A method for filling missing observation data of lithium batteries based on variational autoencoder

Abstract

The present invention discloses a method for filling missing observation data of lithium batteries based on variational autoencoders, the steps of which include: S1, obtaining data observed within a normal period; S2, labeling a data set of normal observation data to determine the period of missing data; S3, constructing a variational autoencoder; S4, using normal data as input to train the variational autoencoder; S5, using an encoder model to obtain a latent vector of normal data; S6, selecting any two-dimensional data of the latent vector, constructing a latent variable coordinate system, and observing the distribution law of the latent vector to determine whether the constructed variational autoencoder meets the requirements; S7, linearly interpolating the latent vector of the missing data according to the latent vector of the normal data; S8, inputting the latent vector corresponding to the missing period into the decoder model to reconstruct the missing data. The method of the present invention is of great significance for improving the accuracy and generalization of lithium-ion battery life estimation.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithium batteries and relates to a method for filling missing observation data of lithium batteries based on a variational autoencoder. Background Art

[0002] Lithium battery is the abbreviation of lithium-ion battery. It has high energy density, long life and good stability. It has become an energy storage tool for major green transportation tools such as electric vehicles. However, the use of lithium batteries requires necessary monitoring and management. The battery management system (BMS) needs to continuously estimate the calendar life of the battery based on the parameters captured by the sensor during the use of the battery, and use this to predict the health state (State-of-Health, SOH) of the lithium battery to ensure the safety and reliability of the system driven by the lithium battery, help improve the charging and discharging strategy, avoid battery abuse, guide scientific battery management methods, and increase the battery life for a longer period of time. In addition, the prediction of the health state of lithium-ion batteries can also provide the most direct reference for battery recycling and cascade utilization.

[0003] At present, the data-driven method for predicting the SOH of lithium batteries has the characteristics of high prediction accuracy and strong generalization, and can quickly complete parameter adjustment under new working conditions. It has become the mainstream method for predicting SOH in the current BMS system. However, the essence of the data-driven method is to learn the mapping of voltage, current, impedance and other parameters to SOH during the battery charging and discharging process offline through the algorithm, and then adjust and use it online. Therefore, the learning data used to build the model is particularly important. For the key parameters of a specific type of battery, it is generally obtained by repeatedly charging and discharging the battery in the laboratory. Data acquisition is difficult, and for a battery throughout its life cycle, the data acquisition time is long. During the battery charging and discharging experiment, due to the instability of the experimental environment or noise interference, it is very easy to cause data missing or contamination in a certain charging cycle of the battery. Therefore, it is particularly important to complete the missing or contaminated cycle data based on similar charging and discharging parameters in adjacent cycles. Complete data can directly improve the prediction accuracy of the data-driven method and reduce the impact of missing or noisy data on the training process. In addition, with the help of data generation methods, potentially unexplored and meaningful training samples can be provided to improve the generalization of the SOH estimation model.

[0004] In summary, the performance of lithium-ion battery life estimation based on data-driven methods is excellent. However, the cost of obtaining lithium-ion battery aging data is relatively expensive. Completing lost or contaminated data can not only significantly reduce the cost of data acquisition, but also provide necessary support for improving SOH estimation performance. Summary of the invention

[0005] The purpose of the present invention is to provide a method for filling missing observation data of lithium batteries based on variational autoencoders, which solves the problem in the prior art that the cost of obtaining lithium-ion battery aging data is relatively expensive and the missing observation data of lithium batteries is difficult to fill.

[0006] The technical solution adopted by the present invention is a method for filling missing observation data of lithium batteries based on a variational autoencoder, which is specifically implemented according to the following steps:

[0007] S1. Obtain the data observed during a normal period;

[0008] S2. Label the data set of normal observation data and determine the period of missing data;

[0009] S3, construct variational autoencoder;

[0010] S4, using normal data as input to train the variational autoencoder;

[0011] S5. Use the encoder model to obtain the latent vector of normal data

[0012] S6. Selecting latent vectors , construct a latent variable coordinate system, and observe the distribution law of the latent vector to determine whether the constructed variational autoencoder meets the requirements;

[0013] S7. Latent vector according to normal data Linearly interpolate the missing data's latent vector Z u ;

[0014] S8. The latent vector Z corresponding to the missing period u Input to the decoder model P θ (Z u |X), reconstruct the missing data

