A Method for Estimating the State of Health of Lithium Batteries Based on a Hybrid Deep Network Incorporating an Attention Mechanism

Abstract

The present invention discloses a method for estimating the state of health (SOH) of a lithium battery using a hybrid deep network integrating an attention mechanism, comprising the following steps: (1) obtaining cyclic aging data of N lithium batteries; (2) calculating the true SOH values of the N lithium batteries and eliminating abnormal cycles to obtain an initial data set D1 to D N ; (3) obtaining the temperature difference curves of the N lithium batteries; (4) extracting the aging characteristics of the N lithium batteries; (5) constructing a training set and a test set; (6) establishing an SOH estimation model for a lithium battery using a hybrid deep network integrating an attention mechanism; (7) training the SOH estimation model for a lithium battery using a hybrid deep network integrating an attention mechanism; (8) obtaining the SOH estimation value. The present invention fully integrates the advantages of three algorithms, namely a CNN network, a GRU network, and an attention mechanism, can achieve accurate and reliable estimation of the SOH of a lithium battery, and has strong robustness and generality to battery inconsistency.

Description

Technical Field

[0001] The present invention relates to the field of battery SOH estimation, and in particular to a lithium battery SOH estimation method using a hybrid deep network integrating an attention mechanism. Background Art

[0002] Accurate and reliable SOH estimation is crucial to the safety and power of electric vehicles. However, high-precision SOH estimation becomes a particularly challenging task due to the complex and irreversible physical or chemical side reactions that occur during battery aging.

[0003] Common SOH methods mainly include experimental-based methods, model-based methods, and data-driven methods. Experimental-based methods have strict requirements on the test environment and equipment and are difficult to apply online. Model-based methods make it difficult to establish a high-fidelity physical model to describe the aging process of batteries under various complex operating conditions. In recent years, data-driven methods have attracted widespread attention because they do not consider the complex aging reactions of batteries and directly use historical battery data to establish SOH prediction models. As we all know, SOH estimation is a time series prediction task. With the rapid development of deep learning, LSTM networks and GRU networks developed based on recurrent neural networks have become important methods for time series prediction tasks and have also been gradually applied to SOH estimation. Although these methods have achieved satisfactory results, there are two problems that need to be further resolved.

[0004] First, the aging feature determination process based on LSTM networks or GRU networks is cumbersome and complex. This process mainly includes, but is not limited to, manually generating candidate features based on experience, evaluating the relationship between the generated features and SOH through correlation analysis, and then selecting features with high correlation as aging features. Undoubtedly, this process is time-consuming. Therefore, how to simplify the aging feature determination process is a problem that needs to be further solved.

[0005] Second, although the SOH estimation methods based on LSTM networks or GRU networks can extract useful time-related information from battery data, they cannot distinguish the importance of each time information to SOH. Specifically, when using LSTM networks or GRU networks to predict SOH, different historical time series data have different effects on the current SOH prediction. As the prediction time increases, the degree of time influence also changes. Needless to say, paying too much attention to unimportant information and ignoring important information are not conducive to SOH estimation. Therefore, how to emphasize the impact of useful information on SOH and reduce the impact of unimportant information on SOH to improve the estimation accuracy of SOH also needs to be further addressed. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a lithium battery SOH estimation method of a hybrid deep network integrating an attention mechanism to solve the problems in the prior art that the aging feature selection process is cumbersome and time-consuming, and the LSTM network and GRU network cannot distinguish the importance of each time information to SOH.

