Method and System for Evaluating State of Health of Lithium-Ion Batteries Based on Time Decomposition and Feature Fusion

Abstract

The present invention discloses a method and system for evaluating the state of health of a lithium-ion battery based on time decomposition and feature fusion, including acquiring lithium-ion battery aging data, performing preprocessing, and extracting key features to form a battery aging data set; serializing the data, and dividing the sequence into a trend term and a seasonal term based on time decomposition; using the trend term and the seasonal term as two components to input into a battery state of health evaluation network; swapping the time and channel dimensions of the sequence, and fusing the time dimension information and the channel feature information in a skip connection manner; performing inverse normalization on the fused data and connecting it to a projection layer composed of a single linear layer to perform mapping from the input sequence to the output sequence; using the trained network to predict the state of health of the lithium-ion battery. This application can solve the problem that it is difficult to accurately predict the details of the local capacity recovery during the battery aging process, realize the accurate evaluation of the state of health of the lithium-ion battery, and ensure the safe operation of the lithium-ion battery.

Description

Technical Field

[0001] The present invention relates to the technical field of battery management, and particularly to a method and system for evaluating the state of health of lithium-ion batteries based on time decomposition and feature fusion. Background Art

[0002] As the core component of the current energy storage system, lithium-ion batteries play a crucial role in many fields such as electric vehicles, portable electronic devices, photovoltaic energy storage systems, and backup power supplies. With the rapid development of new energy technologies and the expansion of the application scale, the stability and safety of battery performance have become the key factors affecting the reliability of the entire system. Therefore, effective evaluation and management of lithium-ion batteries have become an essential part of the battery power supply system. The state of health (SOH) of lithium-ion batteries is an important parameter for measuring battery performance and aging degree, which is directly related to the service life, safety performance, and energy storage efficiency of the battery. The health state of the battery is affected by various factors, including the number of charge and discharge cycles, operating temperature, magnitude of charge and discharge current, depth of charge and discharge, etc. As the usage time increases, the chemical composition, physical structure, and electrochemical performance inside the battery will gradually change, resulting in a decrease in battery capacity, an increase in internal resistance, and even potential safety problems such as thermal runaway.

[0003] Currently, the mainstream research methods mainly include experimental measurement methods, model-based methods, and data-driven methods. The experimental measurement methods mainly include the Coulomb counting method and the open-circuit voltage method. The Coulomb counting method has a simple principle and is easy to use, but due to the inevitable current monitoring error, there is an error accumulation problem in this method. The open-circuit voltage method has the advantages of small calculation amount and no error accumulation, but it is difficult to be actually applied because obtaining the open-circuit voltage of the battery requires long-term static placement. Model-based methods mainly include equivalent circuit models, electrochemical models, and incremental capacity analysis methods. These methods avoid the problems of long time and error accumulation in experimental measurement methods and obtain better SOH estimation results. However, since they calculate SOH through battery models, there are still problems such as relying heavily on the accuracy of the model, difficulties in model construction and solution processes, and relatively poor generalization when dealing with various battery types and usage scenarios.

[0004] In recent years, data-driven methods have been widely applied to battery SOH prediction due to their advantages of not relying on models and being unrestricted by any physical and chemical models, with high adaptability and flexibility. Traditional data-driven methods mainly include probability-based methods such as particle filters and Kalman filters, and machine learning-based methods such as support vector machines and decision trees. Weng et al. combined data-driven methods with model-based methods and proposed a support vector regression algorithm based on incremental capacity analysis for SOH estimation of lithium-ion batteries (Weng C, Cui Y, Sun J, et al. On-board state of health monitoring of lithium-ion batteries using incremental capacity analysis with support vector regression[J]. Journal of Power Sources, 2013, 235: 36-44.). Yang et al. proposed a Gaussian process regression model based on grey relational analysis, using parameters in the charging curve rather than the number of cycles as the input of the model, and obtained better SOH prediction results (Yang D, Zhang X, Pan R, et al. A novel gaussian process regression model for state-of-health estimation of lithium-ion battery using charging curve[J]. Journal of Power Sources, 2018, 384: 387-395). However, traditional data-driven methods are difficult to handle the task of evaluating the health state of lithium-ion batteries on a long time scale due to the relatively simple models used for estimation.

