WO2022198616A1

WO2022198616A1 - Battery life prediction method and system, electronic device, and storage medium

Info

Publication number: WO2022198616A1
Application number: PCT/CN2021/083165
Authority: WO
Inventors: 申文静; 陈长; 周冀; 王红志; 吕启涛
Original assignee: 深圳技术大学
Priority date: 2021-03-26
Filing date: 2021-03-26
Publication date: 2022-09-29

Abstract

Disclosed are a battery life prediction method and system, an electronic device, and a storage medium. The method comprises: obtaining past data concerning battery capacity; preprocessing the past data to obtain principal component data and secondary component data of battery capacity attenuation; inputting the principal component data and the secondary component data into a pre-trained long short-term memory neural network; receiving an output result of the long short-term memory neural network, and processing the output result to obtain an attenuation sequence of the battery capacity; and determining whether a numerical value in the attenuation sequence reaches a preset battery failure threshold so as to predict remaining battery life, and inputting the past data and a prediction result into the long short-term memory neural network for reverse training. Generalized, long-term, and effective battery life prediction can be achieved.

Description

A battery life prediction method, system, electronic device and storage medium

technical field

The present invention relates to the technical field of batteries, and in particular, to a battery life prediction method, system, electronic device and storage medium.

Background technique

With the development of new energy technologies, lithium-ion batteries have been widely used in many important fields. However, lithium-ion batteries still face many challenges, one of which is performance degradation. There are many factors involved in performance degradation. For example, when many chemical side reactions of anode, electrolyte and cathode are affected, the performance of the battery will be degraded, and the capacity of the battery will be partially regenerated, self-charging phenomenon, user habits, and ambient temperature. , road vibration and other factors, the battery capacity may be attenuated, thus affecting the battery life.

Therefore, it is of great significance to predict the remaining life of the battery to ensure the reliable operation and maintenance of the battery management system.

However, the existing battery life prediction methods do not separate the influence of various factors that cause battery capacity fading, or can only achieve short-term prediction, or can only predict the life of a specific battery, so it cannot achieve generalization, long-term and valid forecast.

SUMMARY OF THE INVENTION

The main purpose of the present invention is to provide a battery life prediction method, system, electronic device and storage medium, which can achieve generalized, long-term and effective prediction of battery life.

In order to achieve the above purpose, a first aspect of the present invention provides a battery life prediction method, including: acquiring historical data of battery capacity; preprocessing the historical data to obtain primary component component data and secondary component component data of battery capacity decay data; input the principal component component data and the secondary component component data into a pre-trained long-term and short-term memory neural network; receive the output of the long-term and short-term memory neural network, and process the output to obtain a battery Decay sequence of capacity; determine whether the value in the decay sequence reaches a preset battery failure threshold, so as to predict the remaining life of the battery.

Further, the step of preprocessing the historical data includes: using the method of collective empirical mode decomposition to decompose the historical data of the battery capacity into at least three component data, the at least three components include at least two eigenmodes, one residual Dimensionality reduction is performed on the eigenmodes and the residual by using the principal component analysis method, so as to reduce the component data into two, and the two component data includes the principal component component data and the secondary component component data.

Further, the training method of the long-term and short-term neural network includes: acquiring sample data of battery capacity and establishing an original long-term and short-term neural network; preprocessing the sample data to obtain sample principal component component data and samples of battery capacity decay Secondary component component data; input the sample principal component component data and the sample secondary component component data into the original long-term and short-term neural network for training.

Further, the training method of the long-term and short-term neural network further includes: performing sparse processing on the sample principal component component data and the sample secondary component component data of the sample data; The component component data is input into the original long-term and short-term neural network for training.

Further, the training method of the long-term and short-term neural network further includes: acquiring the type of battery of the sample data, and acquiring auxiliary data of the battery capacity of the same type as the battery of the sample data; preprocessing the auxiliary data to obtain auxiliary principal component components data; input the auxiliary principal component data into the original long-term and short-term neural network, perform auxiliary training, and obtain a long-term and short-term neural network.

Further, the training method for the long-term and short-term neural network further includes: acquiring N auxiliary data with the same type of battery capacity as the sample data battery, where N is an integer greater than 1; preprocessing the N auxiliary data, Obtain N auxiliary principal component data; calculate the average value of the N auxiliary principal component data as an auxiliary sequence; input the auxiliary sequence into the original long-term and short-term neural network, perform auxiliary training, and obtain a long-term and short-term neural network.

Further, the principal component analysis method includes an inverse transformation matrix, and the inverse transformation matrix is used to inversely transform the components; the output result includes a principal component component prediction result and a secondary component prediction result; predict the output result The processing step includes: multiplying the prediction result of the principal component and the prediction result of the secondary component with the inverse transformation matrix to perform inverse transformation; The composition prediction results are superimposed to obtain the predicted battery capacity decay sequence.

