A state of health estimation method for lithium-ion batteries under different fast charging conditions

Abstract

The application discloses a kind of for lithium ion battery different fast charging health state estimation method under condition, comprising: obtaining the full process voltage curve of lithium ion battery under different fast charging condition, based on sampling point to the full process voltage curve is sliced;Voltage curve segment is carried out sample entropy calculation after slicing, and the sample entropy data of voltage curve segment is obtained;The capacity data of lithium ion battery is obtained, and the sample entropy data and the capacity data are carried out correlation coefficient calculation, based on the value of correlation coefficient, select the sample entropy data of highest correlation;Cycle network model is constructed, and the sample entropy data of highest correlation is used as the input data of cycle network model, and the health state of lithium ion battery is predicted to obtain.The application utilizes the historical information of lithium ion battery under different fast charging condition, obtains adaptive battery health state estimation model, reduces the modeling training demand to new data set.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery state estimation technology, and particularly relates to a method for estimating the health status of lithium-ion batteries under different fast charging conditions. Background Technology

[0002] Lithium-ion batteries, characterized by high energy density, long lifespan, and low self-discharge rate, along with rapidly decreasing costs in recent years, have become a primary energy storage component in many industrial sectors, from portable electronics to electric vehicles and power grids. Accurately estimating the battery's State of Health (SOH) allows for timely assessment of battery health, preventing potential failures and ensuring optimal performance and cost-effectiveness while maintaining safety, reliability, and durability. Battery state of health is typically defined as the ratio of current maximum capacity to the nominal capacity of a brand-new battery. However, using current commercial sensing technologies, battery capacity cannot be directly measured physically.

[0003] Multiple degradation mechanisms can occur throughout a battery's lifespan, such as electrode fatigue and delamination, solid-electrolyte interface growth, and electrolyte decomposition. In this context, it is difficult to establish a battery degradation model that encompasses all these nonlinear and coupled dynamics. Current data-driven approaches place high demands on data availability and quality. This is because battery management systems can only measure battery terminal voltage, current, and surface temperature, while the availability of datasets covering various operating conditions, aging stress factors, and failure modes is crucial for accurately estimating battery health. Aging stress and aging trajectories are uncertain in terms of load and temperature distribution across different cells. For example, charging current amplitudes are inconsistent across various fast-charging strategies.

[0004] Therefore, there is currently a lack of an efficient method that can better utilize historical information to obtain an adaptive battery health state estimation model and reduce the need for modeling and training on new datasets. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a health status estimation method for lithium-ion batteries under different fast charging conditions, thereby resolving the issues present in the prior art.

[0006] To achieve the above objectives, the present invention provides a method for estimating the health status of lithium-ion batteries under different fast charging conditions, comprising:

[0007] Obtain the full-process voltage curve of lithium-ion battery under different fast charging conditions, and slice the full-process voltage curve based on the sampling points;

[0008] The sample entropy of the sliced ​​voltage curve segments is calculated to obtain the sample entropy data of the voltage curve segments.

[0009] Obtain the capacity data of lithium-ion batteries, calculate the correlation coefficient between the sample entropy data and the capacity data, and select the sample entropy data with the highest correlation based on the value of the correlation coefficient.

[0010] A recurrent network model is constructed, and the entropy data of the samples with the highest correlation is used as the input data of the recurrent network model to predict the health status of the lithium-ion battery.

[0011] Preferably, the voltage curve segments after slicing are N-dimensional time series sampled at equal time intervals.

[0012] Preferably, the process of calculating the sample entropy of the sliced ​​voltage curve segment is as follows:

[0013] Based on the N-dimensional time series, an m-dimensional vector X is reconstructed. m (1),X m (2),…,X m (n-m+1), where X m (i)=[x(i),x(i+1),…,x(i+m-1)];

[0014] Calculate two corresponding elements X m (i) and X m The absolute value distance d[X] between the maximum differences between (j) m (i),X m (j)];

[0015] Set a threshold r, and calculate d[X] for each value i. m (i),X m (j)] The ratio of the number of items less than r to the total distance Nm B m i (r);

[0016] Find the ratio B i m (r) is the average value B for all i values. m (r);

[0017] Let k = m + 1, and repeat the calculation of the ratio and the average value to obtain the first parameter A. k (r);

[0018] Based on the first parameter A k (r) and the average value B m (r) is used to calculate the sample entropy data of the voltage curve segment.

