Lithium battery performance evaluation system and method based on multi-task deep learning

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses a lithium battery performance evaluation system and method based on multi-task deep learning, and the system comprises a feature generation module, a first multi-task prediction model and a second multi-task prediction model. The feature generation module is used for obtaining an optimal feature set related to the battery performance according to the information data of a plurality of charge-discharge cycle processes of the battery; the optimal feature set comprises cycle features related to the health state and the remaining life and time sequence features related to the state of charge; the first multi-task prediction model is used for predicting the health state and the residual life of the battery according to the cycle characteristics; and the second multi-task prediction model is used for predicting the state of charge of the battery according to the time sequence characteristics and the health state predicted by the first multi-task model. According to the method, the mutual promotion effect among the tasks is enhanced by sharing the feature information of the related tasks, so that the prediction accuracy and robustness are improved, and support is provided for safe use and performance improvement of the lithium battery.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and relates to the research on the prediction of the health state, remaining life and state of charge of lithium batteries, and particularly relates to a lithium battery performance evaluation system and method based on multi-task deep learning. Background Art

[0002] In recent years, lithium-ion batteries (LiBs) have become the main energy storage devices in fields such as portable electronic devices and electric vehicles due to their high energy density, long cycle life and low self-discharge rate. During their life cycle, the state of health (SOH), state of charge (SOC) and remaining useful life (RUL) are key indicators for measuring battery performance and have an important impact on the safety and reliability of the battery. The state of health refers to the ratio of the current available capacity of the battery to the rated capacity, and the remaining useful life refers to the number of cycles required from the current state to the end of life (EOL). The state of charge represents the percentage of the remaining capacity relative to the maximum capacity, and its accurate estimation can effectively prevent the battery from overcharging or over-discharging.

[0003] Traditional methods for estimating the state of health and remaining useful life mainly include direct calculation methods and model-based methods (such as equivalent circuit models and electrochemical models). These methods rely on accurate model parameters and strict hardware conditions, and have certain limitations in practical applications. Similarly, state of charge estimation methods include Coulomb counting method, open circuit voltage (OCV) method and model-based methods, but each has its own deficiencies. For example, the Coulomb counting method is prone to cumulative errors, while the OCV method is not applicable to dynamic load environments.

[0004] In recent years, data-driven estimation methods directly establish non-linear relationships by using machine learning and deep learning models with historical data, without the need for an accurate battery mechanism model, and have high prediction accuracy. For example, linear regression, support vector machine (SVM) and Gaussian process (GP) are applied to battery capacity prediction. Recurrent neural networks (such as RNN) perform excellently in capturing data features and are especially suitable for modeling non-linear high-dimensional systems. Models such as long short-term memory (LSTM) are used for state of charge and remaining useful life prediction with remarkable effects.

[0005] However, existing data-driven methods are mostly single-object models and do not fully consider the mutual relationship between the state of health, state of charge and remaining useful life, resulting in limited prediction accuracy. In addition, traditional models rely on specific charging data and are difficult to generalize to other application scenarios. At the same time, considering that battery charge and discharge data not only have time series characteristics, but there is also a certain spatial relationship between feature variables, and only focusing on the time series relationship is not comprehensive enough, thus affecting the model prediction effect. Summary of the Invention

[0006] The objective of the present invention aims to address the above problems existing in the prior art, and provides a lithium battery performance evaluation system and method based on multi-task deep learning. Based on multi-task learning (MTL), multiple related tasks can be placed in the same model. By establishing the correlation relationship between tasks, the information sharing of the model under different tasks can be enhanced, thereby improving the prediction accuracy and robustness of the state of health (SOH), remaining useful life (RUL), and state of charge (SOC), and further realizing the accurate evaluation of the lithium battery performance.

[0007] To achieve the above objective, the present invention provides a lithium battery performance evaluation system based on multi-task deep learning, which includes:

[0008] A feature generation module, configured to obtain an optimal feature set related to the battery performance according to the information data of several charge and discharge cycles of the battery; the optimal feature set includes cycle features related to the state of health and remaining useful life, and time series features related to the state of charge;

[0009] A first multi-task prediction model, configured to predict the state of health and remaining useful life of the battery according to the cycle features; the first multi-task prediction model includes a first variational autoencoder, a first principal component analysis module, and a first convolutional neural network; the first variational autoencoder is configured to reconstruct the cycle features to obtain corresponding reconstructed features; the first principal component analysis module is configured to perform dimensionality reduction processing on the reconstructed features obtained by the first variational autoencoder; the first convolutional neural network is configured to predict the state of health and remaining useful life of the battery according to the reconstructed features after dimensionality reduction processing by the first principal component analysis module.