A second aspect of the present invention provides a battery life prediction system, comprising: a historical data acquisition module for acquiring historical data of battery capacity; a preprocessing module for preprocessing the historical data to obtain a main indicator of battery capacity decay. Component component data and secondary component component data; a training module for pre-training a long short-term memory neural network; an input module for inputting the principal component component data and the secondary component component data into the training module trained by the training module. long and short-term memory neural network; a receiving module, used for receiving the output result of the long-term and short-term memory neural network, and predicting the output result to obtain a decay sequence of battery capacity; a prediction module, used for judging the decay sequence in the Whether the value reaches the preset battery failure threshold to predict the remaining battery life.

A third aspect of the present invention provides an electronic device, comprising: a memory and a processor, the memory stores a computer program that can run on the processor, and when the processor executes the computer program, the above-mentioned Any one of the battery life prediction methods.

A fourth aspect of the present invention provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, implements any one of the battery life prediction methods described above.

The present invention provides a battery life prediction method, system, electronic device and storage medium, and the beneficial effects are:

By preprocessing the sample data to the main component data and the secondary component data, the complex parameters of the battery capacity decay can be effectively separated, so that the long short-term memory network can better identify the factors that cause the capacity decay, so as to achieve Effective prediction; in addition, pre-training the long short-term memory neural network can learn different sample data during training, and obtain the ability to identify different batteries, thereby improving the generalization ability of prediction; in addition, by predicting battery life Then, using the prediction results to reversely train the long-term and short-term memory neural network can make the long-term and short-term memory neural network identify the battery based on the previous data and prediction results in the middle and late battery life, so as to make the long-term battery life. Predictions are more accurate.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. For those skilled in the art, other drawings can also be obtained according to these drawings without creative efforts.

1 is a schematic flowchart of a battery life prediction method according to an embodiment of the present application;

2 is a schematic diagram of a PCA of a battery life prediction method according to an embodiment of the present application;

3 is a unit structure diagram of a long short-term memory neural network of a battery life prediction method according to an embodiment of the present application;

4 is a schematic diagram of a capacity decay curve of a NASA lithium-ion battery for a battery life prediction method according to an embodiment of the present application;

FIG. 5 is a schematic diagram of a capacity decay curve of a CALCE lithium-ion battery for a battery life prediction method according to an embodiment of the present application;

6 is a schematic diagram of a component curve obtained by EEMD processing of B5 of the battery life prediction method according to an embodiment of the present application;

7 is a schematic diagram of a PCA component curve of B5 of a battery life prediction method according to an embodiment of the present application;

8 is a schematic diagram of a PC curve of a NASA lithium-ion battery according to a battery life prediction method according to an embodiment of the present application;

FIG. 9 is a statistical graph of the correlation coefficient of the PC of the NASA lithium-ion battery in the first 70, 80, and 90 cycles of the battery life prediction method according to the embodiment of the present application;

Figure 10(a) is a schematic diagram of a curve diagram of a NASA lithium-ion battery PC prediction result of a battery life prediction method according to an embodiment of the present application, and Figure 10(b) is a diagrammatic diagram of a curve diagram of the SC1 prediction result;

11 is a schematic diagram of the sparseness and interpolation reconstruction of CALCE data according to the battery life prediction method according to the embodiment of the present application;

Figure 12 (a) is the online prediction result and error lattice and curve diagram of B5 of the battery life prediction method according to the embodiment of the application, and Figure 12 (b) is the lattice and curve diagram of the online prediction result and error of CX2-34;

13 is a schematic structural block diagram of a battery life prediction system according to an embodiment of the present application;

FIG. 14 is a schematic block diagram of the structure of an electronic device according to an embodiment of the present application.

Detailed ways

In order to make the purpose, features and advantages of the present invention more obvious and understandable, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. The embodiments described above are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present invention.

Please refer to FIG. 1, which is a battery life prediction method, including: S1, obtaining historical data of battery capacity; S2, preprocessing the historical data to obtain principal component component data and secondary component component data of battery capacity decay; S3 , Input the principal component component data and the secondary component component data into the pre-trained long-term and short-term memory neural network; S4, receive the output results of the long-term and short-term memory neural network, and process the output results to obtain the battery capacity attenuation sequence; S5, Determine whether the value in the decay sequence reaches the preset battery failure threshold to predict the remaining life of the battery, and input the historical data and prediction results into the long short-term memory neural network for reverse training.

In this embodiment, by preprocessing the sample data to the main component data and the secondary component data, the complex parameters of the battery capacity decay can be effectively separated, so that the long short-term memory network can better identify the resulting capacity In addition, pre-training the long-term and short-term memory neural network can learn different sample data during training to obtain the ability to identify different batteries, thereby improving the generalization ability of prediction; After predicting the life of the battery, using the prediction results to reversely train the long-term memory neural network can make the long-term memory neural network identify the battery based on the previous data and the prediction results in the middle and late battery life. This makes the long-term battery life prediction more accurate.

In one embodiment, the step of preprocessing the historical data includes: using an ensemble empirical mode decomposition method to decompose the historical battery capacity data into at least three component data, where the at least three components include at least two eigenmodes , a margin; using the principal component analysis method to reduce the dimensionality of the eigenmodes and the margin to reduce the component data into two, the two component data include the principal component component data and the secondary component component data.

