SOH estimation method of battery pack

Abstract

The invention provides an SOH (state of health) estimation method for a battery pack, which comprises the following steps of: S1, acquiring a capacity label value required for estimating the SOH of the battery pack based on historical actual operation data of the battery pack, and constructing a sliding window input data matrix related to the SOH; s2, inputting a data matrix based on a sliding window, and automatically extracting a health feature map related to the aging of the battery pack through a neural network; and S3, based on the extracted health feature map, modeling is carried out on battery attenuation dynamics through a PINN neural network, and estimation of the SOH of the battery is realized.

Description

Technical Field

[0001] The present application relates to the field of energy storage, and in particular to a method for estimating the SOH of a battery pack. Background Art

[0002] In the field of power electronics, battery packs are widely used as energy storage components in various scenarios, such as electric vehicles, portable electronic devices, and energy storage applications.

[0003] Among them, the battery pack is composed of multiple single cells connected in series, parallel, or in series-parallel. The battery pack also includes a Battery Management System (BMS) for managing and monitoring the battery pack, such as obtaining the total voltage of the battery pack, the highest single cell voltage, the lowest single cell voltage, the battery pack current, the highest temperature, the lowest temperature, the total charge of the battery pack, and the SOC of the battery pack at a certain sampling interval.

[0004] In actual applications, the battery pack will age during its service life. The aging of the battery pack is mainly manifested as its capacity attenuation and internal resistance increase, which are related to the service time, SOC, depth of discharge, charge-discharge rate, temperature, and other working conditions of the battery pack. To ensure the long-term, safe, and continuous use of the battery pack, it is necessary to accurately evaluate its State of Health (SOH).

[0005] The SOH of a battery pack is defined as the ratio of its current available capacity to the initial capacity, and can be used as a measure of the aging degree of the battery pack. For example, when the SOH of the battery pack drops to 80%, the battery pack reaches its first service life. Therefore, accurately estimating the SOH of the battery pack is crucial.

[0006] In the prior art, the SOH estimation of battery packs mainly includes battery model methods and data-driven methods, but the model method and the data-driven method have problems such as relying on working conditions, poor interpretability, and difficulty in balancing calculation accuracy, calculation efficiency, and robustness. Summary of the Invention

[0007] According to one embodiment, the present application provides a method for estimating the SOH of a battery pack, including: S1: obtaining the capacity tag value required to estimate the SOH of the battery pack based on the historical actual operation data of the battery pack, and constructing a sliding window input data matrix related to SOH; S2: automatically extracting the health feature map related to the aging of the battery pack through a neural network based on the sliding window input data matrix; S3: modeling the battery attenuation dynamics through a PINN neural network based on the extracted health feature map to realize the estimation of the battery SOH.

[0008] Further, step S1 further includes: S11: Obtain the historical actual operation data of the battery pack; S12: Extract the charging data set of the battery pack from the historical actual operation data, and divide the charging data set into multiple charging segments; S13: Calculate the capacity value of each charging segment, the capacity values of multiple charging segments form a capacity value sequence, and obtain the capacity label value of each charging segment; S14: Calculate the capacity increment value at each moment of each charging segment to obtain the capacity increment sequence of each charging segment; S15: Normalize the m variables related to the SOH value and the capacity increment value within each charging segment for normalization; S16: Use the sliding window mechanism to divide the m variables related to the SOH value and the capacity increment value within each charging segment after normalization , and obtain a plurality of n×m sliding window input data matrices under each charging segment, where n is the window size, m is the number of columns, representing the number of variable channels, and both n and m are natural numbers greater than or equal to 1.

[0009] Further, the historical actual operation data in step S11 includes the corresponding database of sampling time and the total voltage of the battery pack, the highest single cell voltage, the lowest single cell voltage, the battery pack current, the highest temperature, the lowest temperature, the cumulative charging power of the battery pack, and the SOC of the battery pack.

[0010] Further, step S12 includes: Determine the state of each frame of data according to the battery pack current and the change of the SOC of the battery pack in the historical actual operation data, and extract the charging data set of the battery pack from the historical actual operation data; For the charging data set, divide it into multiple charging segments according to the change continuity of the parameter representing the battery pack charge.

[0011] Further, step S12 further includes: Retain the charging segments with the lowest SOC less than the first threshold and the highest SOC greater than the second threshold.

