Lithium ion battery health state estimation method and system based on unsupervised transfer learning, and electronic equipment

Abstract

The invention discloses a lithium ion battery health state estimation method and system based on unsupervised transfer learning and electronic equipment, and the method comprises the following steps: selecting two different types of batteries, and determining a source domain battery and a target domain battery; obtaining a loop test data set, and further obtaining a sample set and a label set; performing data processing on the sample set and the label set to obtain a training verification data set; using the training verification data set to train a domain adversarial neural network model; inputting cycle test data of the battery to be tested into the domain adversarial neural network to obtain an estimation result; according to the method, the cycle test data and the calibration test data of the existing active domain battery and a small amount of cycle test data of the target domain battery without calibration test are utilized, a redundant target domain battery cycle test process is omitted, and accurate estimation of the health state of the target domain battery is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithium-ion battery health estimation, and particularly relates to a method, a system and an electronic device for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning. Background Art

[0002] In recent years, with the booming development of the new energy vehicle market, the production and retirement volume of lithium-ion batteries have also increased significantly. As the number of charge and discharge cycles increases, lithium-ion batteries will gradually age, and the change of internal electrochemical properties will lead to the reduction of their capacity, and even cause safety accidents. In the field of electric vehicles, when the capacity of a lithium-ion battery reaches 80% of the rated capacity, it can be considered that the battery has reached the end-of-life point and is no longer suitable as an on-vehicle battery. Accurately estimating the state of health of a lithium-ion battery can ensure the safety, maintenance and replacement of the electric drive system, and ensure the stability of the system.

[0003] In the prior art, generally, multiple lithium-ion battery aging characteristics in a charging curve are obtained, and after calculating the Pearson correlation coefficient between these characteristics and the maximum capacity of the battery, the aging characteristics with high correlation coefficients are selected to construct a feature matrix, and the principal component analysis method is used for feature dimensionality reduction, and then the reduced features are sent into a trained GRU model to estimate the state of health of the lithium-ion battery. Or calculate the incremental energy-state of charge (IES) curve and extract health indicators, establish a state-of-health estimation model based on bidirectional LSTM-reduction, and send the IES features extracted from the lithium-ion battery segment charging data into the LSTM-reduction model to obtain the state-of-health estimation result of the lithium-ion battery.

[0004] However, the above-mentioned methods for estimating the state of health of lithium-ion batteries usually only involve batteries of a certain model. If it is necessary to estimate the state of health of another model of battery, a neural network model needs to be rebuilt. In addition, it usually takes a long time to obtain the cyclic data for training the neural network model. If another model of battery is cyclically tested again, the production efficiency will be greatly reduced, and it is difficult to improve the benefits. Moreover, when applying transfer learning to the field of estimating the state of health of lithium-ion batteries, the distribution difference between the source domain battery and the target domain battery needs to be considered to help the domain adversarial neural network model complete the transition between the source domain battery and the target domain battery. Summary of the Invention

[0005] Aiming at the problems existing in the prior art, the present invention provides a method, system and electronic device for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning, which solves the deficiency that the existing state of health estimation method can only estimate a certain type of battery, and improves the efficiency. In addition, considering the problem of excessive distribution difference between the source domain battery and the target domain battery, it is beneficial to realize the transition between the source domain battery and the target domain battery of the domain adversarial neural network model, and improve the prediction accuracy.

[0006] The technical solution of the present invention is realized as follows:

[0007] A method for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning includes the following steps:

[0008] S100. Select sample batteries of different models, and divide the sample batteries into a source domain battery and a target domain battery;

[0009] S200. Obtain the cyclic test data of the source domain battery and the target domain battery, and collect the sample set and label set of the sample batteries, where the number of samples of the source domain battery is greater than the number of samples of the target domain battery;

[0010] S300. Process the sample sets of the source domain battery and the target domain battery, analyze the processed data by using the dynamic mode decomposition method, obtain the charging capacity change mode sequence in the constant current charging stage of the battery, and construct a binary sequence as the training and verification data set;

[0011] S400. Build an unsupervised transfer learning domain adversarial neural network model with a bridging module, and use the training and verification data set and the label set of the source domain battery to train the domain adversarial neural network model to obtain a trained domain adversarial neural network model;

[0012] S500. Obtain the cyclic test data of the battery to be estimated and perform data processing, where the model of the battery to be estimated is the same as that of the target domain battery, and then input it into the trained domain adversarial neural network model to obtain the state of health estimation value of the battery to be estimated.

[0013] Further, in S200, collecting the sample set of the sample battery specifically includes:

[0014] S210. In the obtained cyclic test data, collect the data of voltage V and charging capacity Q in each cycle of the constant current charging stage, and form the sample set of the source domain battery or the sample set of the target domain battery with all the obtained voltages V and charging capacities Q according to whether the sample battery is a source domain battery or a target domain battery.

[0015] Further, in S200, collecting the label set of the sample battery specifically includes:

[0016] S220. Among the obtained cyclic test data, collect the maximum discharge capacity data in the discharge stage to obtain the health state value SOH corresponding to the battery health, and form the label set of the source domain battery or the label set of the target domain battery with all the obtained health state values SOH; the health state value SOH is specifically expressed as:

[0017]

[0018] where i represents the i-th sample in the source domain battery sample set, and SOH i is the health state value of the i-th sample; j represents the j-th sample in the target domain battery sample set, and SOH j is the health state value of the j-th sample; Q i represents the maximum discharge capacity in the source domain battery sample set, and Q j represents the maximum discharge capacity in the target domain battery sample set; Q S represents the rated capacity of the source domain battery, and Q T represents the rated capacity of the target domain battery.