[0010] A second multi-task prediction model, configured to predict the state of charge of the battery according to the time series features and the state of health predicted by the first multi-task model; the second multi-task prediction model includes a second variational autoencoder, a second principal component analysis module, and a second convolutional neural network; the second variational autoencoder is configured to reconstruct the time series features to obtain corresponding reconstructed features; the second principal component analysis module is configured to perform dimensionality reduction processing on the reconstructed features obtained by the second variational autoencoder; the second convolutional neural network is configured to predict the state of charge of the battery according to the charge state feature matrix constructed by splicing the reconstructed features after dimensionality reduction processing by the second principal component analysis module and the state of health predicted by the first multi-task model.

[0011] In one possible implementation, the information data during the battery charge and discharge cycle includes voltage, current, charging time, discharge time, number of charge and discharge cycles, etc. The cycle characteristics related to the health state and remaining life include the skewness, kurtosis, variance, maximum value, minimum value, average value of the relaxation charge voltage of each cycle, and the initial value and average value of the voltage in the charging stage, etc.; the Pearson correlation coefficient between each cycle characteristic and the health state and remaining life can also be calculated, and the cycle characteristics with an absolute value of the Pearson correlation coefficient greater than 0.7 can be selected. The timing characteristics related to the state of charge include timing data such as current, voltage, and time.

[0012] In one possible implementation, the variational autoencoder is a generative model that can map inputs to latent space representations and then reconstruct new representations to mine deep information within the data. Compared with other deep learning tasks, the battery state prediction task has fewer features. Using the first variational autoencoder and the second variational autoencoder to increase the dimension of the data can deeply mine and fully utilize the existing features, providing a basis for subsequent learning.

[0013] In one possible implementation, principal component analysis is performed on the representation output by the variational autoencoder through a principal component analysis module. Principal component analysis is a widely used data dimensionality reduction method for reducing the dimension of data while retaining the main features of the data. By applying principal component analysis, the reconstructed representation output by the variational autoencoder can be effectively denoised and purified, thereby further improving the quality of the representation.

[0014] In one possible implementation, the first convolutional neural network and the second convolutional neural network have the same structure, and are both convolutional neural networks based on the attention mechanism. Multi-task learning is performed on the features after principal component analysis through the convolutional neural network based on the attention mechanism. This model inherits both the ability of the convolutional neural network to capture local structural features and the ability of the attention mechanism to mine global long-term dependencies, and can more efficiently learn prediction targets that have internal connections and timing characteristics for battery health status, remaining life, and state of charge. Combined with the multi-task learning method, by establishing internal deep connections between the three prediction targets, the model can promote and improve the learning of the three, thereby achieving good prediction results.

[0015] The first convolutional neural network and the second convolutional neural network both include sequentially arranged convolutional layers, one or more moving flip bottleneck convolutional layers, one or more relative self-attention mechanism modules and one or more pooling layers.

[0016] The convolution layer (ConvLayer) is used to perform preliminary extraction of data features. As input, the initial single-layer convolution is defined as in A convolution kernel representing the quantity f and the size L, b q representing the bias, σ(·) being the activation function, is the output feature, m(=m - L + 1).

[0017] The mobile flipped bottleneck convolution layer (MBConv) is mainly used for local structural feature extraction. Compared with the ordinary convolution layer, an important change in the mobile flipped bottleneck convolution is the introduction of a squeeze - excitation module, enabling the model to automatically learn the importance of different feature channels, thereby improving the utilization rate of features according to the importance level.

[0018] The present invention defines the squeeze operator of the squeeze - excitation module in the mobile flipped bottleneck convolution layer as:

[0019]

[0020] where X t represents the feature of the input squeeze layer, and T represents the time step of the feature;

[0021] The activation operator is defined as:

[0022] e = σ(W SE2 ·δ(W SW1 ·s));

[0023] where W SE is the weight matrix, and σ and δ are the Sigmoid and ReLU activation functions respectively. In a specific implementation manner, the mobile flipped bottleneck convolution layer includes a sequentially arranged convolution layer Conv1×1, a depthwise separable convolution layer DepthwiseConv3×3, a squeeze - excitation module SENet, and a convolution layer Conv1×1. At the same time, the input feature of the mobile flipped bottleneck convolution layer is skip - connected to the output; the squeeze - excitation module SENet includes a sequentially arranged squeeze layer, a fully - connected layer, a ReLU activation function, a fully - connected layer, and a Sigmoid function.

[0024] The relative self - attention mechanism module is used to model the long - term dependence relationship globally on the data. The relative self - attention mechanism module includes a relative self - attention layer (Rel - Attention) and a feed - forward layer (Feedforward). Compared with the ordinary attention mechanism, the relative self - attention mechanism used in the present invention can better cooperate with the convolutional neural network to jointly learn local features and global features.

[0025] The input features of the above - mentioned mobile flipped bottleneck convolution layers, relative self - attention layers, and feed - forward layers are simultaneously skip - connected to the outputs of the corresponding layers.

[0026] The above - mentioned pooling layer uses a global pooling layer.

[0027] For the above-mentioned second convolutional neural network, the input includes the charge feature matrix at the current moment, and the output correspondingly is the charge state at the current moment. It also includes the charge feature matrix at the previous moment or / and the next moment, and the output correspondingly is the charge state at the previous moment or / and the next moment. The prediction of the charge state at the current moment is the main task, and the prediction of the charge state at the previous moment or / and the next moment is the auxiliary task.