In this embodiment, the EEMD (Ensemble Empirical Mode Decomposition, ensemble empirical mode decomposition) method is used to decompose the original battery capacity data into multiple IMFs (Intrinsic Mode Functions, eigenmodes) and a Res (Residual, Then use the PCA (Principal Component Analysis, principal component analysis) method to reduce the dimensionality of the multiple components decomposed in the previous step, that is, on the premise of maintaining 99% of the effective information, reduce the number of components to about two , one of the components is called PC (Principal Component, principal component component) is a monotonic decay component that contains most of the effective information, and the other components are called SC (Secondary Component, secondary component component) is a fluctuation component with less information; Finally, a correlation analysis was performed on the component PCs of the same battery capacity series processed by EEMD-PCA, and the similar PCs with the strongest correlation with the predicted battery capacity series and the average value of the same battery PCs were simultaneously used as auxiliary sequences.

Specifically, to extract the overall degradation trend of Li-ion battery capacity fading data, an ensemble empirical mode decomposition (EEMD) method is first employed. EEMD is a method for analyzing nonlinear and non-stationary signals. The most obvious feature of EEMD is that there is no basis function in the decomposition process, so it can adaptively represent the local fluctuation characteristics and global degradation trend of the original signal.

The process of extracting IMF by EEMD is called a screening algorithm, which is an iterative method. The specific decomposition steps are as follows:

1. Add Gaussian white noise _ni (t) to the original capacity sequence ξ(t) to obtain the superimposed sequence ξ _{i, j} (t):

ξ _i,j (t)=ξ(t)+n _i (t) (1)

Here, i is the iterative process of adding Gaussian white noise for the ith time, and j represents the calculation process of the jth IMF component in each iterative process.

2. Determine the extreme values (local minimum and maximum values) of ξ _{i, j} (t), and then use cubic spline interpolation to obtain the upper envelope U(t) and lower envelope of ξ _{i, j} (t) Line L(t).

3. Calculate the mean value m(t) of the upper envelope U(t) and the lower envelope L(t), and subtract the mean value m(t) from ξ _{i, j} (t) to obtain the intermediate quantity h(t):

h(t)=ξi _,j (t)-m(t)

4. Check whether h(t) satisfies the following conditions a and b. If not, replace ξi with h(t) _{, and j} (t) repeats steps 2 and 3 until h(t) satisfies conditions a and b. When the conditions a and b can be satisfied at the same time, h(t) can be regarded as an IMF _i,j .

Condition a: The number of local extreme points is equal to the number of zero-crossing points, or the difference is at most 1;

Condition b: The mean value m(t) of the upper envelope U(t) and the lower envelope L(t) satisfies: m(t)=0. (In practical applications, too much iterative processing will make the IMF become a simple constant amplitude FM signal, thus losing its practical significance. Therefore, when m(t)≤ε, it can be considered that the condition b has been satisfied, here ε is a given positive value close to 0).

5. Subtract IMF _i,j from ξi _,j (t) to obtain the difference ξi _,j+1 (t). Replace ξ _{i, j} (t) with ξ _{i, j+1} (t), and repeat steps 2-5 until the signal ξ _{i, j+1} (t) does not fluctuate more than 2 times. At this time, ξ _{i, j+1} (t) is the margin _ri .

ξi _,j+1 (t)=ξi _,j (t)-IMF _i,j (4)

6. Repeat steps 1 to 5 until the number of iterations i reaches the given value θ, and then average the IMF _i _{, j} and ri obtained by θ iterations:

The original capacity sequence ξ(t) can finally be decomposed into multiple IMFs and Res:

EEMD is a noise-assisted decomposition method that aims to improve the shortcomings of EMD. As shown in Figure 2, EEMD essentially repeats the EMD process for a given number of times on the original signal ξ(t), and then averages the corresponding components resulting from the iterations. In each experiment of EEMD, Gaussian white noise n _i (t) is added to assist the decomposition, so that the noise interference signal not only has a uniform decomposition scale, but also smoothes outliers caused by impulse interference, etc., effectively solving the modal noise mixing question.

However, too many IMF components will lead to a large cumulative calculation error and affect the final prediction accuracy. To solve this problem, principal component analysis (PCA) is employed here. PCA is a dimensionality reduction statistical analysis technique, which reduces the number of components while retaining the effective information of the original data. As shown in Figure 2, PCA can reduce the original correlated components and convert them into fewer, irrelevant components through linear combination. Thus fewer and irrelevant principal components can represent the original sequence and reflect its changes without causing heavy computation and computational errors.

here,

_{Corresponding} to the _IMF ₁ , _IMF ₂ , _.

The specific process of PCA is as follows:

1. Calculate the mean matrix M of m components ζ _j and the matrix Φ after the centering process by formula (8) and formula (9) respectively:

here,

and

are the mean and centering sequence of components ζ _j , respectively, j ∈ [1, m].