[0012] Further, in step S13, according to the formula calculate the capacity value of each charging segment , where is the charging current of the current charging segment, is the first SOC value selected, is the second SOC value selected, is the difference between the second SOC value and the first SOC value.

[0013] Further, step S13 further includes: Based on the locally weighted linear regression algorithm, linearly fit the capacity value sequence with respect to the cumulative charging power of the battery pack to obtain a denoised capacity value sequence.

[0014] Further, according to the formula Obtain the capacity increment value at each moment of each charging segment, where is the sampling interval.

[0015] Furthermore, step S15 is: According to the formula Normalize the m variables related to the SOH value and the capacity increment value within each charging segment , where represents the data of the current normalization object dimension, and represent the maximum and minimum values of the current normalization object in the charging dataset of the battery pack.

[0016] Furthermore, S2 includes: S21: Input each sliding window input data matrix into the global feature extraction network, and output the global features optimized by channel attention and spatial attention; S22: Perform spatial graph feature extraction on the global features through the graph feature extraction network to obtain a health feature map related to the aging of the battery pack.

[0017] Furthermore, step S22 is: Perform spatial graph feature extraction on the global features through the GAT+GCN model to obtain a health feature map related to the aging of the battery pack.

[0018] Furthermore, S3 includes: S31: Input the health feature map into a multi-layer fully connected neural network and output the SOH estimated value; S32: Construct a PINN neural network such that the partial derivative of the function defined by the multi-layer fully connected neural network with respect to time is equal to the output of the PINN neural network; S33: Train the global feature extraction network, the graph feature extraction network, the multi-layer fully connected neural network, and the PINN neural network.

[0019] Furthermore, the input of the PINN neural network is the SOH estimated value , the partial derivative of the SOH estimated value with respect to time , the partial derivative of the SOH estimated value with respect to the health feature map , the health feature map and time , and the output of the PINN neural network is the decay rate of the SOH. The PINN neural network makes the partial derivative of the function defined by the multi-layer fully connected neural network with respect to time equal to the decay rate of the SOH.

[0020] The features and technical advantages of the present disclosure have been outlined quite extensively above so that the following detailed description of the disclosure can be better understood. Additional features and advantages of the present disclosure will be described below, which form the subject matter of the claims of the present disclosure. Those skilled in the art should understand that the disclosed concepts and specific embodiments can be easily used as a basis for modifying or designing other structures or processes for achieving the same purpose of the present disclosure. Those skilled in the art should also recognize that such equivalent structures do not depart from the spirit and scope of the present disclosure as set forth in the appended claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] To more fully understand the present disclosure and its advantages, reference is now made to the following description taken in conjunction with the accompanying drawings, in which: Figure 1 FIG. shows a schematic flow chart of a method for estimating the SOH of a battery pack according to an embodiment of the present application; Figure 2 FIG. shows a schematic flow chart of a data processing stage according to an embodiment of the present application; Figure 3 FIG. shows a schematic flow chart of dividing a charging segment according to an embodiment of the present application; Figure 4 FIG. shows a schematic diagram of the principle for estimating the SOH of a battery pack according to an embodiment of the present application.

[0022] Unless otherwise specified, corresponding numbers and symbols in different drawings generally refer to corresponding parts. These drawings are drawn to clearly illustrate the relevant aspects of various embodiments and are not necessarily drawn to scale. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] The technical solutions in the present application will be clearly and completely described below with reference to the accompanying drawings. Apparently, the described embodiments are some but not all of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without making creative efforts shall fall within the scope of protection of the present application.

[0024] An embodiment of the present application lies in providing a method for estimating the SOH of a battery pack. Please refer to Figure 1 the schematic flow chart of the method for estimating the SOH of a battery pack according to an embodiment of the present application shown in FIG.. The method for estimating the SOH of a battery pack proposed by the present application includes: S1: Obtain the capacity tag value required to estimate the SOH of the battery pack based on the historical actual operation data of the battery pack, and construct a sliding window input data matrix related to the SOH; S2: Based on the sliding window input data matrix, automatically extract a health feature map related to the aging of the battery pack through a neural network; S3: Based on the extracted health feature map, use the PINN neural network to model the battery degradation kinetics to estimate the battery SOH.