[0028] By simultaneously predicting the state of health, remaining useful life, and charge state of the battery, the model can learn the internal relationships among the respective parameters and shared parameters in the three tasks, thereby improving the model performance.

[0029] Based on the prediction objective, let be the data set, and x n represent the input features for sample i, represent the target value for task task∈{SOH, RUL, SOC}. The objective of multi-task learning is to optimize a cross-task loss function, that is, where Θ, Θ task are the shared parameters (in the present invention, they refer to the parameters in the first variational autoencoder / second variational autoencoder) and the parameters for task task (in the present invention, they refer to the parameters in the first convolutional neural network / second convolutional neural network) respectively, and λ task and are the weight parameter of task task and the loss function respectively.

[0030] For the regression prediction tasks of the state of health, remaining useful life, and charge state of the battery, the root mean square error and the mean absolute error are selected as the evaluation indexes to measure whether the prediction results are accurate.

[0031] For the prediction result and the true value y, the root mean square error is defined as:

[0032]

[0033] The mean absolute error is defined as:

[0034]

[0035] where n represents the number of samples.

[0036] Combining the two, the loss function is defined as:

[0037]

[0038] where α and β are hyperparameters, satisfying α + β = 1.

[0039] The present invention also provides a lithium battery performance evaluation method based on multi-task deep learning, which is carried out according to the following steps using the above-mentioned lithium battery performance evaluation system:

[0040] S1. Through the feature generation module, obtain the optimal feature set related to the battery performance based on the information data of several charge and discharge cycles of the battery; the optimal feature set includes cycle features related to the state of health and remaining life, and time series features related to the state of charge; the cycle features and time series features are respectively input into the first multi-task prediction model and the second multi-task prediction model;

[0041] S2. Predict the state of health and remaining life of the battery through the first multi-task prediction model; it includes the following sub-steps:

[0042] S21. Through the first variational autoencoder, reconstruct the cycle features to obtain the corresponding reconstructed features;

[0043] S22. Through the first principal component analysis module, perform dimensionality reduction processing on the reconstructed features;

[0044] S23. Input the reconstructed features after dimensionality reduction into the first convolutional neural network to predict the state of health and remaining life of the battery;

[0045] S3. Predict the state of charge of the battery through the second multi-task prediction model; it includes the following sub-steps:

[0046] S31. Through the second variational autoencoder, reconstruct the time series features to obtain the corresponding reconstructed features;

[0047] S32. Through the second principal component analysis module, perform dimensionality reduction processing on the reconstructed features;

[0048] S33. Concatenate the reconstructed features after dimensionality reduction and the state of health predicted by the first convolutional neural network to construct a state-of-charge feature matrix, and input it into the second convolutional neural network to predict the state of charge of the battery.

[0049] Compared with the prior art, the lithium battery performance evaluation system and method based on multi-task deep learning provided by the present invention have the following beneficial effects:

[0050] 1. The present invention uses variational autoencoders and principal component analysis to perform feature dimension elevation and noise reduction on data; in order to adapt to the characteristics of lithium battery data features, the selected feature set is input into the variational autoencoder for dimension elevation, mapping to a high-dimensional latent space to capture deep features; subsequently, principal component analysis is applied to perform noise reduction processing on the dimension-elevated data, retaining the main features and removing noise interference at the same time to improve the representativeness of the features;

[0051] 2. The present invention enhances the mutual promotion among tasks by sharing the feature information of multiple related tasks, thereby improving the accuracy and robustness of prediction, and providing support for the safe use and performance improvement of lithium batteries.

[0052] 3. The present invention processes the lithium battery charge and discharge cycle data from the perspective of multi-task learning, fully considers the internal relationship among the state of health, remaining life, and state of charge, and introduces a relative self-attention mechanism, so that the mutual influence of different battery features is intuitively reflected, thereby improving the accuracy and interpretability of the model in lithium battery data prediction, and effectively solving the dynamic complexity problem in the process of lithium battery performance degradation. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 Schematic diagram of the lithium battery performance evaluation system framework based on multi-task deep learning provided in Embodiment 1 of the present invention;

[0054] Figure 2 Schematic diagram of the convolutional neural network structure;

[0055] Figure 3 Flow chart of the mobile flip bottleneck convolutional layer of the present invention;

[0056] Figure 4 Voltage data during the lithium battery charge and discharge cycle; where (a) corresponds to the charging process, and (b) corresponds to the relaxation stage process after charging;

[0057] Figure 5 Voltage and current diagram during the lithium battery charge and discharge cycle;

[0058] Figure 6 Variation results of the capacities of four batteries with the number of cycles;

[0059] Figure 7 Analysis relationship diagram of Pearson correlation coefficient;

[0060] Figure 8 Prediction result of the battery state of health;

[0061] Figure 9 Prediction result of the battery remaining life;

[0062] Figure 10 Prediction result of the battery state of charge. DETAILED DESCRIPTION OF THE INVENTION

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope protected by the present invention.