[0005] Recently, with the powerful learning ability of deep networks, data-driven methods based on deep learning have gradually become popular and significantly improved the accuracy of battery SOH estimation. Kaur et al. used three different deep network models, namely FNN, CNN, and LSTM, for battery SOH estimation and explored the weights of different features for battery SOH estimation (Kaur K, Garg A, Cui X, et al. Deep learning networks for capacity estimation for monitoring SOH of Li-ion batteries for electric vehicles[J]. International Journal of Energy Research, 2021, 45(2): 3113-3128.). Xu et al. proposed an improved CNN-LSTM network to improve the prediction accuracy of SOH by adding residual connections to the model and removing useless features to reduce the computational complexity (Xu H, Wu L, Xiong S, et al. An improved CNN-LSTM model-based state-of-health estimation approach for lithium-ion batteries[J]. Energy, 2023, 276: 127585.). Li et al. proposed an SOH and RUL prediction framework called AST-LSTM, which simultaneously obtains new and old information by fixing the connection coupling of the input and forget gates, and then multiplies the new input with the historical cell state unit by unit to filter out more useful information (Li P, Zhang Z, Xiong Q, et al. State-of-health estimation and remaining useful life prediction for the lithium-ion battery based on a variant long short-term memory neural network[J]. Journal of Power Sources, 2020, 459: 228069.). However, the above methods mainly estimate the battery SOH through deep networks such as LSTM and its variants. By virtue of the learning ability of the LSTM network for the time dependence in features, they have obtained higher accuracy than other traditional SOH estimation methods. However, the LSTM network itself still has some room for improvement in processing long-cycle data and is easily affected by noise and sudden fluctuations, resulting in error accumulation.

[0006] Compared with the LSTM series models, the Transformer model can achieve the learning of longer-period data features through the multi-head self-attention mechanism of parallel computing. Some scholars have begun to apply the Transformer model to battery SOH estimation. Gu et al. combined CNN with Transformer for battery SOH estimation, and used principal component analysis to eliminate redundant features to reduce the model calculation burden. (Gu X, See KW, Li P, et al. A novel state-of-health estimation for the lithium-ion battery using a convolutional neural network and transformer model [J]. Energy, 2023, 262: 125501.). Bai et al. proposed a SOH estimation method based on a multi-view information perception framework (MVIP-Trans) based on convolutional transformation, which enhances valuable information and suppresses useless noise through a parallel multi-scale attention local information sensor. (Bai T, Wang H. Convolutional Transformer-Based Multiview Information Perception Framework for Lithium-Ion Battery State-of-Health Estimation[J]. IEEE Transactions on Instrumentation and Measurement, 2023, 72: 1-12.). Although the Transformer model can process long-term series of battery data, parallel computing and self-attention mechanisms also cause it to require more computing resources, and the self-attention mechanism inevitably leads to local information loss due to its arrangement-invariant nature. Therefore, how to simultaneously consider the long-term dependence of the battery aging process and accurately identify the local capacity recovery details during the aging process is crucial to accurately assess the health status of lithium-ion batteries. Summary of the invention

[0007] Therefore, the purpose of the present invention is to provide a lithium-ion battery health status assessment method and system based on time decomposition and feature fusion, which can process battery aging signals on a long time scale while solving the problem that the details of local capacity recovery during battery aging are difficult to accurately predict, thereby achieving accurate assessment of the health status of lithium-ion batteries and ensuring the safe operation of lithium-ion batteries.

[0008] To achieve the above object, a method for evaluating the state of health of a lithium-ion battery based on time decomposition and feature fusion provided by the present invention includes the following steps:

[0009] S1. Obtain voltage, current, temperature, and capacity data during the aging process of the lithium-ion battery; perform preprocessing and extract key features to form a battery aging dataset;

[0010] S2. Serialize the data in the battery aging dataset, and divide the sequence into a trend term and a seasonal term based on time decomposition; use the trend term and the seasonal term as two components to input into the battery state of health evaluation network;

[0011] S3. Exchange the time and channel dimensions of the sequence, and fuse the time dimension information and channel feature information in a skip connection manner;

[0012] Inverse normalize the fused data and connect it to a projection layer composed of a single linear layer to perform the mapping from the input sequence to the output sequence;

[0013] S4. Obtain the real-time aging data of the lithium-ion battery, iteratively train the lithium-ion battery state of health evaluation network, and use the trained network to predict the state of health of the lithium-ion battery.

