Deep Learning-Based Lithium-Ion Battery Life Prediction Method and System

Abstract

The present invention provides a method and system for predicting the lifespan of lithium-ion batteries based on deep learning, which relates to the technical field of battery health prediction. In the present invention, first, the original signal data is decomposed once to obtain an IMF sequence distributed from high to low in frequency; the high and low frequency demarcation points are obtained, and based on this, the IMF sequence is divided to obtain high-frequency data and low-frequency data; the high-frequency data is decomposed twice to obtain denoised high-frequency data; this helps to accurately extract the dynamic characteristics of battery capacity attenuation without being interfered by noise. Then, the denoised high-frequency data is used as the input of the CNN-Transformer network to obtain a first prediction value; and the low-frequency data is used as the input of the Transformer network to obtain a second prediction value; the first and second prediction values are combined as the prediction result of the lithium-ion battery lifespan; the RUL is predicted using a deep learning method, and the prediction accuracy is improved by accurately capturing global and local signal characteristics.

Description

Technical Field

[0001] The present invention relates to the technical field of battery health prediction, and particularly to a method and system for predicting the life of a lithium-ion battery based on deep learning. Background Art

[0002] Lithium-ion batteries have become an effective energy storage solution and are widely used in various fields such as electric vehicles, renewable energy, and grid energy storage. However, these batteries face a major challenge during deployment, which is to predict their remaining useful life (RUL). The remaining useful life refers to the time during which the battery can effectively store and release energy until its capacity drops to a certain percentage (usually 70% to 80%) of its initial capacity. The life of a lithium-ion battery is affected by many factors, including the number of charge-discharge cycles, the battery operating temperature, the depth of discharge (DoD), the state of health (SoH), and the charge-discharge rate (C-rate). Over time, these factors cause the battery capacity to gradually decrease, leading to battery aging and eventually its failure, which may also pose safety hazards to the energy storage system. Therefore, accurately predicting the life of a lithium-ion battery is a complex multi-factor, multi-stage, and dynamic problem.

[0003] Generally, there are three types of methods for RUL prediction, including model-based, data-driven, and hybrid methods. For model-based methods, a degradation model of the battery must be established based on prior knowledge or physical laws; these models can be further divided into equivalent circuit models, filtering models, and stochastic process models. Data-driven methods directly infer the degradation state from battery monitoring data and predict RUL. Therefore, this method does not require electrochemical analysis or prior knowledge, which differentiates it from model-based methods. Recently, various intelligent algorithms, such as statistical techniques, machine learning, and deep learning, etc. Hybrid methods combine different methods with the aim of overcoming the limitations of various single methods and making better use of all available information to improve the accuracy of diagnosis and prediction. Hybrid methods can be divided into two categories: hybrid methods based on filtering techniques (Kalman filter, particle filter, and variants) and hybrid methods based on intelligent algorithms such as machine learning.

[0004] The capacity regeneration (CR) phenomenon refers to the sudden recovery of the degraded capacity of the battery after a long rest period between charge-discharge cycles. However, the above research did not consider the capacity regeneration phenomenon in the battery degradation process, which can change the degradation process and thus change the time to reach the end-of-life, seriously affecting the degradation modeling and RUL prediction. Summary of the Invention

[0005] (I) Technical Problems to be Solved

[0006] In view of the deficiencies of the prior art, the present invention provides a method and system for predicting the life of a lithium-ion battery based on deep learning, which solves the technical problem of ignoring the capacity regeneration phenomenon and affecting the battery degradation process.

[0007] (2) Technical solutions

[0008] To achieve the above objectives, the present invention is realized through the following technical solutions:

[0009] A method for predicting the life of a lithium-ion battery based on deep learning, including:

[0010] Performing a first decomposition on the original signal data to obtain an IMF sequence distributed from high to low in frequency;

[0011] According to the IMF sequence, obtaining a high-low frequency demarcation point, and accordingly dividing and obtaining high-frequency data and low-frequency data;

[0012] Performing a second decomposition on the high-frequency data to obtain denoised high-frequency data;

[0013] Using the denoised high-frequency data as the input of the CNN-Transformer network to obtain a first prediction value; and using the low-frequency data as the input of the Transformer network to obtain a second prediction value;

[0014] Combining the first and second prediction values as the prediction result of the lithium-ion battery life.