[0025] Please refer to Figure 2 the schematic diagram of the data processing stage flow of an embodiment of the present application shown in the figure. Specifically, step S1 includes: S11: Obtain the historical actual operation data of the battery pack; S12: Extract the charging data set of the battery pack from the historical actual operation data, and divide the charging data set into multiple charging segments; S13: Calculate the capacity value of each charging segment, and the capacity values of multiple charging segments form a capacity value sequence, and obtain the capacity label value of each charging segment; S14: Calculate the capacity increment value at each moment of each charging segment to obtain the capacity increment sequence of each charging segment; S15: Normalize the m variables related to the SOH value and the capacity increment value within each charging segment; S16: Use the sliding window mechanism to divide the m variables related to the SOH value and the capacity increment value within each charging segment after normalization , to obtain multiple n×m sliding window input data matrices under each charging segment, where n is the window size, m is the number of columns, representing the number of variable channels, and n and m are natural numbers greater than or equal to 1.

[0026] For step S11, specifically, during the actual operation of the battery pack, the battery pack total voltage, the highest single cell voltage, the lowest single cell voltage, the battery pack current, the highest temperature, the lowest temperature, the battery pack cumulative charge and the battery pack SOC, etc. are obtained at a certain sampling interval (such as 10s, but not limited to this, which can be determined according to different application scenarios). Therefore, the historical actual operation data in step S1 is the actual data formed in the real world of the battery pack during historical operation. For example, the historical actual operation data includes the corresponding database of sampling time and the battery pack total voltage, the highest single cell voltage, the lowest single cell voltage, the battery pack current, the highest temperature, the lowest temperature, the battery pack cumulative charge and the battery pack SOC. The historical actual operation data includes the battery pack cumulative charge, as well as the highest single cell voltage, the lowest single cell voltage, the highest temperature and the lowest temperature reflecting the inconsistency of the battery cells, which can improve the accuracy of the SOH estimation of the battery pack.

[0027] Furthermore, in step S11, when there are outliers or missing values in the historical actual operation data, correct the historical actual operation data. Specifically, fill in the missing values by linear interpolation when the number of missing time frames is small, the SOC is monotonic or constant, or the current is constant to correct the historical actual operation data; when the number of missing time frames is large (such as the data missing time is greater than 50s) and the missing segment is in the middle of the charging process, discard the charging cycle data to correct the historical actual operation data.

[0028] For step S12, specifically, it includes: determining the state of each frame of data according to the battery pack current and the change of the battery pack SOC in the historical actual operation data, and extracting the charging data set of the battery pack from the historical actual operation data; then, for the charging data set, different charging cycles are divided according to the continuity of the change of the parameter (such as SOC) representing the battery pack charge, and one charging cycle can be called a charging segment. For example, if the SOC of the battery pack continuously increases during the period from time t1 to time t2 and then decreases, it is considered that the period from time t1 to time t2 is a charging segment. Specifically, please refer to Figure 3 the flow schematic diagram of dividing charging segments in an embodiment of the present application shown in Figure 3 As shown, dividing the charging data set into multiple charging segments includes: S121: receiving the charging data set; S122: defining start = 1, t = 1; S123: judging whether the SOCt at the previous moment is greater than the SOCt+1 at the current moment. If yes, enter S124; if not, enter S126; S124: dividing the charging data from the start moment to the t moment in the charging data set into a charging segment; S125: updating start = t + 1 and entering S127; S126: updating t = t + 1 and entering step S127; S127: judging whether t + 1 is greater than the length of the charging data set. If so, end; if not, enter step S123. In this way, if the SOCt at the previous moment is greater than the SOCt+1 at the current moment, for example, if the SOC suddenly drops from 90% to 10%, then the current charging cycle ends, and the charging data from the start moment (start) to the current moment t can be divided into a charging segment, and the start moment (start) is updated to t + 1, and continue to judge the charging data at subsequent moments; if the SOCt at the previous moment is less than the SOCt+1 at the current moment, for example, if the SOC increases from 88% to 90%, then the current charging cycle is still in progress, update the moment t to t + 1, and continue to judge the charging data at subsequent moments until the last moment of the charging data set. In this way, the charging data set is divided into multiple charging segments.

