Energy storage battery health state estimation method based on multi-feature fusion and Swin Transform model

Abstract

The invention relates to the technical field of battery management, and particularly discloses an energy storage battery health state estimation method based on multi-feature fusion and a Swin Transform model, and the method comprises the steps: obtaining battery test data under different test conditions; selecting health features related to capacity recession and carrying out data preprocessing; performing modal decomposition on the preprocessed health features to obtain residual components; utilizing Pearson and Spearman correlation analysis methods to select health features and residual components highly related to capacity recession, and performing feature fusion to obtain input feature data; inputting the input feature data into a pre-constructed energy storage battery SOH estimation model based on multi-feature fusion and a Swin Transform model for iterative training, and obtaining a pre-trained SOH estimation model; and finally, an estimated SOH value is obtained. The method not only improves the accuracy of SOH estimation of the energy storage battery, but also has high adaptability and stability, and provides powerful support for practical application of a battery management system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery management, and in particular to a method for estimating the state of health of energy storage batteries based on multi-feature fusion and a SwinTransformer model. BACKGROUND

[0002] In an energy storage power station, the state of health (SOH) of the battery is a core indicator reflecting the degree of aging of battery capacity attenuation, internal resistance degradation, cycle life loss, etc., and is also a key prerequisite for ensuring the efficient and reliable operation of the energy storage power station.

[0003] The current mainstream method for estimating the state of health (SOH) of lithium batteries has significant limitations in feature extraction and modeling. Specifically, data-driven models usually directly input current, voltage and other raw parameters into neural networks for prediction, ignoring the multi-scale health information (such as internal resistance fluctuations, capacity attenuation trends, etc.) implied in each parameter during the charging and discharging cycle, resulting in a disconnect between feature engineering and health state modeling, which limits the estimation accuracy of the model. In addition, existing methods rely on traditional empirical mode decomposition (EMD) or variational mode decomposition (VMD) in the modal decomposition stage, but such methods are susceptible to noise interference and have mode mixing problems, making it difficult to stably extract residual components reflecting battery degradation patterns. At the same time, the cost of obtaining large-scale high-quality data sets is high, and the data labeling and preprocessing process is time-consuming, further increasing the time and economic burden of model training. The above problems jointly constrain the adaptability and generalization ability of existing SOH estimation methods in actual scenarios.

[0004] Therefore, there is an urgent need for an SOH estimation method that can fuse multi-scale features, overcome the defects of modal decomposition, and have strong generalization ability to achieve accurate modeling and prediction of battery capacity degradation. SUMMARY

[0005] To solve the problems in the prior art, the present application provides a method for estimating the state of health of energy storage batteries based on multi-feature fusion and a SwinTransformer model, which solves the problems mentioned in the background art.

[0006] To achieve the above purpose, the present application provides the following technical solution: a method for estimating the state of health of energy storage batteries based on multi-feature fusion and a SwinTransformer model, characterized in that it comprises the following steps:

[0007] S1, obtaining battery test data under different test conditions;

[0008] S2, selecting health features related to capacity degradation from the test data;

[0009] S3, data preprocessing is performed on the extracted health characteristics to obtain complete and valid feature data;

[0010] S4, modal decomposition is performed on the preprocessed health characteristics by using a completely self-adaptive noise ensemble empirical mode decomposition method to obtain residual components;

[0011] S5, health characteristics and residual components highly correlated with capacity degradation are selected by using Pearson and Spearman correlation analysis methods, and input feature data is obtained after feature fusion;

[0012] S6, the input feature data is divided into training set, validation set and test set, and is input into a pre-constructed energy storage battery SOH estimation model based on multi-feature fusion and Swin Transformer model for iterative training to obtain a pre-trained SOH estimation model;

[0013] S7, the test set is used in the pre-trained SOH estimation model to obtain the estimated SOH value.

[0014] Preferably, in step S1, test data of the operation condition of the energy storage battery under the constant current and constant voltage charging and constant current discharging strategy at 24°C and 0°C is collected to obtain training samples; the collected data includes the current, voltage and temperature data of the battery during operation.

[0015] Preferably, in step S2, the health characteristics of constant current charging time CCCT, constant voltage charging time CVCT, constant current discharging time DCCT and temperature peak reaching time RTPT are extracted; the extraction formula of the health characteristics is as follows:

[0016] CCCT=T cc2 -T cc1

[0017] CVCT=T cv2 -T cv1

[0018] DCCT=T dc2 -T dc1

[0019] RTPT=T pk2 -T pk1

[0020] In the formula, T cc2 represents the time of constant current charging stage to cut-off voltage; T cc1 represents the start time of constant current charging stage; T cv2 represents the time of constant voltage charging stage to cut-off current; T cv1 represents the start time of constant voltage charging stage; Tdc2 is the time of constant current discharge phase, discharge to the cut-off voltage time; T dc1 is the time of constant current discharge phase, discharge start time; T pk2 is the time required for the whole charging and discharging process to reach the highest temperature; T pk1 is the running start time.

[0021] Preferably, in step S3, the data preprocessing includes removing outliers in health features and data normalization operation, specifically: using the isolation forest outlier detection method to detect outliers in the extracted health features, and using linear interpolation method, spline interpolation method and local boosting regression method to interpolate and complete the marked outliers to obtain complete and effective health features.

[0022] Preferably, in step S5, the Pearson and Spearman correlation analysis method is used to analyze the correlation between the battery health features and the residual components and the battery SOH, and the health features highly correlated with capacity degradation are selected for feature fusion to obtain input feature data.

[0023] Preferably, the energy storage battery SOH estimation model based on multi-feature fusion and SwinTransformer model includes an input layer, a convolution layer, a SwinTransformer layer, an average pooling layer, a linear projection layer and a full connection layer.

