Lithium ion battery health state estimation method based on Bayesian optimization CNN-Transformer

Abstract

The invention relates to the technical field of lithium ion battery state-of-health estimation, and discloses a Bayesian optimization CNN-Transform-based lithium ion battery state-of-health estimation method, which is used for obtaining lithium ion battery aging data: battery charge and discharge voltage, current, time and capacity data, building a CNN-Transform hybrid model in combination with a CNN model and a Transform model, performing hyper-parameter optimization, and estimating the state-of-health of a lithium ion battery based on the Bayesian optimization CNN-Transform hybrid model. And training the CNN-Transformer hybrid model, inputting the equal interval voltage charging time amount, and outputting the health state of the battery. According to the method, a battery charging voltage curve is analyzed, interval division is carried out on the curve, part of voltage charging time is extracted to serve as health characteristics, the SOH is estimated through CNN-Transformer, optimal CNN-Transformer algorithm hyper-parameters are searched through a Bayesian optimization method, and therefore the hyper-parameter space is explored efficiently under the condition that the number of evaluation times is smaller, and the evaluation efficiency is improved. Compared with a traditional hyper-parameter optimization method, hyper-parameters can be selected more intelligently, and a globally optimal solution can be found more quickly.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium-ion battery health state estimation, and in particular to a lithium-ion battery health state estimation method based on Bayesian optimization CNN-Transformer. Background Art

[0002] Lithium-ion batteries are widely used in the electric vehicle industry due to their high energy density, long life cycle, and no memory effect. However, as the battery is continuously charged and discharged, irreversible chemical reactions inside the battery continue to accumulate, causing the battery to age. In addition, as the usage time increases, side reactions also intensify, resulting in growth at the electrolyte interface and lithium plating, which leads to a decrease in battery performance and ultimately affects the reliability of the battery. Therefore, accurate estimation of the battery state of health (SOH) is an important guarantee for the safe use of batteries.

[0003] The data-driven method extracts features from the collected voltage, current, temperature and other data through charge and discharge experiments, and then establishes a mapping relationship with SOH to achieve SOH estimation of lithium-ion batteries. This method requires almost no basic theory and prior knowledge of batteries, and has high estimation accuracy, so it is widely used in battery SOH estimation. This method first needs to extract reasonable health features. The extraction of these features requires a complete charging curve. It is difficult to obtain complete charging curve data in practical applications because car owners prefer to charge the vehicle before the battery is fully discharged. Therefore, it is necessary to find a simple feature that is strongly correlated with SOH. On the other hand, in deep learning neural networks, convolutional neural networks (CNN), long short-term memory neural networks (LSTM) and GRU have received great attention in lithium-ion battery SOH estimation because of their excellent nonlinear mapping capabilities for complex systems. Although the deep learning-based methods in existing studies can guarantee high estimation accuracy, there are still problems in practical applications such as cumbersome feature extraction, complex model training process, huge computational load, and weak correlation and spatiality between features. Therefore, it is important to choose a multi-dimensional and multi-space neural network. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] In view of the shortcomings of the prior art, the present invention provides a lithium-ion battery health status estimation method based on Bayesian optimization CNN-Transformer, which has the advantages of accurate prediction and solves the above-mentioned technical problems.

[0006] (II) Technical solution

[0007] To achieve the above object, the present invention provides the following technical solution: a method for estimating the health status of a lithium-ion battery based on Bayesian optimization CNN-Transformer, comprising the following steps:

[0008] S1. Obtaining lithium-ion battery aging data: Performing cyclic charge and discharge tests on lithium-ion batteries until the discharge capacity is less than 80% of the nominal capacity, and recording the battery charge and discharge voltage, current, time and capacity data in real time;

[0009] S2, data processing: pre-process the voltage, current, time and capacity data obtained in S1 to obtain equal interval voltage charging time and health status;

[0010] S3. CNN-Transformer hybrid model construction: Combine CNN and Transformer models to build a CNN-transformer hybrid model;