[0015] Preferably, using the adaptive noise complete ensemble empirical mode decomposition algorithm to perform a first decomposition on the original signal data, including:

[0016] S11. Adding white noise to the original signal data, and performing empirical mode decomposition to obtain the first IMF1;

[0017] Y i (t) = Y(t) + ε0ω i (t), i = 1, 2,..., N

[0018]

[0019] Among them, N is the sample sequence length, Y(t) is the original signal data, ω i (t) is Gaussian white noise that satisfies the standard normal distribution, ε0 is the initial weight coefficient for adding white noise, and Y i (t) is the data obtained by adding white noise ω i (t) and a signal-to-noise ratio of ε0 to the original sequence Y(t), and the operator E i (·) is to solve the i-th eigenmode component of the empirical mode decomposition;

[0020] S12. Obtain the first residual component r1(t);

[0021] r1(t) = Y(t) - IMF1

[0022] S13. Add the adaptive white noise to the first residual component, and perform empirical mode decomposition to obtain the second IMF2;

[0023]

[0024] where ε1 is the weight coefficient of the white noise added for the first time;

[0025] S14. By analogy with S12 and S13, for j = 1, 2,..., J, the j-th residual component is expressed as:

[0026] r j = r j-1 (t) - IMF j

[0027] Add the adaptive white noise to the j-th residual component, and perform empirical mode decomposition to obtain the (j + 1)-th IMF j+1 ;

[0028]

[0029] S15. Repeat S14 until the residual component is less than the preset threshold and cannot be further decomposed, then stop the algorithm decomposition and output the IMF sequence distributed from high frequency to low frequency;

[0030]

[0031] where r(t) is the final residual component.

[0032] Preferably, the high and low frequency demarcation point is obtained by using the continuous mean square error algorithm.

[0033] Preferably, variational mode decomposition is used to perform secondary decomposition on the high-frequency data, including:

[0034] S31. Assume that the high-frequency data is decomposed into K modes, with the goal of minimizing the sum of the estimated bandwidths of each mode, and the optimal solution is obtained by using the alternating direction multiplier; the corresponding constrained variational expression is:

[0035]

[0036] where Y HF (t) is the high-frequency data, δ is the Dirac function, * is the convolution operation; u k is the mode function, and {u k}: = {u1, u2,..., uK}; ω k is the center frequency of each modal component, and {ω k} := {ω1, ω2, …, ω K};

[0037] S32. Transform the above equality-constrained optimization problem into an unconstrained optimization problem through the augmented Lagrangian function:

[0038]

[0039] where λ is the Lagrange multiplier and α is the penalty factor;

[0040] S33. Use the alternating direction multiplier method to solve S32 to optimize the (n + 1)-th iteration step in S31:

[0041]

[0042] where argmin is the variable value when the objective function takes the minimum value, and p is the noise tolerance; S34. Use Parseval's theorem to solve S33 in the frequency domain:

[0043]

[0044] Take the finally obtained z0 = {u1, u2, …, u K} as the high-frequency data after noise reduction.

[0045] Preferably, the process of obtaining the first predicted value includes:

[0046] S411. Use a double convolutional layer to perform local feature extraction on the high-frequency data after noise reduction;

[0047] z1 = ReLU(W1 * z0 + b)

[0048] z2 = LPPooling(ReLU(W2 * z1 + b))

[0049] where ReLU represents the non-linear activation function, LPPooling is the pooling operation, W1 and W2 represent the filter weights, * represents the convolution operation, z0 is the high-frequency data after noise reduction, z1 is the output of the first convolutional layer and serves as the input to the second convolutional layer, z2 is the output of the second convolutional layer after pooling, and b is the bias value;

[0050] S412. Transmit the result after convolution and pooling to the encoding layer of the Transformer to obtain the first predicted value, where the encoding layer includes a multi-head attention layer, two normalization layers, and a feed-forward neural network; the specific process of obtaining includes:

[0051] S4121. In the multi-head attention layer, each single-head attention represents the input matrix z2 with Q1, K1, and V1 obtained from three linear mapping layers;

[0052]

[0053] where attention is the attention mechanism, is the transposed matrix of K1, is the scaling factor for normalization, and softmax is the activation function;

[0054] Combine the output results of each self-attention to obtain the final output of the multi-head attention layer;

[0055]

[0056] where, are all trained weight matrices, is the model weight matrix for joint training, is the output result of the multi-head attention, head i is the output result of the i-th self-attention;

[0057] S4122. After passing the output result of the multi-head attention layer through the first normalization layer, feed it to the feed-forward neural network for processing; where the feed-forward neural network sequentially includes a first linear transformation layer, a linear unit activation function, and a second linear transformation layer for processing;

[0058]

[0059] where, is the normalized output of the multi-head attention layer in the j-th encoder, LN represents the normalization process, W3 and W4 are the weight parameters of the FFN layer, and b2 is the bias value;

[0060] S4123. After passing the output result of the feed-forward neural network through the second normalization layer, transmit it to the linear regression layer for processing;

[0061]

[0062] where, z output_1 is the first predicted value; σ is the sigmoid activation function, z LN is the input after normalization, and are the training parameters of the regressor.