2. Calculate the sample covariance matrix R:

Perform eigendecomposition on the covariance matrix R= ^UΛUT , where Λ=diag{λ ₁ , λ ₂ , ..., λ _m } is the main diagonal matrix composed of the eigenvalues of R, satisfying λ ₁ ≥λ ₂ ≥...≥ λ _m ≥ 0, U is an orthogonal matrix composed of eigenvectors of R, and satisfies U ^-1 = ^UT , so the principal components can be calculated by the following formula:

3. The size of the eigenvalue λ _j of the covariance matrix R reflects the size of the effective information of the principal component component η _j , that is, the larger λ _j is, the more effective information η _j contains. The percentage of valid information for η _j is calculated by:

Finally, by calculating the cumulative contribution rate

We can determine how many components to keep, and the small contribution can be ignored as noise.

4. The k components η determined to be retained can be calculated by the following formula:

here,

is a matrix consisting of the first k columns reserved by U ^T.

Preprocessing the data through the EEMD-PCA combination can improve the prediction performance of the subsequent neural network prediction model.

In one embodiment, the training method of the long-term and short-term neural network includes: acquiring sample data of battery capacity and establishing an original long-term and short-term neural network; preprocessing the sample data to obtain sample principal component component data and sample secondary data of battery capacity decay Component component data is required; the sample principal component component data and the sample secondary component component data are input into the original long-term and short-term neural network for training.

The excellent performance of the deep learning neural network model in feature extraction and time series analysis, the present invention improves the LSTM (Long and Short-Term Memory, long short-term memory) neural network, respectively, the PC and SC obtained by the previous module Two types of data are modeled, analyzed and predicted. In order to further improve the prediction accuracy, the PC of the same lithium-ion battery capacity sequence is also used as an auxiliary quantity to train the LSTM network.

In one embodiment, the training method for the long-term and short-term neural network further includes: acquiring the battery type of the sample data, and acquiring auxiliary data of the battery capacity of the same type as the sample data battery; preprocessing the auxiliary data to obtain auxiliary principal component component data ; Input the auxiliary principal component data into the original long-term and short-term neural network, and perform auxiliary training to obtain the long-term and short-term neural network.

In one embodiment, the training method for the long-term and short-term neural network further includes: acquiring N auxiliary data with the same type of battery capacity as the sample data battery, where N is an integer greater than 1; preprocessing the N auxiliary data to obtain N auxiliary data Auxiliary principal component data; calculate the average value of N auxiliary principal component data as an auxiliary sequence; input the auxiliary sequence into the original long-term and short-term neural network, perform auxiliary training, and obtain a long-term and short-term neural network.

After the experimental data sequence is preprocessed, it is still highly nonlinear and long-term dependent. To address these issues, this study adopts an LSTM neural network. The LSTM neural network structure is a special kind of recurrent neural network, which is usually used to solve long-term dependency problems.

The unit structure of the LSTM network is shown in Figure 3, which consists of a forget gate, an input gate and an output gate. LSTM networks can both preserve meaningful information and forget useless data. Also, it can decide what information to output. These properties can make LSTMs more effective in handling long-term correlated and highly nonlinear sequences. The calculation formula of each gate is as follows:

Forgotten Gate:

f _t =σ(W _f ·[y _t-1 , x _t ]+b _f ) (14)

2) Input gate:

3) Output gate:

where x is the input data, y is the output data, and i, f, O, and C are the input gate, forget gate, output gate, and cell state, respectively. The matrices W and b represent the weights and biases to be trained,

σ(·) is the sigmoid function, and tanh(·) is the hyperbolic tangent function. These gates of LSTMs work together to effectively capture both long-term and short-term features of the input time series data, preventing vanishing and exploding gradients during information transfer.

In one embodiment, the training method for the long-term and short-term neural network further includes: performing sparse processing on the sample principal component data and the sample secondary component data of the sample data; The component component data is fed into the original long- and short-term neural network for training.

By sparse combing, data points can be reduced and the identification ability of the original input original long-term and short-term neural network can be improved.

In one embodiment, the principal component analysis method includes an inverse transformation matrix, and the inverse transformation matrix is used to inversely transform the components; the output result includes the principal component component prediction result and the secondary component prediction result; the step of performing prediction processing on the output result Including: multiplying the prediction results of the principal component components and the prediction results of the secondary components with the inverse transformation matrix to perform inverse transformation; superimposing the prediction results of the principal component components and the prediction results of the secondary components after the inverse transformation to obtain the predicted battery capacity decay sequence .

This embodiment also uses the technical solutions described above to verify two industry-recognized lithium-ion battery data sets. In addition, further experiments were conducted to evaluate the online prediction performance, and 100 repeated experiments were performed to reduce the effect of model randomness, resulting in more robust and accurate prediction results in cloud computing.

The first lithium battery dataset was released by NASA-Ames (National Aeronautics and Space Administration-Ames) research center, and the second dataset was from CALCE (Center for Advanced Life Cycle Engineering (CALCE) of the University of Maryland, University of Maryland Advanced Lifecycle Engineering Center). CALCE batteries last almost 10 times longer than NASA batteries, so they represent two different types of lithium-ion batteries, long-life batteries and short-life batteries. It is very challenging to validate the proposed method with these two sets of widely different battery data, but it is also more able to demonstrate the generality of the method.