[0019] Preferably, the formula for calculating the correlation coefficient between the sample entropy data and the volume data is as follows:

[0020]

[0021] Where ρ is the Pearson correlation coefficient, X is the sample entropy sequence, and X = {x} i} N i＝1 Let Y be a sequence of capacity, Y = {y} i} N i-1 , and These are the means of X and Y, respectively.

[0022] Preferably, the recurrent network model includes: a bidirectional long short-term memory network and an attention layer;

[0023] The bidirectional long short-term memory network consists of two long short-term memory networks. One network processes the input data in a forward direction to obtain a first output, and the other network processes the data in a reverse direction to obtain a second output. The first output and the second output are connected to obtain the output vector of the bidirectional long short-term memory network.

[0024] Based on the weights of the attention layer, the output vector is weighted and summed to generate a context vector.

[0025] Preferably, the long short-term memory network includes: a forget gate, an input gate, and an output gate;

[0026] The calculation formula for the forget gate is as follows:

[0027] f t =σ(W fh h t-1 +W fx x t +b f )

[0028] Where σ is the activation function of the forget gate, and W fh b is the weight of fh f For the bias of f;

[0029] The calculation formula for the input gate is as follows:

[0030] i t =σ(W ih h t-1 +W ix x t +b i )

[0031]

[0032] Among them, W ih W ix W Ch and W Cxb represents the weight of the corresponding layer. i and b f For bias, This represents the candidate state of the current unit;

[0033]

[0034] Among them, C t-1 For the previous internal state, C t This is the current state;

[0035] Use sigmoid layer t and tanh layer h t The formula for generating the output value of the Long Short-Term Memory network is as follows:

[0036] o t =σ(W oh h t-1 +W ox x t +b o )

[0037] h t =o t tanh(C t ).

[0038] Preferably, the formula for generating the context vector is:

[0039]

[0040] Where c is the context vector, α t,i For the combination of input position i and output position t, v a and W a This is the weight matrix that needs to be learned.

[0041] Preferably, the method further includes: training the recurrent network model based on fine-tuned transfer learning, first freezing the weight coefficients of the initial seven layers of the recurrent network model, and then modifying the weight coefficients of the last four layers of the recurrent network model.

[0042] Compared with the prior art, the present invention has the following advantages and technical effects:

[0043] This invention provides a method for estimating the health status of lithium-ion batteries under different fast-charging conditions, comprising: firstly, acquiring the full-process voltage curve of the lithium-ion battery, and slicing the full-process voltage curve based on sampling points; then, calculating the sample entropy of the sliced ​​voltage curve segments to obtain sample entropy data of the voltage curve segments; next, acquiring the capacity data of the lithium-ion battery, calculating the correlation coefficient between the sample entropy data and the capacity data, and selecting the sample entropy data with the highest correlation based on the value of the correlation coefficient; finally, constructing a recurrent network model, using the sample entropy data with the highest correlation as the input data of the recurrent network model to predict the health status of the lithium-ion battery.

[0044] This invention first considers the various non-standard fast charging protocols currently on the market and uses feature engineering to extract a robust feature information related to battery aging under different fast charging strategies. Then, considering the temporal characteristics of battery aging information, it proposes a method to capture long-term dependency information of battery aging time series data using a deep recurrent network structure. Attached Figure Description

[0045] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0046] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0047] Figure 2 This is a voltage and current curve diagram of the entire charging and discharging process of a battery using an example one-step fast charging strategy according to an embodiment of the present invention.

[0048] Figure 3 This is a voltage and current curve diagram of the entire charging and discharging process of a battery using an example two-step fast charging strategy according to an embodiment of the present invention.

[0049] Figure 4 This is a schematic diagram of the voltage curve of the battery throughout its entire charge and discharge process according to an embodiment of the present invention;

[0050] Figure 5 This is a schematic diagram of the first voltage curve segment of the battery's full-life charge-discharge process according to an embodiment of the present invention;

[0051] Figure 6 This is a schematic diagram of the second voltage curve segment during the full life charge-discharge process of a battery according to an embodiment of the present invention;

[0052] Figure 7 This is a schematic diagram of the third voltage curve segment of the battery's full-life charge and discharge process according to an embodiment of the present invention, namely, the voltage curve segment of the CC charging and CC-CV charging conversion stage;

[0053] Figure 8 This is a schematic diagram of the fourth voltage curve segment during the full life charge-discharge process of a battery according to an embodiment of the present invention;

[0054] Figure 9 This is a schematic diagram of the fifth voltage curve segment during the full lifespan charge-discharge process of a battery according to an embodiment of the present invention.