[0019] Further, in S300, data processing is performed on the sample sets of the source domain battery and the target domain battery, which specifically includes:

[0020] S310. Determine the voltage window of the source domain battery or the target domain battery; where the voltage window of the source domain battery is expressed as W S , and the voltage window of the target domain battery is expressed as W T , that is:

[0021] W S =[V S start , V S start +ΔV1, V S start +2ΔV1,..., V S end -ΔV1, V S end ;

[0022] W T =[V T start , V T start +ΔV2, V T start +2ΔV2,..., V T end -ΔV2, V T end ;

[0023] where V Sstart is the starting voltage of the source - domain battery voltage window, V T start is the starting voltage of the target - domain battery voltage window; V S end is the ending voltage of the source - domain battery voltage window, V T end is the ending voltage of the target - domain battery voltage window; ΔV1 is the voltage interval of the source - domain battery voltage window, and ΔV2 is the voltage interval of the target - domain battery voltage window; the number of voltage points in the voltage window is determined by the starting voltage, ending voltage, and voltage interval, that is:

[0024]

[0025] where N represents the number of voltage points in the voltage window;

[0026] S320. Use the linear interpolation method to interpolate the data corresponding to the voltage V and charge capacity Q of each sample in the sample sets of the source - domain battery and the target - domain battery, obtain the capacity data at each voltage point in the voltage window, form a charge - capacity sequence, and obtain a charge - capacity sequence sample set;

[0027] The charge - capacity sequence of the source - domain battery is:

[0028] P i =[p i1 ,p i2 ,…,p iN-1 ,p iN ;

[0029] The charge - capacity sequence of the target - domain battery is:

[0030] P j =[p j1 ,p j2 ,…,p jN-1 ,p jN ;

[0031] where P i represents the charge - capacity sequence obtained from the i - th sample in the source - domain battery sample set, and p ix (x = 1, 2,…,N - 1,N) represents the charge capacity corresponding to the voltage point at that place; P j represents the charge - capacity sequence obtained from the j - th sample in the target - domain battery sample set, and p jy (y = 1, 2,…,N - 1,N) represents the charge capacity corresponding to the voltage point at that place;

[0032] S330. Normalize the obtained charging capacity sequence sample set by using the rated capacities of the source domain battery and the target domain battery respectively, to obtain the normalized charging capacity sequence sample set, that is:

[0033]

[0034] Among them, S i represents the normalized charging capacity sequence obtained from the i-th sample in the source domain battery sample set, and q ix (x = 1, 2, …, N−1, N) represents the normalized charging capacity corresponding to the voltage point here; S j represents the normalized charging capacity sequence obtained from the j-th sample in the target domain battery sample set, and q jy (y = 1, 2, …, N−1, N) represents the normalized charging capacity corresponding to the voltage point here.

[0035] Furthermore, in S300, the processed data is analyzed by using the dynamic mode decomposition method, specifically including:

[0036] S340. Process the normalized charging capacity sequence sample set by using the dynamic mode decomposition method to obtain the charging capacity change mode sequence in the constant current charging stage;

[0037] The charging capacity change mode sequence of the source domain battery is:

[0038] M i = [m i2 , m i3 , …, m iN-1 , m iN ;

[0039] The charging capacity change mode sequence of the target domain battery is:

[0040] M j = [m j2 , m j3 , …, m jN-1 , m jN ;

[0041] Among them, M i represents the charging capacity change mode sequence obtained from the i-th sample in the source domain battery charging capacity sequence sample set, and m ix (x = 2, 3, …, N−1, N) represents the change mode corresponding to the source domain battery; M j represents the charging capacity change mode sequence obtained from the j-th sample in the target domain battery charging capacity sequence sample set, and m jy (y = 2, 3, …, N−1, N) represents the change mode corresponding to the target domain battery;

[0042] S350. Remove the charging capacity at the first voltage point, obtain the normalized charging capacity sequence, and form a binary sequence with the charging capacity change pattern sequence to obtain the binary sequence sample sets of the source domain battery and the target domain battery. The corresponding binary sequences are expressed as:

[0043] INPUT i =[(q i2 , m i2 ), (q i3 , m i3 ), …, (q iN-1 , m iN-1 ), (q iN , m iN )];

[0044] INPUT j =[(q j2 , m j2 ), (q j3 , m j3 ), …, (q jN-1 , m jN-1 ), (q jN , m jN )];

[0045] Among them, INPUT i represents the binary sequence obtained by processing the i-th sample of the source domain battery sample set, and INPUT j represents the binary sequence obtained by processing the j-th sample of the target domain battery sample set.

[0046] Furthermore, in S400, it specifically includes:

[0047] S410. The constructed domain adversarial neural network model includes a feature extractor, a predictor, and a discriminator; the feature extractor is composed of a convolutional neural network and a long short-term memory neural network, and both the predictor and the discriminator are composed of stacked fully connected layers;

[0048] S420. Add a bridging module between the feature extractor and the predictor;

[0049] S430. Input the binary sequence sample set of the source domain battery into the neural network containing only the feature extraction module and the prediction module to establish the mapping relationship between the binary sequence of the source domain battery and its corresponding label set;

[0050] After training, the parameters of the feature extraction module and the prediction module will be transferred to the feature extractor and the predictor in the domain adversarial neural network respectively as initial parameters;

[0051] S440. Use the binary sequence sample sets of the source domain battery and the target domain battery as the training and validation data sets, and input them into the constructed domain adversarial neural network. Compare the output of the predictor with the label set of the source domain battery, calculate the loss value of the predictor using the mean squared error, and backpropagate the loss value to optimize and adjust the weights of the feature extractor and the predictor. The discriminator receives the features output by the feature extractor and outputs the source discrimination of the corresponding features. Calculate the discriminant error loss value using the sparse multi-class cross-entropy loss function, and backpropagate the loss value through the gradient reversal layer to optimize and adjust the weights of the feature extractor and the discriminator.