[0024] The first layer input layer includes a fully connected network, which is further enhanced after ReLu nonlinear processing to enhance the expression ability of the features;

[0025] The second layer convolution layer includes 16 one-dimensional convolution kernels, the convolution kernel size is 1x3, the convolution step is 1, the same convolution padding padding='same' is adopted to ensure that the sequence length of the output is consistent with the input; then an average pooling layer is used, the pooling window size is 2x1, and the pooling step is 2, which changes the feature dimension from BxLx3 to BxLx16, B is the batch size, and L is the sequence length.

[0026] The third layer Swin Transformer layer includes four stages of 1D Swin Transformer Block blocks, each 1D Swin Transformer Block block is composed of 1D W-MSA, 1D SW-MSA, LN, multi-layer perception MLP, Gaussian error linear unit GELU activation function and residual connection module; the calculation formula of a single 1D Swin Transformer Block block is as follows:

[0027]

[0028] In the formula, and z l denotes the output feature; LN is a LayerNorm operation; 1D W-MSA is a window multi-head self-attention mechanism operation; 1D SW-MSA is a shifted window multi-head self-attention mechanism operation;

[0029] The fourth layer is an average pooling layer, which is used to compress the time dimension and reserve global features;

[0030] The fifth layer is a linear projection layer and a fully connected layer, which is used to project the global features to 1 dimension and output the SOH estimation value.

[0031] Preferably, in the energy storage battery SOH estimation model based on multi-feature fusion and Swin Transformer model,

[0032] The input layer inputs the tensor X of the battery feature data set in ∈R B×L×f , B = 64 is the batch size, L = 128 is the sequence length, and f = 3 is the number of features per sequence length;

[0033] The convolutional layer is responsible for dimension reduction and feature extraction of data;

[0034] In the Swin Transformer layer, the data is divided into non-overlapping patches by the Patchpartition operation, each patch is regarded as a token, the patch size is 4x1, and the feature dimension is 64. Then, the input features are projected to the C dimension. Next, multiple 1D Swin Transformer Block blocks process the input data in turn. Each Block transforms the features in each Patch through the multi-head self-attention mechanism Self-Attention to capture long-distance dependencies and local features.

[0035] In the Stage1 stage, N 1D SwinTransformer blocks keep the token number as L / 4, and together with the linear embedding layer, they constitute the first stage. As the network depth increases, the adjacent two Patches are merged into larger Patches, reducing the token number and speeding up the training speed.

[0036] In the Stage2 stage, after the Patches are merged, a 4C-dimensional spliced feature map is formed, and then the output dimension is adjusted to 2C through 2 times down-sampling and linear projection operation. Next, N Swin Transformer blocks perform feature conversion, and the resolution is kept as L / 8.

[0037] In the Stage3 stage, the patch is merged to form a spliced feature map with a dimension of 8C, and then through 2 times down-sampling and linear projection operation, the output dimension is adjusted to 4C; then, N Swin Transformer blocks are used for feature conversion, and the resolution is kept as L / 16;

[0038] In the Stage4 stage, the patch is merged to form a spliced feature map with a dimension of 16C, and then through 2 times down-sampling and linear projection operation, the output dimension is adjusted to 8C; then, N Swin Transformer blocks are used for feature conversion, and the resolution is kept as L / 32; through the four stages of Stage1-Stage4, a hierarchical feature map with gradually reduced resolution is gradually constructed;

[0039] Finally, the output dimension is converted to Bx1 through an average pooling layer, a linear projection layer and a fully connected layer, and the SOH estimation value is obtained.

[0040] On the other hand, in order to achieve the above-mentioned purpose, the application also provides the following technical scheme: an energy storage battery state of health estimation device based on multi-feature fusion and Swin Transformer model, the device comprises the following modules:

[0041] Test data acquisition module: acquiring battery test data under different test conditions;

[0042] Data feature extraction module: selecting health features related to capacity degradation from test data;

[0043] Data preprocessing module: data preprocessing of the extracted health features to obtain complete and effective feature data; using the completely adaptive noise ensemble empirical mode decomposition method to decompose the preprocessed health features to obtain residual components;

[0044] Data feature fusion module: using Pearson and Spearman correlation analysis method to select health features and residual components highly related to capacity degradation, and obtaining input feature data after feature fusion;

[0045] Pre-training module: dividing the input feature data into training set, validation set and test set, and inputting into the pre-constructed energy storage battery SOH estimation model based on multi-feature fusion and Swin Transformer model for iterative training to obtain the pre-trained SOH estimation model;

[0046] SOH estimation module: using the test set for the pre-trained SOH estimation model to obtain the estimated SOH value.

[0047] In another aspect, to achieve the above object, the present application also provides the following technical solution: an electronic device includes a processor and a memory for storing one or more programs; when the one or more programs are executed by the processor, the processor executes the energy storage battery state of health estimation method based on multi-feature fusion and Swin Transformer model.

[0048] In another aspect, to achieve the above object, the present application also provides the following technical solution: a computer readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the energy storage battery state of health estimation method based on multi-feature fusion and Swin Transformer model.

