relationship in the data, while the battery system contains numerous nonlinear factors, such as the nonlinear effect in the charging and discharging process, the present application is suitable for capturing these complex nonlinear relationships, thereby improving the prediction accuracy and stability.

[0054] Figure 3 and Figure 4 In the Oxford dataset, 8 batteries are used to verify the comparison of the present application and several existing SOH prediction models (LSTM, RNN, CNN). It can be seen that the MAE (mean absolute error) and RMSE (root mean square error) of the present application are the smallest, and the error is less than one percent, so the existing prediction model may be more difficult to maintain robustness when facing abnormal situations in the battery system, and the present application can better adapt and handle abnormal situations in the data by modeling two uncertainties, thereby improving the robustness of the model.

[0055] The above description is only a preferred embodiment of the present application, and therefore cannot limit the scope of the present application. Any equivalent changes and modifications made in accordance with the scope of the patent application and the content of the specification shall still be within the scope of the present application.

Claims

1. A method for estimating the state of health of a lithium battery based on uncertainty quantification, characterized in that: Includes the following steps: Step S1: During the lithium battery charge-discharge cycle, acquire the voltage, time, and temperature data of the lithium battery; Step S2: Extract multiple features based on voltage, time, and temperature data, and select the optimal feature combination to form a training dataset; Step S3: Construct a Bayesian neural network model that integrates cognitive uncertainty and random uncertainty, specifically including: Step S31: Place a prior distribution W~N(0,1) on the weights of the neural network, and train the neural network using MCdropout to obtain a Bayesian neural network model. Step S32, Quantifying Cognitive Uncertainty: In MC dropout training, the parameter distribution of the neural network is sampled H times to obtain H different models. These H models are then used to predict the same sample, and the cognitive uncertainty is quantified as follows: ,in, This represents the inherent noise of the input data, i.e., the variance of the input data. This represents the predicted mean. Let represent the model parameters obtained from the h-th sampling, and x represent the sample used for prediction, which is obtained from the training dataset obtained in step S2. H represents the predicted value obtained from the h-th prediction, where H is an integer, 1 ≤ h ≤ H; Step S33, Quantify random uncertainty: Use a loss function Establish a random uncertainty model, where θ represents the parameters of the random uncertainty model, N represents the number of training samples used for the model, and y i Let x represent the i-th true value. i This represents the i-th training sample, which is obtained from the training dataset obtained in step S2. Indicates the variance of the input data; Step S34: Integrate cognitive uncertainty and quantitative uncertainty into the Bayesian neural network model obtained in step S31, using the loss function... Training the fused model, where, K represents the number of training samples used for the model, which are obtained from the training dataset obtained in step S2. This represents the variance of the model's output. Indicates the predicted value. Represents the actual value; Step S4: Train the fused model using the training dataset obtained in step S2.

2. The method for estimating the state of health of a lithium battery based on uncertainty quantification according to claim 1, characterized in that: Step S2 specifically includes: Step S21: Extract multiple features based on the voltage, time, and temperature data described in step S1; Step S22: Using the formula Calculate the correlation coefficient between the M-th type feature and the SOH estimate. ,in, For the l-th data point of the extracted M-th feature, This represents the actual SOH value of the lithium battery corresponding to the Mth type of characteristic. Let L be the mean of the data points. The average of L actual SOH values; Step S23: Using the formula Calculate the correlation coefficient between the M-th class feature and features from other classes. ,in, The determination coefficients of the multiple linear regression obtained when using features from other classes to predict features from the Mth class; Step S24: Obtain features whose correlation coefficient with the SOH estimate is greater than a set first threshold and whose correlation coefficient with other types of features is less than a set second threshold, and select at least two types of features from these features as the optimal feature combination.

3. The method for estimating the state of health of a lithium battery based on uncertainty quantification according to claim 2, characterized in that: The first threshold is set to 0.8, and the second threshold is set to 10.

4. A method for estimating the state of health of a lithium battery based on uncertainty quantification according to claim 1, 2, or 3, characterized in that: In step S31, during MCdropout training, the minimization objective is... ,in, , Let p be the parameter distribution to be optimized, and p be the random kill rate in MCdropout. The number of training samples. For the first The parameters obtained from the second sampling. For the first The training samples are obtained from the training dataset obtained in step S2. For the first A true value, This represents the variance of the input data.

5. A method for estimating the state of health of a lithium battery based on uncertainty quantification according to claim 1, 2, or 3, characterized in that: The optimal combination of features is the peak data of the capacity increment curve and the left-side peak data of the temperature change rate curve.

6. A method for estimating the state of health of a lithium battery based on uncertainty quantification according to claim 2 or 3, characterized in that: Step S21 specifically involves: obtaining the battery capacity increment curve and temperature change rate curve of the lithium battery based on the voltage, time, and temperature data; extracting the extreme points, slopes, and mean values ​​of each battery capacity increment curve and each temperature change rate curve to obtain the multi-type features.

7. A method for estimating the state of health of a lithium battery based on uncertainty quantification according to claim 1, 2, or 3, characterized in that: Using formula Define SOH, where, This is the current battery capacity data. This is the maximum capacity data during this charge-discharge cycle.

8. A method for estimating the state of health of a lithium battery based on uncertainty quantification according to claim 1, 2, or 3, characterized in that: The charge-discharge cycle refers to the first 60% charge-discharge cycle of the lithium battery capacity.

Images