[0063] Preferably, the process of obtaining the second predicted value includes:

[0064] Transmit the low-frequency data to the encoding layer of the Transformer to obtain a second predicted value, where the encoding layer includes a multi-head attention layer, two normalization layers, and a feed-forward neural network; the obtaining process specifically includes:

[0065] S421. In the multi-head attention layer, each single-head attention uses the low-frequency data Y LF (t) is represented by Q2, K2, and V2 obtained from three linear mapping layers;

[0066]

[0067] where attention is the attention mechanism, is the transposed matrix of K2, is the scaling factor for normalization, and softmax is the activation function;

[0068] Combine the output results of each self-attention to obtain the final output of the multi-head attention layer;

[0069]

[0070] where, are all trained weight matrices, is the jointly trained model weight matrix, is the output result of the multi-head attention, and head i is the output result of the i-th self-attention;

[0071] S422. After passing the output result of the multi-head attention layer through the first normalization layer, feed it to the feed-forward neural network for processing; where the feed-forward neural network sequentially includes a first linear transformation layer, a linear unit activation function, and a second linear transformation layer for processing;

[0072]

[0073] where, is the normalized output of the multi-head attention layer in the j-th encoder, LN represents the normalization process, W3 and W4 are the weight parameters of the FFN layer, and b2 is the bias value;

[0074] S423. After passing the output result of the feed-forward neural network through the second normalization layer, transmit it to the linear regression layer for processing;

[0075]

[0076] where, z output_2 is the second predicted value; σ is the sigmoid activation function, and z LN is the input after normalization, and They are the training parameters of the regressor.

[0077] Preferably, when training the CNN-Transformer network, the mean absolute value error L MSE_1 is selected as the first loss function;

[0078]

[0079] where, L MSE_1 is the first loss function, is the first predicted value at the i-th charge and discharge cycle point, is the true value at the i-th charge and discharge cycle point, and N is the sequence length of the high-frequency data;

[0080] Preferably, when training the Transformer network, the mean absolute value error L MSE_2 is selected as the second loss function;

[0081]

[0082] where, L MSE_2 is the first loss function, is the second predicted value at the i-th charge and discharge cycle point, is the true value at the i-th charge and discharge cycle point, and N is the sequence length of the low-frequency data.

[0083] A lithium-ion battery life prediction system based on deep learning, comprising:

[0084] A primary decomposition module, configured to perform primary decomposition on the original signal data to obtain an IMF sequence distributed from high to low in frequency;

[0085] A division module, configured to obtain a high-low frequency demarcation point according to the IMF sequence, and divide and obtain high-frequency data and low-frequency data based on this;

[0086] A secondary decomposition module, configured to perform secondary decomposition on the high-frequency data to obtain denoised high-frequency data;

[0087] A prediction module, configured to use the denoised high-frequency data as the input of the CNN-Transformer network to obtain a first predicted value; and use the low-frequency data as the input of the Transformer network to obtain a second predicted value;

[0088] A merging module, configured to merge the first and second predicted values as the lithium-ion battery life prediction result.

[0089] A storage medium stores a computer program for predicting the lifespan of a lithium-ion battery based on deep learning. Among them, the computer program enables a computer to execute the lithium-ion battery lifespan prediction method described above.

[0090] An electronic device includes:

[0091] One or more processors; a memory; and one or more programs, where the one or more programs are stored in the memory and configured to be executed by the one or more processors. The programs include those for executing the lithium-ion battery lifespan prediction method described above.

[0092] (III) Beneficial effects

[0093] The present invention provides a method and system for predicting the lifespan of a lithium-ion battery based on deep learning. Compared with the prior art, it has the following beneficial effects:

[0094] In the present invention, first, the original signal data is decomposed once to obtain an IMF sequence distributed from high to low in frequency; the high and low frequency demarcation points are obtained, and based on this, the IMF sequence is divided to obtain high-frequency data and low-frequency data; the high-frequency data is decomposed twice to obtain denoised high-frequency data; this helps to accurately extract the dynamic characteristics of battery capacity attenuation without being interfered by noise. Then, the denoised high-frequency data is used as the input of the CNN-Transformer network to obtain a first prediction value; and the low-frequency data is used as the input of the Transformer network to obtain a second prediction value; the first and second prediction values are combined as the prediction result of the lithium-ion battery lifespan; the RUL prediction is carried out using a deep learning method, and the prediction accuracy is improved by accurately capturing global and local signal characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0096] Figure 1 It is a block diagram of a method for predicting the lifespan of a lithium-ion battery based on deep learning provided by an embodiment of the present invention;

[0097] Figure 2 It is a flowchart of an adaptive noise complete ensemble empirical mode decomposition provided by an embodiment of the present invention;

[0098] Figure 3Schematic diagram of a structure of an improved Transformer network provided by an embodiment of the present invention;

[0099] Figure 4 Schematic diagram of a structure of a lithium battery life prediction system based on data denoising and deep learning provided by an embodiment of the present invention. Detailed implementation manners

[0100] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Apparently, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0101] The embodiments of the present application solve the technical problem of ignoring the capacity regeneration phenomenon and affecting the battery degradation process by providing a lithium-ion battery life prediction method and system based on deep learning.