[0049] The present application has the beneficial effect that: to solve the problems of the existing energy storage system, such as the deficiency of the SOH estimation method of the energy storage battery (including but not limited to flow battery, lithium ion battery, sodium ion battery, sodium sulfur battery, lead carbon battery, lead acid battery, etc.) in local and global feature capture and noise interference in modal decomposition, a health state estimation method based on multi-feature fusion and Swin Transformer fusion model is proposed, by extracting the health factors related to capacity degradation, combining CEEMDAN modal decomposition, correlation analysis and 1D CNN-Swin Transformer fusion model, the present application can effectively improve the accuracy of SOH estimation. The experimental results verify the superiority and stability of the present application under complex working conditions, and the present application has high estimation accuracy, saves data collection and training cost, and provides reliable support for the battery management system. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 The figure is a flow chart of the energy storage battery state of health estimation method based on multi-feature fusion and Swin Transformer model in the embodiment of the present application;

[0051] Figure 2 The figure is a framework diagram of the energy storage battery state of health estimation method based on multi-feature fusion and Swin Transformer model in the embodiment of the present application;

[0052] Figure 3a The figure is a structure diagram of 1D CNN-Swin Transformer model;

[0053] Figure 3b The figure is a structure diagram of 1D Swin TransformerBlock block;

[0054] Figure 4a The figure is a comparison diagram of SOH estimation results of B0005 battery under different methods at an ambient temperature of 24℃;

[0055] Figure 4b Figure 6 is a comparison chart of SOH estimation results of different methods for battery B0006 at an ambient temperature of 24°C;

[0056] Figure 4c Figure 7 is a comparison chart of SOH estimation results of different methods for battery B0007 at an ambient temperature of 24°C,

[0057] Figure 4d Figure 8 is a comparison chart of SOH estimation results of different methods for battery B0018 at an ambient temperature of 24°C;

[0058] Figure 5a Figure 9 is a comparison chart of SOH estimation results of different methods for battery CS2_35 at an ambient temperature of 1°C;

[0059] Figure 5b Figure 10 is a comparison chart of SOH estimation results of different methods for battery CS2_36 at an ambient temperature of 1°C;

[0060] Figure 5c Figure 11 is a comparison chart of SOH estimation results of different methods for battery CS2_37 at an ambient temperature of 1°C;

[0061] Figure 5d Figure 12 is a comparison chart of SOH estimation results of different methods for battery CS2_38 at an ambient temperature of 1°C;

[0062] Figure 6 Figure 13 is a schematic diagram of a device module for estimating the state of health of an energy storage battery based on multi-feature fusion and a SwinTransformer model according to an embodiment of the present application;

[0063] Figure 7 Figure 14 is a schematic diagram of an electronic device according to an embodiment of the present application;

[0064] In the figure, 110 is a test data acquisition module, 120 is a data feature extraction module, 130 is a data preprocessing module, 140 is a data feature fusion module, 150 is a pre-training module, 160 is an SOH estimation module, 210 is a processor, and 220 is a memory. DETAILED DESCRIPTION

[0065] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0066] Embodiment 1

[0067] The application provides a technical solution: a method for estimating the state of health of energy storage batteries based on multi-feature fusion and a Swin Transformer model, as shown in Figure 1 and Figure 2 Specifically, the method comprises the following steps:

[0068] S1, obtaining battery test data under different test conditions;

[0069] The test data of the operating conditions of the energy storage battery under the constant-current constant-voltage charging and constant-current discharging strategies at room temperature 24 DEG C and 0 DEG C are collected to obtain training samples. The data mainly collected include the current, voltage and temperature data of the battery during the operation of the battery.

[0070] The test data includes the current, voltage and battery shell temperature collected during the operation of the energy storage battery, which are collected by high-precision sensors. At the same time, the SOH label during model training is defined in the form of capacity, and the calculation method is as follows:

[0071]

[0072] In the formula, SOH is the state of health of the lithium battery; Q cur is the current actual capacity of the battery; Q init is the initial rated capacity of the battery.

[0073] Through the analysis of current, voltage, temperature and other data, the battery capacity, service life, safety and possible problems can be comprehensively evaluated.

[0074] It can be understood that the SOH estimation device includes a car control platform, a portable notebook computer, a smart phone or a tablet computer and other electronic devices relying on battery power supply.

[0075] S2, selecting health features related to capacity degradation from the test data;

[0076] The health features such as constant-current charging time CCCT, constant-voltage charging time CVCT, constant-current discharging time DCCT and temperature peak reaching time RTPT are extracted from the battery test data; the extraction formula of part of the health features is as follows:

[0077] CCCT=T cc2 -T cc1

[0078] CVCT=T cv2 -T cv1

[0079] DCCT=T dc2 -T dc1

[0080] RTPT = T pk2 -T pk1

[0081] In the formula, T cc2 is the time of constant current charging phase to the cut-off voltage; T cc1 is the start time of constant current charging phase; T cv2 is the time of constant voltage charging phase to the cut-off current; T cv1 is the start time of constant voltage charging phase; T dc2 is the time of constant current discharging phase to the cut-off voltage; T dc1 is the start time of constant current discharging phase; T pk2 is the time required for the entire charging and discharging process to reach the maximum temperature; T pk1 is the start time of operation. These features can effectively reflect the performance changes of the battery during the charging and discharging process, especially the behaviors closely related to capacity degradation. By analyzing these health features, the degradation patterns of the battery can be more accurately identified, the health status can be evaluated, and valuable input data can be provided for SOH estimation. The trend of these features is of great significance for predicting battery life, detecting potential failures in advance, and optimizing the management of energy storage systems.

[0082] S3, data preprocessing is performed on the extracted health features to obtain complete and valid feature data.

[0083] During data collection, the battery management system (BMS) may produce abnormal data points, including zero value abnormalities and data significantly deviating from adjacent samples, due to factors such as delay and noise. Therefore, these abnormal values must be detected, removed and interpolated to ensure the accuracy and reliability of subsequent battery data analysis.

[0084] The present application uses the Isolation Forest outlier detection method to detect outliers in the extracted health features, and the marked outliers are interpolated and completed using the linear interpolation method (Linear), the spline interpolation method (Spine) and the locally weighted regression method (LOWESS) to obtain complete and valid health features. For slowly changing feature data, linear interpolation can effectively fill in abnormal values, and the calculation formula is as follows:

[0085]

[0086] In the formula, (x0, y0) and (x1, y1) are known data points; x is the interpolation point; y(x) is the estimated interpolation effect.