[0064] Embodiment 1

[0065] This embodiment provides a lithium battery performance evaluation system based on multi-task deep learning, as Figure 1 shown, which includes a feature generation module, a first convolutional neural network, and a second convolutional neural network.

[0066] The feature generation module is used to obtain an optimal feature set related to the battery performance based on the information data of several charge and discharge cycles of the battery. The optimal feature set includes cycle features related to the state of health and remaining life, and temporal features related to the state of charge.

[0067] By performing charge and discharge cycle tests on the battery, the information data during the charge and discharge cycle of the battery, including relevant parameters such as voltage, current, charge time, discharge time, and number of charge and discharge cycles, are obtained to deeply understand the degradation of the battery during its entire service life.

[0068] According to the above information data during the charge and discharge cycle of the battery, calculate the skewness, kurtosis, variance, maximum value, minimum value, and average value of the relaxation charging voltage of each cycle, as well as the initial value and average value of the voltage during the charging stage, etc. These can all be used as cycle features during the charge and discharge cycle. According to the historical data of battery aging research, calculate the Pearson correlation coefficients between the above cycle features and the state of health and remaining life, and select the cycle features with the absolute value of the Pearson correlation coefficient greater than 0.7 as the final cycle features.

[0069] The temporal features related to the state of charge include temporal data such as current, voltage, and time.

[0070] The first multi-task prediction model is used to predict the state of health and remaining life of the battery based on the cycle features; the first multi-task prediction model includes a first variational autoencoder, a first principal component analysis module, and a first convolutional neural network. The first variational autoencoder is used to reconstruct the cycle features to obtain corresponding reconstructed features; the first principal component analysis module is used to perform dimensionality reduction processing on the reconstructed features obtained by the first variational autoencoder; the first convolutional neural network is used to predict the state of health and remaining life of the battery based on the reconstructed features after dimensionality reduction processing by the first principal component analysis module.

[0071] The second multi-task prediction model is used to predict the state of charge of the battery based on the temporal features and the health state predicted by the first multi-task model; the second multi-task prediction model includes a second variational autoencoder, a second principal component analysis module, and a second convolutional neural network. The second variational autoencoder is used to reconstruct the temporal features to obtain corresponding reconstructed features; the second principal component analysis module is used to perform dimensionality reduction on the reconstructed features obtained by the second variational autoencoder; the second convolutional neural network is used to predict the state of charge of the battery based on the reconstructed features after dimensionality reduction by the second principal component analysis module and the state-of-charge feature matrix constructed by splicing the health state predicted by the first multi-task model.

[0072] The first variational autoencoder and the second variational autoencoder have the same structure, both of which are conventional variational autoencoders.

[0073] As Figure 1 shown, the variational autoencoder (VAE) is a generative model that can map the input (x) to the latent space representation (z) through an encoder and then reconstruct a new representation through a decoder so as to mine the deep information inside the data. Compared with other deep learning tasks, the number of features in the battery state prediction task is relatively small. Using a variational autoencoder to perform dimensionality increase on the data can deeply mine and make full use of the existing features, providing a basis for subsequent learning.

[0074] In this embodiment, since the health state and the remaining life are relative to the entire cycle, while the state of charge is targeted at temporal features. Based on this, this embodiment provides two convolutional neural networks based on the attention mechanism with the same structure, namely the first convolutional neural network and the second convolutional neural network. Among them, the first convolutional neural network takes the cyclic features as the input to predict the prediction results of the health state and the remaining life of the battery; the second convolutional neural network takes the state-of-charge feature matrix constructed by splicing the temporal features and the health state predicted by the first convolutional neural network as the input to predict the state of charge of the battery. Especially for the second convolutional neural network, the input includes the state-of-charge feature matrices at the previous moment (t - 1), the current moment (t), and the next moment (t + 1), and the output is the state of charge at the previous moment, the current moment, and the next moment. The prediction of the state of charge at the current moment is the main task, and the prediction of the state of charge at the previous moment and the next moment is the auxiliary task.

[0075] The above-mentioned first convolutional neural network and second convolutional neural network both include a convolutional layer, two mobile flipped bottleneck convolutional layers, two relative self-attention mechanism modules, and two pooling layers arranged in sequence; the relative self-attention mechanism module includes a relative self-attention layer (Rel-Attention) and a feedforward layer (Feedforward). The input features of each mobile flipped bottleneck convolutional layer, relative self-attention layer, and feedforward layer are simultaneously skip-connected to the output of the corresponding layer.

[0076] The above convolutional layer (ConvLayer) is used for preliminary extraction of data features. Let be the input, and the preliminary single-layer convolution is defined as where represents a convolution kernel with quantity f and size L, and b q represents the bias, σ(·) is the activation function, is the output feature, and m′ = m - L + 1.