[0055] Figure 10 This is a schematic diagram of the sixth voltage curve segment during the full life charge-discharge process of a battery according to an embodiment of the present invention;

[0056] Figure 11 This is a schematic diagram of the seventh voltage curve segment during the full life charge-discharge process of a battery according to an embodiment of the present invention.

[0057] Figure 12 This is a heatmap showing the correlation coefficients between the entropy of seven voltage slice samples and battery capacity in an embodiment of the present invention.

[0058] Figure 13 This is a block diagram of the BiLSTM-Att neural network according to an embodiment of the present invention;

[0059] Figure 14 This is a comparison chart of the SOH prediction results of test battery 1 in this embodiment of the invention with and without transfer learning.

[0060] Figure 15 This is a comparison chart of the SOH prediction results of test battery 2 in this embodiment of the invention, with and without transfer learning. Detailed Implementation

[0061] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0062] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0063] Example 1

[0064] This embodiment provides a method for estimating the health status of lithium-ion batteries under different fast charging conditions, such as... Figure 1 As shown, it includes:

[0065] Step 1: Based on the full life cycle charge and discharge voltage data of lithium-ion batteries under different fast charging conditions, slice the voltage curve of the whole process based on the sampling points.

[0066] The voltage and current curves of the battery during the entire charging and discharging process under the one-step fast charging strategy and the two-step fast charging strategy are as follows: Figures 2-3 As shown, the difference between the two operating conditions lies in the number of CC charging stages.

[0067] Figure 4 These are the voltage curves throughout the entire charge-discharge process of a battery's lifespan. Each curve represents one charge-discharge cycle, and the number of curves equals the total number of charge-discharge cycles in the battery's lifespan. Using this battery as an example, we perform segmentation. Specifically, given an arbitrary cycle voltage curve as V(t), and the sampling time t0 of the last constant current (CC) charging and constant current-constant voltage (CC-CV) charging transition phase, then V(t), t∈(t0-T,t0+T), is the voltage curve segment of the transition phase, with a length of 2T. The remaining voltage curves are then sliced ​​from beginning to end with a length of 2T, discarding curves shorter than 2T at the end, resulting in num voltage curve segments. These num segments serve as a control group, yielding a total of num+1 voltage segments. Each curve is sliced ​​separately, and the results are as follows: Figures 5-11 As shown.

[0068] Step 2: Based on the sliced ​​voltage curve segments, calculate using the sample entropy algorithm.

[0069] The sample entropy algorithm is as follows:

[0070] S21. A voltage curve segment is an N-dimensional time series D = {x1, x2, ..., xN} obtained by sampling at equal time intervals;

[0071] S22. Reconstruct the m-dimensional vector X m (1),X m (2),…,X m (n-m+1), where X m (i)=[x(i),x(i+1),…,x(i+m-1)];

[0072] S23. Define X m (i) and X m The distance d[X] between (j) m (i),X m [j] is the absolute value of the maximum difference between two corresponding elements, i.e.

[0073] d[X m (i),X m (j)]=max[|x(i+k)-x(j+k)|];

[0074] S24. Given a threshold r, calculate d[X] for each value i. m (i),X m The ratio of the number of items less than r to the total distance Nm is denoted as B.m i (r), i.e.

[0075]

[0076] S25. Find B i m (r) is the average of all values ​​of i, denoted as B. m (r), that is

[0077]

[0078] S26. Let k = m + 1, repeat steps S24-S25, and obtain:

[0079]

[0080] S27. Sample entropy (sampEn) is

[0081]

[0082] Since N cannot be infinite in practical calculations, when N takes finite values, the sample entropy is approximately...

[0083] SampEn = —ln[A k (r) / B m (r)];

[0084] Step 3: Calculate the Pearson correlation coefficient based on the sample entropy and capacity data of each voltage curve segment.

[0085] For two given sequences, the sample entropy sequence X = {xi} N i = 1, capacity sequence Y = {yi} N i-1, the Pearson correlation coefficient ρ(X,Y) can be calculated as:

[0086]

[0087] Where ρ is the Pearson correlation coefficient. and These are the means of X and Y, respectively. The entropy data of the samples with the highest correlation are used as input features. The correlation coefficient results between voltage segment sample entropy and battery capacity are as follows: Figure 4 As shown, HI1-7 represent the sample entropy features of the seven voltage segments, respectively.

[0088] Step 4: Construct the BiLSTM-Att network structure.