[0087] The spline interpolation method fills in abnormal values by constructing a smooth curve, which is especially suitable for complex data changes and smooth transitions, and the calculation formula is as follows:

[0088] S(x)=a i (x-x i ) 3 +b i (x-x i ) 2 +c i (x-x i )+d i

[0089] In the formula, S(x) represents the interpolation result; x is the position that needs to be interpolated; a i ,b i ,c i ,d i are coefficients obtained by solving linear equations.

[0090] The LOWESS method smoothes the data and effectively removes the influence of outliers by weighted regression on the data points in the neighborhood, and its calculation formula is as follows:

[0091]

[0092] In the formula, x i is the current data point; x j is other data points in the neighborhood; h is a smoothing parameter for controlling the neighborhood size.

[0093] S4, decomposing the health feature mode by using complete adaptive noise set empirical mode decomposition, to obtain a residual component.

[0094] The obtained health factor data includes multi-dimensional characteristics such as voltage and current of the battery, and reflects the running state of the battery. In order to improve the SOH estimation accuracy, the CEEMDAN mode decomposition method is used to extract the residual component of the health feature, because the residual component can effectively remove high-frequency noise, focus on the long-term trend of battery health, and accurately capture performance degradation or recovery.

[0095] Therefore, the residual component of the health factor can be used as a key feature to input the subsequent SOH estimation model to improve the estimation accuracy of the SOH estimation model. For health factor data x(t), the steps of using the CEEMDAN method to obtain the residual component can be briefly described as three processes: preparing health factor data; adding white noise; EMD decomposition.

[0096] Step 1: Prepare health factor data. The health factor data after outlier processing and interpolation completion is used as the input signal x(t), wherein t is the index of the time sequence, and x(t) is the health factor of the battery.

[0097] Step 2: Add white noise. To improve the robustness of decomposition, different amplitudes of white noise ∈(t) are added to the input signal x(t) to generate multiple noisy signal copies:

[0098] x noisy (t) = x(t) + ∈(t)

[0099] where ∈(t) is the added white noise with zero mean and known variance. By adding different noise multiple times and performing decomposition, clearer modal components can be obtained.

[0100] Step 3: EMD decomposition. The CEEMDAN method performs EMD decomposition on each noisy signal x noisy (t) to obtain multiple intrinsic mode functions (IMFs) and a residual component, as shown in the following process:

[0101]

[0102] where k is the number of IMFs; R(t) is the residual component, representing the low-frequency trend component in the health factor.

[0103] S5, select the health features and residual components highly correlated with capacity degradation using Pearson and Spearman correlation analysis methods, and obtain input feature data after feature fusion.

[0104] The Pearson and Spearman correlation coefficients of different health factors under different interpolation methods are shown in Table 1. The Pearson correlation analysis method is used to measure the degree of linear relationship between two variables, which evaluates the correlation between them by calculating the product of the covariance and the standard deviation of the two variables, as shown in equation (2):

[0105]

[0106] where x i represents the reference value of the health factor; y i represents the reference value of SOH; represents the average value of the health factor; represents the average value of SOH.

[0107] Spearman correlation analysis is a non-parametric test method used to measure the degree of non-linear relationship between two variables. It calculates the correlation between two variables based on the ranking of variable values, as shown in equation (2):

[0108]

[0109] where d i is the difference between the observed values of the health factor and SOH, and m is the total number of observed values.

[0110] Table 1 Pearson and Spearman correlation coefficients of different health factors under different interpolation methods

[0111]

[0112] S6, the input feature data is divided into training set, validation set and test set, and input into the pre-constructed SOH estimation model of energy storage battery based on multi-feature fusion and Swin Transformer model for iterative training, to obtain a pre-trained SOH estimation model;

[0113] In the embodiment, when the SOH estimation model estimates the SOH at the i th moment, the model input includes the sliding window test data used to estimate the SOH at the moment, that is:

[0114] SOH i = FusionModel ([f i ,f i-1 ,f i-2 ,...,f i-n ])

[0115] Wherein, FusionModel represents the SOH estimation model based on multi-feature fusion and Swin Transformer model, SOH i represents the SOH at the i th moment, [f i ,f i-1 ,f i-2 ,...,f i-n ] represents n historical acquisition data and extracted health features from the i th moment.

[0116] The pre-constructed SOH estimation model of energy storage battery based on multi-feature fusion and Swin Transformer model (SOH estimation model) has a model structure as shown in Figure 3aDropout operation and linear transformation operation between some modules are omitted. The SOH estimation model based on multi-feature fusion and Swin Transformer model mainly includes one-dimensional convolution module, one-dimensional stacked Swin Transformer module (containing self-attention mechanism module), average pooling module, linear projection module and full connection layer; the one-dimensional SwinTransformer module thereof is composed of Swin Transformer Block, and each SwinTransformer Block includes multi-head self-attention mechanism and shift multi-head self-attention mechanism operation, the purpose of the two operations is to enhance the model's ability to capture long-range dependencies, and further improve the model's sensitivity when processing different local features through shift operation. The multi-head self-attention mechanism can allow the model to process input features in different subspaces in parallel, and dig out more rich context information, while the shift multi-head self-attention mechanism can process data by segmentation, reduce the computational complexity, and at the same time maintain the efficient modeling of multi-scale features. These mechanisms work together not only to improve the model's expression ability in different feature scales, but also to better capture the complex rules of battery state of health changes.

[0117] With the help of the one-dimensional convolution module, the model can effectively extract local features in the input signal, and after processing by the stacked SwinTransformer module, these local features can be more deeply understood and fused to provide more accurate SOH estimation. The average pooling module further reduces the data dimension through weighted average of multiple feature channels, while maintaining the transmission of important information. The linear projection module maps the features to a new space, ensuring that the subsequent full connection layer can process and utilize more compact and information-rich feature representations, and the mapping generates the final SOH estimation.