[0052] Further, in S430, the structures of the feature extraction module and the prediction module of the neural network are respectively the same as the structures of the feature extractor and the predictor in the domain adversarial neural network. Transfer the parameters of the feature extraction module and the prediction module to the feature extractor and the predictor in the domain adversarial neural network respectively to learn the degradation information of the source domain battery.

[0053] Further, in S500, it specifically includes:

[0054] S510. Obtain the battery to be estimated with the same model as the target domain battery;

[0055] S520. Obtain the cyclic test data of the battery to be estimated, and process it according to S200 - S300 to obtain the binary sequence and the label set of the battery to be estimated. Input the binary sequence of the battery to be estimated into the trained domain adversarial neural network model, and output and obtain the health state estimation value of the battery to be estimated.

[0056] A lithium-ion battery health state estimation system based on unsupervised transfer learning applies an unsupervised transfer learning-based lithium-ion battery health state estimation method as described in any one of the above, including:

[0057] Data acquisition module: used to acquire the data of each cycle of the source domain battery and the target domain battery for cyclic testing; including the data for training the domain adversarial neural network model and the data of the battery to be estimated;

[0058] Data processing module: used to preprocess the acquired cyclic test data, including: collecting the voltage V and charging capacity Q data in the constant current charging stage; collecting the maximum charging capacity data in the discharging stage to obtain the label set; obtaining the sample set according to the voltage window and normalizing the sample set; processing the normalized sequence data using the dynamic mode decomposition method to obtain the binary sequence sample set; dividing the binary sequence sample set according to a ratio to generate the training and validation data sets used when training the domain adversarial neural network; the data processing module simultaneously processes the data of the battery to be estimated;

[0059] Model training module: Use the obtained training and validation data and the label set to train the domain adversarial neural network model, and establish the mapping relationship between the input feature INPUT i and the output label SOH i to obtain the trained domain adversarial neural network model;

[0060] Prediction module: Input the processed data of the battery to be estimated into the trained domain adversarial neural network model to obtain the estimated value of the state of health of the battery to be estimated.

[0061] An electronic device for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning, including a processor, a memory, and a bus;

[0062] The memory stores machine-readable instructions executable by the processor. When the electronic device runs, the processor communicates with the memory through the bus. When the machine-readable instructions are run by the processor, they execute the method for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning described in any one of the above.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] 1. The present invention can complete the estimation of the state of health of the battery to be tested with the same model as the target domain battery by using the cyclic test data and calibration test data of the existing source domain battery and the cyclic test data of a small amount of target domain batteries that do not require calibration tests.

[0065] 2. The present invention analyzes the charging capacity sequence through dynamic mode decomposition, can capture the change characteristics of the charging capacity in the constant current charging stage, and has a high correlation with the label data; by building a domain adversarial neural network to establish the mapping relationship between the binary sequence sample set and the label set, and adding a transfer learning strategy, the knowledge transfer from the source domain battery to the target domain battery is realized, and the estimation of the state of health of the battery with the same model as the target domain battery is completed without additional calibration tests for the target domain battery, improving the production efficiency.

[0066] 3. The present invention uses simple data and simple data processing. The input features of the neural network are provided by the charging capacity and the change pattern obtained by its dynamic mode decomposition, avoiding the complex feature extraction process.

[0067] 4. The present invention takes into account the large distribution difference between the source domain battery and the target domain battery, and adds a bridging module to enable the model to complete the slow migration from the source domain battery to the target domain battery, so that the feature extractor can better extract the domain-invariant features, improving the accuracy of the state of health estimation. Description of the Drawings

[0068] Figure 1It is a flowchart of a lithium-ion battery health state estimation method based on unsupervised transfer learning provided in Embodiment 1 of the present invention;

[0069] Figure 2 It is a flowchart of data processing in a lithium-ion battery health state estimation method based on unsupervised transfer learning provided in Embodiment 1 of the present invention;

[0070] Figure 3 It is a schematic diagram of the neural network structure in a lithium-ion battery health state estimation method based on unsupervised transfer learning provided in Embodiment 1 of the present invention;

[0071] Figure 4 It is a schematic diagram of the domain adversarial neural network structure in a lithium-ion battery health state estimation method based on unsupervised transfer learning provided in Embodiment 1 of the present invention;

[0072] Figure 5 It is a test result graph of the source domain battery correlation in a lithium-ion battery health state estimation method based on unsupervised transfer learning provided in Embodiment 1 of the present invention;

[0073] Figure 6 It is a test result graph of the target domain battery correlation in a lithium-ion battery health state estimation method based on unsupervised transfer learning provided in Embodiment 1 of the present invention;

[0074] Figure 7 It is a test result graph of the target domain battery correlation in a lithium-ion battery health state estimation method based on unsupervised transfer learning provided in Embodiment 2 of the present invention;

[0075] Figure 8 It is a schematic diagram of the functional modules of a lithium-ion battery health state estimation system based on unsupervised transfer learning provided in Embodiment 1 of the present invention;

[0076] Figure 9 It is a schematic diagram of the structure of an electronic device for lithium-ion battery health state estimation based on unsupervised transfer learning provided in Embodiment 1 of the present invention. Detailed implementation manners

[0077] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without any creative work shall fall within the protection scope of the present invention.