[0014] Further preferably, in S1, the obtaining of voltage, current, temperature, and capacity data during the aging process of the lithium-ion battery and performing preprocessing includes:

[0015] Select the time when the battery voltage drops to the discharge threshold voltage, voltage slope, time required for the temperature to reach the maximum value, temperature slope, maximum temperature, constant voltage charging time, constant current charging time, and internal resistance as input features;

[0016] Calculate the state of health of the current battery according to the battery capacity data, and use the state of health value of the battery as the label of the dataset;

[0017] Perform outlier processing on the input features and the dataset label, and jointly form a battery aging dataset.

[0018] Further preferably, the preprocessing further includes removing outliers from the obtained data based on the 3-sigma principle before extracting the input features, and normalizing the data to have zero mean and variance.

[0019] Further preferably, in S2, serializing the data in the battery aging dataset and dividing the sequence into a trend term and a seasonal term based on time decomposition includes:

[0020] Divide the battery aging dataset into multiple data sequences where X represents an input sequence containing sequences of length and number of channels (features) set;

[0021] For the input sequence at any moment Average pooling is used to eliminate periodic fluctuations and highlight the long-term trend, and at the same time, padding operations are used to keep the length of the input sequence unchanged. The specific formula is

[0022]

[0023] where are the decomposed trend term and seasonal term respectively; represents the padding operation on and represents average pooling.

[0024] Further preferably, inputting the trend term and the seasonal term as two components into the battery health state evaluation network includes respectively inputting the two decomposed components into the MLP-Block of the lithium-ion battery health state evaluation network for time feature extraction. The MLP-Block includes two fully connected layers and is represented by the following formula:

[0025]

[0026] where are the corresponding weights and biases of the trend term MLP-Block, and are the corresponding weights and biases of the seasonal term MLP-Block, is the output of the time decomposition module.

[0027] Further preferably, the activation function GELU is represented by the following formula (4):

[0028]

[0029] where represents the input of the network at the current layer.

[0030] Further preferably, in S2, it also includes serializing the data in the battery aging dataset and then using instance normalization to remove non-stationary information to obtain a stationary sequence and non-stationary attribute parameters and affine transformation parameters. The specific steps are as follows:

[0031] Divide the battery aging dataset into multiple data sequences, calculate the average value and standard deviation of each data sequence, and use the average value and standard deviation to normalize the instances in each sequence for normalization;

[0032]

[0033] where is the affine parameter vector, is the normalized instance, To perform the operation of calculating the average value of data, to perform the operation of calculating the standard deviation of data, and to be a fixed minimum value to prevent the denominator from being 0.

[0034] Further preferably, in S3, the fused data is inversely normalized and then connected to a projection layer composed of a single linear layer; where the inverse normalization of the fused data is to perform inverse instance normalization after combining the fused result with the normalization related parameters of S2, including:

[0035] In the battery health state evaluation network, the symmetric network output layer performs an inverse normalization operation on the output; explicitly returns the non-stationary attributes deleted from the input data and the affine transformation parameters learned by the network to the model output, that is

[0036]

[0037] Use to replace as the output of the network.

[0038] The present invention also provides a lithium-ion battery health state evaluation system based on time decomposition and feature fusion, which is used to implement the steps of the above-mentioned lithium-ion battery health state evaluation method based on time decomposition and feature fusion, including a data acquisition module, a data processing module, and a battery health state evaluation network;

[0039] The data acquisition module is used to acquire voltage, current, temperature, and capacity data during the aging process of the lithium-ion battery; perform preprocessing, and extract key features to form a battery aging data set;

[0040] The data processing module is used to serialize the data in the battery aging data set, and divide the sequence into a trend term and a seasonal term based on time decomposition; use the trend term and the seasonal term as two components to input into the battery health state evaluation network;

[0041] The battery health state evaluation network includes a time decomposition module and a channel feature fuser; the time decomposition module is used to exchange the time and channel dimensions of the sequence, and the channel feature fuser uses a skip connection method to fuse the time dimension information and the channel feature information; inversely normalize the fused data and then connect it to a projection layer composed of a single linear layer to perform the mapping from the input sequence to the output sequence;

[0042] Acquire the real-time aging data of the lithium-ion battery, perform iterative training on the lithium-ion battery health state evaluation network, and use the trained network to predict the health state of the lithium-ion battery.