[0118] The SOH estimation model is built based on the deep learning Pytorch framework, specifically including:

[0119] Input layer: tensor X of input battery feature dataset in ∈R B×L×f (B = 64 is the batch size, L = 128 is the sequence length, and f = 3 is the number of features per sequence length).

[0120] Convolution layer: the convolution layer is responsible for data dimension reduction and feature extraction, thereby providing more representative and discriminative feature representations for subsequent layers. The size of the one-dimensional convolution kernel is set to K = 16, and its calculation formula is as follows:

[0121]

[0122] In the formula, X convis the output feature of one-dimensional convolution; X in is the input feature; W conv is the weight parameter; b conv is the bias value.

[0123] Swin Transformer layer: The 1D Swin Transformer layer includes four stages, and each stage is composed of a certain number of 1D Swin Transformer blocks, such as Figure 3bAs shown, each 1D Swin Transformer Block is composed of 1D Window Multi-Head Self-Attention (1D W-MSA), 1D Shifted Window Multi-Head Self-Attention (1D SW-MSA), LayerNorm (LN), Multi-Layer Perceptron (MLP), Gaussian Error Linear Unit (GELU) activation function and residual connection module. The data is divided into non-overlapping Patches by the Patch partition operation, and each Patch is regarded as a token. The Patch size is 4x1, and the feature dimension is 64 (4x1x16). Then, the input features are projected to the C dimension (C is set to 96) by using the linear embedding module. Next, multiple 1D Swin Transformer Blocks (the number is N) process the input data in turn. Each Block transforms the features in each Patch by the multi-head self-attention mechanism (Self-Attention), so as to capture long-distance dependencies and local features. In the Stage1 stage, the N 1D Swin Transformer blocks keep the token number as L / 4, and together with the linear embedding layer, they constitute the first stage. As the network layer deepens, the adjacent two Patches are merged into a larger Patch, aiming to reduce the token number and speed up the training speed. In the Stage2 stage, after the Patches are merged, a splicing feature map with a 4C dimension is formed, and then the output dimension is adjusted to 2C by 2 times down-sampling and linear projection operation; then, N Swin Transformer blocks are used for feature conversion, and the resolution is kept as L / 8. In the Stage3 stage, after the Patches are merged, a splicing feature map with an 8C dimension is formed, and then the output dimension is adjusted to 4C by 2 times down-sampling and linear projection operation; then, N Swin Transformer blocks are used for feature conversion, and the resolution is kept as L / 16. In the Stage4 stage, after the Patches are merged, a splicing feature map with a 16C dimension is formed, and then the output dimension is adjusted to 8C by 2 times down-sampling and linear projection operation; then, N Swin Transformer blocks are used for feature conversion, and the resolution is kept as L / 32. In this way, the four stages gradually construct hierarchical feature maps with gradually reduced resolutions.

[0124] Finally, the output dimension is converted to Bx1 by the average pooling layer, the linear projection layer and the fully connected layer, and the SOH estimation value is obtained.

[0125] The SOH estimation model parameter based on the multi-feature fusion and the Swin Transformer model, specifically comprises:

[0126] The first layer is an input layer including a fully connected network, which is further enhanced by ReLu non-linear processing to enhance the expression ability of features and improve the recognition of input data by the network.

[0127] The second layer is a convolution layer including 16 one-dimensional convolution kernels, with a convolution kernel size of 1x3, a convolution step of 1, and same convolution padding (padding='same') to ensure that the output sequence length is consistent with the input. Then, an average pooling layer is used with a pooling window size of 2x1 and a pooling step of 2. After these operations, the feature dimension changes from BxLx3 (B is the batch size, and L is the sequence length) to BxLx16.

[0128] The third layer is a Swin Transformer layer including four stages of 1D Swin Transformer Block blocks. The calculation formula of a single 1D Swin Transformer Block block is as follows:

[0129]

[0130] In the formula, and z l represent the output features; LN is the LayerNorm operation; 1DW-MSA is the window multi-head self-attention mechanism operation; and 1D SW-MSA is the shift window multi-head self-attention mechanism operation.

[0131] The stacking number parameters of the four-stage 1D Swin Transformer Block block are set to (N0, N1, N2, N3) with values of (2, 2, 4, 2) or (2, 2, 6, 2). The number of attention heads is set to (h0, h1, h2, h3) = (2, 4, 8, 16). The shift window size is Window size = [16, 32, 64]. The Dropout is set to [0.1, 0.2]. The Learning rate is set to [0.0001, 0.001]. The Training epoch is 200. The Early stopping patience is 50.

[0132] The fourth layer is an average pooling layer that compresses the time dimension and retains global features.

[0133] The fifth layer is a linear projection layer and a fully connected layer that projects global features to 1D and outputs SOH estimates.

[0134] The specific parameter settings of the FusionModel are shown in Table 2:

[0135] Table 2 FusionModel parameter setting table

[0136]

[0137]

[0138] For a lithium battery health feature sequence where N is the sequence length, d is the feature dimension; assuming there are h attention heads, then the query vector of the i-th head is Q i , the key value vector is K i , and the value vector is V i , then where and are the learnable weight matrices of the i-th head. The attention weight of the i-th head is where d k is the dimension of the attention head, and the output of the i-th head is Output i = A i V i , and the outputs of all heads are concatenated and linearly transformed to obtain the final output result as follows:

[0139] Output = concat (Output1Output2, …, Output h ) W o .