[0029] In actual operation, furthermore, step S12 further includes: screening charging segments. In order to cover as many charging segments as possible while ensuring that each charging segment includes the non - plateau region of the battery pack voltage, specifically: retaining the charging segments with the lowest SOC less than the first threshold (such as 40%) and the highest SOC greater than the second threshold (such as 70%), and removing the charging segments that do not meet this condition, which are considered inappropriate charging segments.

[0030] For step S13, specifically, according to the formula calculate the capacity value of each charging segment , where is the charging current of the current charging segment, is the selected first SOC value, is the selected second SOC value. Generally, it is desired that the voltage platform region of the battery pack is included from the first SOC value to the second SOC value, and a part of the non-voltage platform region is also included. For example, in one embodiment, the first SOC value can be taken as 40% and the second SOC value as 70%. is the difference between the second SOC value and the first SOC value.

[0031] In actual operation, further, step S13 further includes: capacity value sequence noise processing. Specifically: linearly fitting the capacity value sequence with respect to the cumulative charge of the battery pack based on the locally weighted linear regression algorithm to obtain the denoised capacity value sequence. This is because the cumulative cycle count cannot accurately reflect the usage degree of the battery pack.

[0032] In practical applications, the denoised capacity value sequence can be used as the label value for subsequent model training of estimating SOC. The capacity value sequence obtained in step S13 can also be directly used as the label value for subsequent model training of estimating SOC.

[0033] For step S14, specifically, according to the formula the capacity increment value at each moment of each charging segment is obtained, where is the sampling interval. Among them, at the start time of , . The capacity increment can represent the characteristics of the battery pack.

[0034] For step S15, specifically, according to the formula normalize the m variables related to the SOH value within each charging segment, such as the charging current , the total voltage of the battery pack , the highest single-cell voltage , the lowest single-cell voltage , the highest temperature , the lowest temperature , and the capacity increment value . Among them, represents the data of the dimension of the current normalization object, and represent the maximum and minimum values of the current normalization object in the charging dataset of the battery pack. For example, when normalizing the charging current , represents the maximum value of the charging current in the charging dataset of the battery pack, represents the minimum value of the charging current in the charging dataset of the battery pack.

[0035] For step S16, specifically, in order to enhance the highly relevant part of the data reading information and better capture the temporal correlation of the data, a sliding window mechanism is introduced. Taking 7 variables related to the SOH value as the charging current , the total voltage of the battery pack , the highest single cell voltage , the lowest single cell voltage , the highest temperature , the lowest temperature as an example, in an embodiment of actual application, the window size n can take an initial value of 500, then a sliding window input data matrix can be expressed as . The number of windows can take an initial value of 3, that is, 3 sliding window input data matrices of 500×7 are taken under each charging segment. It can be seen that 500 is the window size, that is, the number of rows of the sliding window input data matrix, 7 is the number of variable channels, and one column represents one variable channel. In a subsequent embodiment of actual application, the trainable verification pair of parameters can be optimized based on grid search within the range of window sizes {300, 400, 500} and window numbers {3, 5, 7}, and the validation set loss is used as the evaluation index.

[0036] Thus, for an embodiment of the present application, the sliding window input data matrix in step S1 is the sliding window input data matrix formed in step S16. Taking the number of variable channels of the sliding window input data matrix of the present application as 7 as an example, of course, the number of variables can also be increased or decreased according to different requirements or application scenarios, and different variables can be adopted.

[0037] Where step S2 includes: S21: Input each sliding window input data matrix into the global feature extraction network to output the global features optimized by channel attention and spatial attention; S22: Perform spatial graph feature extraction on the global features through the graph feature extraction network to obtain a health feature map related to the aging of the battery pack.

[0038] In an embodiment, the global feature extraction network in step S21 is a CNN-LSTM+CBAM hybrid module. Specifically, please refer to Figure 4 the schematic diagram of the SOH estimation principle of the battery pack in an embodiment of the present application shown. In actual application, the importance of each parameter in the sliding window input data matrix for the final SOH is different. Therefore, each sliding window input data matrix Input into the CNN-LSTM+CBAM hybrid module, which is a deep learning hybrid model that combines a Convolutional Neural Network (CNN), Long Short-Term Memory Network (LSTM), and Convolutional Block Attention Module (CBAM). Its task is to obtain weights, and through the deep learning of this CNN-LSTM+CBAM hybrid module, global features corresponding to each sliding window input data matrix are obtained. The global feature is a weight matrix with the same dimensions (number of rows and columns) as the sliding window input data matrix, and its value ranges from 0 to 1. In this way, features with high correlation with SOH in the sliding window input data matrix can be enhanced, and variable channels with high correlation with SOH in the sliding window input data matrix can also be enhanced. At the same time, features with low correlation with SOH in the sliding window input data matrix are weakened, and variable channels with low correlation with SOH in the sliding window input data matrix are also weakened. In this way, through the CNN-LSTM+CBAM hybrid module, the attention can be focused on features with high correlation with SOH, and features with low correlation with SOH are suppressed.