[0043] The lithium-ion battery state of health assessment method and system based on time decomposition and feature fusion disclosed in this application have the following beneficial effects compared with the prior art:

[0044] 1. The present invention performs data processing and feature extraction on the battery aging process data, extracts features that can be used for battery state of health estimation from different perspectives such as voltage, current, temperature, and time, and reduces the interference of outlier data and noise;

[0045] 2. Through the time decomposition module, the global and local information is perceived from the time perspective, which can solve the long-term dependence problem in the battery aging process and accurately identify the local capacity recovery details during the aging process, improving the prediction accuracy; Subsequently, from the perspective of channel features, the combined influence of different features on the aging data is further fused. Compared with using a single feature, the robustness and generalization of the battery state of health assessment are improved; Finally, by adding reversible instance normalization at both the input and output ends, the non-stationary problem of the data statistical attributes changing over time during the battery aging process is alleviated, reducing the distribution difference between data, and further improving the accuracy of health assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a block diagram of the steps of the lithium-ion battery state of health assessment method based on time decomposition and feature fusion of the present invention.

[0047] Figure 2 Battery voltage, current, temperature and capacity curves during the battery aging process.

[0048] Figure 3 Comparison chart of the state of health assessment results of different numbered batteries in the NASA dataset by multiple methods.

[0049] Figure 4 Comparison chart of the state of health assessment results of different numbered batteries in the CALCE dataset by multiple methods.

[0050] Figure 5 Comparison chart of the error distribution of the state of health assessment of different numbered batteries in the NASA dataset by multiple methods.

[0051] Figure 6 Comparison chart of the error distribution of the state of health assessment of different numbered batteries in the CALCE dataset by multiple methods. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] The present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0053] As Figure 1 shown, the lithium-ion battery state of health assessment method based on time decomposition and feature fusion provided by an embodiment of the present invention on the one hand includes the following steps:

[0054] S1. Obtain the voltage, current, temperature, and capacity data during the aging process of the lithium-ion battery; perform preprocessing, extract key features, and form a battery aging dataset;

[0055] Perform preprocessing on the data. First, remove outliers based on the 3-sigma principle, then standardize the data to have zero mean and variance, and then extract aging features based on this.

[0056] Extract the voltage, current, temperature, and capacity data of the lithium-ion battery during charge and discharge throughout the entire battery life cycle. The entire charge-discharge cycle includes multiple charge-discharge cycles, and each cycle contains sequences of battery voltage, current, and temperature changing over time. The capacity is the battery capacity value measured after the discharge ends. Select the time when the battery voltage drops to the discharge threshold voltage, voltage slope, time required for the temperature to reach the maximum value, temperature slope, maximum temperature, constant voltage charging time, constant current charging time, and internal resistance as input features. Calculate the state of health of the current battery based on the battery capacity data, use the state of health value of the battery as the label of the dataset, combine the obtained input features, remove outliers based on the 3-sigma principle, and standardize the data to have zero mean and variance to form a battery aging dataset available for network training. The state of health of the battery is defined by the capacity of the current cycle of the battery, and the specific formula is:

[0057]

[0058] where and are the capacity and nominal capacity of the battery in the current state, respectively.

[0059] In the preferred embodiment of the present invention, the application objects come from the lithium-ion battery accelerated aging experiments conducted by the NASA Ames Center of Excellence in Prediction and the Center for Advanced Life Cycle Engineering (CALCE) at the University of Maryland. A total of two groups of experiments are set up, and the experimental datasets are the NASA dataset and the CALCE dataset. Each group of experiments has four batteries. The battery numbers in the NASA dataset are B05, B06, B07, and B18, and the battery numbers in the CALCE dataset are CS2_35, CS2_36, CS2_37, and CS2_38. The curves of battery voltage, current, temperature, and capacity during the battery aging process are as Figure 2As shown, the colors of all curves gradually deepen as the number of battery cycles increases, that is, the worse the state of health (SOH) of the battery, the darker the color. In the capacity-cycle number scatter plot, the capacity data at each cycle number corresponds to the voltage, current, and temperature change curves of the same color. It can be seen from the figure that as the number of battery cycles increases, the overall capacity of the battery shows a downward trend, and there is a phenomenon of capacity recovery locally. At the same time, as the degree of aging deepens, the time for the battery to reach the discharge threshold voltage during discharge gradually becomes smaller, the time change rate gradually increases, and the highest temperature and the time when the temperature reaches the peak also gradually change.