[0140] S7, use the test set for the pre-trained SOH estimation model to obtain the estimated SOH value. The pre-training and testing process of the model is as follows:

[0141] First, obtain the battery sample: the battery sample uses the public CALCE and NASA data set, which collects current, voltage and temperature data tested at 1℃ and 24℃ ambient temperature. Perform steps S1 to S5 on the data set, then divide the data set into training set and test set. Among them, the first 60% of the data of each data set is used as the training set, the middle 10% of the data is used as the validation set to verify the effectiveness of the model, and the remaining data is used as the test set to evaluate the performance of the model.

[0142] Subsequently, the loss function of the SOH estimation model training is set to mean square error (MSE), and the calculation formula is as follows:

[0143]

[0144] In the formula, is the SOH estimation value, and y i is the SOH true value.

[0145] The learning rate adopts a decay strategy. Specifically, the initial learning rate is set to 0.001, and then it is decayed every 50 cycles with a decay factor of 0.9. This setting allows the model to make smoother adjustments in the solution space, avoiding drastic fluctuations in the early stages of training while maintaining fine-tuned adjustments in later stages. By gradually reducing the learning rate, the model can gradually converge to a better local minimum, improving the stability and accuracy of model convergence. This decay strategy helps to avoid premature convergence and ensures the efficiency of the training process.

[0146] During training, the Early Stopping strategy is also adopted, which is a simple and effective regularization technique. By monitoring the performance changes of the validation set in real time, it can timely detect the overfitting phenomenon in the model training process and stop training in advance when appropriate. In this way, the model avoids overfitting on training data, thereby improving the generalization ability of the model on unknown data. Specifically, the patience value is set to 50, which means that if the performance of the validation set does not improve significantly within 50 cycles, the training will be stopped. This strategy effectively improves the performance of the model while also saving computational resources.

[0147] The model optimization algorithm is set to Adam, which is an adaptive learning rate optimization algorithm that combines the advantages of momentum method and RMSProp. It can adaptively adjust the learning rate for each parameter, thereby accelerating the convergence process of the model and reducing the training time. Through the Adam optimization algorithm, the model can continuously update and optimize parameters through the backpropagation mechanism, so that the loss function gradually decreases, and the prediction ability of the model is significantly improved.

[0148] During training, the Adam optimization algorithm adjusts the update step of each parameter by calculating the mean and variance of the gradient, making the model more flexible and efficient when dealing with different features. After multiple training iterations, the model gradually reduces the prediction error by continuously adjusting the parameters, so that the predicted value of the model gradually approaches the true value, and finally realizes high-precision prediction of the target task. Through this optimization process, the model can better capture the complex patterns in the data, improving the accuracy and stability of the prediction.

[0149] After the model is trained, it can be used to estimate the battery SOH. The comparison chart of the estimation results of different methods is shown in Figures 4 and 5. Figure 4a 、 Figure 4b 、 Figure 4c 、 Figure 4d Figures 4 and 5 respectively represent the comparison of estimation results of B0005, B0006, B0007 and B0018 batteries using different methods on the NASA dataset (ambient temperature 24℃); Figure 5a 、 Figure 5b 、 Figure 5c ,Figure 5d respectively represent the comparison of the estimation results of CS2_35, CS2_36, CS2_37 and CS2_38 batteries by different methods on the CALCE data set (ambient temperature is 1℃); it can be seen from the figure that after the fusion of one-dimensional convolution and one-dimensional SwinTransformer model, the SOH estimation accuracy is improved.

[0150] The energy storage battery health state estimation method based on multi-feature fusion and Swin Transformer model in the embodiment of the application can obtain good estimation accuracy under the ambient temperature of 24℃ and 1℃ by extracting the health features of the battery charging and discharging process and fusing one-dimensional convolution and one-dimensional Swin Transformer model, and has excellent performance in MAE, RMSE and MAPE error indicators in the estimation results, and well solves the shortcomings of the existing energy storage battery SOH estimation method in capturing local and global feature information, and the mode aliasing and noise interference problems existing in the classical modal decomposition method.

[0151] Based on the same inventive concept as the above method embodiment, the embodiment of the application also provides an energy storage battery health state estimation device based on multi-feature fusion and Swin Transformer model, which can realize the functions provided by the above method embodiment, as shown in the figure. Figure 6 As shown in the figure, the device includes the following modules:

[0152] The test data acquisition module 110 acquires battery test data under different test conditions;

[0153] The data feature extraction module 120 selects health features related to capacity degradation from the test data;

[0154] The data preprocessing module 130 performs data preprocessing on the extracted health features to obtain complete and valid feature data; and performs modal decomposition on the preprocessed health features by using the completely adaptive noise ensemble empirical mode decomposition method to obtain residual components;

[0155] The data feature fusion module 140 selects health features and residual components highly related to capacity degradation by using Pearson and Spearman correlation analysis method, and obtains input feature data after feature fusion;

[0156] The pre-training module 150 divides the input feature data into training set, validation set and test set, inputs it into the pre-constructed energy storage battery SOH estimation model based on multi-feature fusion and Swin Transformer model for iterative training, and obtains the pre-trained SOH estimation model;

[0157] The SOH estimation module 160: uses the test set for the pre-trained SOH estimation model to obtain an estimated SOH value, and verifies the effectiveness of the model.

[0158] Based on the same inventive concept as the above method embodiments, the embodiments of the present application also provide an electronic device, as shown in the figure. Figure 7 The device includes a processor 210 and a memory 220 for storing one or more programs.

[0159] When the one or more programs are executed by the processor 210, the processor performs the energy storage battery state of health estimation method based on multi-feature fusion and Swin Transformer model.