[0015] The beneficial effect of the present invention is that a variational autoencoder is constructed only through unsupervised learning of normal charge and discharge data; a latent vector of missing data is constructed by interpolating the latent vector output by the encoder; and the missing data is reconstructed using the constructed decoder. The present invention can directly reconstruct the missing data from the normal data in the normal lithium-ion battery charge and discharge process, which is of great significance to improving the accuracy and generalization of lithium-ion battery life estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a simplified flow chart of the method of the present invention;

[0017] Figure 2 is a simplified diagram for marking missing data in the method of the present invention;

[0018] Figure 3 It is a simplified structural diagram of the variational autoencoder in the method of the present invention;

[0019] Figure 4 It is a simplified structural diagram of the encoder model in the method of the present invention;

[0020] Figure 5 It is a simplified structural diagram of the decoder model in the method of the present invention. DETAILED DESCRIPTION

[0021] The present invention is based on the variational autoencoder lithium battery missing observation data filling method, the whole process is referred to Figure 1 , follow the steps below to implement it:

[0022] S1. Obtain the data observed during the normal period.

[0023] The specific process is: in the experimental environment, a specific battery is tested for cyclic charge and discharge, and the normal observation data of each sensor is recorded during the test to obtain the corresponding data set. Where c∈{1…n}, n is the number of charging cycles; k represents the parameter type, k∈{V,I,T}, where V represents voltage data, I represents current data, and T represents temperature data;

[0024] S2. Label the data set of normal observation data and determine the period of missing data.

[0025] Reference Figure 2 The specific process is: Screening, according to experimental records and data rules, artificially determine the missing data period The rest of the data is classified into the normal data set Recorded as normal data set

[0026] S3. Construct a variational autoencoder.

[0027] The overall structure of the variational autoencoder is shown in Figure 3 , the specific process is:

[0028] S3-1. Construct the encoder model P in the variational autoencoder φ (Z|X), the encoder model is mainly composed of convolutional layer, Flatten layer, and fully connected layer, see Figure 4 The output of the encoder model is to fit the mean μ and variance log(σ) of the latent vector through the fully connected layer, and then reparameterize it to get the latent vector Z∈{Z1,Z2,…,Z n};

[0029] S3-2. Construct the decoder model P in the variational autoencoder θ (X|Z), the decoder model is mainly composed of a fully connected layer, an Unflatten layer, and a deconvolution layer. Figure 5 , used to reconstruct the potential vector Z established in step S3-1 into the original data

[0030] S4. Use normal data as input to train the variational autoencoder.

[0031] The specific process is: set an appropriate data batch size b, use the normal data set confirmed in step S2 The variational autoencoder constructed in step S3 is trained; the total loss function used in training is the reconstruction loss function and KL divergence The accumulation expression is as follows:

[0032]

[0033] Among them, the expression of KL divergence is:

[0034]

[0035] Among them, J represents the number of latent vectors, σ j represents the variance, μ j Represents the mean.

[0036] Reconstruction loss function The expression is:

[0037]

[0038] Among them, x i is the training sample, is the reconstruction value of the variational autoencoder.

[0039] S5. Use the encoder model to obtain the latent vector of normal data

[0040] The specific process is as follows: After completing the training in step S4, use the normal data set Input the trained encoder P φ (Z|X), get the potential feature vector of normal samples

[0041] S6. Selecting latent vectors , construct a potential variable coordinate system and observe the distribution law to determine whether the constructed variational autoencoder meets the requirements.

[0042] The specific process is: Select the potential feature vector Any two dimensions and p1 and p2 are the two selected dimensions, and a latent variable coordinate system is constructed and a distribution diagram of the two-dimensional latent feature vector is drawn; if the distribution diagrams show strong clustering characteristics, it means that the established variational autoencoder meets the reconstruction requirements; otherwise, return to step S4 for retraining.

[0043] S7, in the latent feature vector In the space formed by linear interpolation, the latent vector Z of missing data u ;

[0044] The specific process is as follows: for a single missing period t', the corresponding neighborhood periods within the range of the normal data set are t'-1 and t'+1; the potential feature vector obtained from step S6 In the example above, we obtain the potential vector corresponding to the normal data. and

[0045] Then use linear interpolation to obtain the potential vector of missing data Where interp represents the linear interpolation operation.