[0060] S2. Serialize the data in the battery aging dataset, and divide the sequence into a trend term and a seasonal term based on time decomposition; input the trend term and the seasonal term as two components into the battery state of health assessment network;

[0061] Specifically as follows:

[0062] Divide the dataset into multiple input sequences as the input of the lithium-ion battery state of health assessment network That is, a set containing input sequences with a length of and a channel (feature) number of. For the input sequence at a certain moment Adopt average pooling to eliminate periodic fluctuations and highlight long-term trends, and at the same time use padding operations to keep the length of the input sequence unchanged. The specific formula is

[0063]

[0064] where are the decomposed trend term and seasonal term. The overall aging trend of the battery is learned from the trend term, and the local fluctuation characteristics and capacity recovery phenomenon of the battery during the aging process are learned from the seasonal term. Finally, the two decomposed components are respectively input into the lithium-ion battery state of health assessment network.

[0065] Among them, time feature extraction is performed in the MLP-Block of the battery state of health assessment network. The MLP-Block consists of two fully connected layers and a GELU activation function, that is

[0066]

[0067] where are the corresponding weights and biases of the trend term MLP-Block, and are the corresponding weights and biases of the seasonal term MLP-Block, is the output of the time decomposition module, and the activation function GELU is defined as

[0068]

[0069] ​In the preferred embodiment of the present invention, the sliding window of average pooling is 25, and the number of hidden layer neurons in the MLP-Block of the trend term and the seasonal term is 64. The default length of the input sequence is 96. It should be noted that the above parameters are all non-restrictive and relatively optimal choices, and can be adjusted according to the number of samples in the dataset. The above time decomposition and feature extraction both operate on the time dimension of the sequence, that is, the rows of the sequence while the subsequent channel feature fuser operates on the channel dimension of the sequence, that is, the columns of the sequence of.

[0070] S3. Swap the time and channel dimensions of the sequence, and use skip connections to fuse the time dimension information and channel feature information;

[0071] Inverse normalize the fused data and connect it to a projection layer composed of a single linear layer to perform the mapping from the input sequence to the output sequence;

[0072] The result of the time decomposition module's perception of the time dimension information is skip-connected with the input of the time decomposition module. The result of the skip connection is layer-normalized, then the time and channel dimensions of the sequence are swapped, and a channel feature fuser composed of a single MLP-Block is used to learn the relationship between multiple different aging features of the battery in the spatial dimension. And the learning result of the channel feature fuser is skip-connected with the learning result of the time decomposition module to fuse the perceived time dimension information and channel feature information. Finally, after the fused result is further layer-normalized, a projection layer composed of a single linear layer is connected subsequently to complete the mapping learning from the input sequence to the output sequence.

[0073] In the preferred embodiment of the present invention, the MLP-Block structure of the channel feature fuser is the same as that in step 2, but the number of hidden layer neurons is different. The number of hidden layer neurons in the MLP-Block of the channel feature fuser is 32. At the same time, the entire encoder module that fuses the time dimension information and channel feature information is not single, but consists of 4 parallel structures with the same structure. The weights obtained by training each individual encoder module are different. This parallel structure can learn the differences in the battery aging process from different angles and improve the prediction accuracy.

[0074] It should be noted that the instance normalization in step 2 and the inverse instance normalization in step 3 are as follows:

[0075] In step 2, for the input sequence at a certain moment to solve the distribution problem of the corresponding prediction sequence , where the length of the input sequence and the length of the network prediction can be different.

[0076] First, the input data is normalized using the instance mean and standard deviation of the input data, and the input sequence for each instance The mean and standard deviation are calculated as

[0077]

[0078] Furthermore, these statistics are used to normalize the instance to

[0079]

[0080] where is the learnable affine parameter vector of the network, is the normalized instance. The sequence formed by has more consistent mean and variance at different times, reducing the difference between different data distributions by reducing non-stationary information.

[0081] In step 3, at the symmetric network output layer position, the output is de-normalized, explicitly returning the non-stationary attributes deleted from the input data and the affine transformation parameters learned by the network to the model output, that is

[0082]

[0083] Replace with as the output of the network, effectively alleviating the distribution difference of the data while ensuring that the original distribution information is not lost, enabling the battery health state assessment network to better learn the mapping relationship between the battery aging sequence and the health state sequence, and improving the accuracy of battery health state assessment by correcting the model prediction error caused by the battery data distribution difference.