[0039] For step S22, as Figure 4 shown, the global features are subjected to spatial graph feature extraction through the GAT+GCN model to obtain a health feature map related to the aging of the battery pack. The GAT+GCN model combines the Graph Attention Network (GAT) and the Graph Convolutional Network (GCN). Specifically, the Graph Attention Network (GAT) uses each variable channel of the global feature as a graph node, and each node is connected to its nearest neighbor nodes to form edges, and an adjacency matrix is constructed based on the information of the nodes and edges , and the calculation formula is:

[0040] where the node is the connection relationship between node and node , is the node in the nearest neighborhood of node . And through the attention mechanism of GAT, the updated node is obtained, and the formula is: where is the global feature (i.e., the weight matrix), and are the node features in the global feature, is the updated node feature, representing each node and its attention weight coefficient for each neighbor node When equals it represents the attention weight coefficient of the node to itself. The formula is:

[0041] where is a non - linear activation function, is the weight vector used to represent attention, is the concatenation operation, is the node 's neighbor. That is, GAT introduces an attention mechanism, enabling the model to automatically learn the importance weights of different neighbor nodes for the current node, thus more effectively capturing the complex relationships between nodes. The output features of GAT are continuously input into the Graph Convolutional Network (GCN) to obtain the final health feature map related to the aging of the battery pack. The health feature map can be expressed as , where is the node set, is the edge set connecting any two nodes, is the adjacency matrix.

[0042] In this way, the data processing, cleaning, and filtering are completed, and the health feature map related to the aging of the battery pack is extracted using neural network layers.

[0043] Step S3 includes: S31: Input the health feature map into a multi - layer fully connected neural network to output the SOH estimation value; S32: Construct a PINN neural network such that the partial derivative of the function defined by the multi - layer fully connected neural network with respect to time is equal to the output of the PINN neural network; S33: Train the global feature extraction network, the graph feature extraction network, the multi - layer fully connected neural network, and the PINN neural network.

[0044] For step S31, specifically, based on the health feature map extracted and fused by CNN - LSTM + CBAM + GAT + GCN, the multi - layer fully connected neural network (Fully Connected Neural Network, abbreviated as FCNN or FC) learns to extract the non - linear relationship between the health feature map (i.e., aging feature) and SOH, and outputs the SOH estimation value.

[0045] The multi - layer fully connected neural network (FC) can be defined as having learnable parameters of , used to characterize the battery aging model , where the battery aging trajectory is modeled as: Where, represents time, represents the extracted health feature map mentioned above.

[0046] Physics-Informed Neural Networks (PINN) is a new type of machine learning model that combines deep learning and physical laws. It embeds physical laws (usually represented in the form of partial differential equations or ordinary differential equations) into the loss function of the neural network, enabling the network to not only rely on data but also satisfy physical laws during training. For step S32, specifically, the input of the PINN neural network is the SOH estimate , the partial derivative of the SOH estimate with respect to time , the partial derivative of the SOH estimate with respect to the health feature map , the health feature map and time . The output of the PINN neural network is the decay rate of SOH. The PINN neural network makes the partial derivative of the function defined by the multi-layer fully connected neural network (FC) with respect to time equal to the decay rate of SOH. Specifically, the decay rate of SOH is expressed as , is the battery degradation dynamic equation of the PINN neural network, are the learnable parameters of the PINN neural network. Learn the battery aging mechanism from the given battery decay aging dataset to explicitly construct the non-linear relationship between battery SOH and various influencing factors. The partial derivative of the function defined by the multi-layer fully connected neural network (FC) with respect to time is characterizes the decay trajectory of the battery pack, that is . In this way, learn the decay kinetic characteristics of the battery pack through the established PINN neural network, and make the learning result closer to the actual situation of the battery pack to achieve the purpose of training the PINN neural network.