[0011] S4, CNN-Transformer hybrid model optimization: Use the Bayesian optimization algorithm to optimize the hyperparameters of the CNN-Transformer hybrid model obtained in S3 to obtain the optimized CNN-Transformer hybrid model;

[0012] S5, CNN-Transformer hybrid model training: Take the training set formed by the equally spaced voltage charging time data obtained in step S2 to run the optimized CNN-Transformer hybrid model, find the correlation between the equally spaced voltage charging time and the health status, and obtain the trained CNN-Transformer hybrid model;

[0013] S6. Obtaining the battery health status: Based on the trained CNN-Transformer hybrid model obtained in step S5, inputting the equal-interval voltage charging time, outputting the health status, and obtaining the battery health status.

[0014] As a preferred technical solution of the present invention, the specific process of step S2 is as follows:

[0015] S2.1. Extract the charging voltage range of each cycle obtained from the test in S1 as [V start ,V stop ] charging time t, where V start Indicates the starting voltage, V stop Indicates stop voltage;

[0016] S2.2, each cycle charging time constitutes a model input {x};

[0017] S2.3. Extract the discharge capacity value of each cycle in S1 and calculate the state of health SOH:

[0018]

[0019] Among them, SOH i is the health status of the ith cycle, C n is the nominal capacity of the battery, C i is the actual capacity at the i-th cycle;

[0020] S2.4, input {x} in the model of S2.2 and the corresponding health status SOH in S2.3 to form a data set, and divide the data set into training set D according to the cross-validation method. train and the test set D test , the training set and test set are as follows:

[0021] Training set:

[0022]

[0023] Test set:

[0024]

[0025] Among them, x i ,y i are the i-th model input and the corresponding health status in the training set, respectively. are the i-th model input and the corresponding health status in the test set, and h represents the dimension.

[0026] As a preferred technical solution of the present invention, the step S3 of establishing a CNN-Transformer hybrid model includes the following steps:

[0027] S3.1. Construct a CNN network to extract spatial features. The CNN network includes convolutional layer, pooling layer, and fully connected layer. The corresponding expressions are as follows:

[0028]

[0029] Among them, h i represents the features extracted by the i-th convolution layer, x i represents the input data of the i-th layer convolution, σ represents the relu activation function, represents the convolution operation, w i represents the convolution kernel weight matrix in the i-th convolution layer, b i Represents the bias vector in the i-th convolutional layer;

[0030] S3.2. Use Transformer to extract temporal features from the output of the CNN network, which specifically includes the following steps:

[0031] S3.2.1, Transformer encoder converts the input data into different vectors and encodes each data;

[0032] S3.2.2, the encoder analyzes the dependencies in the sequence through the self-attention mechanism and generates an encoding vector containing feature information;

[0033] S3.2.3. The Transformer decoder uses the features extracted by the encoder and combines its own input to gradually decode and generate predicted time series data through self-attention and encoder-decoder attention mechanisms.

[0034] As a preferred technical solution of the present invention, the calculation expression of the encoder is as follows:

[0035]

[0036] Among them, PE(p,2i dim ) and PE(p,2i dim +1) represent two position encoding vectors, sin(*) represents the sine function, cos(*) represents the cosine function, p represents the position, i dim Denotes dimension, d model Represents the embedding dimension.

[0037] As a preferred technical solution of the present invention, the specific expression of using the Bayesian optimization algorithm in step S4 to optimize the hyperparameters of the CNN-Transformer hybrid model obtained in S3 is as follows:

[0038] f(x)~gp(m(x),k(x,x'))

[0039] p(f(x * )|D,x * )=N(μ(x * ),σ 2 (x * ))

[0040]

[0041] EI(x)=E[max(0,f(x)-f(X + ))]

[0042] UCB(x)=μ(X)+κσ(X)