[0084] S4. Obtain the real-time aging data of the lithium-ion battery, perform iterative training on the lithium-ion battery health state assessment network, and use the trained network to predict the health state of the lithium-ion battery.

[0085] The battery health state evaluation network selects 70% of the data as the training set and 30% of the data as the test set. The mean squared error (MSE) is selected as the loss function, and the Adam optimizer is used for optimization. The parameters β1 and β2 of Adam adopt the default values of 0.9 and 0.999 respectively. The initial learning rate is set to 0.005 and is halved at each epoch. The batch size is set to 16, and an early stopping strategy is adopted to prevent overfitting. The input sequence length of the model is 96, and the prediction sequence length is 3. The hyperparameters are optimized using grid search, and all data is normalized by Z-Score. The battery health state evaluation network iterates for 50 epochs, compares the predicted battery health state values with the actual battery health state values, calculates the error loss, and updates the network parameters according to the loss value. Finally, the trained network is used to predict the health state of the lithium-ion battery.

[0086] It should be noted that in the preferred embodiment of the present invention, the training method of the network and the parameter configuration of the learning rate optimization strategy are a non-restrictive and relatively optimal choice. Those skilled in the art can select the network training method and parameter configuration according to various indicators such as the effect and accuracy of health assessment.

[0087] In the preferred embodiment of the present invention, the early stopping strategy is judged by the validation set loss. When the validation set loss does not decrease for 10 consecutive epochs, it is judged that the network has overfitted, and the network with the minimum validation set loss is saved as the optimal network for prediction.

[0088] The present invention also provides a lithium-ion battery health state evaluation system based on time decomposition and feature fusion, which is used to implement the steps of the above-mentioned lithium-ion battery health state evaluation method, including a data acquisition module, a data processing module, and a battery health state evaluation network;

[0089] The data acquisition module is used to acquire voltage, current, temperature, and capacity data during the aging process of the lithium-ion battery; perform preprocessing, and extract key features to form a battery aging data set;

[0090] The data processing module is used to serialize the data in the battery aging data set, and based on time decomposition, divide the sequence into a trend term and a seasonal term; input the trend term and the seasonal term as two components into the battery health state evaluation network;

[0091] The battery health state evaluation network includes a time decomposition module and a channel feature fuser; the time decomposition module is used to exchange the time and channel dimensions of the sequence, and the channel feature fuser uses a skip connection method to fuse the time dimension information and the channel feature information; the fused data is de-normalized and then connected to a projection layer composed of a single linear layer to perform the mapping from the input sequence to the output sequence;

[0092] Obtain real-time aging data of lithium-ion batteries, iteratively train the health state assessment network of lithium-ion batteries, and use the trained network to predict the health state of lithium-ion batteries.

[0093] In the preferred embodiment of the present invention, the software configuration for model construction and testing is python3.9, pytorch2.1.1, cuda12.1, visual studio code2023; the hardware configuration is an inter core i5-8300H 8-core 16-thread CPU, an Nvidia GTX-1050Ti 4GB GPU, and 16G of memory.

[0094] For the lithium-ion battery health state assessment network obtained in the preferred embodiment of the present invention, use the test set data divided by the NASA dataset and the CALCE dataset for testing. In order to verify the effectiveness of the method proposed in the present invention, compare this method with other existing mainstream methods such as CNN-LSTM, CNN-Transformer, and MVIP-Trans. The experimental results of the NASA dataset are as Figure 3 shown, and the experimental results of the CALCE dataset Figure 4 are shown.

[0095] From Figure 3 and Figure 4 it can be seen that under battery experiments with different numbers, the method proposed in the present invention is superior to other mainstream methods, indicating that the proposed method can well adapt to different battery health state assessment tasks. In the Figure 3 NASA dataset, this method can well predict the local capacity recovery details, and at the same time can avoid affecting the overall battery aging trend due to local mutations. When conducting battery health assessment, CNN-LSTM and CNN-Transformer cannot accurately predict the capacity recovery details, resulting in error accumulation, and thus showing large fluctuations at the 150-170 cycles of batteries numbered B05 and B06. For the more data-intensive CALCE dataset, as Figure 4As shown, the local details during its aging process are more abundant. The traditional CNN-LSTM model can only judge the general aging trend of the battery. Although the CNN-Transformer and MVIP-Trans methods based on Transformer can roughly predict the local change details, due to the interference of position encoding and the capture of time dependence by CNN, the prediction of local changes is not accurate enough. Compared with other existing methods, the method proposed in the present invention benefits from its time decomposition and feature fusion architecture, and can learn the trend, local details and different features respectively from the whole, and thus achieves better prediction results in the task of battery health state assessment on a long time scale.