[0078] Embodiment 1

[0079] As Figures 1 to 9 , a method for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning includes the following steps:

[0080] S100. Select sample batteries of different models and divide the sample batteries into source domain batteries and target domain batteries;

[0081] The source domain batteries and target domain batteries selected in this embodiment are shown in Table 1:

[0082] Table 1

[0083]

[0084] S200. Obtain the cyclic test data of the source domain batteries and target domain batteries, and collect the sample set and label set of the sample batteries, where the number of samples of the source domain batteries is greater than the number of samples of the target domain batteries. Specifically, obtain the cyclic test data of multiple source domain batteries and a small number of target domain batteries;

[0085] S210. In the obtained cyclic test data, collect the data of voltage V and charge capacity Q in each cyclic constant current charging stage, and form the sample set of the source domain batteries or the sample set of the target domain batteries with all the obtained voltages V and charge capacities Q according to whether the sample battery is a source domain battery or a target domain battery.

[0086] In this embodiment, the number of source domain battery samples obtained is 4262, and the number of samples of the target domain batteries (batteries to be estimated) is 1733. Select 20% of the samples in the target domain battery sample set as part of the training and validation data set of the domain adversarial neural network model, and the remaining 80% of the samples do not need to be tested.

[0087] S220. In the obtained cyclic test data, collect the maximum discharge capacity data in the discharge stage to obtain the health state value SOH corresponding to the battery health degree, and form the label set of the source domain batteries or the label set of the target domain batteries with all the obtained health state values SOH; The health state value SOH is specifically expressed as:

[0088]

[0089] where i represents the i-th sample in the source domain battery sample set, and SOH i is the health state value of the i-th sample; j represents the j-th sample in the target domain battery sample set, and SOH j is the health state value of the j-th sample; Q i represents the maximum discharge capacity in the source domain battery sample set, and Q j represents the maximum discharge capacity in the target domain battery sample set; Q S represents the rated capacity of the source domain battery, and QT Represents the rated capacity of the target domain battery.

[0090] S300. Perform data processing on the sample sets of the source domain battery and the target domain battery, analyze the processed data using the dynamic mode decomposition method, obtain the charging capacity change mode sequence in the constant current charging stage of the lithium battery, and construct a binary sequence as the training and validation data set; specifically including:

[0091] S310. Manually determine the voltage window of the source domain battery or the target domain battery; where the voltage window of the source domain battery is denoted as W S , and the voltage window of the target domain battery is denoted as W T , that is:

[0092] W S = [V S start , V S start + ΔV1, V S start + 2ΔV1,..., V S end - ΔV1, V S end ;

[0093] W T = [V T start , V T start + ΔV2, V T start + 2ΔV2,..., V T end - ΔV2, V T end ;

[0094] Where, V S start is the starting voltage of the voltage window of the source domain battery, V T start is the starting voltage of the voltage window of the target domain battery; V S end is the ending voltage of the voltage window of the source domain battery, V T end is the ending voltage of the voltage window of the target domain battery; ΔV1 is the voltage interval of the voltage window of the source domain battery, ΔV2 is the voltage interval of the voltage window of the target domain battery; the number of voltage points in the voltage window is determined by the starting voltage, the ending voltage, and the voltage interval, that is:

[0095]

[0096] Wherein, N represents the number of voltage points in the voltage window;

[0097] In this embodiment, V S start , V S end and ΔV1 are 3.2V, 3.6V, 0.04V respectively, V T start , V T end and ΔV2 are 3.75V, 4.15V and 0.04V respectively, and the number of voltage points N is 11.

[0098] S320. Use the linear interpolation method to interpolate the data corresponding to the voltage V and the charging capacity Q of each sample in the sample sets of the source domain battery and the target domain battery, obtain the capacity data at each voltage point in the voltage window, form a charging capacity sequence, and obtain a charging capacity sequence sample set;

[0099] The charging capacity sequence of the source domain battery is:

[0100] P i = [p i1 , p i2 , …, p iN-1 , p iN ;

[0101] The charging capacity sequence of the target domain battery is:

[0102] P j = [p j1 , p j2 , …, p jN-1 , p jN ;

[0103] Wherein, P i represents the charging capacity sequence obtained from the i-th sample in the source domain battery sample set, and p ix (x = 1, 2, …, N - 1, N) represents the charging capacity corresponding to the voltage point at that place; P j represents the charging capacity sequence obtained from the j-th sample in the target domain battery sample set, and p jy (y = 1, 2, …, N - 1, N) represents the charging capacity corresponding to the voltage point at that place;

[0104] S330. Respectively use the rated capacities of the source domain battery and the target domain battery to normalize the obtained charging capacity sequence sample set to obtain a normalized charging capacity sequence sample set, that is:

[0105]

[0106] Wherein, S iDenote the normalized charging capacity sequence obtained from the \(i\)-th sample in the source domain battery sample set as \(q\). ix (\(x = 1, 2, \ldots, N - 1, N\)) represents the normalized charging capacity corresponding to the voltage point at this location; \(S\). j Denote the normalized charging capacity sequence obtained from the \(j\)-th sample in the target domain battery sample set as \(q\). jy (\(y = 1, 2, \ldots, N - 1, N\)) represents the normalized charging capacity corresponding to the voltage point at this location.

[0107] S340. Process the normalized charging capacity sequence sample set using the dynamic mode decomposition method to obtain the charging capacity change mode sequence in the constant current charging stage.

[0108] Dynamic mode decomposition (DMD) is a data-driven analysis method mainly used to extract the key dynamic modes of a system from the time series data of a dynamic system. Dynamic mode decomposition identifies the dynamic modes of the system by constructing a data matrix, singular value decomposition, and eigenvalue decomposition. These modes reflect the main behavioral characteristics of the system. Dynamic mode decomposition does not depend on a specific physical model and only relies on data for analysis, and can reveal the dominant dynamic behavior of the system.