[0007] In order to solve the above technical problems, the technical solution of the present invention is to provide a lithium battery SOH estimation method of a hybrid deep network integrating an attention mechanism, and its innovation lies in that it includes the following steps:

[0008] (1) Obtaining cycle aging data of N lithium batteries: performing cycle charge and discharge aging tests on N lithium batteries, and collecting aging data of the N lithium batteries respectively;

[0009] (2) Calculate the actual SOH values ​​of N lithium batteries and eliminate abnormal cycles to obtain the initial data set D1~D N : According to the aging data in step (1), the real SOH value of each lithium battery in each discharge cycle is calculated respectively, and the abnormal cycle is eliminated according to the real SOH value of each lithium battery in each discharge cycle to obtain the initial data set D1~D N ;

[0010] (3) Obtaining the temperature difference curves of N lithium batteries: Extract the initial data sets D1 to D2 obtained in step (2) respectively. N The temperature and time collected in each charging cycle are calculated, the temperature difference value of each charging cycle is calculated, and then the KF filter is used for smoothing, and the temperature difference curve of N lithium batteries is obtained with voltage as the horizontal axis and temperature difference value as the vertical axis;

[0011] (4) extracting aging characteristics of N lithium batteries: according to the temperature difference curves of the N lithium batteries obtained in step (3), extracting multiple aging characteristics from the peaks and troughs of each temperature difference curve of the N lithium batteries in the voltage range of 3.3 V to 3.9 V, wherein the aging characteristics include temperature difference value, voltage and time;

[0012] (5) Constructing training and test sets: The initial data set D1 to D2 of N lithium batteries obtained in step (2) is constructed. N The SOH values ​​in the above equations are matched one by one with the aging characteristics of the N lithium batteries extracted in step (4) to form N new data sets A1~A N , from A1 to A N Randomly select one data set as the training set, and the remaining N-1 data sets as the test set;

[0013] (6) Establish a lithium battery SOH estimation model with a hybrid deep network integrating an attention mechanism: Build a lithium battery SOH estimation model using an input layer, a CNN layer, a GRU layer, an attention mechanism layer, a Flatten layer, a Dense layer, and an output layer, and use the trial-and-error method to determine the hyperparameters of the lithium battery SOH estimation model;

[0014] (7) Train the lithium battery SOH estimation model with a hybrid deep network integrating an attention mechanism: Use the training set determined in step (5) to train the lithium battery SOH estimation model established in step (6) to determine the lithium battery SOH estimation model with a hybrid deep network integrating an attention mechanism;

[0015] (8) Obtain the SOH estimation value: Input the test set determined in step (5) into the lithium battery SOH estimation model determined in step (7) to obtain the SOH estimation value of the lithium battery.

[0016] Further, the aging data of the N lithium batteries collected in step (1) includes voltage, current, temperature, capacity, and time data.

[0017] Further, the specific process of obtaining the initial datasets D1 to D N in step (2) is as follows: Calculate the maximum capacity that can be discharged in each discharge cycle of each lithium battery, and take the maximum capacity divided by the nominal capacity of the battery and multiplied by 100% as the true SOH value of each lithium battery. Then, calculate the SOH difference between adjacent two discharge cycles of each lithium battery, and regard the discharge cycles and subsequent charge-discharge cycles with an SOH difference greater than 5% as abnormal cycles. The charge-discharge cycle data of the N lithium batteries after excluding the abnormal cycles is used as the initial datasets D1 to D N .

[0018] Further, the calculation formula for calculating the temperature difference value of each charging cycle in step (3) is:

[0019]

[0020] where l represents the time step, and l is set to 20 by the trial-and-error method.

[0021] Further, the lithium battery SOH estimation model of the hybrid deep network with a fusion attention mechanism established in step (6) includes an input layer, a CNN layer, a GRU layer, an attention mechanism layer, a Flatten layer, a Dense layer, and an output layer. Its hyperparameters include the number of CNN layers, the number of convolutional kernels, the size of the convolutional kernels, the number of GRU layers, the number of GRU units, the number of Dense layers, and the number of neurons in the Dense layers. The number of CNN layers, the number of convolutional kernels, the size of the convolutional kernels, the stride, the number of GRU layers, the number of GRU units, the number of Dense layers, and the number of neurons in the Dense layers are determined to be 1, 32, 1, 1, 60, 1, 20 respectively by the trial-and-error method.