[0096] In a preferred embodiment of the present invention, the mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE) and coefficient of determination R 2 are used as evaluation indicators to evaluate the accuracy of predicting the health state of lithium-ion batteries, and the results are shown in Tables 1 and 2.

[0097] Table 1 Battery health state assessment results of NASA dataset

[0098]

[0099]

[0100] As can be seen from Table 1, the battery health state assessment network in this embodiment is tested on the NASA dataset, and the average values of MSE, MAE, RMSE and R 2 under different numbered batteries are 0.00003, 0.0039, 0.0047 and 0.9859. Compared with other existing assessment methods, the average values of MAE and RMSE of this method are 27.9%, 24.5%, 29.3% and 30.1%, 22.5%, 28.8% of CNN-LSTM, CNN-Transformer and MVIP-Trans. As can be seen from Table 2, the battery health state assessment network in this embodiment is tested on the CALCE dataset, and the average values of MSE, MAE, RMSE and R 2 under different numbered batteries are 0.0038, 0.0345, 0.0617 and 0.9956. Compared with other existing assessment methods, the average values of MAE and RMSE of this method are 27.4%, 26.2%, 38.2% and 37.3%, 32.0%, 50.0% of CNN-LSTM, CNN-Transformer and MVIP-Trans. The results show that the health state assessment error of the method proposed in the present invention is lower, and there is a large improvement compared with the existing methods, and it has a good effect on learning the global trend and local details of the battery aging process on a long time scale.

[0101] Table 2 Battery health state assessment results of the CALCE dataset

[0102]

[0103] To more intuitively compare the error distribution of the method proposed in the present invention with that of other existing methods, Figure 5 and Figure 6 the absolute error box plots of different methods for batteries with different numbers in the NASA and CALCE datasets are plotted. The black horizontal line is the median of the error, the upper and lower boundaries of the gray box line are the first quartile and the third quartile respectively, the black boundary is the maximum and minimum values except for the outliers, and the circles are the outliers. It can be seen from the figure that compared with other methods, the median error of the method proposed in the present invention is almost the lowest. Although there are a small number of outliers like other methods, the error range is more concentrated and the overall error distribution range is lower. This proves that the method we proposed can adapt to the task of lithium-ion battery health state assessment on a long time scale, has less prediction fluctuations for different battery cycles, and has better robustness and accuracy.

[0104] Obviously, the above embodiments are only examples for clear illustration and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or variations can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. And the obvious changes or variations derived therefrom are still within the protection scope of the present invention.

Claims

1. A lithium-ion battery health status assessment method based on time decomposition and feature fusion, characterized in that: The following steps are involved: S1. Obtaining voltage, current, temperature and capacity data of lithium-ion battery aging process; Perform preprocessing and extract key features to form a battery aging data set; S2, serialize the data in the battery aging dataset, and divide the sequence into trend items and seasonal items based on time decomposition; The trend term and the season term are input into the battery health status assessment network as two components; including: Divide the battery aging dataset into multiple data series , where X indicates inclusion The length is , the number of channels is The input sequence A collection of; For any input sequence at any time , average pooling is used to eliminate periodic fluctuations and highlight long-term trends, while padding operations are used to keep the input sequence length unchanged. The specific formula is: Formula (2) in are the trend term and seasonal term obtained by decomposition respectively; Express Perform the filling operation. represents average pooling; S3, exchange the time and channel dimensions of the sequence, and use skip connections to fuse the time dimension information and channel feature information; The fused data is denormalized and then connected to the projection layer consisting of a single linear layer to map the input sequence to the output sequence; The result of the time dimension information perception of the time decomposition module is jump-connected with the input of the time decomposition module, and the result of the jump connection is normalized through the layer to exchange the time and channel dimensions of the sequence. The channel feature fusion device composed of a single MLP-Block is used to learn the relationship between multiple different aging features of the battery in the spatial dimension, and the learning result of the channel feature fusion device is jump-connected with the learning result of the time decomposition module to fuse the perceived time dimension information and channel feature information; after the fusion result is further normalized through the layer, it is subsequently connected to a projection layer composed of a single linear layer to complete the mapping learning from the input sequence to the output sequence; S4. Acquire real-time aging data of lithium-ion batteries, iteratively train a lithium-ion battery health status assessment network, and use the trained network to predict the health status of lithium-ion batteries.