[0102] The general idea of the technical solutions in the embodiments of the present application to solve the above technical problems is as follows:

[0103] Since the capacity of a lithium-ion battery will rapidly decrease and enter a slow decay stage until it is completely degraded after the capacity increases. Therefore, the embodiments of the present invention recognize that in order to accurately predict the remaining service life of the battery, two problems need to be solved: one is to identify the capacity regeneration phenomenon, and the other is to capture the capacity degradation trend. Although this phenomenon makes life prediction complicated, the research on understanding and suppressing the regeneration phenomenon is important to improve the life and reliability of lithium-ion batteries.

[0104] Specifically: to solve the battery capacity regeneration problem and improve the prediction accuracy of RUL, the embodiments of the present invention propose a prediction method based on the combination of complete ensemble empirical mode decomposition with adaptive white noise (CEEMDAN), variational mode decomposition (VMD) and an improved Transformer. Among them, the CEEMDAN method and the VMD method are used to separate the CR problem and denoise the data, which is convenient for subsequent use of deep learning methods for RUL prediction.

[0105] In addition, using only a single neural network to extract global degradation and local volatility features for training and prediction in the prior art will lead to a decrease in accuracy and is prone to falling into local minima. The embodiments of the present invention also consider the problem that the Transformer model has weak ability to capture local signal features, and improve the Transformer model so that it can accurately capture global and local signal features to improve the accuracy of prediction results. Correspondingly, the Transformer model is designed as a lightweight deep neural network, aiming to achieve a balance between model complexity and performance.

[0106] To better understand the above technical solutions, the above technical solutions will be described in detail below in conjunction with the accompanying drawings of the specification and specific embodiments.

[0107] Embodiment:

[0108] As Figure 1 shown, the embodiments of the present invention provide a method for predicting the life of a lithium-ion battery based on deep learning, including:

[0109] S1. Perform a first decomposition on the original signal data to obtain an IMF sequence distributed from high to low in frequency;

[0110] S2. According to the IMF sequence, obtain the high and low frequency demarcation points, and divide and obtain high-frequency data and low-frequency data accordingly;

[0111] S3. Perform a second decomposition on the high-frequency data to obtain denoised high-frequency data;

[0112] S4. Use the denoised high-frequency data as the input of the CNN-Transformer network to obtain a first prediction value; and use the low-frequency data as the input of the Transformer network to obtain a second prediction value;

[0113] S5. Combine the first and second prediction values as the prediction result of the life of the lithium-ion battery.

[0114] In the embodiments of the present invention, the CEEMDAN method and the VMD method are used to separate the CR problem and denoise the data, which is convenient for subsequent use of deep learning methods for RUL prediction; the Transformer model is improved so that it can accurately capture global and local signal features to improve the accuracy of prediction results.

[0115] Next, each step of the above solution will be introduced in detail:

[0116] In step S1, perform a first decomposition on the original signal data to obtain an IMF sequence distributed from high to low in frequency.

[0117] Empirical Mode Decomposition (EMD) is an adaptive data analysis method for signal decomposition, which can decompose any signal into a set of intrinsic modes called Intrinsic Mode Functions (IMFs). CEEMDAN is an improvement of the EMD algorithm, which introduces the concepts of ensemble and adaptive noise suppression, making the decomposition more efficient.

[0118] Therefore, in this step, the adaptive noise complete ensemble empirical mode decomposition algorithm as shown in Figure 2 is used to perform one decomposition on the original signal data, including:

[0119] S11. Add white noise to the original signal data and perform empirical mode decomposition to obtain the first IMF1;

[0120] Y i (t) = Y(t) + ε0ω i (t), i = 1, 2,..., N

[0121]

[0122] where N is the length of the sample sequence, Y(t) is the original signal data, ω i (t) is Gaussian white noise that satisfies the standard normal distribution, ε0 is the initial weight coefficient for adding white noise, and Y i (t) is the data obtained by adding white noise ω i (t) and a signal-to-noise ratio of ε0 to the original sequence Y(t). Define the operator E i (·) as solving the i-th intrinsic mode component of the empirical mode decomposition;

[0123] S12. Obtain the first residual component r1(t);

[0124] r1(t) = Y(t) - IMF1

[0125] S13. Add adaptive white noise to the first residual component and perform empirical mode decomposition to obtain the second IMF2;

[0126]

[0127] where ε1 is the weight coefficient for adding white noise for the first time;

[0128] S14. By analogy with S12 and S13, for j = 1, 2,..., J, the j-th residual component is expressed as:

[0129] r j = r j-1 (t) - IMF j

[0130] Add adaptive white noise to the j-th residual component and perform empirical mode decomposition to obtain the (j + 1)-th IMF j+1 ;

[0131]

[0132] S 15. Repeat S 14 until the residual component is less than the preset threshold and cannot be further decomposed, then stop the algorithm decomposition and output the IMF sequence distributed from high to low in frequency;

[0133]

[0134] where r(t) is the final residual component.