[0022] Compared with the prior art, the lithium battery SOH estimation method of the hybrid deep network with a fusion attention mechanism of the present invention has the following advantages:

[0023] (1) The present invention uses the CNN layer with a powerful automatic feature extraction ability to automatically extract the potential features in the aging features, avoiding the re-selection of features after various correlation analyses and simplifying the determination process of the aging features.

[0024] (2) Use the GRU layer with strong time modeling ability to capture the time correlation between different time series of the cyclic aging data of the lithium battery.

[0025] (3) Use the attention mechanism layer to assign different weights to the hidden states of the GRU layer to express the influence of the hidden states on the SOH at different times, highlighting the important time information and weakening the unimportant time information.

[0026] (4) The present invention fully integrates the advantages of the three algorithms of the CNN network, the GRU network, and the attention mechanism, can achieve accurate and reliable estimation of the lithium battery SOH, and has strong robustness and generality to battery inconsistency. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0028] Figure 1 It represents a flowchart of a lithium battery SOH estimation method of a hybrid deep network with a fusion attention mechanism.

[0029] Figure 2 It represents the SOH graph after removing abnormal cycles of B1 to B8.

[0030] Figure 3 The temperature difference curve diagram representing B1.

[0031] Figure 4 The schematic diagram representing the extraction of aging characteristics.

[0032] Figure 5 The structural schematic diagram of the hybrid deep network integrating the attention mechanism.

[0033] Figure 6 The schematic diagram representing the SOH estimation result of B2. Specific implementation manners

[0034] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention will be clearly described below in conjunction with the accompanying drawings and implementation manners.

[0035] The present invention provides a method for estimating the SOH of a lithium battery by using a hybrid deep network integrating the attention mechanism. The innovation lies in the following steps:

[0036] (1) Obtain the cyclic aging data of N lithium batteries: Perform cyclic charge and discharge aging tests on N lithium batteries, and respectively collect the aging data of N lithium batteries. The collected aging data of N lithium batteries includes voltage, current, temperature, capacity, and time data.

[0037] (2) Calculate the true SOH values of N lithium batteries and eliminate abnormal cycles to obtain the initial datasets D1 to D N : Calculate the true SOH value of each discharge cycle of each lithium battery according to the aging data in step (1), and eliminate abnormal cycles through the true SOH values of each discharge cycle of each lithium battery to obtain the initial datasets D1 to D N .

[0038] The specific process is as follows: Calculate the maximum capacity that can be discharged in each discharge cycle of each lithium battery, and multiply the maximum capacity by 100% after dividing it by the nominal capacity of the battery as the true SOH value of each lithium battery. Then, calculate the SOH difference between two adjacent discharge cycles of each lithium battery respectively. Consider the discharge cycles with an SOH difference greater than 5% and the subsequent charge and discharge cycles as abnormal cycles. Take the charge and discharge cycle data of N lithium batteries after eliminating abnormal cycles as the initial datasets D1 to D N .

[0039] (3) Obtain the temperature difference curves of N lithium batteries: Respectively extract the temperature and time collected in each charge cycle from the initial datasets D1 to D N obtained in step (2), and calculate the temperature difference value of each charge cycle. The calculation formula is:

[0040]

[0041] Among them, l represents the time step, and l is set to 20 through the trial-and-error method. Then, KF filtering is used for smoothing processing, and a temperature difference curve of N lithium batteries is obtained with voltage as the abscissa and the temperature difference value as the ordinate.

[0042] (4) Extract the aging characteristics of N lithium batteries: According to the temperature difference curves of N lithium batteries obtained in step (3), multiple aging characteristics are extracted from the peaks and valleys of each temperature difference curve of N lithium batteries within the voltage range of 3.3V to 3.9V. The aging characteristics include temperature difference value, voltage, and time.

[0043] (5) Construct a training set and a test set: The SOH values in the initial data sets D1 to D N of the N lithium batteries obtained in step (2) are put into one-to-one correspondence with the aging characteristics of N lithium batteries extracted in step (4) to form N new data sets A1 to A N , and any one of the data sets from A1 to A N is randomly selected as the training set, and the remaining N - 1 data sets are used as the test set.