[0077] The above-mentioned mobile inverted bottleneck convolutional layer (MBConv) is mainly used for local structural feature extraction. Compared with the ordinary convolutional layer, an important change in the mobile inverted bottleneck convolution is the introduction of a squeeze-excitation module, enabling the model to automatically learn the importance of different feature channels, thereby improving the utilization rate of features according to the importance. In the specific implementation, as Figure 2 shown, the mobile inverted bottleneck convolutional layer includes a sequentially arranged convolutional layer Conv1×1, a depthwise separable convolutional layer Depthwise Conv3×3, a squeeze-excitation module SENet, and a convolutional layer Conv1×1. At the same time, the input feature of the mobile inverted bottleneck convolutional layer is skip-connected to the output; the squeeze-excitation module SENet includes a sequentially arranged squeeze layer, a fully connected layer, a ReLU activation function, a fully connected layer, and a Sigmoid function.

[0078] In this embodiment, the squeeze operator of the squeeze-excitation module in the mobile inverted bottleneck convolutional layer is defined as:

[0079]

[0080] where X t represents the feature input to the squeeze layer, and T represents the time step of the feature;

[0081] The activation operator is defined as:

[0082] e = σ(W SE2 ·δ(W SE1 ·s));

[0083] where W SE is the weight matrix, and σ and δ are the Sigmoid and ReLU activation functions respectively.

[0084] The above relative self-attention mechanism module is used to model the long-term dependence relationship globally in the data. Compared with the ordinary attention mechanism, the relative self-attention mechanism is used here to better cooperate with the convolutional neural network to complete the joint learning of local and global features.

[0085] The relative self-attention mechanism based on the relative self-attention layer is defined as follows:

[0086] Assume the input sequence is \(x=(x_1,\ldots,x_{ n )\), where The attention score between the feature at position \(i\) and the feature at position \(j\) is as follows:

[0087]

[0088] where, \(x_{ i \) and \(x_{ j \) are the features at positions \(i\) and \(j\), \(e_{ i,j \) is the attention score, indicating the degree of attention of \(x_{ i \) to \(x_{ j \), \(W_{ Q}\), \(W_{ K}\), \(W_{ V}\) are the parameter matrices of the attention mechanism, and \(d_{ z}\) represents the vector dimension;

[0089] Then, through the softmax function, the attention scores are converted into values between 0 and 1 and sum to 1, and then the corresponding attention weights \(\alpha_{ ij}\) are obtained:

[0090]

[0091] Finally, multiply the value vector of each element by its corresponding attention weight and then sum to obtain the final output of the relative self-attention layer:

[0092]

[0093] The boundary between the input elements \(x_{ i}\) and \(x_{ j}\) is represented by the vectors \(_{ }\) and \(_{ }\). For a linear sequence, the boundary can capture information about the relative position differences between input elements. The maximum relative position considered is clipped to the maximum absolute value \(k\) (\(2\leq k\leq n - 4\)). Assume that the exact relative position information is useless after a certain distance. Clipping the maximum distance also enables the model to generalize to sequence lengths not seen during training. Therefore, consider \(2k + 1\) unique boundary labels:

[0094]

[0095] clip(r,k)=\(\max(-k,\min(k,r))\);

[0096] where, \(\omega_{ K}\), \(\omega_{ V}\) are learnable parameters and are representations of relative positions,

[0097] The above pooling layer uses a global pooling layer.

[0098] The following details the construction and effectiveness of the lithium battery performance evaluation system based on multi-task deep learning provided in this embodiment in combination with the charge and discharge cycle aging process information data of four different lithium batteries (CY25-0.5_1#5, CY25-0.5_1#8, CY25-0.5_1#15, CY25-0.5_1#16) collected.

[0099] The construction process of the above lithium battery performance evaluation system based on multi-task deep learning includes the following steps:

[0100] Step 1: Obtain the optimal feature set related to battery performance based on the information data of several charge and discharge cycles of the battery.

[0101] In this step, the information data of the aging process of four lithium batteries is collected.

[0102] All the lithium batteries use lithium cobalt oxide as the cathode material, with an initial capacity of 3.2 Ah and an initial voltage of 4.2 V; the information data of the aging process includes: voltage, current, charging time, discharging time, number of charge and discharge cycles, etc., as Figure 3 , Figure 4 described. Figure 5 The change results of the capacities of the four batteries with the number of cycles are given.

[0103] Check for obvious outliers in the data, mainly the outlier values that are too different from other data, and delete them. Then calculate the skewness (V r_Ske ), kurtosis (V r_kur ), variance (V r_Var ), maximum value (V r_Max ), minimum value (V r_Min ) and average value (V r_ave ) of the relaxation charging voltage for each cycle, and the initial value (V ini ) and average value (V ave ) of the voltage in the charging stage as cycle features. Calculate the Pearson correlation coefficients of the above cycle features with the state of health (SOH) and remaining useful life (RUL). The results are as Figure 6 shown. It can be seen from the figure that the skewness (V r_Ske ), kurtosis (V r_kur ), variance (V r_Var ), maximum value (V r_Max ), minimum value (V r_Min ), average value (V r_ave ), and the initial value (V ini)The absolute value of the Pearson correlation coefficient with the state of health (SOH) and remaining useful life (RUL) is greater than 0.7, which is a cyclic feature for predicting the state of health (SOH) and remaining useful life (RUL).