[0135] In the embodiment of the present invention, aiming at the battery capacity regeneration problem, CEEMDAN can adaptively suppress the residual white noise in the signal and decompose the non-stationary signal into multiple IMFs and residual components.

[0136] In step S2, according to the IMF sequence, obtain the high-frequency and low-frequency demarcation points, and divide and obtain high-frequency data and low-frequency data accordingly.

[0137] After decomposition by CEEMDAN, an IMF sequence distributed from high to low in frequency is obtained. Since the high-frequency signal is the part dominated by noise, that is, the capacity regeneration part, in this step, the high-frequency and low-frequency demarcation points are calculated by using the continuous mean square error (CMES) algorithm. The specific calculation formula is as follows:

[0138]

[0139] After obtaining the mean square error between the continuously reconstructed signals by using the above formula, the reconstructed signal corresponding to the global minimum value is the best estimate of the effective signal. That is, the IMF component here can be used as the demarcation point K between the high-frequency noise interference and the low-frequency signal, and the high-frequency data and low-frequency data are respectively expressed as follows:

[0140]

[0141]

[0142] In step S3, perform secondary decomposition on the high-frequency data to obtain the denoised high-frequency data.

[0143] Although the decomposed high-frequency data and low-frequency data can be obtained through steps S1 and S2, since there are still a few noise situations in the high-frequency data, in this step, variational mode decomposition (VMD) is used to perform secondary decomposition on the high-frequency data. The specific steps are as follows:

[0144] S31. Assume that the high-frequency data is decomposed into K modes. With the goal of minimizing the sum of the estimated bandwidths of each mode, the optimal solution is solved by the alternating direction multiplier method; the corresponding constrained variational expression is:

[0145]

[0146] Among them, Y HF (t) is the high-frequency data, δ is the Dirac function, and * is the convolution operation; u k is the mode function, and {u k}: = {u1, u2, …, u K}; ω k is the center frequency of each mode component, and {ω k}: = {ω1, ω2, …, ω K};

[0147] S32. Transform the above equality-constrained optimization problem into an unconstrained optimization problem through the augmented Lagrangian function:

[0148]

[0149]

[0150] Among them, λ is the Lagrange multiplier, and α is the penalty factor;

[0151] S33. Solve S32 using the alternating direction multiplier method to optimize the (n + 1)-th iteration step in S31:

[0152]

[0153] Among them, argmin is the variable value when the objective function takes the minimum value, and p is the noise tolerance; S34. Solve S33 in the frequency domain using Parseval's theorem:

[0154]

[0155] Take the finally obtained z0 = {u1, u2, …, u K} as the denoised high-frequency data.

[0156] In the embodiment of the present invention, VMD is used for secondary decomposition of high-frequency data with high information complexity to solve the problems that the subsequences generated by a single decomposition method have irregular frequency fluctuation ranges and dynamic complexity.

[0157] In step S4, use the denoised high-frequency data as the input of the CNN-Transformer network to obtain the first predicted value; and use the low-frequency data as the input of the Transformer network to obtain the second predicted value.

[0158] The denoised high-frequency data and low-frequency data are obtained through steps S2 and S3. Since the high-frequency data has large fluctuations, in this step, an improved Transformer model is used to train the battery degradation data and the predicted capacity recovery part, as Figure 3 shown; considering that the low-frequency data represents the global degradation trend of the battery and has low data fluctuation frequency, in this step, the Transformer model is used to directly capture the global dependence features of the low-frequency data to improve the accuracy of the model. Specifically as follows:

[0159] (1) Based on the high-frequency data, the process of obtaining the first predicted value includes:

[0160] S411. Use a double convolutional layer to perform local feature extraction on the denoised high-frequency data;

[0161] z1 = ReLU(W1 * z0 + b)

[0162] z2 = LPPooling(ReLU(W2 * z1 + b))

[0163] where ReLU represents a non-linear activation function, LPPooling is a pooling operation, W1 and W2 represent filter weights, * represents a convolutional operation, z0 is the denoised high-frequency data, z1 is the output of the first convolutional layer and serves as the input to the second convolutional layer, z2 is the output of the second convolutional layer after pooling, and b is the bias value;

[0164] S412. Transmit the result after convolution and pooling to the encoding layer of the Transformer to obtain the first predicted value, where the encoding layer includes a multi-head attention layer, two normalization layers, and a feed-forward neural network; the obtaining process specifically includes:

[0165] S4121. In the multi-head attention layer, each single-head attention represents the input matrix z2 by Q1, K1, and V1 obtained from three linear mapping layers;

[0166]