In the NASA dataset, four groups of 18650 lithium-ion batteries (rated capacity of 2Ah) were selected: B5, B6, B7 and B18, which were obtained by charging and discharging the batteries at room temperature (24°C) and EIS (Electrochemical Impedance Spectroscopy, electrochemical impedance spectroscopy). The specific experimental process is as follows:

1) The battery is first charged with a constant current of 0.75C (1.5A) to a cut-off voltage of 4.2V, and then charged with a constant voltage of 4.2V until the cut-off current drops below 0.02A;

2) Then the battery is discharged at a constant current of 1C (2A) until the voltages of B5, B6, B7, and B18 drop to 2.7V, 2.5V, 2.2V, and 2.5V, respectively;

3) Repeat steps 1 and 2 to accelerate the aging process of the lithium battery while recording impedance data.

Figure 4 shows the capacity degradation curves of NASA lithium-ion batteries measured by the above experiments.

Four sets of data were selected from the CALCE lithium-ion battery dataset: CX2-34, CX2-36, CX2-37, and CX2-38, which are prismatic cells with LiCoO2 cathodes with a nominal capacity of 1.35Ah.

CALCE's lithium-ion battery capacity decay data is obtained through the ArbinBT2000 battery experimental test system at room temperature (25-30 °C), and the experimental process is similar to NASA:

1) The battery is first charged at a constant current rate of 0.5C (0.675A) until the voltage reaches 4.2V, and then maintained at 4.2V until the charging current drops below 0.05A.

2) The battery was then discharged at a constant current rate of 0.5C (0.675A) until the voltage dropped to 2.7V.

3) Repeat steps 1 and 2 and perform an EIS test to measure impedance and obtain internal parameters reflecting capacity degradation after each cycle.

For some obvious outliers in the CALCE lithium-ion battery dataset, they are removed to reduce noise. According to the 3σ criterion, the difference between two adjacent battery capacity cycles was calculated, and the mean μ and standard deviation σ of all the differences were calculated. Replace capacity data points with differences outside the interval (μ-3σ, μ+3σ) with a linear interpolation of two adjacent capacity values. The capacity decay curve of the treated CALCE lithium-ion battery is shown in Figure 5.

Referring to the experimental data standards of NASA and CALCE respectively, when the discharge capacities of B5, B6, B7, B18 and CX2-34, CX2-36, CX2-37 and CX2-38 batteries drop to 1.4Ah and 0.945Ah respectively, it is considered that Lithium batteries have reached their failure threshold (70% of rated capacity).

Data preprocessing:

The data preprocessing process included EEMD, PCA and auxiliary sequence analysis.

Due to various interference errors in the measurement process, as well as the complex physical and chemical characteristics of the battery, there are local fluctuations in the capacity decay curve of Li-ion batteries, which will greatly affect the predicted performance. To solve these problems, EEMD is used to decompose the original sequence, taking the decomposition result of group B5 as an example, as shown in Figure 6. The original sequence is decomposed according to formulas (1)-(7) to obtain 7 components, which are respectively expressed as IMF1, IMF2, IMF3, IMF4, IMF5, IMF6 and Res.

If all these components were used directly for subsequent neural network processing, the computation would be very time-consuming and would introduce many accumulated errors. In order to avoid these problems, PCA is adopted to retain the effective components: under the precursor that retains more than 99% of the effective information of the original signal, at least two components are retained, and the components with less than 1% of the effective information are discarded as noise. As shown in Figure 7, after PCA processing, two components are obtained, which contain more than 99% of the information of the original seven EEMD components. The first component with simpler decay contains 98.89% valid information and is called the principal component (PC). The second fluctuation contains 0.36% of the valid information called the secondary component (SC1).

Two simple components were extracted from the raw experimental data by joint processing of EEMD and PCA. The overall degradation trend component PC showed a good monotonic downward trend without fluctuation. The local fluctuation component SC1 mainly contains information on battery capacity regeneration and self-charging phenomena. The proposed EEMD-PCA decomposition method can effectively separate the local fluctuation and global degradation trend of battery capacity degradation data, which is helpful to improve the performance of subsequent deep learning methods.

The two components in Figure 7 are divided into two segments, and the neural network is trained using the data of the previous segment respectively to predict the data of the latter segment. Since the training data and prediction data of the PC are two disjoint intervals, large deviations are prone to occur in the prediction. Therefore, in order to obtain more accurate and stable prediction results, similar other battery sequences are introduced for auxiliary training instead of relying on this single sequence. Figure 8 illustrates the PC of each NASA Li-ion battery after EEMD-PCA pretreatment. Each PC contains 96%-99% of the original capacity sequence. It can be seen that the main trend of battery capacity decline has a high similarity. This paper uses correlation analysis (CA) to analyze the similarity between these PCs, it should be noted that since B18 has only 132 data points, the first 132 data points of the other three NASA lithium-ion batteries are also analyzed. related analysis. The correlation coefficients for the four Li-ion batteries were calculated, as shown in Table 1. It can be observed that the correlation coefficient of any two battery capacity series is very close to 1. This also shows that the degradation curves of the four cells are highly correlated.