[0109] The charging capacity change mode sequence of the source domain battery is:[[]]

[0110] M i = [m i2 , m i3 , \ldots, m iN-1 , m iN ;

[0111] The charging capacity change mode sequence of the target domain battery is:[[]]

[0112] M j = [m j2 , m j3 , \ldots, m jN-1 , m jN ;

[0113] Among them, \(M\) i represents the charging capacity change mode sequence obtained from the \(i\)-th sample in the source domain battery charging capacity sequence sample set, and \(m\) ix (\(x = 2, 3, \ldots, N - 1, N\)) represents the change mode corresponding to the source domain battery; \(M\) j represents the charging capacity change mode sequence obtained from the \(j\)-th sample in the target domain battery charging capacity sequence sample set, and \(m\) jy (\(y = 2, 3, \ldots, N - 1, N\)) represents the change mode corresponding to the target domain battery.

[0114] It should be noted that the length of the charging capacity change pattern sequence is the length of the charging capacity sequence minus 1. Here, both x and y start from 2 for the convenience of representing the subsequent binary sequence formed.

[0115] S350. Remove the charging capacity at the first voltage point, obtain the normalized charging capacity sequence, and form a binary sequence with the charging capacity change pattern sequence to obtain the binary sequence sample sets of the source domain battery and the target domain battery. The corresponding binary sequences are expressed as:

[0116] INPUT i =[(q i2 ,m i2 ),(q i3 ,m i3 ),…,(q iN-1 ,m iN-1 ),(q iN ,m iN )];

[0117] INPUT j =[(q j2 ,m j2 ),(q j3 ,m j3 ),…,(q jN-1 ,m jN-1 ),(q jN ,m jN )];

[0118] Among them, INPUT i represents the binary sequence obtained by processing the i-th sample of the source domain battery sample set, and INPUT j represents the binary sequence obtained by processing the j-th sample of the target domain battery sample set.

[0119] In this embodiment, the length of each sequence in the binary sequence is 10.

[0120] In this embodiment, the dimension of each binary sequence is (10, 2), representing the input features with a time step of 10 and a feature dimension of 2.

[0121] As Figure 5 shown, in this embodiment, the Pearson correlation results of the two dimensions in the binary sequence of a source domain battery with the label set are both above 0.9, showing a very strong correlation. Among them Figure 5 (a) is the first dimension, Figure 5 (b) is the second dimension.

[0122] As Figure 6As shown, in this embodiment, the Pearson correlation results between the two dimensions in the binary sequence of a target domain battery and the label set are both around 0.8, showing a very strong correlation, where Figure 6 (a) is the first dimension, Figure 6 (b) is the second dimension.

[0123] S400. Build an unsupervised transfer learning domain adversarial neural network model with a bridging module as shown in Figure 3 . Use the training and validation data set and the label set of the source domain battery to train the domain adversarial neural network model to obtain a trained domain adversarial neural network model;

[0124] S410. The built domain adversarial neural network model includes a feature extractor, a predictor, and a discriminator; the feature extractor is composed of a convolutional neural network (CNN) and a long short-term memory neural network (LSTM); both the predictor and the discriminator are composed of stacked fully connected layers;

[0125] The feature extractor is composed of an input layer, two CNN layers, and two LSTM layers. The output of the feature extractor is the abstract features extracted from the binary sequence INPUT i or INPUT j . Convolutional neural networks are a type of deep neural network widely used to process data with a lattice structure. CNNs use convolutional kernels to slide on the input data through convolutional layers to extract local features and gradually learn abstract features from low-level to high-level, thus avoiding the complexity of manual feature engineering; Long short-term memory networks are a special type of recurrent neural network (RNN) designed to solve the problems of vanishing gradients and exploding gradients faced by traditional RNNs when processing long-time sequence data. The core advantage of LSTM is that it can effectively retain dependency information on a long time scale and overcome the short-term memory limitations of traditional RNNs;

[0126] The predictor and the discriminator are composed of two fully connected layers and an output layer, receive the abstract features from the feature extractor, the predictor outputs the health state estimation value, and the discriminator discriminates the source of the abstract features.

[0127] S420. Add a bridging module between the feature extractor and the predictor; the position of the bridging module in the neural network is as shown in Figure 4 .

[0128] The detailed explanation of the bridging module is as follows:

[0129] Assume that the source domain feature is represented as: S ∈ R n×d, the target domain feature is represented as: T ∈ R m×d , where n and m respectively represent the number of samples in the source domain and the target domain sample set, d is the feature dimension. In this embodiment, the feature dimension is two-dimensional, and the value of d is 2.

[0130] In the bridging module, the features of the source domain and the target domain are fused by weighted averaging to obtain a new feature representation:

[0131] B = (1 - α)·S + α·T

[0132] where, B ∈ R n×d is the fused feature, α ∈ [0, 1] is the dynamic smoothing factor, and the calculation formula is as follows:

[0133]

[0134] where, epoch is the current training round, and total_epochs is the total number of training rounds.

[0135] α will gradually increase the contribution of the target domain feature to the fused feature in a smooth manner through the control of the sigmoid function during the training process of the domain adversarial neural network model.

[0136] Through the bridging module, it helps to achieve the gradual adaptation between the source domain and the target domain and promote the effect of transfer learning.

[0137] S430. Input the binary sequence sample set of the source domain battery into the neural network that only contains the feature extraction module and the prediction module, and establish the mapping relationship between the binary sequence of the source domain battery and its corresponding label set;

[0138] After training, the parameters of the feature extraction module and the prediction module will be transferred to the feature extractor and the predictor in the domain adversarial neural network respectively as initial parameters;

[0139] The structures of the feature extraction module and the prediction module in this neural network are respectively the same as the structures of the feature extractor and the predictor in the domain adversarial neural network. Transferring the parameters of the feature extraction module and the prediction module to the feature extractor and the predictor in the domain adversarial neural network respectively can enable the model to first learn the degradation information of the source domain battery.