[0160] The energy storage battery state of health estimation method based on multi-feature fusion and Swin Transformer model specifically includes the following:

[0161] Obtain battery test data under different test conditions;

[0162] Select health features related to capacity degradation from the test data;

[0163] Perform data preprocessing on the extracted health features to obtain complete and valid feature data;

[0164] Perform modal decomposition on the preprocessed health features using the fully adaptive noise ensemble empirical mode decomposition method to obtain residual components;

[0165] Select health features and residual components highly correlated with capacity degradation using Pearson and Spearman correlation analysis methods, and obtain input feature data after feature fusion;

[0166] Divide the input feature data into training set, validation set and test set, and input it into the pre-constructed energy storage battery SOH estimation model based on multi-feature fusion and Swin Transformer model for iterative training to obtain a pre-trained SOH estimation model;

[0167] Use the test set for the pre-trained SOH estimation model to obtain an estimated SOH value.

[0168] Based on the same inventive concept as the above method embodiments, the embodiments of the present application also provide a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by the processor 210 to implement the energy storage battery state of health estimation method based on multi-feature fusion and Swin Transformer model.

[0169] The energy storage battery state of health estimation method based on multi-feature fusion and Swin Transformer model specifically comprises the following steps.

[0170] Obtain battery test data under different test conditions;

[0171] Select health features related to capacity degradation from the test data;

[0172] Perform data preprocessing on the extracted health features to obtain complete and valid feature data;

[0173] Perform modal decomposition on the preprocessed health features using the fully adaptive noise ensemble empirical mode decomposition method to obtain residual components;

[0174] Select health features and residual components highly correlated with capacity degradation using Pearson and Spearman correlation analysis methods, and obtain input feature data after feature fusion;

[0175] Divide the input feature data into training set, validation set and test set, and input them into the pre-constructed energy storage battery SOH estimation model based on multi-feature fusion and Swin Transformer model for iterative training to obtain a pre-trained SOH estimation model;

[0176] Use the test set to the pre-trained SOH estimation model to obtain the estimated SOH value.

[0177] In several embodiments provided in the embodiments of the present application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device and method embodiments described above are only illustrative, for example, the flowchart and block diagram in the drawings show the possible implementation architecture, function and operation of the device, method and computer program product according to the embodiments of the present application. In this regard, each block in the flowchart or block diagram can represent a module, program segment or part of code containing one or more executable instructions for implementing the specified logic function. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur in different orders than those noted in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, and sometimes they can be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and the combination of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system that performs the specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions.

[0178] In addition, each functional module in various embodiments of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0179] If the functions are implemented in the form of software functional modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for making a computer device (which can be a personal computer, an electronic device, or a network device, etc.) execute all or part of the steps of the methods described in various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes. It should be noted that in this paper, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the processes, methods, articles or devices that include a series of elements not only include those elements, but also include other elements that are not explicitly listed, or include elements inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "includes a" does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0180] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. The singular forms "a", "an" and "the" used in the embodiments of the present application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0181] It should be understood that the term "and / or" used herein is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper generally represents that the front and rear associated objects are a "or" relationship.

[0182] Depending on the context, the word "if" as used herein can be interpreted to mean "when" or "while" or "in response to determining" or "in response to detecting." Similarly, the phrase "if it is determined" or "if [a stated condition or event] is detected" can be interpreted to mean "when it is determined" or "in response to determining" or "when [the stated condition or event] is detected" or "in response to detecting [the stated condition or event]."

[0183] The "first / second" mentioned in the embodiments are only to distinguish similar objects, and do not represent a specific order of the objects. Understandably, the "first / second" can be interchanged in a specific order or sequence as permitted. It should be understood that the objects distinguished by the "first / second" can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0184] Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can modify the technical solutions recorded in the foregoing embodiments or make equivalent replacements to some of the technical features, and any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for estimating the state of health of an energy storage battery based on multi-feature fusion and a Swin Transformer model, characterized in that, The method comprises the following steps: S1, acquiring battery test data under different test conditions; S2, selecting health characteristics related to capacity degradation from the test data; S3, performing data preprocessing on the extracted health characteristics to obtain complete and valid feature data; S4, performing modal decomposition on the preprocessed health characteristics using the completely adaptive noise ensemble empirical modal decomposition method to obtain residual components; S5, selecting health characteristics and residual components highly correlated with capacity degradation using Pearson and Spearman correlation analysis methods, and obtaining input feature data after feature fusion; S6, dividing the input feature data into training set, validation set and test set, and inputting them into the pre-constructed energy storage battery SOH estimation model based on multi-feature fusion and Swin Transformer model for iterative training to obtain a pre-trained SOH estimation model; S7, using the test set to the pre-trained SOH estimation model to obtain the estimated SOH value.

2. The method of claim 1, wherein the method is based on multi-feature fusion and a Swin Transformer model. In step S1, the test data of the energy storage battery under the operating conditions of constant current and constant voltage charging and constant current discharging strategy at 24°C and 0°C temperature are collected to obtain training samples; the collected data include the current, voltage and temperature data of the battery during operation.

3. The method of claim 1, wherein the method is based on multi-feature fusion and a Swin Transformer model. In step S2, the health characteristics of constant current charging time CCCT, constant voltage charging time CVCT, constant current discharging time DCCT and temperature peak time RTPT are extracted; the extraction formula of the health characteristics is as follows: CCCT = T cc2 -T cc1 CVCT = T cv2 -T cv1 DCCT = T dc2 -T dc1 RTPT = T pk2 - T pk1 In the formula, T cc2 is the time for constant current charging phase to the cut-off voltage; T cc1 is the constant current charging phase, the charging start time; T cv2 is the constant voltage charging phase, charging to the cut-off current time; T cv1 is the constant voltage charging phase, the charging start time; T dc2 is the time of the constant current discharge phase, discharge to the cut-off voltage; T dc1 is the discharge start time; T pk2 is the time required to reach the maximum temperature during the entire charging and discharging process; T pk1 is the time of the start of the operation.