[0089] The Long Short-Term Memory (LSTM) network mainly consists of a forget gate, an input gate, and an output gate. The forget gate determines whether data from the current cell should be discarded.

[0090] ft=σ(Wfhht -1 +Wfxxt+b f )

[0091] Where σ is the activation function of the forget gate. W fh It is the weight of fh, b f It is the bias of f.

[0092] The input gate determines which new data should be updated. The determination of new information is achieved by activating a sigmoid layer. Then, a tanh layer generates a vector to store the candidate states of the current unit. It can be represented as:

[0093] i t =σ(W ih h t-1 +W ix x t +b i )

[0094]

[0095] Among them W ih W ix W Ch and W Cx b represents the weight of the corresponding layer. i and b f It is a bias.

[0096] The aforementioned gate can be used to retrieve the previous internal state C. t-1 Update to current state C t .

[0097]

[0098] Finally, the sigmoid layer is used. t and tanh layer h t To generate the output value of the LSTM, it can be represented as:

[0099] o t =σ(W oh h t-1 +W ox x t +b o )

[0100] h t =o t tanh(C t )

[0101] A Bidirectional Long Short-Term Memory (BiLSTM) network consists of two LSTM components. One processes the input sequence in the forward direction, while the other processes it in the reverse direction. After processing, the outputs of the two LSTMs are concatenated to obtain the final BiLSTM output. Consider the input sequence X = {x...} i} n i＝1 To the target sequence Y = {x i} n i＝1 The mapping task, where n is the length of the sequence. The encoder state can be defined based on the hidden layers of BiLSTM:

[0102]

[0103] in, and These represent the forward hidden state and the reverse hidden state of BiLSTM, respectively.

[0104] The weights calculated by the Attention layer are used to perform a weighted summation on the output vector of the BiLSTM layer. The specific network structure is as follows: Figure 13 As shown. This will generate a context vector, which is a weighted representation of all elements in the input sequence. The decoder network contains the hidden state s of the output at position t. t =f(s) t-1 ,y t -1,c t ), where the context vector c is the sum of the hidden states of the input sequence.

[0105]

[0106] α t,i The score assigned to the combination of input position i and output position t depends on the degree of match between them. t ,i} is a set of weights that represents how many source hidden states should be considered for each output.

[0107] α is parameterized by a single-hidden-layer feedforward neural network, which is trained along with the rest of the model. Therefore, the score function takes the following form (tanh as the non-linear activation function):

[0108]

[0109] Among them, v a and W a It is the weight matrix that needs to be learned.

[0110] Step 5: Fine-tune transfer learning based on the pre-trained model.

[0111] The BiLSTM-ATt model structure constructed in this embodiment is shown in Table 1, consisting of 11 layers. The pre-trained network used is quite complex, so training from scratch would take a significant amount of time. The fine-tuning strategy employed in this embodiment includes freezing the weight coefficients of the initial seven layers of the network. During transfer training, only the weight coefficients of the last four layers of the network are modified.

[0112] Table 1

[0113]

[0114] Based on one-step fast charging strategy data and two-step fast charging strategy data, the model was trained and tested respectively to verify the accuracy of the improved deep recurrent network proposed in this embodiment in predicting battery health status. Two battery data points were taken for each of the two fast charging conditions, and four network models were used for verification. The SOH prediction results were quantified using four evaluation indicators, as shown in Table 2. Among them, MAE is the mean absolute error loss, MAPE is the mean absolute percentage error, and RMSE is the root mean square error.

[0115] Table 2

[0116]

[0117] Based on the model pre-trained with data from the one-step fast charging strategy, fine-tuning transfer learning was performed. The number of fine-tuning layers was verified through experiments, and the remaining network layers were frozen.

[0118] Based on the fine-tuned transfer learning model and two-step fast charging strategy data, fine-tuned transfer learning training was performed, and tested on the two-step fast charging strategy data test set to verify the accuracy, efficiency, and robustness of the proposed transfer learning method in predicting battery health status across charging conditions. Two randomly selected test batteries under two-step fast charging conditions were used to verify the transfer model and the non-transfer model. The results are as follows: Figures 14-15 As shown in Table 3, the training time is as follows, where BiLSTM-Att-TL is the transfer model and the other is the non-transfer model.

[0119] Table 3

[0120]

[0121] This embodiment proposes a transfer learning algorithm to handle knowledge transfer between different fast charging modes, thereby reducing the data requirements and model training time costs for different fast charging modes.