Table 1 Correlation coefficient of NASA lithium-ion battery PC

However, in real-world situations, the capacity degradation data of the battery to be predicted are incomplete until the failure threshold is reached. In order to find the sequence with the highest correlation as an auxiliary sequence, we can compare the existing capacity data of the battery under test with other complete battery capacity sequences in the dataset through correlation analysis. As shown in Figure 9, the PC of B7 has the strongest correlation with the PC of B5. Therefore, the PC of B7 was used as the helper sequence, and the average PC value of B6 and B7 could be chosen as another helper sequence for generality (excluded because the sequence of B18 was too short). These auxiliary sequences with full-cycle degradation information of the battery and the existing historical data of the battery are used as the input LSTM neural network for the training set. In the prediction stage, the known historical data of the sequence to be predicted is input into the training model to obtain the prediction result.

LSTM prediction and data post-processing:

The LSTM network algorithm in the present invention is implemented in MATLAB, and the model parameters are set as shown in Table 2 after repeated experiments. Use the historical data of PC and SC1 to train the designed network model, and add auxiliary sequence training to the network model of PC to improve its prediction accuracy. Then use the trained LSTM network model to make predictions. First, the preprocessed capacity value of the current cycle is input into the neural network to predict the capacity value of the next cycle. Then, the predicted value is used as the input for the next iteration. The capacity of the battery in the future is predicted by the extrapolation method, and the prediction result can be expressed by equation (17):

Q _t+h = LSTM(Q _t+h-1 ) (17)

where LSTM( ) is the LSTM neural network prediction model, t is the starting point of the prediction, h is the number of cycles after the starting point, and Q _t+h is the available capacity of the battery at the t+hth cycle.

Table 2 Parameter settings of LSTM

Figure 10 shows the PC and SC1 prediction results for NASA B5 with the previous 90 cycles of data as the training set. It can be seen that the prediction result of the principal component PC is basically consistent with the decreasing trend of the actual capacity. Although the prediction accuracy of the secondary component SC1 is obviously inferior to that of PC due to fluctuations, its contribution and effective information are also much smaller than that of PC, which has little effect on the superimposed results. Post-processing (represented as Eq. (18)) is applied to process the result output by the LSTM. Multiply the prediction results of each component by the inverse transformation coefficient matrix of PCA to obtain the final capacity prediction sequence:

here,

is the predicted battery capacity sequence,

Represents the predicted PC and SC components, k is the number of reserved components, U _coeff is the inverse transformation coefficient matrix of PCA, U _coeff =U _k can be obtained according to formulas (11) and (13), and the matrix M _k is defined by the following formula :

here,

is a matrix composed of M reserved first k columns in formula (8), and n is the capacity sequence

length.

In particular, the battery life used by CALCE is much longer than the battery life of NASA, so the capacity data points in the CALCE battery data set sequence are denser, and the variation of two adjacent capacity data is also smaller. In order to improve the prediction performance and running speed of the LSTM network model on the CALCE ultra-long dataset, further appropriate processing methods need to be taken. As shown in Figure 11, after the training data is preprocessed by EEMD-PCA, it is sparsed, that is, the capacity data of every 8 adjacent cycles is averaged to reduce the data volume of the original sequence, and then the average value is calculated. Value series regression forecast. The predicted components are inversely transformed and then subjected to interpolation and reconstruction processing, that is, the cubic spline interpolation method is used to interpolate 7 data points between the inversely transformed data points to obtain a complete capacity prediction sequence.

In order to verify the effectiveness of this method on battery datasets with different life scales, it is applied to both NASA datasets and CALCE datasets. Figure 14 shows different prediction results for B5 and CX2-34 at different prediction starting points. The forecast starting point for B5 is 90, 80, and 70 periods, and the forecast starting point for CX2-34 is 801, 705, and 601 periods. The predicted curve is basically consistent with the original capacity decay curve, and there are local fluctuations, showing the phenomenon of capacity regeneration. Compared to these predicted series, the data points are relatively scattered due to the relatively short cycle life of B5. Obviously, the predicted results fluctuate locally, which is in good agreement with the actual degradation curve; while CX2-34 has a longer cycle life and less data variation, and some predicted data points are obtained by cubic spline interpolation. Therefore, the capacity regeneration phenomenon causes The local fluctuation of , is not fully reflected, but the main degradation trend of the prediction curve is basically satisfied, and the prediction accuracy is also guaranteed. To sum up, the validity and robustness of the lithium battery life prediction method based on the fusion of pre-decomposition and deep learning proposed in the present invention has been verified for RUL prediction of lithium batteries with different life scales.

To quantitatively evaluate the performance of the model, the following four classical evaluation criteria are adopted.

(1) root mean square error (root mean square error, RMSE)

(2) Mean absolute percent error (MAPE)

(3) Absolute error (accuracy error, AE)

(4) Relative error (relative error, RE)

Here, n is the data length, Ti is the measured value for the _ith cycle, and _Pi is the corresponding predicted value for the ith cycle. The smaller the value of these four criteria, the higher the prediction accuracy.