[0167] where attention is the attention mechanism, is the transpose matrix of K1, is the scaling factor for normalization, and softmax is the activation function;

[0168] Combine the output results of each self-attention to obtain the final output of the multi-head attention layer;

[0169]

[0170] where, All are trained weight matrices, is the weight matrix of the jointly trained model, is the output result of the multi-head attention, head i is the output result of the i-th self-attention;

[0171] S4122. After passing the output result of the multi-head attention layer through the first normalization layer, it is fed to the feed-forward neural network for processing; where the feed-forward neural network sequentially includes a first linear transformation layer, a linear unit activation function, and a second linear transformation layer for processing;

[0172]

[0173] Among them, is the normalized output of the multi-head attention layer in the j-th layer encoder, LN represents the normalization process, W3 and W4 are the weight parameters of the FFN layer, and b2 is the bias value;

[0174] S4123. After passing the output result of the feed-forward neural network through the second normalization layer, it is transmitted to the linear regression layer for processing;

[0175]

[0176] Among them, z output_1 is the first predicted value; σ is the sigmoid activation function, z LN is the input after normalization, and are the training parameters of the regressor.

[0177] Specifically, in the stage of training the CNN-Transformer network, the mean absolute squared error L MSE_1 is selected as the first loss function;

[0178]

[0179] Among them, L MSE_1 is the first loss function, is the first predicted value at the i-th charge and discharge cycle point, is the true value at the i-th charge and discharge cycle point, and N is the sequence length of the high-frequency data.

[0180] (2) The process of obtaining the second predicted value based on the low-frequency data includes:

[0181] Transmit the low-frequency data to the encoding layer of the Transformer to obtain the second predicted value, where the encoding layer includes a multi-head attention layer, two normalization layers, and a feed-forward neural network; the specific obtaining process includes:

[0182] S421. In the multi-head attention layer, each single-head attention takes the low-frequency data Y LF (t) is represented by Q2, K2, and V2 obtained from three linear mapping layers;

[0183]

[0184] where attention is the attention mechanism, is the transpose matrix of K2, is the scaling factor for normalization, and softmax is the activation function;

[0185] Combine the output results of each self-attention to obtain the final output of the multi-head attention layer;

[0186]

[0187] where, are all trained weight matrices, is the model weight matrix for joint training, is the output result of the multi-head attention, and head i is the output result of the i-th self-attention;

[0188] S422. After passing the output result of the multi-head attention layer through the first normalization layer, feed it to the feed-forward neural network for processing; where the feed-forward neural network sequentially includes a first linear transformation layer, a linear unit activation function, and a second linear transformation layer for processing;

[0189]

[0190] where, is the normalized output of the multi-head attention layer in the j-th encoder, LN represents the normalization process, W3 and W4 are the weight parameters of the FFN layer, and b2 is the bias value;

[0191] S423. After passing the output result of the feed-forward neural network through the second normalization layer, transmit it to the linear regression layer for processing;

[0192]

[0193] where z output_2 is the second predicted value; σ is the sigmoid activation function, and z LN is the input after normalization, and are the training parameters of the regressor.

[0194] Specifically, in the stage of training the Transformer network, the mean absolute squared error L is also selectedMSE_2 as the second loss function;

[0195]

[0196] where L MSE_2 is the first loss function, is the second predicted value at the i-th charge-discharge cycle point, is the true value at the i-th charge-discharge cycle point, and N is the sequence length of the low-frequency data.

[0197] The embodiment of the present invention adopts a lightweight Transformer model, combines it with other deep learning architectures, and can accurately capture the spatio-temporal characteristics of the preprocessed data when predicting the RUL of the battery.

[0198] In step S5, the first and second predicted values are combined as the prediction result of the lithium-ion battery life.

[0199] As Figure 4 shown, the embodiment of the present invention provides a lithium-ion battery life prediction system based on deep learning, including:

[0200] A primary decomposition module for performing primary decomposition on the original signal data to obtain an IMF sequence distributed from high to low in frequency;

[0201] A partitioning module for obtaining a high-low frequency demarcation point according to the IMF sequence, and partitioning and obtaining high-frequency data and low-frequency data accordingly;

[0202] A secondary decomposition module for performing secondary decomposition on the high-frequency data to obtain denoised high-frequency data;

[0203] A prediction module for using the denoised high-frequency data as the input of the CNN-Transformer network to obtain a first predicted value; and using the low-frequency data as the input of the Transformer network to obtain a second predicted value;

[0204] A merging module for merging the first and second predicted values as the prediction result of the lithium-ion battery life.

[0205] The embodiment of the present invention provides a storage medium storing a computer program for predicting the life of a lithium-ion battery based on deep learning, wherein the computer program causes a computer to execute the above-mentioned lithium-ion battery life prediction method.