[0044] (6) Establish a lithium battery SOH estimation model with a hybrid deep network integrating an attention mechanism: Use an input layer, a CNN layer, a GRU layer, an attention mechanism layer, a Flatten layer, a Dense layer, and an output layer to build a lithium battery SOH estimation model. Establishing a lithium battery SOH estimation model with a hybrid deep network integrating an attention mechanism includes an input layer, a CNN layer, a GRU layer, an attention mechanism layer, a Flatten layer, a Dense layer, and an output layer. Its hyperparameters include the number of CNN layers, the number of convolutional kernels, the size of convolutional kernels, the number of GRU layers, the number of GRU units, the number of Dense layers, and the number of neurons in the Dense layers. And the number of CNN layers, the number of convolutional kernels, the size of convolutional kernels, the stride, the number of GRU layers, the number of GRU units, the number of Dense layers, and the number of neurons in the Dense layers are determined to be 1, 32, 1, 1, 60, 1, 20 respectively through the trial-and-error method.

[0045] (7) Train the lithium battery SOH estimation model with a hybrid deep network integrating an attention mechanism: Use the training set determined in step (5) to train the lithium battery SOH estimation model established in step (6) to determine the lithium battery SOH estimation model with a hybrid deep network integrating an attention mechanism.

[0046] (8) Obtain the SOH estimated value: Input the test set determined in step (5) into the lithium battery SOH estimation model determined in step (7) to obtain the SOH estimated value of the lithium battery.

[0047] Based on the above technical solution, the present invention further describes the technical solution through the following specific implementation process. Implementation method The cycle aging data of 8 lithium batteries with a rated capacity of 0.74Ah are obtained. The 8 lithium batteries are named B1 to B8 respectively. The specific method flow chart is as follows Figure 1 As shown, the following steps are included:

[0048] (1) Obtaining cycle aging data of 8 lithium batteries: Performing cycle charge and discharge aging tests on 8 lithium batteries with a rated capacity of 0.74Ah, and collecting aging data such as voltage, current, temperature, capacity and time parameters of the 8 lithium batteries.

[0049] (2) Calculate the true SOH values ​​of the eight lithium batteries and eliminate abnormal cycles to obtain initial data sets D1 to D8: Calculate the true SOH value of each lithium battery for each discharge cycle based on the aging data in step (1), and eliminate abnormal cycles based on the true SOH value of each lithium battery for each discharge cycle to obtain initial data sets D1 to D8.

[0050] The specific process is: calculate the maximum capacity that can be discharged in each discharge cycle of each lithium battery, and divide the maximum capacity by the nominal capacity of the battery and multiply it by 100% as the true SOH value of each lithium battery, and then calculate the SOH difference between two adjacent discharge cycles of each lithium battery respectively, and regard the discharge cycle and the subsequent charge and discharge cycle with an SOH difference greater than 5% as abnormal cycles, and use the charge and discharge cycle data of the 8 lithium batteries after the abnormal cycles are eliminated as the initial data sets D1~D8.

[0051] like Figure 2 The figure shows the SOH graphs of eight lithium batteries B1 to B8 drawn using the SOH values ​​in the initial data set D1 to D8. It can be seen that there are certain differences in the initial SOH of the cells, and as the number of cycles exceeds 1000, the aging trend of the cells tends to gradually disperse. These phenomena indicate that as the battery performance deteriorates, the consistency between the eight batteries becomes worse and worse.

[0052] (3) Obtaining the temperature difference curves of the eight lithium batteries: extracting the temperature and time collected in each charging cycle in the initial data set D1 to D8 obtained in step (2), respectively, and calculating the temperature difference value of each charging cycle. The calculation formula is:

[0053]

[0054] Among them, l represents the time step, and l is set to 20 by trial and error. Then KF filtering is used for smoothing, and the temperature difference curve of N lithium batteries is obtained with voltage as the horizontal axis and temperature difference as the vertical axis.