2. According to claim 1, a lithium-ion battery health status assessment method based on time decomposition and feature fusion is characterized in that: In S1, the voltage, current, temperature and capacity data of the lithium-ion battery aging process are obtained and pre-processed, including: Select the time for the battery voltage to drop to the discharge threshold voltage, the voltage slope, the time required for the temperature to reach the maximum value, the temperature slope, the maximum temperature, the constant voltage charging time, the constant current charging time and the internal resistance as input characteristics; Calculate the current battery health status based on the battery capacity data, and use the battery health status value as the label of the data set; The input features and the data set labels are processed for outliers to form a battery aging data set.

3. The method for evaluating the health status of a lithium-ion battery based on time decomposition and feature fusion according to claim 1, characterized in that: The preprocessing also includes removing outliers from the acquired data based on the 3-sigma principle and standardizing the data to have zero mean and variance before extracting the input features.

4. The method for evaluating the health status of a lithium-ion battery based on time decomposition and feature fusion according to claim 1, characterized in that: The trend term and the season term are input as two components into the battery health status assessment network, including respectively inputting the two decomposed components into the MLP-Block of the lithium-ion battery health status assessment network for time feature extraction. The MLP-Block includes two fully connected layers, which are expressed by the following formula: Formula (3) in, is the corresponding weight and bias of the trend term MLP-Block, are the weights and biases corresponding to the seasonal term MLP-Block, It is the output of the time decomposition module.

5. The method for evaluating the health status of a lithium-ion battery based on time decomposition and feature fusion according to claim 1, characterized in that: The activation function GELU is also expressed by the following formula (4): Formula (4) in, Represents the input of the network at the current layer.

6. The method for evaluating the health status of a lithium-ion battery based on time decomposition and feature fusion according to claim 1, characterized in that: S2 also includes serializing the data in the battery aging data set, using instance normalization to remove non-stationary information, and obtaining stationary sequences and non-stationary attribute parameters and affine transformation parameters, which specifically includes the following steps: The battery aging data set is divided into multiple data sequences, the mean and standard deviation of each data sequence are calculated, and the instances in each sequence are separated by the mean and standard deviation. Perform normalization processing; Formula(6) in, is the affine parameter vector, is the normalized instance, To calculate the average value of the data, To obtain the standard deviation of the data, is a fixed minimum value to prevent the denominator from being zero.

7. A lithium-ion battery health status assessment method based on time decomposition and feature fusion according to claim 6, characterized in that: In S3, the fused data is inverse normalized and then connected to a projection layer consisting of a single linear layer; The fused data is denormalized, which is to combine the fused result with the S2 normalization related parameters and then perform denormalization, including: In the battery health status assessment network, the symmetrical network output layer outputs Perform a denormalization operation; explicitly remove non-stationary properties from the input data And the affine transformation parameters learned by the network Returning to the model output, Formulas (7) Will replace as the output of the network.

8. A lithium-ion battery health status assessment system based on time decomposition and feature fusion, characterized in that: The steps for implementing any one of the above-mentioned lithium-ion battery health status assessment methods based on time decomposition and feature fusion in claims 1-7 include a data acquisition module, a data processing module and a battery health status assessment network The data acquisition module is used to acquire the voltage, current, temperature and capacity data of the lithium-ion battery during the aging process; Perform preprocessing and extract key features to form a battery aging data set; The data processing module is used to serialize the data in the battery aging data set, and divide the sequence into trend items and seasonal items based on time decomposition; and input the trend items and seasonal items as two components into the battery health status assessment network; The battery health status assessment network includes a time decomposition module and a channel feature fusion module; The time decomposition module is used to exchange the time and channel dimensions of the sequence, and the channel feature fuser fuses the time dimension information and the channel feature information by means of skip connection; the fused data is inversely normalized and then connected to the projection layer composed of a single linear layer to map the input sequence to the output sequence; Acquire real-time aging data of lithium-ion batteries, iteratively train the lithium-ion battery health status assessment network, and use the trained network to predict the health status of lithium-ion batteries.

Images