[0043] Where f(x) represents the target function, x represents the input vector, k(x,x') represents the kernel function, m(x) represents the mean function, gp(m(x),k(x,x')) represents the Gaussian process, and p(f(x * )|D,x* ) means that given the data D and the new input vector x * When f(x * ), μ(x * ) represents the predicted mean, σ 2 (x * ) represents the new input vector x * The prediction variance, N(*) represents the normal distribution, k represents the target vector of the training set, k * Represents the new input vector x * The similarity vector with all training samples, Indicates k * The transpose of , K represents the kernel matrix, I represents the identity matrix, EI(x) represents the expected improvement at x, E[*] represents the mean calculation, max(0,f(x)-f(X + )) means taking the maximum value of two elements, f(X + ) represents the optimal value of the objective function in the known data, X + represents the current optimal point, X represents the current input point of interest, μ(X) represents the predicted mean at x, σ(X) represents the predicted standard deviation of the Gaussian process at x, UCB(x) represents the upper confidence bound at x, and κ represents the parameter that controls the trade-off between exploration and exploitation. represents the noise variance.

[0044] Compared with the prior art, the present invention provides a lithium-ion battery health status estimation method based on Bayesian optimization CNN-Transformer, which has the following beneficial effects:

[0045] The present invention analyzes the battery charging voltage curve, divides the curve into intervals, extracts partial voltage charging time as health features, estimates SOH using CNN-Transformer, and uses the Bayesian optimization method to search for the optimal CNN-Transformer algorithm parameters, thereby efficiently exploring the hyperparameter space with fewer evaluation times. Compared with traditional hyperparameter optimization methods, it can select hyperparameters more intelligently and find the global optimal solution faster. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a schematic diagram of SOH estimation results of different batteries of the present invention;

[0047] Figure 2 This is a schematic diagram of SOH estimation results in different voltage ranges of the present invention;

[0048] Figure 3 This is a schematic diagram comparing the SOH estimation results using different methods of the present invention;

[0049] Figure 4It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] See also Figure 1 - Figure 4 , a lithium-ion battery health state estimation method based on Bayesian optimization CNN-Transformer, comprising the following steps:

[0052] S1. Obtaining lithium-ion battery aging data: Performing cyclic charge and discharge tests on lithium-ion batteries until the discharge capacity is less than 80% of the nominal capacity, and recording the battery charge and discharge voltage, current, time and capacity data in real time;

[0053] S2, data processing: pre-process the voltage, current, time and capacity data obtained in S1 to obtain equal interval voltage charging time and health status;

[0054] S2.1. Extract the charging voltage range of each cycle obtained from the test in S1 as [V start ,V stop ] charging time t, where V start Indicates the starting voltage, V stop Indicates stop voltage;

[0055] S2.2, each cycle charging time constitutes a model input {x};

[0056] S2.3. Extract the discharge capacity value of each cycle in S1 and calculate the state of health SOH:

[0057]

[0058] Among them, SOH i is the health status of the ith cycle, C n is the nominal capacity of the battery, C i is the actual capacity at the i-th cycle;

[0059] S2.4, input {x} in the model of S2.2 and the corresponding health status SOH in S2.3 to form a data set, and divide the data set into training set D according to the cross-validation method. train and the test set D test , the training set and test set are as follows:

[0060] Training set:

[0061]

[0062] Test set:

[0063]

[0064] Among them, x i ,y i are the i-th model input and the corresponding health status in the training set, respectively. are the i-th model input and the corresponding health status in the test set, respectively, and h represents the dimension;

[0065] S3. CNN-Transformer hybrid model construction: Combine CNN and Transformer models to build a CNN-Transformer hybrid model;

[0066] CNN is a feedforward neural network with deep structure and convolution operation. It is often used to process image classification, image segmentation, target inspection, natural language and other fields. CNN has the advantages of local perception and weight sharing, which reduces the number of model parameters and improves the learning efficiency of the model. Through one-dimensional convolution operation, CNN can extract local features in time series and capture the trend and periodic changes of data. In addition, the parameter sharing mechanism of CNN makes it very efficient when processing one-dimensional time series such as battery data. By performing convolution operation on time series data, CNN can capture the correlation and change trend between different time points, so as to better understand the operating status and performance changes of the battery. CNN mainly includes convolution layer, pooling layer and fully connected layer. The convolution layer contains multiple convolution kernels, which are used to extract features of different time and space dimensions. The health feature data of lithium-ion batteries are calculated by multiple convolution kernels to obtain strongly correlated features, and then the output of the convolution layer is nonlinearly transformed by the activation function to improve the nonlinear expression ability of the model, as shown in the formula:

[0067]

[0068] Among them, h i represents the features extracted by the i-th convolution layer, x i represents the input data of the i-th layer convolution, σ represents the relu activation function, represents the convolution operation, w i represents the convolution kernel weight matrix in the i-th convolution layer, b iRepresents the bias vector in the i-th convolutional layer. Usually, convolution maps low-dimensional data to high-dimensional space for feature extraction. The extracted features are often of high dimension. In order to reduce the computational cost and prevent network overfitting, a pooling layer is usually added after the convolution, and finally the extracted features are output through the fully connected layer.

[0069] After the CNN layer extracts spatial features, Transformer is used to extract temporal features from the CNN layer output:

[0070] Unlike traditional models such as RNN and CNN, Transformer adds a self-attention mechanism and does not use the RNN sequential structure, so that the model can be trained in parallel and can draw the global dependency between input and output, which enables it to capture long-term dependencies more effectively. By paying attention to different parts of the input sequence, Transformer can well understand the relationship between the previous and subsequent sequences and generate accurate predictions. In processing time series data, the encoder first converts the input data into different vectors and encodes each data to capture the time series feature information. Subsequently, the encoder analyzes the dependencies in the sequence through the self-attention mechanism and generates an encoding vector containing feature information. Finally, the decoder uses the features extracted by the encoder and combines its own input (previous time prediction or additional input) to gradually decode and generate predicted time series data through self-attention and encoder-decoder attention mechanisms. Throughout the process, the encoder is responsible for extracting features from the time series, while the decoder generates accurate prediction results based on these features.

[0071] Typically, position encoders calculate position using sine and cosine functions at different frequencies, as shown below:

[0072]

[0073]

[0074] Among them, PE(p,2i dim ) and PE(p,2i dim +1) represent two position encoding vectors, sin(*) represents the sine function, cos(*) represents the cosine function, p represents the position, i dim Denotes dimension, d model represents the embedding dimension;

[0075] The self-attention mechanism is the core of Transformer. Each element of the sequence is compared with other elements in the sequence to obtain the attention weight between each data and other data. Multiply the input sequence with different weight matrices to obtain the matrix corresponding to each data:

[0076] q=xj W Q ,k=x j W K ,v=x j W V

[0077] Q represents the input information of the current position, which is used to calculate the attention weights of the current position and other positions. K represents the correlation of each position in the sequence, which is used to calculate the attention weights of the current position and other positions. V contains the information of each position in the sequence. The output of the current position is obtained by weighted summing the input information according to the calculated attention weights. W Q , W K , W V represents the weight matrix that transforms the input matrix into the query vector matrix q, the weight matrix that transforms the input matrix into the key vector matrix k, and the weight matrix that transforms the input vector into the value vector matrix v, x j represents the input data matrix;

[0078] To prevent the dot product result from being too large, you need to divide by As shown below:

[0079]

[0080] Among them, softmax represents the activation function The values ​​are normalized to a probability distribution so that they are all positive and sum to 1, d k represents the dimension of the key vector, It means to reduce the inner product value of the query vector and the key vector, so that the attention score of the corresponding position is calculated.

[0081] When processing battery sequence data, in order to extract multiple features and relationships in the sequence and enhance the model's ability to express data, a multi-head attention mechanism is often used. The multi-head self-attention mechanism consists of multiple self-attention heads, which are connected in series and then linearly transformed to finally obtain a multi-head self-attention output.

[0082] S4, CNN-Transformer hybrid model optimization: Use the Bayesian optimization algorithm to optimize the hyperparameters of the CNN-Transformer hybrid model obtained in S3 to obtain the optimized CNN-Transformer hybrid model;

[0083] The Bayesian optimization algorithm is used to optimize the number of neurons, batch size, learning rate, and number of iterations in the CNN-Transformer parallel hybrid model to improve the hyperparameter effect, reduce the amount of training calculations, and improve the model accuracy and adaptability.