[0104] The time-series features related to the state of charge include current, voltage, and time, and a charge feature matrix is constructed by combining them with the state of health (SOH).

[0105] Therefore, a cyclic feature set is constructed using the skewness, kurtosis, variance, maximum value, minimum value, average value of the relaxation charging voltage of each cycle of the above four batteries, and the initial value of the voltage in the charging stage. The current, voltage, and time of the battery are used to construct a time-series feature set. The cyclic feature set and the time-series feature set constitute the optimal feature set.

[0106] The optimal feature sets of three batteries are combined with the corresponding state of health (SOH), remaining useful life (RUL), and state of charge (SOC) labels as the training set, and the optimal feature set of another battery is combined with the corresponding state of health (SOH), remaining useful life (RUL), and state of charge (SOC) labels as the test set.

[0107] Step 2, train the first multi-task prediction model, including the following sub-steps:

[0108] Step 21, reconstruct the cyclic features through the first variational autoencoder to obtain the corresponding reconstructed features.

[0109] In this step, the cyclic features of each sample in the training set are input into the first variational encoder to obtain the reconstructed features of each sample.

[0110] Step 22, perform dimensionality reduction processing on the reconstructed features through the first principal component analysis module.

[0111] In this step, the dimensionality reduction processing of the reconstructed features of each sample is performed through the first principal component analysis module.

[0112] The steps of performing dimensionality reduction processing on the reconstructed features of each sample using the first principal component analysis module are as follows:

[0113] (1) Normalize the sample data of the same feature parameter to obtain a normalized dataset X.

[0114] (2) Based on the normalized sample data of each sample, construct a covariance matrix N represents the number of samples in the dataset.

[0115] (3) Perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalues of the covariance matrix and the eigenvectors corresponding to the eigenvalues; and arrange them in descending order of the eigenvalues;

[0116] (4) Based on the obtained eigenvectors, obtain their corresponding orthonormal vectors;

[0117] (5) Multiply the obtained orthonormal vectors by the dataset obtained in step (1) to obtain the corresponding principal components, and calculate the contribution rate of the corresponding principal components; The contribution rate is calculated by the proportion of each eigenvalue in all eigenvalues;

[0118] (6) Use the top 24 principal components ranked by contribution rate as the reconstructed features after dimensionality reduction.

[0119] Step 23: Input the reconstructed features after dimensionality reduction into the first convolutional neural network to obtain the health state and remaining life of the battery.

[0120] In this step, input the reconstructed features after dimensionality reduction of each sample into the first convolutional neural network to predict the health state and remaining life of the battery. And based on the predicted health state, remaining life and their corresponding labels, calculate according to the root mean square error and mean absolute error given above to obtain the loss function of the corresponding task, and then obtain the cross-task loss function. Then use the cross-task loss function to optimize the parameters of the first variational autoencoder and the first convolutional neural network through the gradient descent optimization algorithm.

[0121] Repeat the above steps S21 - S23 until the set iteration number threshold (such as 500 epochs) is reached, and the first multi-task prediction model converges.

[0122] Step 3: Train the second multi-task prediction model, including the following sub-steps:

[0123] Step 31: Reconstruct the time series features through the second variational autoencoder to obtain the corresponding reconstructed features.

[0124] In this step, input the time series features of each sample in the training set into the second variational encoder to obtain the reconstructed features of each sample.

[0125] Step 32: Perform dimensionality reduction processing on the reconstructed features through the second principal component analysis module.

[0126] In this step, perform dimensionality reduction processing on the reconstructed features of each sample through the second principal component analysis module. The specific operation is the same as the processing process of the first principal component analysis module and will not be elaborated here.

[0127] Step 33: Concatenate the reconstructed features after dimensionality reduction and the health state predicted by the first convolutional neural network to construct a charge feature matrix, and input it into the second convolutional neural network to obtain the state of charge of the battery.

[0128] In this step, a charge feature matrix is constructed from the reconstructed features after dimensionality reduction of the samples at the previous moment, the current moment, and the next moment, and the state of health (SOH) predicted by the first convolutional neural network for the corresponding charge cycle, and is input into the second convolutional neural network to obtain the predicted state of charge at the corresponding moment. Then, based on the predicted state of charge at each moment and the corresponding labels, the root mean square error and the mean absolute error given above are calculated to obtain the loss function of the corresponding task, and further the cross-task loss function is obtained. Then, the parameters of the second variational autoencoder and the second convolutional neural network are optimized using the cross-task loss function through the gradient descent optimization algorithm.

[0129] Repeat the above steps S31 - S33 until the set iteration number threshold (e.g., 500 epochs) is reached and the second multi-task prediction model converges.