[0047] For step S33, specifically, according to the formula: Calculate the SOH label value for each charging segment of the battery pack, where represents the capacity label value for the charging segment, represents the rated capacity.

[0048] For characterizing the battery decay trajectory in And aging attenuation kinetics in Trained by minimizing the mean squared error loss, the loss function is: where, is the PDE loss to be followed in the PINN neural network optimization process The embodied PDE loss has the formula: where the subscript represents the charging segment, represents the health feature map under the charging segment.

[0049] is the data fitting loss, and the formula is:

[0050] where the subscript represents the charging segment, represents the number of charging segments, represents the label value under the charging segment, represents the estimated value under the charging segment.

[0051] is the monotonicity loss of the SOH trajectory, and the formula is: where the monotonicity loss is based on the physical characteristics of battery degradation, that is, the SOH of the next cycle should be less than or equal to the SOH of the previous cycle (unless capacity regeneration occurs), where represents the label value under the charging segment, represents the label value under the charging segment.

[0052] Thus, the SOH value output by the neural network is compared with the label value. If the loss is too large, the parameters and in the neural network will be updated by gradient so that the SOH value output by the neural network in each segment is closer to the label value.

[0053] The SOH estimation method provided in this application is based on the operating data of the battery pack in the real world. The basic data is cleaned and the capacity label value is calculated to obtain the basic input data of the neural network model. In order to make the neural network model have better working condition adaptability, based on the end-to-end idea, features are initially extracted through CNN-LSTM and the Convolutional Block Attention Module (CBAM), and then a multi-layer Graph Neural Network (GNN) is constructed to capture the complex node and edge relationships in the graph and autonomously learn and extract aging-related features. Finally, the previous network and the physics-informed neural network are fused to train the model to achieve the SOH estimation of the battery pack. It can be seen that the method provided in this application does not rely on the constant current charging working condition requirements, and at the same time fully combines the advantages of the battery pack physical information model and the data-driven method to estimate the SOH of the battery pack, solving the imbalance problems of traditional model methods and data-driven algorithms in terms of working condition adaptability, calculation efficiency, and calculation accuracy.

[0054] The above battery pack can be a lithium iron phosphate battery, a ternary lithium battery, etc., and this application does not make any limitations in this regard.

[0055] The above battery pack can be applied in any occasion, such as actual usage scenarios of lithium-ion batteries in new energy vehicles, photovoltaic energy storage, and wind power energy storage.

[0056] In practical applications, the SOH estimation method of the battery pack provided in this application is executed by a cloud computing platform. The memory and computing efficiency of the cloud computing platform are relatively high, which is more suitable for model training of large batches of data. In practical applications, during the operation of the battery pack, the battery management system collects the operating data of the battery pack. Therefore, the cloud computing platform obtains the above historical actual operating data from the battery management system of the battery pack. In practical applications, the model trained by the cloud computing platform can also be transmitted to the battery management system, and the battery management system obtains the SOH estimation value of the battery pack.

[0057] Although the embodiments of the present disclosure and their advantages have been described in detail, it should be understood that various changes, substitutions, and alterations can be made herein without departing from the spirit and scope of the present disclosure as defined by the appended claims.

[0058] Moreover, the scope of the present application is not intended to be limited to the specific embodiments of the processes, machines, manufactures, compositions of matter, devices, methods, and steps described in the specification. As will be readily understood by one of ordinary skill in the art from the disclosure of the present disclosure, processes, machines, manufactures, compositions of matter, means, methods, or steps that perform substantially the same function, whether currently existing or developed or implemented in the future, produce substantially the same results as the corresponding embodiments described herein that are available according to the present disclosure. Accordingly, the appended claims are intended to include within their scope such processes, machines, manufactures, compositions of matter, devices, methods, or steps.

Claims

1. A method for estimating the SOH of a battery pack, characterized in that, Including: S1: Obtain the capacity label values required to estimate the SOH of the battery pack based on the historical actual operation data of the battery pack, and construct a sliding window input data matrix related to the SOH; S2: Based on the sliding window input data matrix, automatically extract the health feature map related to the aging of the battery pack through a neural network; S3: Based on the extracted health feature map, model the battery decay kinetics through the PINN neural network to realize the estimation of the battery SOH.