[0084] In order to enable CNN-Transformer to achieve the best performance, the model parameters need to be tuned. Traditional hyperparameter optimization methods include grid search and random search. Although these methods are intuitive, they require a lot of model training and evaluation, especially in high-dimensional space, the computational cost and time will increase significantly. Bayesian optimization is a systematic and efficient method that can be used to automatically select the best hyperparameter configuration for machine learning models. Its core is to build a Gaussian model to approximate the objective function and use acquisition functions (such as expected improvement, confidence limit) to determine the next location to evaluate the hyperparameter. The calculation process can be expressed as:

[0085] f(x)~gp(m(x),k(x,x′))

[0086] p(f(x * )|D,x * )=N(μ(x * ),σ 2 (x * ))

[0087]

[0088] EI(x)=E[max(0,f(x)-f(X + ))]

[0089] UCB(x)=μ(X)+κσ(X)

[0090] Where f(x) represents the target function, x represents the input vector, k(x,x') represents the kernel function, m(x) represents the mean function, gp(m(x),k(x,x')) represents the Gaussian process, and p(f(x * )|D,x * ) means that given the data D and the new input vector x * When f(x * ), μ(x * ) represents the predicted mean, σ 2 (x * ) represents the new input vector x * The prediction variance, N(*) represents the normal distribution, k represents the target vector of the training set, k * Represents the new input vector x * The similarity vector with all training samples, Indicates k * The transpose of , K represents the kernel matrix, I represents the identity matrix, EI(x) represents the expected improvement at x, E[*] represents the mean calculation, max(0,f(x)-f(X + )) means taking the maximum value of two elements, f(X +) represents the optimal value of the objective function in the known data, X + represents the current optimal point, X represents the current input point of interest, μ(X) represents the predicted mean at x, σ(X) represents the predicted standard deviation of the Gaussian process at x, UCB(x) represents the upper confidence bound at x, and κ represents the parameter that controls the trade-off between exploration and exploitation. represents the noise variance;

[0091] S5, CNN-Transformer hybrid model training: Take the training set formed by the equally spaced voltage charging time data obtained in step S2 to run the optimized CNN-Transformer hybrid model, find the correlation between the equally spaced voltage charging time and the health status, and obtain the trained CNN-Transformer hybrid model;

[0092] S6. Obtaining the battery health status: Based on the trained CNN-Transformer hybrid model obtained in step S5, inputting the equal-interval voltage charging time, and outputting the battery health status.

[0093] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for estimating the health status of lithium-ion batteries based on Bayesian optimization CNN-Transformer, Its feature extraction and SOH estimation include the following steps: S1. Obtaining lithium-ion battery aging data: Performing cyclic charge and discharge tests on lithium-ion batteries until the discharge capacity is less than 80% of the nominal capacity, and recording the battery charge and discharge voltage, current, time and capacity data in real time; S2, data processing: pre-process the voltage, current, time and capacity data obtained in S1 to obtain equal interval voltage charging time and health status; S3. CNN-Transformer hybrid model construction: Combine CNN and Transformer models to build a CNN-transformer hybrid model; S4, CNN-Transformer hybrid model optimization: Use the Bayesian optimization algorithm to optimize the hyperparameters of the CNN-Transformer hybrid model obtained in S3 to obtain the optimized CNN-Transformer hybrid model; S5, CNN-Transformer hybrid model training: Take the training set formed by the equally spaced voltage charging time data obtained in step S2 to run the optimized CNN-Transformer hybrid model, find the correlation between the equally spaced voltage charging time and the health status, and obtain the trained CNN-Transformer hybrid model; S6. Obtaining the battery health status: Based on the trained CNN-Transformer hybrid model obtained in step S5, inputting the equal-interval voltage charging time, and outputting the battery health status.

2. The method for estimating the health status of a lithium-ion battery based on Bayesian optimization CNN-Transformer according to claim 1, characterized in that: The specific process of step S2 is as follows: S2.