Table 3 shows the prediction results of B5 and CX2-34 at different starting point of prediction (SPP). AE and RE of actual life (end of life, EoL) and predicted end of life (PEoL) under different SPPs were calculated. Under different SPPs, RMSE and MAPE are used to evaluate the capacity prediction results, and AE and RE are used to evaluate the RUL prediction accuracy.

Table 3 RUL prediction results

To further evaluate the online prediction performance of the proposed method, the RUL predictions of B5 and CX2-34 were calculated and illustrated in 12. Among them, Fig. 12(a) is the RUL prediction of 70 cycles B5, and Fig. 12(b) is the RUL prediction of 601 cycles of CX2-34. It can be observed that B5 and CX2-34 may have relatively large errors in early predictions, but are acceptable under certain circumstances. With the accumulation of historical data, the accuracy gradually improves, and more accurate RUL prediction results can be obtained in the middle and late stages. Online RUL prediction can be used as an important reference for BMS to monitor the future health of batteries.

In Fig. 12, the RUL prediction AE of B5 and CX2-34 does not decrease monotonically with the prediction time point, but decreases fluctuatingly, which is caused by the uncertainty of the prediction algorithm. Uncertainty management has important implications for health forecasting because it provides decision makers with statistical information on forecasting rules. Therefore, to obtain their PEoL distributions, 100 prediction experiments were repeated on B5 and CX2-34 based on the first 90 cycles and the first 801 cycles, respectively. In addition, the mean and median of its PEoL were also calculated for performance evaluation. According to the statistical results, the absolute uncertainty error between the mean and median PEoL values of B5 and CX2-34 and the true value is no more than 2 cycles. Therefore, a single PEoL can be replaced by the mean or median of multiple PEoLs, which can effectively deal with the adverse effects of uncertainty factors and maintain the stability and accuracy of RUL prediction for lithium batteries. With the popularization and application of cloud computing and IoT technologies, computing and storage are no longer a problem. The battery-related data of different vehicles can be measured on-board and seamlessly uploaded to the cloud, so the method can make more accurate RUL predictions based on the data collected by the cloud-based battery system.

Referring to FIG. 13, an embodiment of the present application also provides a battery life prediction system, including: a historical data acquisition module 1, a preprocessing module 2, a training module 3, an input module 4, a receiving module 5, and a prediction module 6; historical data acquisition Module 1 is used to obtain historical data of battery capacity; preprocessing module 2 is used to preprocess historical data to obtain principal component data and secondary component data of battery capacity decay; training module 3 is used to pre-train long short-term memory Neural network; input module 4 is used for inputting principal component component data and secondary component component data into the long short-term memory neural network trained by training module 3; receiving module 5 is used for receiving the output result of the long short-term memory neural network, and is used for the output result. Perform prediction to obtain a decay sequence of battery capacity; the prediction module 6 is used to judge whether the value in the decay sequence reaches a preset battery failure threshold, so as to predict the remaining life of the battery.

The division of each module in the above battery life prediction system is only for illustration. In other embodiments, the battery life prediction system can be divided into different modules as required to complete all or part of the functions of the above battery life prediction system.

For the specific limitation of the battery life prediction system, reference may be made to the definition of the battery life prediction method above, which will not be repeated here.

An embodiment of the present application provides an electronic device, please refer to FIG. 14, the electronic device includes: a memory 601, a processor 602, and a computer program stored in the memory 601 and running on the processor 602, and the processor 602 executes the computer When the program is executed, the battery life prediction method described in the preceding paragraph is implemented.

Further, the electronic device further includes: at least one input device 603 and at least one output device 604 .

The above-mentioned memory 601 , processor 602 , input device 603 and output device 604 are connected through a bus 605 .

The input device 603 may specifically be a camera, a touch panel, a physical button, a mouse, or the like. The output device 604 may specifically be a display screen.

The memory 601 may be a high-speed random access memory (RAM, Random Access Memory) memory, or may be a non-volatile memory (non-volatile memory), such as a disk memory. Memory 601 is used to store a set of executable program codes, and processor 602 is coupled to memory 601 .

Further, an embodiment of the present application further provides a computer-readable storage medium, which may be provided in the electronic device in each of the foregoing embodiments, and the computer-readable storage medium may be the foregoing memory 601. A computer program is stored on the computer-readable storage medium, and when the program is executed by the processor 602, the battery life prediction method described in the foregoing embodiment is implemented.

Further, the computer-storable medium may also be a U disk, a removable hard disk, a read-only memory 601 (ROM, Read-Only Memory), a RAM, a magnetic disk or an optical disk and other mediums that can store program codes.

In the several embodiments provided in this application, it should be understood that the disclosed apparatus and method may be implemented in other manners. For example, the apparatus embodiments described above are only illustrative. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods. For example, multiple modules or components may be combined or Can be integrated into another system, or some features can be ignored, or not implemented. On the other hand, the shown or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or modules, and may be in electrical, mechanical or other forms.