2. The method for estimating the SOH of the battery pack according to claim 1, wherein Step S1 further includes: S11: Obtain the historical actual operation data of the battery pack; S12: Extract the charging data set of the battery pack from the historical actual operation data, and divide the charging data set into multiple charging segments; S13: Calculate the capacity value of each charging segment, and the capacity values of multiple charging segments form a capacity value sequence, and obtain the capacity label value of each charging segment; S14: Calculate the capacity increment value at each moment of each charging segment to obtain the capacity increment sequence of each charging segment; S15: Normalize the m variables related to the SOH value and the capacity increment value within each charging segment ; S16: Divide the m variables related to the SOH value and the capacity increment value within each normalized charging segment by using the sliding window mechanism , and obtain multiple n×m sliding window input data matrices under each charging segment, where n is the window size, m is the number of columns, representing the number of variable channels, and both n and m are natural numbers greater than or equal to 1.

3. The method for estimating the SOH of the battery pack according to claim 2, characterized in that, The historical actual operation data in step S11 includes the corresponding database of the sampling time and the total voltage of the battery pack, the highest single-cell voltage, the lowest single-cell voltage, the battery pack current, the highest temperature, the lowest temperature, the total charging power of the battery pack, and the SOC of the battery pack.

4. The method for estimating the SOH of the battery pack according to claim 3, wherein Step S12 includes: Determine the state of each frame of data according to the battery pack current and the change of the SOC of the battery pack in the historical actual operation data, and extract the charging data set of the battery pack from the historical actual operation data; For the charging data set, divide it into multiple charging segments according to the change continuity of the parameters representing the state of charge of the battery pack.

5. The method for estimating the SOH of the battery pack according to claim 4, characterized in that, Step S12 further includes: retaining the charging segments with the lowest SOC less than the first threshold and the highest SOC greater than the second threshold.

6. The method for estimating the SOH of the battery pack according to claim 4, wherein, In step S13, according to the formula calculate the capacity value of each charging segment , where is the charging current of the current charging segment, is the selected first SOC value, is the selected second SOC value, is the difference between the second SOC value and the first SOC value.

7. The method for estimating the SOH of the battery pack according to claim 6, characterized in that, Step S13 further includes: linearly fitting the capacity value sequence with respect to the total charging power of the battery pack based on the locally weighted linear regression algorithm to obtain the denoised capacity value sequence.

8. The method for estimating the SOH of the battery pack according to claim 7, wherein According to the formula obtain the capacity increment value at each moment of each charging segment, where is the sampling interval.

9. The method for estimating the SOH of the battery pack according to claim 6, wherein Step S15 is: According to the formula normalize the m variables related to the SOH value and the capacity increment value in each charging segment, where represents the data of the dimension of the current normalization object, and represent the maximum and minimum values of the current normalization object in the charging dataset of the battery pack.

10. The method for estimating the SOH of the battery pack according to claim 1, wherein S2 Including: S21: Input each sliding window input data matrix into the global feature extraction network, and output the global features optimized by channel attention and spatial attention; S22: Perform spatial graph feature extraction on the global features through the graph feature extraction network to obtain the health feature map related to the aging of the battery pack.

11. The method for estimating the SOH of the battery pack according to claim 10, wherein, Step S22 is: perform spatial graph feature extraction on the global features through the GAT+GCN model to obtain the health feature map related to the aging of the battery pack.

12. The method for estimating the SOH of the battery pack according to claim 10, wherein S3 Including: S31: Input the health feature map into a multi-layer fully connected neural network, and output the SOH estimation value; S32: Construct a PINN neural network such that the partial derivative of the function defined by the multi-layer fully connected neural network with respect to time is equal to the output of the PINN neural network; S33: Train the global feature extraction network, the graph feature extraction network, the multi-layer fully connected neural network, and the PINN neural network.

13. The method for estimating the SOH of the battery pack according to claim 12, characterized in that, The input of the PINN neural network is the SOH estimation value , the partial derivative of the SOH estimation value with respect to time , the partial derivative of the SOH estimation value with respect to the health feature map , the health feature map and time , the output of the PINN neural network is the attenuation rate of SOH, and the PINN neural network makes the partial derivative of the function defined by the multi-layer fully connected neural network with respect to time equal to the attenuation rate of SOH.

Images