[0206] The embodiment of the present invention provides an electronic device, including:

[0207] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including those for executing the lithium-ion battery life prediction method as described above

[0208] It can be understood that the lithium-ion battery life prediction system, storage medium and electronic device provided by the embodiments of the present invention based on deep learning correspond to the lithium-ion battery life prediction method provided by the embodiments of the present invention. For the explanations, examples, beneficial effects and other parts of the relevant content, reference can be made to the corresponding parts in the lithium-ion battery life prediction method, which will not be elaborated here.

[0209] In summary, compared with the prior art, the following beneficial effects are achieved:

[0210] 1. CEEMDAN is used to decompose the battery data, which helps to accurately extract the dynamic characteristics of battery capacity attenuation without being disturbed by noise.

[0211] 2. By performing secondary decomposition and noise reduction processing on the high-frequency data after the first decomposition through variational mode decomposition, the prediction accuracy of the high-frequency data sequence can be significantly improved.

[0212] 3. Without using a decoder, the model based on the Transformer encoder will be simpler and lighter than other models, saving time and cost.

[0213] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including the element.

[0214] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the lifespan of a lithium-ion battery based on deep learning, characterized in that, Including: Performing a first decomposition on the original signal data to obtain an IMF sequence distributed from high to low in frequency; According to the IMF sequence, obtaining a high-low frequency demarcation point, and accordingly dividing to obtain high-frequency data and low-frequency data; Performing a second decomposition on the high-frequency data to obtain denoised high-frequency data; Using the denoised high-frequency data as the input of the CNN-Transformer network to obtain a first prediction value; and using the low-frequency data as the input of the Transformer network to obtain a second prediction value; Combining the first and second prediction values as the lithium-ion battery life prediction result; Performing a second decomposition on the high-frequency data using variational mode decomposition, including: S31. Assume that the high-frequency data is decomposed into K modes, aiming to minimize the sum of the estimated bandwidths of each mode, and solving the optimal solution through the alternating direction multiplier method; the corresponding constrained variational expression is: Among them, Y HF (t) is high-frequency data, δ is the Dirac function, and * is the convolution operation; u k is a modal function, and {u k}: = {u1, u2, …, u K}; ω k is the central frequency of each modal component, and {ω k}: = {ω1, ω2, …, ω K}; S32. Transforming the above equality-constrained optimization problem into an unconstrained optimization problem through the augmented Lagrangian function: where λ is the Lagrange multiplier and α is the penalty factor; S33. Using the alternating direction multiplier method to solve S32 to optimize the (n + 1)-th iteration step in S31: where argmin is the variable value when the objective function takes the minimum value, and p is the noise tolerance; S34. Using Parseval's theorem to solve S33 in the frequency domain: Take the finally obtained z0 = {u1, u2,..., u K} as the high-frequency data after noise reduction.

2. The method for predicting the service life of a lithium-ion battery according to claim 1, wherein Performing a first decomposition on the original signal data using the adaptive noise complete ensemble empirical mode decomposition algorithm, including: S11. Adding white noise to the original signal data and performing empirical mode decomposition to obtain the first IMF1; Y i (t) = Y(t) + ε0ω i (t), i = 1, 2, …, N Among them, N is the length of the sample sequence, Y(t) is the original signal data, ω i (t) is Gaussian white noise that satisfies the standard normal distribution, ε0 is the initial weight coefficient for adding white noise, Y i (t) is the data obtained by adding white noise ω i (t) and a signal-to-noise ratio of ε0 to the original sequence Y(t), and the operator E i (·) is to solve the i-th eigenmode component of the empirical mode decomposition; S12. Obtaining the first residual component r1(t); r1(t) = Y(t) - IMF1 S13. Adding adaptive white noise to the first residual component and performing empirical mode decomposition to obtain the second IMF2; where ε1 is the weight coefficient of the white noise added for the first time; S14. By analogy with S12 and S13, for j = 1, 2,..., J, the j-th residual component is expressed as: r j = r j-1 (t) - IMF j Add adaptive white noise to the j-th residual component and perform empirical mode decomposition to obtain the (j + 1)-th IMF j+1 ; S15. Repeating S14 until the residual component is less than the preset threshold and cannot be further decomposed, stopping the algorithm decomposition and outputting the IMF sequence distributed from high to low in frequency; where r(t) is the final residual component.

3. The method for predicting the lifespan of a lithium-ion battery according to claim 1, wherein Using the continuous mean square error algorithm to obtain the high-low frequency demarcation point.