[0140] S440. Use the binary sequence sample sets of the source domain battery and the target domain battery as the training and validation data sets, and input them into the constructed domain adversarial neural network. Compare the output of the predictor with the label set of the source domain battery, calculate the loss value of the predictor using the mean square error (MSE), and backpropagate the loss value to optimize and adjust the weights of the feature extractor and the predictor. The discriminator receives the features output from the feature extractor and outputs the discrimination of the source of the features. Use the sparse categorical cross entropy loss function (SCCE) to calculate the discrimination error loss value, and backpropagate the loss value through the gradient reversal layer to optimize and adjust the weights of the feature extractor and the discriminator.

[0141] Specifically, the calculation formulas of the mean square error (MSE) and the sparse categorical cross entropy (SCCE) loss functions are as follows:

[0142]

[0143] Among them, is the model prediction value, y i is the true label, and G is the number of samples.

[0144]

[0145] Among them, z i is the unnormalized predicted value vector output by the model, a i is the true class label of the i-th sample, C is the total number of classes, and G is the number of samples.

[0146] S500. Obtain the cyclic test data of the battery to be estimated and perform data processing. Among them, the model of the battery to be estimated is the same as that of the target domain battery, and then input it into the trained domain adversarial neural network model to obtain the health state estimation value of the battery to be estimated.

[0147] S510. Obtain the battery to be estimated with the same model as the target domain battery; that is, the model of the battery to be estimated needs to be the same as that of the target domain battery;

[0148] S520. Obtain the cyclic test data of the battery to be estimated, and process it according to S200 - S300 to obtain the binary sequence and label set of the battery to be estimated; input the binary sequence of the battery to be estimated into the trained domain adversarial neural network model, and output and obtain the health state estimation value of the battery to be estimated.

[0149] In this embodiment, the trained domain adversarial neural network is used to estimate the health state of the remaining 80% of the batteries to be tested. The estimation results are mainly evaluated using the mean absolute error (MAE), the maximum absolute error (Max AE), and the root mean square error (RMSE), as shown in Table 2:

[0150] Table 2

[0151] MAE (%) Max AE (%) RMSE 1.781 5.263 0.0222

[0152] As can be seen from the data in the table, in the absence of the target domain battery label data, the MAE of the model is 1.781%, and the maximum absolute error is only 5.263%. The health state of the batteries to be tested can be estimated relatively accurately.

[0153] Such as Figure 8 , a lithium-ion battery health state estimation system based on unsupervised transfer learning, which applies a lithium-ion battery health state estimation method based on unsupervised transfer learning as described above, includes:

[0154] Data acquisition module: used to acquire the data of each cycle of the source domain battery and the target domain battery for the cyclic test; including the data for training the domain adversarial neural network model and the data of the battery to be estimated;

[0155] Data processing module: used to preprocess the acquired cyclic test data, including: collecting the voltage V and the charge capacity Q data in the constant current charging stage; collecting the maximum charge capacity data in the discharging stage to obtain the label set; obtaining the sample set according to the voltage window and normalizing the sample set; processing the normalized sequence data using the dynamic mode decomposition method and obtaining the binary sequence sample set; dividing the binary sequence sample set according to a certain ratio to generate the training and validation data set used when training the domain adversarial neural network; the data processing module also processes the data of the battery to be estimated;

[0156] Model training module: using the obtained training and validation data and the label set to train the domain adversarial neural network model, and establishing the mapping relationship between the input feature INPUT i and the output label SOH i to obtain the trained domain adversarial neural network model;

[0157] Prediction module: inputting the processed data of the battery to be estimated into the trained domain adversarial neural network model to obtain the health state estimation value of the battery to be estimated.

[0158] Such as Figure 9, an electronic device for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning, comprising a processor, a memory, and a bus;

[0159] The memory stores machine-readable instructions executable by the processor. When the electronic device runs, the processor communicates with the memory through the bus. When the machine-readable instructions are run by the processor, the method for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning as described above is executed.

[0160] A computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, the method for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning as described above is executed.

[0161] Embodiment 2

[0162] As Figures 1 to 9 , in this embodiment, the detailed information of the source domain battery and the target domain battery selected is shown in Table 3:

[0163] Table 3

[0164]

[0165] In this embodiment, V S start , V S end and ΔV1 are 3.2V, 3.6V, and 0.04V respectively, V T start , V T end and ΔV2 are 3.75V, 4.15V, and 0.04V respectively, and the number of voltage points N is 11.

[0166] As Figure 7 shown, in this embodiment, the Pearson correlation results of the two dimensions in the binary sequence of a target domain battery with the label set are both above 0.9, showing a very strong correlation, where Figure 7 (a) is the first dimension, Figure 7 (b) is the second dimension.

[0167] Except for the above description, other data processing, the construction of the domain adversarial neural network, and other settings in this embodiment are the same as those in Embodiment 1, and will not be repeated here.

[0168] In this embodiment, the trained domain adversarial neural network is used to estimate the state of health of the remaining 80% of the samples of the batteries to be tested, and the estimation results are shown in Table 4:

[0169] Table 4

[0170] MAE (%) Max AE (%) RMSE 1.326 3.701 0.0157

[0171] As can be seen from the data in the table, in the case of no target domain battery label data, the MAE of the model is 1.326%, and the maximum absolute error is only 3.701%, which can relatively accurately estimate the health state of the battery to be tested.

[0172] Next, by comparing the neural network model of the domain adaptation method with the bridging module added with the ordinary transfer learning neural network model without the above method, the beneficial effects of the technical solution of the present disclosure are further demonstrated.