[0122] Beneficial effects of this embodiment:

[0123] This embodiment proposes a method for estimating the health status of lithium-ion batteries under different fast-charging conditions. It can extract robust features highly correlated with battery aging from partial voltage curves of batteries under different fast-charging strategies; accurately predict battery health status using these aging features; and accurately predict battery health status using pre-trained models for transfer learning under different fast-charging conditions. The transfer learning method proposed in this embodiment solves the problem of limited data volume in different fast-charging datasets and reduces model training time by 64%, saving computational resources.

[0124] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for estimating the health status of lithium-ion batteries under different fast charging conditions, characterized in that, Includes the following steps: Obtain the full-process voltage curve of lithium-ion battery under different fast charging conditions, and slice the full-process voltage curve based on the sampling points; The sample entropy of the sliced ​​voltage curve segments is calculated to obtain the sample entropy data of the voltage curve segments. Obtain the capacity data of lithium-ion batteries, calculate the correlation coefficient between the sample entropy data and the capacity data, and select the sample entropy data with the highest correlation based on the value of the correlation coefficient. The formula for calculating the correlation coefficient between the sample entropy data and the volume data is as follows: Where ρ is the Pearson correlation coefficient, X is the sample entropy sequence, and X = {x} i } N i＝1 Let Y be a sequence of capacity, Y = {y} i } N i-1 , and These are the means of X and Y, respectively; A recurrent network model is constructed, and the entropy data of the samples with the highest correlation is used as the input data of the recurrent network model to predict the health status of the lithium-ion battery. The recurrent network model includes: a bidirectional long short-term memory network and an attention layer; The bidirectional long short-term memory network consists of two long short-term memory networks. One network processes the input data in a forward direction to obtain a first output, and the other network processes the data in a reverse direction to obtain a second output. The first output and the second output are connected to obtain the output vector of the bidirectional long short-term memory network. Based on the weights of the attention layer, the output vector is weighted and summed to generate a context vector; The long short-term memory network includes: a forget gate, an input gate, and an output gate; The calculation formula for the forget gate is as follows: f t =σ(W fh h t-1 +W fx x t +b f ) Where σ is the activation function of the forget gate, and W fh b is the weight of fh f For the bias of f; The calculation formula for the input gate is as follows: i t =σ(W ih h t-1 +W ix x t +b i ) Among them, W ih W ix W Ch and W Cx b represents the weight of the corresponding layer. i and b C For bias, This represents the candidate state of the current unit; Among them, C t-1 For the previous internal state, C t This is the current state; Use sigmoid layer t and tanh layer h t The formula for generating the output value of the Long Short-Term Memory network is as follows: o t =σ(W oh h t-1 +W ox x t +b o ) h t =o t fishy(C) t ) 2. The method for estimating the health status of lithium-ion batteries under different fast charging conditions according to claim 1, characterized in that, The voltage curve segments after slicing are N-dimensional time series sampled at equal time intervals.

3. The method for estimating the health status of lithium-ion batteries under different fast charging conditions according to claim 2, characterized in that, The process of calculating the sample entropy of the sliced ​​voltage curve segments: Based on the N-dimensional time series, an m-dimensional vector X is reconstructed. m (1),X m (2),…,X m (n-m+1), where X m (i)=[x(i),x(i+1),…,x(i+m-1)]; Calculate two corresponding elements X m (i) and X m The absolute value distance d[X] between the maximum differences between (j) m (i),X m (j)]; Set a threshold r, and calculate d[X] for each value i. m (i),X m (j)] The ratio of the number of items less than r to the total distance Nm B m i (r); Find the ratio B i m (r) is the average value B for all i values. m (r); Let k = m + 1, and repeat the calculation of the ratio and the average value to obtain the first parameter A. k (r); Based on the first parameter A k (r) and the average value B m (r) is used to calculate the sample entropy data of the voltage curve segment.

4. The method for estimating the health status of lithium-ion batteries under different fast charging conditions according to claim 1, characterized in that, The formula for generating the context vector is: Where c is the context vector, α t,i For the combination of input position i and output position t, v a and W a This is the weight matrix that needs to be learned.

5. The method for estimating the health status of lithium-ion batteries under different fast charging conditions according to claim 1, characterized in that, Also includes: The recurrent network model is trained based on fine-tuned transfer learning. First, the weight coefficients of the initial seven layers of the recurrent network model are frozen, and then the weight coefficients of the last four layers of the recurrent network model are modified.

Images