[0055] like Figure 3As shown, it is the temperature difference curve graph of B1. It can be clearly seen that as the number of cycles of the B1 battery increases, the peak and valley positions of the curve change regularly, indicating that there is a correlation between the peak and valley positions of the temperature difference curve and battery aging.

[0056] (4) Extract the aging characteristics of 8 lithium batteries: According to the temperature difference curves of the 8 lithium batteries obtained in step (3), in a specific voltage range of 3.3V to 3.9V, extract multiple aging characteristics from the peaks and valleys of each temperature difference curve of the 8 lithium batteries, including temperature difference values, voltages, and times.

[0057] As Figure 4 shown, it is a schematic diagram of aging characteristic extraction. It can be seen that in this embodiment, 11 aging characteristics are extracted from the peaks and valleys of the temperature difference curve, denoted as F1 to F11. F1, F2, and F3 respectively represent the voltages of peak 1, valley 2, and peak 2 of the temperature difference curve; F4 represents the voltage between valley 2 and peak 1 of the temperature difference curve; F5 represents the voltage between peak 2 and valley 2 of the temperature difference curve; F6, F7, and F8 respectively represent the temperature difference values of peak 1, valley 2, and peak 2 of the temperature difference curve; F9 represents the temperature difference value between the peak and peak 1 and valley 2 of the temperature difference curve; F10 represents the time required for the peak 1 of the temperature difference curve to change to valley 2, and F11 represents the time required for valley 1 of the temperature difference curve to change to peak 2.

[0058] (5) Construct a training set and a test set: Correlate the SOH values in the initial data sets D1 to D8 of the 8 lithium batteries obtained in step (2) with the aging characteristics of the 8 lithium batteries extracted in step (4) one by one to form 8 new data sets A1 to A8. In this embodiment, select data set A1 from A1 to A8 as the training set, and the remaining 7 data sets A2 to A8 as the test set.

[0059] (6) Establish a lithium battery SOH estimation model with a hybrid deep network integrating an attention mechanism: Use an input layer, a CNN layer, a GRU layer, an attention mechanism layer, a Flatten layer, a Dense layer, and an output layer to build an SOH estimation model and determine the hyperparameters of the model. The hyperparameters include the number of CNN layers, the number of convolutional kernels, the size of the convolutional kernels, the number of GRU layers, the number of GRU units, the number of Dense layers, and the number of neurons in the Dense layers. The determined values of the number of CNN layers, the number of convolutional kernels, the size of the convolutional kernels, the stride, the number of GRU layers, the number of GRU units, the number of Dense layers, and the number of neurons in the Dense layers are 1, 32, 1, 1, 60, 1, 20 respectively; As Figure 5 shown, it is a schematic diagram of the structure of the hybrid deep network integrating an attention mechanism used in the present invention.

[0060] (7) Training the SOH estimation model of the hybrid deep network integrating the attention mechanism: Use the training set determined in step (5) to train the SOH estimation model, and determine the SOH estimation model of the hybrid deep network integrating the attention mechanism.

[0061] (8) Obtaining the SOH estimation value: Input the test set determined in step (5) into the SOH estimation model determined in step (7) to obtain the SOH estimation value. Table 1 shows the aging characteristics F1 - F11 of B2 and the SOH estimation value. As Figure 6 shown, it is the schematic diagram of the SOH estimation result of B2. It can be seen that the error between the SOH estimation value and the true SOH value is controlled within 1%, indicating the accuracy and effectiveness of the method proposed by the present invention.