[0046] S8. The latent vector Z corresponding to the missing period u Input to the decoder model P θ (Z u |X), reconstruct the missing data

[0047] The interpolated potential vector Z obtained in step S7 is u , input into the constructed variational autoencoder decoder model to obtain the reconstructed missing data Then use the Gaussian filter to reconstruct the missing data Perform smoothing filtering to obtain the final reconstructed data and complete data filling.

Claims

1. A method for filling missing observation data of lithium batteries based on variational autoencoder, characterized in that: Follow these steps to implement it: S1. Obtain the data observed during a normal period; S2. Label the data set of normal observation data and determine the period of missing data; S3. Construct a variational autoencoder. The specific process is as follows: S3-1. Constructing the encoder model in the variational autoencoder The encoder model consists of a convolutional layer, a flatten layer, and a fully connected layer. The output of the encoder model is the mean of the latent vector fitted by the fully connected layer. and variance , after reparameterization, we get the latent vector ; S3-2. Constructing the decoder model in the variational autoencoder The decoder model consists of a fully connected layer, an Unflatten layer, and a deconvolution layer, which is used to transform the potential vector established in step S3-1 Reconstruct the original data ; S4. Use normal data as input to train the variational autoencoder. The specific process is as follows: Set an appropriate batch size , using the normal data set confirmed in step S2 , train the variational autoencoder constructed in step S3; The total loss function used during training is the reconstruction loss function and KL divergence The accumulation expression is as follows: （1） Among them, the expression of KL divergence is: （2） in, represents the number of latent vectors, represents the variance, Mean Reconstruction loss function The expression is: （3） in, is the training sample, is the reconstruction value of the variational autoencoder; S5. Use the encoder model to obtain the latent vector of normal data ; S6. Selecting latent vectors , construct a latent variable coordinate system, and observe the distribution law of the latent vector to determine whether the constructed variational autoencoder meets the requirements; S7. Latent vector according to normal data , the latent vector of the missing data is linearly interpolated ; S8. The latent vector corresponding to the missing cycle Input to the decoder model Reconstruct missing data .

2. The method for filling missing observation data of lithium batteries based on variational autoencoder according to claim 1 is characterized in that: The specific process of step S1 is as follows: In the experimental environment, a specific battery is subjected to a cycle charge and discharge test. During the test, the normal observation data of each sensor is recorded to obtain the corresponding data set. , among which , is the number of charging cycles; Indicates the parameter type, , among which Indicates voltage data, Indicates current data, Represents temperature data.

3. The method for filling missing observation data of lithium batteries based on variational autoencoder according to claim 1 is characterized in that: The specific process of step S2 is: Screening, according to experimental records and data rules, artificially determine the missing data period , the rest of the data is classified into the normal data set , recorded as a normal data set .

4. The method for filling missing observation data of lithium batteries based on variational autoencoder according to claim 1 is characterized in that: The specific process of step S5 is as follows: after completing the training in step S4, use the normal data set , input the trained encoder , get the potential feature vector of normal samples .

5. The method for filling missing observation data of lithium batteries based on variational autoencoder according to claim 1 is characterized in that: The specific process of step S6 is: select potential feature vector Any two dimensions and , For the selected two dimensions, construct a latent variable coordinate system and draw a distribution diagram of the two-dimensional latent feature vector; If both distribution graphs show strong clustering characteristics, it means that the established variational autoencoder meets the reconstruction requirements; Otherwise, return to step S4 for retraining.

6. The method for filling missing observation data of lithium batteries based on variational autoencoder according to claim 1 is characterized in that: The specific process of step S7 is as follows: for a single missing cycle , the corresponding neighborhood period within the range of the normal data set is and ; The potential feature vector obtained from step S6 In the example above, we obtain the potential vector corresponding to the normal data. and ; Then use linear interpolation to obtain the potential vector of missing data ,in Represents a linear interpolation operation.

7. The method for filling missing observation data of lithium batteries based on variational autoencoder according to claim 1 is characterized in that: The specific process of step S8 is: the potential vector interpolated from step S7 is , input into the decoder model of the variational autoencoder constructed in step S5 to obtain the reconstructed missing data ; Then use the Gaussian filter to reconstruct the missing data Perform smoothing filtering to obtain the final reconstructed data and complete data filling.

Images