[0173] The three types of battery information used in the comparative experiment are shown in Table 5:

[0174] Table 5

[0175]

[0176]

[0177] The results of the comparative experiment are shown in Table 6, where AD represents the neural network model of the domain adaptation method with the bridging module added, and BD represents the ordinary transfer learning neural network model without the above method. To ensure fairness, the hyperparameter settings of the two methods are the same.

[0178] Table 6

[0179]

[0180] As can be seen from the results in Table 6, the accurate values of the health state estimation of the neural network model of the domain adaptation method with the bridging module added have been generally improved.

[0181] According to the disclosure and teachings of the above specification, those skilled in the art to which the present invention pertains can also make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and some modifications and changes to the present invention should also fall within the protection scope of the claims of the present invention. In addition, although some specific terms are used in this specification, these terms are only for convenience of description and do not constitute any limitation to the present invention.

Claims

1. A method for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning, characterized in that, It includes the following steps: S100. Select sample batteries of different models, and divide the sample batteries into a source domain battery and a target domain battery; S200. Obtain the cyclic test data of the source domain battery and the target domain battery, and collect the sample set and label set of the sample batteries, where the number of samples of the source domain battery is greater than that of the target domain battery; S300. Process the sample sets of the source domain battery and the target domain battery, analyze the processed data by using the dynamic mode decomposition method, obtain the charging capacity change mode sequence in the constant current charging stage of the battery, and construct a binary sequence as the training and validation data set; S400. Build an unsupervised transfer learning domain adversarial neural network model with a bridging module, and use the training and validation data set and the label set of the source domain battery to train the domain adversarial neural network model to obtain a trained domain adversarial neural network model; S500. Obtain the cyclic test data of the battery to be estimated and perform data processing. The model of the battery to be estimated is the same as that of the target domain battery, and then input it into the trained domain adversarial neural network model to obtain the health state estimation value of the battery to be estimated.

2. The method for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning according to claim 1, wherein In S200, collecting the sample set of the sample battery specifically includes: S210. In the obtained cyclic test data, collect the data of voltage V and charging capacity Q in each cycle of the constant current charging stage, and according to whether the sample battery is a source domain battery or a target domain battery, form the sample set of the source domain battery or the sample set of the target domain battery with all the obtained voltages V and charging capacities Q.

3. A method for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning according to claim 2, wherein In S200, collecting the label set of the sample battery specifically includes: S220. In the obtained cyclic test data, collect the maximum discharge capacity data in the discharge stage to obtain the health state value SOH corresponding to the battery health degree, and form the label set of the source domain battery or the label set of the target domain battery with all the obtained health state values SOH; the health state value SOH is specifically expressed as: where \(i\) represents the \(i\)-th sample in the source domain battery sample set, and SOH i is the state of health value of the \(i\)-th sample; \(j\) represents the \(j\)-th sample in the target domain battery sample set, and SOH j is the state of health value of the \(j\)-th sample; \(Q\ i represents the maximum discharge capacity in the source domain battery sample set, \(Q\ j represents the maximum discharge capacity in the target domain battery sample set; \(Q\ S represents the rated capacity of the source domain battery, \(Q\ T represents the rated capacity of the target domain battery.

4. A method for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning according to claim 3, characterized in that In S300, processing the sample sets of the source domain battery and the target domain battery specifically includes: S310. Determine the voltage window of the source-domain battery or the target-domain battery; where the voltage window of the source-domain battery is denoted as W S , and the voltage window of the target-domain battery is denoted as W T , that is: W S = [V Sstart , V Sstart + ΔV1, V Sstart + 2ΔV1, …, V Send - ΔV1, V Send ; W T = [V Tstart , V Tstart +ΔV2, V Tstart +2ΔV2, …, V Tend -ΔV2, V Tend ; Among them, V Sstart is the starting voltage of the source domain battery voltage window, V Tstart is the starting voltage of the target domain battery voltage window; V Send is the ending voltage of the source domain battery voltage window, V Tend is the ending voltage of the target domain battery voltage window; ΔV1 is the voltage interval of the source domain battery voltage window, and ΔV2 is the voltage interval of the target domain battery voltage window; the number of voltage points in the voltage window is determined by the starting voltage, the ending voltage, and the voltage interval, that is: where N represents the number of voltage points in the voltage window; S320. Use the linear interpolation method to perform interpolation processing on the data corresponding to the voltage V and the charging capacity Q of each sample in the sample sets of the source domain battery and the target domain battery to obtain the capacity data at each voltage point in the voltage window, form the charging capacity sequence, and obtain the charging capacity sequence sample set; The charging capacity sequence of the source domain battery is: P i = [p i1 , p i2 , …, p iN-1 , p iN ; The charging capacity sequence of the target domain battery is: P j = [p j1 , p j2 , …, p jN-1 , p jN ; Among them, P i represents the charging capacity sequence obtained from the i-th sample in the source domain battery sample set, and p ix (x = 1, 2, …, N - 1, N) represents the charging capacity corresponding to the voltage point here; P j represents the charging capacity sequence obtained from the j-th sample in the target domain battery sample set, and p jy (y = 1, 2, …, N - 1, N) represents the charging capacity corresponding to the voltage point here; S330. Respectively perform normalization processing on the charging capacity sequence sample sets obtained by using the rated capacities of the source domain battery and the target domain battery to obtain the normalized charging capacity sequence sample sets, that is: Among them, S i represents the normalized charging capacity sequence obtained from the i-th sample in the source domain battery sample set, and q ix (x = 1, 2, …, N - 1, N) represents the normalized charging capacity corresponding to the voltage point here; S j represents the normalized charging capacity sequence obtained from the j-th sample in the target domain battery sample set, and q jy (y = 1, 2, …, N - 1, N) represents the normalized charging capacity corresponding to the voltage point here.