[0130] After the system training is completed, the collected battery data is input into the trained multi-task learning model (lithium battery performance evaluation system) to synchronously predict the state of health, remaining life, and state of charge of the battery, and the corresponding prediction results are output. The specific steps are as follows:

[0131] S1, through the feature generation module, obtain the optimal feature set related to the battery performance based on the information data of several charge and discharge cycles of the battery.

[0132] In this step, the battery to be tested undergoes several charge and discharge cycles to obtain the information data of several charge and discharge cycles. Then, according to the steps given above, the cycle features related to the state of health and remaining life, and the time series features related to the state of charge are obtained. The cycle features here are the final cycle features determined by the Pearson correlation coefficient mentioned above. And the determined cycle features and time series features are used to construct the optimal feature set, which is used as the optimal feature set of the test battery in this embodiment.

[0133] S2, predict the state of health and remaining life of the battery through the first multi-task prediction model; including the following sub-steps:

[0134] S21, through the first variational autoencoder, reconstruct the cycle features to obtain the corresponding reconstructed features.

[0135] In this step, the cycle features of each sample in the test set are input into the first variational autoencoder to obtain the reconstructed features of each sample.

[0136] S22, through the first principal component analysis module, perform dimensionality reduction on the reconstructed features.

[0137] In this step, the first principal component analysis module performs dimensionality reduction on the reconstructed features of each sample. Refer to the specific operations given above.

[0138] S23, inputting the reconstructed features after dimensionality reduction into the first convolutional neural network to predict the health status and remaining life of the battery.

[0139] In this step, the reconstructed features after dimensionality reduction are input into the first convolutional neural network, and the health status and remaining life of the battery can be obtained at the same time.

[0140] S3, predicting the state of charge of the battery by using a second multi-task prediction model; comprising the following sub-steps:

[0141] S31, reconstructing the time series features through the second variational autoencoder to obtain corresponding reconstructed features.

[0142] In this step, the time series features of each sample in the test set are input into the second variational autoencoder to obtain the reconstructed features of each sample.

[0143] S32, performing dimensionality reduction processing on the reconstructed features through a second principal component analysis module.

[0144] In this step, the second principal component analysis module is used to perform dimensionality reduction processing on the reconstructed features of each sample. Please refer to the specific operations given above.

[0145] S33, concatenating the reconstructed features after dimensionality reduction and the health status predicted by the first convolutional neural network to construct a charge feature matrix, and inputting it into the second convolutional neural network to predict the charge status of the battery.

[0146] In this step, the reconstructed features of the samples at the previous moment, the current moment, and the next moment after dimensionality reduction and the state of health (SOH) predicted by the first convolutional neural network at the corresponding moment of the charging cycle are concatenated to construct the charge feature matrix at the corresponding moment, and input into the second convolutional neural network to obtain the predicted state of charge at the corresponding moment.

[0147] Generally, the end of life of a lithium battery is defined as a capacity lower than 80% of the initial capacity, and the expected life is 250 cycles of charge and discharge. The above-mentioned lithium battery performance evaluation method based on multi-task deep learning predicts that a positive electrode material is LiNi 0.83 Co 0.11 Mn 0.06 The health status, remaining life and state of charge of O2 (NCM) commercial batteries, the charging process includes a combination of constant current constant voltage (CC and CV) modes at a charging rate of 0.5C, followed by a 30-minute relaxation period and a 1C discharge, with an ambient temperature of 25°C. Then the prediction results obtained by the steps S1-S3 given above are as follows Figure 8 , Figure 9 and Figure 10As shown. Among them, the MAE of the battery health state prediction reaches 0.00497, and the RMSE reaches 0.00732; the MAE of the battery remaining life prediction reaches 1.43, and the RMSE reaches 1.87; the MAE of the battery state of charge prediction reaches 0.00341, and the RMSE reaches 0.00438. Through experimental verification, the multi-task learning model provided by the present invention has strong fitting ability and stronger generalization ability, has strong adaptability to multiple key parameters of lithium batteries, and can accurately complete the prediction of the health state, remaining life and state of charge of lithium batteries, making innovations in key technologies such as battery state management and battery intelligent prediction.

[0148] Those of ordinary skill in the art will realize that the embodiments here are to help readers understand the principles of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of the present invention according to these technical revelations disclosed in the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. A lithium battery performance evaluation system based on multi-task deep learning, characterized in that: include: A feature generation module, used to obtain an optimal feature set related to battery performance based on information data of several charge and discharge cycle processes of the battery; the optimal feature set includes cycle features related to health status and remaining life, and time series features related to charge status; A first multi-task prediction model is used to predict the health state and remaining life of the battery according to the cycle characteristics; the first multi-task prediction model includes a first variational autoencoder, a first principal component analysis module and a first convolutional neural network; the first variational autoencoder is used to reconstruct the cycle characteristics to obtain corresponding reconstructed features; the first principal component analysis module is used to perform dimensionality reduction processing on the reconstructed features obtained by the first variational autoencoder; the first convolutional neural network is used to predict the health state and remaining life of the battery according to the reconstructed features after the dimensionality reduction processing by the first principal component analysis module; The second multi-task prediction model is used to predict the state of charge of the battery based on the time series features and the health state predicted by the first multi-task model; the second multi-task prediction model includes a second variational autoencoder, a second principal component analysis module and a second convolutional neural network; the second variational autoencoder is used to reconstruct the time series features to obtain corresponding reconstructed features; the second principal component analysis module is used to perform dimensionality reduction processing on the reconstructed features obtained by the second variational autoencoder; the second convolutional neural network is used to predict the state of charge of the battery based on the charge feature matrix constructed by splicing the reconstructed features after the dimensionality reduction processing by the second principal component analysis module and the health state predicted by the first multi-task model.