4. The method for predicting the lifespan of a lithium-ion battery according to claim 1, wherein The process of obtaining the first prediction value includes: S411. Using a double convolutional layer to perform local feature extraction on the denoised high-frequency data; z1 = ReLU(W1 * z0 + b) z2 = LPPooling(ReLU(W2 * z1 + b)) where ReLU represents the non-linear activation function, LPPooling is the pooling operation, W1 and W2 represent the filter weights, * represents the convolutional operation, z0 is the denoised high-frequency data, z1 is the output of the first convolutional layer and serves as the input of the second convolutional layer, z2 is the output of the second convolutional layer after pooling, and b is the bias value; S412. Transmit the result after convolution and pooling to the encoding layer of the Transformer to obtain a first prediction value, where the encoding layer includes a multi-head attention layer, two normalization layers, and a feed-forward neural network; the obtaining process specifically includes: S4121. In the multi-head attention layer, each single-head attention represents the input matrix z2 by Q1, K1, and V1 obtained from three linear mapping layers; where attention is the attention mechanism, is the transposed matrix of K1, is the scaling factor for normalization, and softmax is the activation function; Combine the output results of each self-attention to obtain the final output of the multi-head attention layer; Among them, are all trained weight matrices, is the model weight matrix obtained by joint training, is the output result of multi-head attention, where head i is the output result of the i-th self-attention; S4122. After passing the output result of the multi-head attention layer through the first normalization layer, feed it to the feed-forward neural network for processing; where the feed-forward neural network sequentially includes a first linear transformation layer, a linear unit activation function, and a second linear transformation layer for processing; Among them, is the normalized output of the multi-head attention layer in the j-th layer encoder, LN represents normalization processing, W3 and W4 are the weight parameters of the FFN layer, and b2 is the bias value; S4123. After passing the output result of the feed-forward neural network through the second normalization layer, transmit it to the linear regression layer for processing; where z output_1 is the first predicted value; σ is the sigmoid activation function, and z LN is the input after normalization, and are the training parameters of the regressor.

5. The method for predicting the lifespan of a lithium-ion battery according to claim 1, wherein The obtaining process of the second prediction value includes: Transmit the low-frequency data to the encoding layer of the Transformer to obtain a second prediction value, where the encoding layer includes a multi-head attention layer, two normalization layers, and a feed-forward neural network; the obtaining process specifically includes: S421. In the multi-head attention layer, each single-head attention represents the low-frequency data Y LF (t) with Q2, K2, and V2 obtained from three linear mapping layers; where attention is the attention mechanism, is the transposed matrix of K2, is the scaling factor for normalization, and softmax is the activation function; Combine the output results of each self-attention to obtain the final output of the multi-head attention layer; Among them, are all trained weight matrices, is the weight matrix of the jointly trained model, is the output result of the multi-head attention, and head i is the output result of the i-th self-attention; S422. After passing the output result of the multi-head attention layer through the first normalization layer, feed it to the feed-forward neural network for processing; where the feed-forward neural network sequentially includes a first linear transformation layer, a linear unit activation function, and a second linear transformation layer for processing; Among them, is the normalized output of the multi-head attention layer in the j-th layer encoder, LN represents normalization processing, W3 and W4 are the weight parameters of the FFN layer, and b2 is the bias value; S423. After passing the output result of the feed-forward neural network through the second normalization layer, transmit it to the linear regression layer for processing; where z output_2 is the second predicted value; σ is the sigmoid activation function, and z LN is the input after normalization, and are the training parameters of the regressor.

6. The method for predicting the life of a lithium-ion battery according to claim 5, wherein When training the CNN-Transformer network, the mean absolute value error L MSE _1 is selected as the first loss function; Among them, L MSE _1 is the first loss function, is the first predicted value at the i-th charge-discharge cycle point, is the true value at the i-th charge-discharge cycle point, and N is the sequence length of the high-frequency data; When training the Transformer network and / or, select the mean absolute square error L MSE _2 as the second loss function; Among them, L MSE _2 is the first loss function, is the second predicted value at the i-th charge-discharge cycle point, is the true value at the i-th charge-discharge cycle point, and N is the sequence length of the low-frequency data.

7. A lithium-ion battery life prediction system based on deep learning, characterized in that, For executing the method for predicting the life of a lithium-ion battery according to claim 1, it includes: A primary decomposition module for performing primary decomposition on the original signal data to obtain an IMF sequence distributed from high to low in frequency; A partitioning module for obtaining a high-low frequency demarcation point according to the IMF sequence and partitioning to obtain high-frequency data and low-frequency data accordingly; A secondary decomposition module for performing secondary decomposition on the high-frequency data to obtain denoised high-frequency data; A prediction module for using the denoised high-frequency data as the input of the CNN-Transformer network to obtain a first prediction value; and using the low-frequency data as the input of the Transformer network to obtain a second prediction value; A merging module for merging the first and second prediction values as the prediction result of the life of the lithium-ion battery.

8. A storage medium, characterized in that, It stores a computer program for predicting the life of a lithium-ion battery based on deep learning, wherein the computer program causes the computer to execute the method for predicting the life of a lithium-ion battery according to any one of claims 1 to 6.

9. An electronic device, characterized in that, It includes: One or more processors; A memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs include those for executing the lithium-ion battery life prediction method according to any one of claims 1 to 6.

Images