[0062] Table 1

[0063] F1 F2 F3 F4 F5 F6 …… F10 F11 SOH Estimation Value 597 1949 3.5042 0.0001 3.4749 -0.0018 …… 0.0009 0.1430 0.9826 579 1811 3.5015 0.0002 3.4571 -0.0018 …… 0.0009 0.1247 0.9758 574 1859 3.5120 0.0001 3.4661 -0.0019 …… 0.0010 0.1291 0.9686 …… …… …… …… …… …… …… …… …… …… 499 1569 3.5428 0.0003 3.4402 -0.0018 …… 0.0010 0.1186 0.8632 497 1619 3.5408 0.0003 3.4341 -0.0017 …… 0.0011 0.1521 0.8619 468 1619 3.5541 0.0002 3.4230 -0.0018 …… 0.0011 0.1660 0.8543 458 1628 3.5558 0.0002 3.4167 -0.0018 …… 0.0012 0.1911 0.8537 …… …… …… …… …… …… …… …… …… …… 289 282 1485 3.6194 -0.0007 3.2957 …… 0.0012 0.4415 0.8274 214 239 1486 3.6453 -0.0011 3.2341 …… 0.0013 0.5085 0.8213 252 271 1449 3.6267 -0.0009 3.2961 …… 0.0013 0.4360 0.8288 256 270 1441 3.6290 -0.0010 3.3048 …… 0.0013 0.4323 0.8225

[0064] In the present invention, four indicators, RMSE, MAE, MAX, and R 2 are used to quantitatively evaluate the SOH estimation performance of the proposed method, and the calculation formulas are as follows:

[0065]

[0066] where L represents the length of the test data, t represents the serial number of the test data, y t represents the true SOH value, represents the SOH estimation value, represents the average value of the SOH estimation values.

[0067] In addition, in order to evaluate the overall SOH prediction performance of the proposed method, the present invention also calculates the average values of RMSE, MAE, MAX, and R 2 .

[0068] The SOH estimation results using B1 as the training set and the remaining 7 lithium batteries (B2 - B8) as the test sets respectively are shown in Table 2:

[0069] Table 2

[0070]

[0071]

[0072] From the estimation results of this embodiment, it can be seen that the optimal RMSE, MAE, MAX, and R 2 appear in B2 at the same time, which are 0.323%, 0.246%, 0.854%, and 0.9971 respectively; the worst RMSE, MAE, MAX, and R 2Appearing in B6 simultaneously, they are 0.519%, 0.413%, 1.247% and 0.9930 respectively. From the overall results, the average values of RMSE, MAE, MAX and R 2 are 0.448%, 0.365%, 1.091% and 0.9952 respectively.

[0073] In summary, the proposed method avoids the re-selection of features after correlation analysis of the extracted aging features by using the CNN layer, simplifies the process of determining aging features; uses the GRU layer to capture the temporal correlation between different time series of the cyclic aging data of lithium batteries; utilizes the attention mechanism layer to assign different weights to the hidden states of the GRU layer to express the influence of the hidden states on the SOH at different moments, highlighting important time information and weakening unimportant time information. From all the above estimation results, it can be shown that the present invention fully integrates the advantages of the three algorithms of the CNN network, the GRU network and the attention mechanism. All the estimation errors are within 1.3%, achieving accurate and reliable estimation of the SOH of lithium batteries, and having strong robustness and generality to battery inconsistency.

[0074] The above-described embodiments are only described as the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various variations and improvements made by ordinary engineering and technical personnel in the field to the technical solutions of the present invention should fall within the protection scope of the present invention. The technical content claimed by the present invention has been fully recorded in the technical requirements.