2. The lithium battery performance evaluation system based on multi-task deep learning according to claim 1 is characterized in that: The information data during the battery charge and discharge cycle includes voltage, current, charging time, discharging time, and number of charge and discharge cycles; the cycle characteristics related to the health status and remaining life include the skewness, kurtosis, variance, maximum value, minimum value, average value of the relaxation charging voltage of each cycle, and the initial value and average value of the voltage in the charging stage; the timing characteristics related to the state of charge include current, voltage, and time.

3. The lithium battery performance evaluation system based on multi-task deep learning according to claim 2 is characterized in that: By calculating the Pearson correlation coefficient between each cycle characteristic and the health status and remaining life span, the cycle characteristics with an absolute value of the Pearson correlation coefficient greater than 0.7 were selected.

4. The lithium battery performance evaluation system based on multi-task deep learning according to claim 1, characterized in that: The first convolutional neural network and the second convolutional neural network have the same structure, both of which include sequentially arranged convolutional layers, one or more moving flip bottleneck convolutional layers, one or more relative self-attention mechanism modules and one or more pooling layers.

5. The lithium battery performance evaluation system based on multi-task deep learning according to claim 4 is characterized in that: The mobile flip bottleneck convolution layer includes a convolution layer Conv1×1, a depthwise separable convolution layer DepthwiseConv3×3, a compression-activation module SENet and a convolution layer Conv1×1 arranged in sequence. At the same time, the input features of the mobile flip bottleneck convolution layer are jump-connected to the output; the compression-activation module SENet includes a compression layer, a fully connected layer, a ReLU activation function, a fully connected layer and a Sigmoid function arranged in sequence.

6. The lithium battery performance evaluation system based on multi-task deep learning according to claim 5, characterized in that: The compression operator of the compression-activation module in the mobile flip bottleneck convolutional layer is defined as: Among them, X t represents the feature of the input compression layer, and T represents the time step of the feature; The activation operator is defined as: e=σ(W SE2 ·δ(W SE1 ·s)); Among them, W SE is the weight matrix, σ and δ are Sigmoid and ReLU activation functions respectively.

7. The lithium battery performance evaluation system based on multi-task deep learning according to claim 4, characterized in that: The relative self-attention mechanism module includes a relative self-attention layer and a feedforward layer.

8. The lithium battery performance evaluation system based on multi-task deep learning according to claim 7, characterized in that: The input features of each moving flip bottleneck convolution layer, relative self-attention layer, and feed-forward layer are simultaneously jump-connected to the output of the corresponding layer.

9. The lithium battery performance evaluation system based on multi-task deep learning according to claim 8, characterized in that: For the second convolutional neural network, the input includes the charge feature matrix at the current moment, and the output corresponds to the charge state at the current moment. It also includes the charge feature matrix at the previous moment and / or the next moment, and the output corresponds to the charge state at the previous moment and / or the next moment.

10. A lithium battery performance evaluation method based on multi-task deep learning, characterized in that: The lithium battery performance evaluation system according to any one of claims 1 to 9 is used in accordance with the following steps: S1, obtaining an optimal feature set related to battery performance according to information data of several charge and discharge cycle processes of the battery through a feature generation module; the optimal feature set includes cycle features related to health status and remaining life, and time series features related to charge status; the cycle features and time series features are respectively input into a first multi-task prediction model and a second multi-task prediction model; S2, predicting the health status and remaining life of the battery through the first multi-task prediction model; The following steps are included: S21, reconstructing the cyclic features through the first variational autoencoder to obtain corresponding reconstructed features; S22, performing dimensionality reduction processing on the reconstructed features through a first principal component analysis module; S23, inputting the reconstructed features after dimensionality reduction into a first convolutional neural network to predict the health status and remaining life of the battery; S3, predicting the state of charge of the battery by using a second multi-task prediction model; The following steps are included: S31, reconstructing the time series features through a second variational autoencoder to obtain corresponding reconstructed features; S32, performing dimensionality reduction processing on the reconstructed features through a second principal component analysis module; S33, concatenating the reconstructed features after dimensionality reduction and the health status predicted by the first convolutional neural network to construct a charge feature matrix, and inputting it into the second convolutional neural network to predict the charge status of the battery.

Images