5. The method for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning according to claim 4, wherein In S300, analyzing the processed data by using the dynamic mode decomposition method specifically includes: S340. Use the dynamic mode decomposition method to process the normalized charging capacity sequence sample set to obtain the charging capacity change mode sequence in the constant current charging stage; The charging capacity change mode sequence of the source domain battery is: M i = [m i2 , m i3 , …, m iN-1 , m iN ; The charging capacity change mode sequence of the target domain battery is: M j = [m j2 , m j3 , …, m jN-1 , m jN ; Among them, M i represents the charging capacity change pattern sequence obtained from the i-th sample in the source domain battery charging capacity sequence sample set, and m ix (x = 2, 3, …, N - 1, N) represents the change pattern corresponding to the source domain battery; M j represents the charging capacity change pattern sequence obtained from the j-th sample in the target domain battery charging capacity sequence sample set, and m jy (y = 2, 3, …, N - 1, N) represents the change pattern corresponding to the target domain battery; S350. Remove the charging capacity at the first voltage point, obtain the normalized charging capacity sequence, and combine it with the charging capacity change pattern sequence to form a binary sequence, obtaining the binary sequence sample sets of the source domain battery and the target domain battery. The corresponding binary sequence is expressed as: INPUT i = [(q i2 , m i2 ), (q i3 , m i3 ), …, (q iN-1 , m iN-1 ), (q iN , m iN )]; INPUT j = [(q j2 , m j2 ), (q j3 , m j3 ), …, (q jN-1 , m jN-1 ), (q jN , m jN )]; Among them, INPUT i represents the binary sequence obtained by processing the i-th sample of the source domain battery sample set, INPUT j represents the binary sequence obtained by processing the j-th sample of the target domain battery sample set.

6. The method for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning according to claim 1, wherein In S400, it specifically includes: S410. The constructed domain adversarial neural network model includes a feature extractor, a predictor, and a discriminator; the feature extractor is composed of a convolutional neural network and a long short-term memory neural network, and both the predictor and the discriminator are composed of stacked fully connected layers; S420. Add a bridging module between the feature extractor and the predictor; S430. Input the binary sequence sample set of the source domain battery into a neural network that only contains a feature extraction module and a prediction module, and establish a mapping relationship between the binary sequence of the source domain battery and its corresponding label set; After training is completed, the parameters of the feature extraction module and the prediction module will be transferred to the feature extractor and the predictor in the domain adversarial neural network respectively as initial parameters; S440. Use the binary sequence sample sets of the source domain battery and the target domain battery as the training and validation data sets and input them into the constructed domain adversarial neural network; compare the output of the predictor with the label set of the source domain battery, calculate the loss value of the predictor using the mean squared error, and backpropagate the loss value to optimize and adjust the weights of the feature extractor and the predictor; the discriminator receives the features output by the feature extractor and outputs the source discrimination of the corresponding features, calculates the discrimination error loss value using the sparse multi-class cross-entropy loss function, and backpropagates the loss value through the gradient reversal layer to optimize and adjust the weights of the feature extractor and the discriminator.

7. A method for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning according to claim 6, wherein In S430, the structures of the feature extraction module and the prediction module of the neural network are the same as those of the feature extractor and the predictor in the domain adversarial neural network respectively; transfer the parameters of the feature extraction module and the prediction module to the feature extractor and the predictor in the domain adversarial neural network respectively to learn the degradation information of the source domain battery.

8. A method for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning according to claim 1, wherein, In S500, it specifically includes: S510. Obtain a battery to be estimated with the same model as the target domain battery; S520. Obtain the cyclic test data of the battery to be estimated, and process it according to S200 - S300 to obtain the binary sequence and label set of the battery to be estimated; input the binary sequence of the battery to be estimated into the trained domain adversarial neural network model, and output and obtain the health state estimation value of the battery to be estimated.

9. A lithium-ion battery state of health estimation system based on unsupervised transfer learning, characterized in that, An estimation method for the health state of a lithium-ion battery based on unsupervised transfer learning as claimed in any one of claims 1 to 8, comprising: A data acquisition module: used to acquire the data of each cycle of the source domain battery and the target domain battery undergoing cyclic tests; including the data for training the domain adversarial neural network model and the data of the battery to be estimated; Data processing module: used to preprocess the acquired cyclic test data, including: collecting voltage V and charging capacity Q data during the constant current charging stage; collecting the maximum charging capacity data during the discharging stage to obtain a label set; obtaining a sample set according to the voltage window and normalizing the sample set; processing the normalized sequence data using the dynamic mode decomposition method and obtaining a binary sequence sample set; dividing the binary sequence sample set according to a ratio to generate a training and validation data set used when training the domain adversarial neural network; the data processing module simultaneously processes the data of the battery to be estimated; Model training module: Use the obtained training and validation data and the label set to train a domain adversarial neural network model, and establish the mapping relationship between the input feature INPUT i and the output label SOH i to obtain a trained domain adversarial neural network model; Prediction module: Input the processed data of the battery to be estimated into the trained domain adversarial neural network model to obtain an estimated value of the state of health of the battery to be estimated.

10. An electronic device for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning, characterized in that, It includes a processor, a memory, and a bus; The memory stores machine-readable instructions executable by the processor. When the electronic device runs, the processor communicates with the memory through the bus. When the machine-readable instructions are run by the processor, they execute a method for estimating the state of health of a lithium-ion battery based on unsupervised transfer learning as described in any one of claims 1 to 8.

Images