The modules described as separate components may or may not be physically separated, and components shown as modules may or may not be physical modules, that is, may be located in one place, or may be distributed to multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this embodiment.

In addition, each functional module in each embodiment of the present invention may be integrated into one processing module, or each module may exist physically alone, or two or more modules may be integrated into one module. The above-mentioned integrated modules can be implemented in the form of hardware, and can also be implemented in the form of software function modules.

If the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they may be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, Read-Only Memory (ROM, Read-Only Memory), Random Access Memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes .

It should be noted that, for the convenience of description, the foregoing method embodiments are all expressed as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described action sequence. As in accordance with the present invention, certain steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily all necessary to the present invention.

In the above-mentioned embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.

The above is a description of a battery life prediction method, system, electronic device and storage medium provided by the present invention. For those skilled in the art, according to the ideas of the embodiments of the present invention, there will be specific implementation methods and application scopes. Changes, in conclusion, the content of this specification should not be construed as a limitation to the present invention.

Claims

A battery life prediction method, comprising:

Get historical data of battery capacity;

Performing preprocessing on the historical data to obtain principal component component data and secondary component component data of battery capacity decay;

Inputting the principal component component data and the secondary component component data into a pre-trained long short-term memory neural network;

receiving the output result of the long short-term memory neural network, and processing the output result to obtain a decay sequence of battery capacity;

Determine whether the value in the decay sequence reaches a preset battery failure threshold, so as to predict the remaining life of the battery, and input the historical data and the prediction result into a long short-term memory neural network for reverse training.
The battery life prediction method according to claim 1, wherein,

The steps for preprocessing historical data include:

Using the method of collective empirical mode decomposition to decompose the historical data of the battery capacity into at least three component data, the at least three components include at least two eigenmodes and a margin;

A principal component analysis method is used to reduce the dimension of the eigenmode and the residual, so as to reduce the component data into two, and the two component data includes the principal component component data and the secondary component component data.
The battery life prediction method according to claim 1, wherein,

The training method of the long-term and short-term neural network includes:

Obtain sample data of battery capacity and build original long-term and short-term neural network;

Preprocessing the sample data to obtain sample principal component component data and sample secondary component component data of battery capacity decay;

The sample principal component component data and the sample secondary component component data are input into the original long-term and short-term neural network for training.
The battery life prediction method according to claim 3, wherein,

The training method of the long-term and short-term neural network further includes:

performing sparse processing on the sample principal component component data and the sample secondary component component data of the sample data;

Input the sparse principal component data and sample secondary component data into the original long-term and short-term neural network for training.
The battery life prediction method according to claim 3, wherein,

The training method of the long-term and short-term neural network further includes:

Obtain the battery type of the sample data, and obtain auxiliary data of the battery capacity of the same type as the sample data battery;

Preprocessing the auxiliary data to obtain auxiliary principal component component data;

The auxiliary principal component component data is input into the original long-term and short-term neural network, and auxiliary training is performed to obtain a long-term and short-term neural network.
The battery life prediction method according to claim 5, wherein,

The training method of the long-term and short-term neural network further includes:

The acquiring N auxiliary data of the battery capacity of the same type as the sample data battery, where N is an integer greater than 1;

Preprocessing the N pieces of auxiliary data to obtain N pieces of auxiliary principal component component data;

Calculate the average value of N auxiliary principal component component data as an auxiliary sequence;

The auxiliary sequence is input into the original long-term and short-term neural network, and auxiliary training is performed to obtain a long-term and short-term neural network.
The battery life prediction method according to claim 2, wherein,

The principal component analysis method includes an inverse transformation matrix, and the inverse transformation matrix is used to inversely transform the components;

The output result includes a principal component component prediction result and a secondary component prediction result;

The step of performing prediction processing on the output result includes:

Multiplying the principal component prediction result and the secondary component prediction result with the inverse transformation matrix to perform inverse transformation;

The inverse-transformed prediction result of the principal component and the prediction result of the secondary component are superimposed to obtain a predicted battery capacity decay sequence.
A battery life prediction system, characterized in that it includes:

Historical data acquisition module, used to acquire historical data of battery capacity;

a preprocessing module, configured to preprocess the historical data to obtain principal component component data and secondary component component data of battery capacity decay;

Training module for pre-training long short-term memory neural network;

an input module for inputting the principal component component data and the secondary component component data into a long short-term memory neural network trained by the training module;

a receiving module, configured to receive the output result of the long-term and short-term memory neural network, and predict the output result to obtain a decay sequence of battery capacity;

The prediction module is used for judging whether the value in the decay sequence reaches a preset battery failure threshold, so as to predict the remaining life of the battery.
An electronic device, comprising: a memory and a processor, wherein a computer program that can be run on the processor is stored in the memory, and characterized in that, when the processor executes the computer program, claims 1 to The method of any one of 7.
A computer-readable storage medium on which a computer program is stored, characterized in that, when the computer program is executed by a processor, the method of any one of claims 1 to 7 is implemented.