Claims

1. A method for estimating the state of health (SOH) of a lithium battery using a hybrid deep network integrating an attention mechanism, characterized in that, It includes the following steps: (1) Obtain the cyclic aging data of N lithium batteries: conduct cyclic charge and discharge aging tests on N lithium batteries, and respectively collect the aging data of N lithium batteries; (2) Calculate the true SOH values of N lithium batteries and eliminate abnormal cycles to obtain the initial datasets D1 to D N : Calculate the true SOH value of each discharge cycle of each lithium battery based on the aging data in step (1), and eliminate abnormal cycles through the true SOH values of each discharge cycle of each lithium battery to obtain the initial datasets D1 to D N ; (3) Obtain the temperature difference curves of N lithium batteries: Extract the temperature and time collected in each charge cycle from the initial data sets D1 to D obtained in step (2) respectively, calculate the temperature difference value of each charge cycle, then perform smoothing processing using KF filtering, and use voltage as the abscissa and the temperature difference value as the ordinate to obtain the temperature difference curves of N lithium batteries; N ​ (4) Extract the aging characteristics of N lithium batteries: according to the temperature difference curves of N lithium batteries obtained in step (3), in the voltage range of 3.3V - 3.9V, extract multiple aging characteristics from the peaks and troughs of each temperature difference curve of N lithium batteries respectively. The aging characteristics include temperature difference value, voltage, and time; (5) Construct a training set and a test set: Correlate the SOH values in the initial data sets D1 to D of the N lithium batteries obtained in step (2) with the aging characteristics of the N lithium batteries extracted in step (4) to form N new data sets A1 to A N , and arbitrarily select one data set from A1 to A N as the training set, and the remaining N - 1 data sets as the test set; N ​ (6) Establish a lithium battery SOH estimation model with a hybrid deep network integrating an attention mechanism: use an input layer, a CNN layer, a GRU layer, an attention mechanism layer, a Flatten layer, a Dense layer, and an output layer to build a lithium battery SOH estimation model and use the trial - and - error method to determine the hyperparameters of the lithium battery SOH estimation model; (7) Train the lithium battery SOH estimation model with a hybrid deep network integrating an attention mechanism: use the training set determined in step (5) to train the lithium battery SOH estimation model established in step (6) to determine the lithium battery SOH estimation model with a hybrid deep network integrating an attention mechanism; (8) Obtain the SOH estimation value: input the test set determined in step (5) into the lithium battery SOH estimation model determined in step (7) to obtain the SOH estimation value of the lithium battery.

2. The method for estimating the state of health (SOH) of a lithium battery using a hybrid deep network integrating an attention mechanism according to claim 1, wherein, The aging data of N lithium batteries collected in step (1) includes voltage, current, temperature, capacity, and time data.

3. The method for estimating the state of health (SOH) of a lithium battery using a hybrid deep network integrating an attention mechanism according to claim 1, wherein, In the step (2), the initial data sets D1 to D N are obtained through the following specific process: calculate the maximum capacity that can be discharged in each discharge cycle of each lithium battery, divide the maximum capacity by the nominal capacity of the battery and multiply by 100% to obtain the true SOH value of each lithium battery, then calculate the SOH difference between two adjacent discharge cycles of each lithium battery respectively, regard the discharge cycles with an SOH difference greater than 5% and the subsequent charge and discharge cycles as abnormal cycles, and use the charge and discharge cycle data after excluding the abnormal cycles of N lithium batteries as the initial data sets D1 to D N .

4. The method for estimating the state of health (SOH) of a lithium battery using a hybrid deep network integrating an attention mechanism according to claim 1, characterized in that, The calculation formula for calculating the temperature difference value of each charging cycle in step (3) is: where l represents the time step, and l is set to 20 through the trial - and - error method.

5. The method for estimating the state of health (SOH) of a lithium battery using a hybrid deep network integrating an attention mechanism according to claim 1, wherein, The lithium battery SOH estimation model with a hybrid deep network integrating an attention mechanism established in step (6) includes an input layer, a CNN layer, a GRU layer, an attention mechanism layer, a Flatten layer, a Dense layer, and an output layer. Its hyperparameters include the number of CNN layers, the number of convolutional kernels, the size of convolutional kernels, the number of GRU layers, the number of GRU units, the number of Dense layers, and the number of neurons in the Dense layers. And the number of CNN layers, the number of convolutional kernels, the size of convolutional kernels, the stride, the number of GRU layers, the number of GRU units, the number of Dense layers, and the number of neurons in the Dense layers are respectively determined to be 1, 32, 1, 1, 60, 1, 20 through the trial - and - error method.

Images