4. The method of claim 1, wherein the method is based on multi-feature fusion and a SwinTransformer model. In step S3, the data preprocessing includes removing outliers in the health characteristics and data normalization operation, specifically: using the isolated forest outlier detection method to detect outliers in the extracted health characteristics, and using linear interpolation method, spline interpolation method and local boosting regression method to interpolate and complete the marked outliers to obtain complete and valid health characteristics.

5. The method of claim 1, wherein the method is based on multi-feature fusion and a SwinTransformer model. In step S5, the Pearson and Spearman correlation analysis method is used to analyze the correlation between the battery health characteristics and the residual components and the battery SOH, and the health characteristics highly correlated with capacity degradation are selected for feature fusion to obtain input feature data.

6. The method of claim 1, wherein the method is based on multi-feature fusion and a SwinTransformer model. The energy storage battery SOH estimation model based on multi-feature fusion and Swin Transformer model comprises an input layer, a convolution layer, a Swin Transformer layer, an average pooling layer, a linear projection layer and a full connection layer; The first layer input layer includes a full connection network, which is further enhanced after ReLu nonlinear processing to enhance the expression ability of the features; The second layer convolution layer includes 16 one-dimensional convolution kernels, the convolution kernel size is 1x3, the convolution step is 1, the same convolution padding padding='same' is adopted to ensure that the sequence length of the output is consistent with the input; then an average pooling layer is used, the pooling window size is 2x1, and the pooling step is 2, the feature dimension is changed from BxLx3 to BxLx16, B is the batch size, and L is the sequence length; The third layer Swin Transformer layer includes four stages of 1D Swin Transformer Block blocks, each 1D Swin Transformer Block being composed of a 1D W-MSA, a 1D SW-MSA, an LN, a multi-layer perception MLP, a Gaussian error linear unit GELU activation function and a residual connection module; the calculation formula of a single 1D Swin Transformer Block is as follows: In the formulae, and z l denotes the output characteristic; LN is a LayerNorm operation; 1D W-MSA is a window multi-head self-attention mechanism operation; 1D SW-MSA is a shift window multi-head self-attention mechanism operation; The fourth layer is an average pooling layer, used for compressing the time dimension and retaining global features; The fifth layer is a linear projection layer and a fully connected layer, used for projecting the global features to 1 dimension and outputting an SOH estimation value.

7. The method of claim 1 or 6, wherein the method is based on multi-feature fusion and a SwinTransformer model. In the SOH estimation model of the energy storage battery based on multi-feature fusion and the SwinTransformer model, Input layer input tensor X of battery characteristic data set in ∈ R B×L×f B = 64 is batch size, L = 128 is sequence length, f = 3 is the number of features per sequence length; The convolutional layer is responsible for dimension reduction and feature extraction of data; In the Swin Transformer layer, the data is divided into non-overlapping patches through the Patchpartition operation, each patch is regarded as a token, the patch size is 4x1, and the feature dimension is 64, then the input features are projected to the C dimension; then, a plurality of 1D Swin Transformer Block blocks process the input data in turn; each Block transforms the features in each Patch through the multi-head self-attention mechanism Self-Attention to capture long-distance dependencies and local features; In the Stage1 stage, N 1D SwinTransformer blocks maintain the token number as L / 4, and together with the linear embedding layer, they constitute the first stage; as the network layer deepens, two adjacent patches are combined into larger patches, reducing the token number and speeding up the training speed; In the Stage2 stage, after the patch is combined, a 4C-dimensional splicing feature map is formed, and then the output dimension is adjusted to 2C through 2 times down-sampling and linear projection operation; then, N Swin Transformer blocks perform feature conversion, and the resolution is maintained as L / 8; In the Stage3 stage, after the patch is combined, an 8C-dimensional splicing feature map is formed, and then the output dimension is adjusted to 4C through 2 times down-sampling and linear projection operation; then, N Swin Transformer blocks perform feature conversion, and the resolution is maintained as L / 16; In the Stage4 stage, after the patch is combined, a 16C-dimensional splicing feature map is formed, and then the output dimension is adjusted to 8C through 2 times down-sampling and linear projection operation; then, N Swin Transformer blocks perform feature conversion, and the resolution is maintained as L / 32; The hierarchical feature maps with gradually reduced resolutions are constructed through four stages of Stage1-Stage4; Finally, the output dimension is converted to Bx1 through an average pooling layer, a linear projection layer and a fully connected layer, and the SOH estimation value is obtained.

8. An energy storage battery state of health estimation device based on the energy storage battery state of health estimation method based on multi-feature fusion and a Swin Transformer model according to any one of claims 1-7, characterized in that: The device comprises the following modules: A test data acquisition module (110) acquires battery test data under different test conditions; A data feature extraction module (120) selects health features related to capacity degradation from the test data; A data preprocessing module (130) preprocesses the extracted health features to obtain complete and valid feature data; The health features after preprocessing are mode decomposed by using a fully adaptive noise ensemble empirical mode decomposition method to obtain residual components; A data feature fusion module (140) selects health features and residual components highly related to capacity degradation by using Pearson and Spearman correlation analysis methods, and obtains input feature data after feature fusion; A pre-training module (150) divides the input feature data into a training set, a validation set and a test set, and inputs them into a pre-constructed SOH estimation model of energy storage batteries based on multi-feature fusion and a Swin Transformer model for iterative training to obtain a pre-trained SOH estimation model; An SOH estimation module (160) uses the test set in the pre-trained SOH estimation model to obtain an estimated SOH value.

9. An electronic device, comprising: The electronic device comprises a processor (210) and a memory (220) for storing one or more programs; When the one or more programs are executed by the processor (210), the processor performs the energy storage battery health state estimation method based on multi-feature fusion and a Swin Transformer model according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program is executed by the processor (210) to implement the energy storage battery health state estimation method based on multi-feature fusion and a Swin Transformer model according to any one of claims 1-7.

Images