1. Extract the charging voltage range of each cycle obtained from the test in S1 as [V start ,V stop ] charging time t, where V start Indicates the starting charging voltage, V stop Indicates stop charging voltage; S2.2, each cycle charging time constitutes a model input {x}; S2.

3. Extract the discharge capacity value of each cycle in S1 and calculate the state of health SOH: Among them, SOH i is the health status of the ith cycle, C n is the nominal capacity of the battery, C i is the actual capacity at the i-th cycle; S2.4, input {x} in the model of S2.2 and the corresponding health status SOH in S2.3 to form a data set, and divide the data set into training set D according to the cross-validation method. train and the test set D test , the training set and test set are as follows: Training set: Test set: Among them, x i ,y i are the i-th model input and the corresponding health status in the training set, respectively. are the i-th model input and the corresponding health status in the test set, and h represents the dimension.

3. The method for estimating the health status of a lithium-ion battery based on Bayesian optimization CNN-Transformer according to claim 2, characterized in that: The step S3 of establishing the CNN-Transformer hybrid model includes the following steps: S3.

1. Construct a CNN network to extract spatial features. The CNN network includes convolutional layer, pooling layer, and fully connected layer. The corresponding expressions are as follows: Among them, h i represents the features extracted by the i-th convolution layer, x i represents the input data of the i-th layer convolution, σ represents the relu activation function, represents the convolution operation, w i represents the convolution kernel weight matrix in the i-th convolution layer, b i Represents the bias vector in the i-th convolutional layer; S3.

2. Use Transformer to extract temporal features from the output of the CNN network, which specifically includes the following steps: S3.2.1, Transformer encoder converts the input data into different vectors and encodes each data; S3.2.2, the encoder analyzes the dependencies in the sequence through the self-attention mechanism and generates an encoding vector containing feature information; S3.2.

3. The Transformer decoder uses the features extracted by the encoder and combines its own input to gradually decode and generate predicted time series data through self-attention and encoder-decoder attention mechanisms.

4. The method for estimating the health status of a lithium-ion battery based on Bayesian optimization CNN-Transformer according to claim 3, characterized in that: The calculation expression of the encoder is as follows: Among them, PE(p,2i dim ) and PE(p,2i dim +1) represent two position encoding vectors, sin(*) represents the sine function, cos(*) represents the cosine function, p represents the position, i dim Denotes dimension, d model Represents the embedding dimension.

5. The method for estimating the health status of a lithium-ion battery based on Bayesian optimization CNN-Transformer according to claim 1, characterized in that: The specific expression for performing hyperparameter optimization on the CNN-Transformer hybrid model obtained in S3 using the Bayesian optimization algorithm in step S4 is as follows: f(x)~gp(m(x),k(x,x′)) p(f(x * )|D,x * )=N(μ(x * ),s 2 (x * )) EI(x)=E[max(0,f(x)-f(X + ))] UCB(*)=μ(X)+kσ(X) Where f(x) represents the target function, x represents the input vector, k(x,x') represents the kernel function, m(x) represents the mean function, gp(m(x),k(x,x')) represents the Gaussian process, and p(f(x * )|D,x * ) means that given data D and a new input vector x * When f(x * ), μ(x * ) represents the predicted mean, σ 2 (x * ) represents the new input vector x * The prediction variance, N(*) represents the normal distribution, k represents the target vector of the training set, k * Represents the new input vector x * The similarity vector with all training samples, Indicates k * The transpose of , K represents the kernel matrix, I represents the identity matrix, EI(x) represents the expected improvement at x, E[*] represents the mean calculation, max(0,f(x)-f(X + )) means taking the maximum value of two elements, f(X + ) represents the optimal value of the objective function in the known data, X + represents the current optimal point, X represents the current input point of interest, μ(X) represents the predicted mean at x, σ(X) represents the predicted standard deviation of the Gaussian process at x, UCB(x) represents the upper confidence bound at x, and κ represents the parameter that controls the trade-off between exploration and exploitation. represents the noise variance.

Images