A method for estimating the health status of lithium-ion batteries under fast charging

Abstract

The present invention discloses a method for estimating the health state of a lithium-ion battery under fast charging, which relates to the technical field of lithium batteries, including: collecting short-time charging voltage data fragments containing current switching points, and processing them into short-time charging voltage time series data; converting the short-time charging voltage time series data into graph structure data; inputting the graph structure data into a GCN-LSTM model constructed and trained based on a graph convolutional network and a long short-term memory network, using the graph convolutional network to capture local features and topological relationships in the battery aging process from the graph structure data, using the long short-term memory network to extract and learn the time series evolution law of the battery aging features from the local features and topological relationships in the battery aging process, and estimating the health state of the battery according to the key local and time series features of the battery in the entire aging cycle. The present invention improves the quality of feature data and the generalization ability between different fast charging strategies, and achieves a balance between high estimation accuracy and low data acquisition difficulty.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium batteries, and in particular to a method for estimating the health state of a lithium-ion battery under rapid charging based on short charging data. Background Art

[0002] Lithium-ion batteries are widely used in electric vehicles, rechargeable consumer electronics, aerospace, energy storage systems, and wearable devices due to their advantages such as high energy density, long life, and low self-discharge rate. However, batteries need to be charged regularly during the cycle to maintain energy supply. In order to alleviate the inconvenience caused by the extended charging time, many battery fast charging strategies have been proposed and applied. However, although the fast charging strategy can shorten the charging time, it also accelerates the battery degradation process, significantly shortens the battery life, and increases the risk of short circuits in the battery. Therefore, it is urgent to develop a high-precision health state estimation method for lithium-ion batteries during the fast charging process.

[0003] In order to shorten the battery charging time, many fast charging strategies have been proposed, including pulse charging, constant current charging, and multi-stage constant current and constant voltage charging. Among them, the multi-stage constant current and constant voltage strategy can maximize the charging rate without exceeding the safety limit of the battery, and is therefore widely used. However, many current battery health estimation methods are only oriented to low-rate constant current and constant voltage charging strategies, and cannot meet the requirements of high-rate fast charging strategies.

[0004] Methods for evaluating battery health mainly include quantitative calculation-based, model-based, and data-driven methods. However, due to the highly coupled electrochemical reactions inside the battery, its attenuation process is difficult to calculate and quantify. In addition, as the battery ages, the changes in the electrochemical model parameters will change, resulting in large errors in the estimation results. In contrast, data-driven methods are widely used because they do not require the construction of physical models. Data-driven methods require the development of advanced algorithms and feature factor extraction methods for specific applications, so as to achieve the mapping between feature factors and health status through data analysis. Existing data-driven methods often require manual feature extraction based on experience, which inevitably has subjectivity and reduces applicability in other data. Battery aging is a typical time series process, and its accurate estimation requires in-depth analysis of the time series features and local features in the aging data. However, existing data-driven methods have not yet fully explored the key features of the battery aging process. In order to improve the accuracy of battery health state estimation, it is necessary to adopt advanced algorithms that are more suitable for processing time series data to more effectively identify and utilize these key features. Summary of the invention

[0005] The present invention aims to provide a method for estimating the health state of a lithium-ion battery under rapid charging based on short-charge data, so as to solve the limitations of existing health estimation methods, such as difficulty in applying to high-rate rapid charging scenarios, large data requirements, subjective feature extraction and low estimation accuracy.

[0006] The technical solution adopted by the present invention is as follows:

[0007] The present invention provides a method for estimating the health state of a lithium-ion battery under rapid charging, comprising:

[0008] Collecting short-time charging voltage data segments of the lithium-ion battery including current switching points, and processing the short-time charging voltage data segments into short-time charging voltage time series data based on an interpolation method;

[0009] Convert the short-time charging voltage time series data into graph structure data, wherein the voltage values ​​and corresponding time points in the short-time charging voltage time series data are represented as nodes and edges of the graph respectively;

[0010] The graph structure data is input into the GCN-LSTM model built and trained based on the graph convolutional network and the long short-term memory network. The graph convolutional network is used to capture the local features and topological relationships in the battery aging process from the graph structure data. The long short-term memory network is used to extract and learn the temporal evolution of the battery aging characteristics from the local features and topological relationships in the battery aging process, and the key local and temporal features of the battery in the entire aging cycle are obtained. Then, the health state of the battery is estimated based on the key local and temporal features of the battery in the entire aging cycle.

[0011] As a further description of the above technical solution: the short-time charging voltage data segment is a voltage segment of four minutes.

[0012] As a further description of the above technical solution: based on the linear interpolation method, the short-time charging voltage data fragments are processed into short-time charging voltage time series data, and the formula is:

[0013]

[0014] Where j is the voltage point number; v k and v k+1 They are t k and t k+1 The voltage value at the moment; △t is the time interval; t' j is the jth interpolation time point, v' j is the jth interpolation voltage.

[0015] As a further description of the above technical solution: the method for converting short-time charging voltage time series data into graph structure data includes:

[0016] The short-time charging voltage time series data is evenly divided into five segments, each of which represents the feature vector of a node in the graph;

[0017] The nodes in the graph are interconnected according to the similarity between the nodes. The interconnected edges between the nodes represent the correlation between the nodes. The calculation formula is:

[0018]

[0019] Among them, S pq represents the similarity between the pth and qth nodes; x p and x q are the feature vectors of the pth and qth nodes respectively; n is the total number of nodes;

[0020] The similarities of all nodes are normalized to a uniform scale; the adjacency matrix A of the graph is expressed as:

[0021]

[0022] Among them, A pq represents the element in the p-th row and q-th column of A.

[0023] As a further description of the above technical solution: In the kth convolutional layer of the graph convolutional network, an aggregation function is first used to aggregate information from adjacent nodes of the graph structure data, and then the aggregated information is integrated and updated with the current state of the central node. The formula is as follows:

[0024]

[0025] Among them, f a and f c are the aggregation function and update function respectively, N(i) is the node adjacent to i, and σ is the activation function.

[0026] The propagation rule of the kth convolutional layer is explained as:

[0027]

[0028] Among them, Z k is the node layer output vector, ω k is the weight matrix;

[0029] Long short-term memory network is based on Z k As input, the temporal characteristics of node features are captured.

[0030] As a further description of the above technical solution: battery capacity is used as the label value of health state estimation. When obtaining the label value of the sample for training the GCN-LSTM model, the following modified ampere-hour integral formula is used to calculate the battery capacity data:

[0031]

[0032] Among them, C a is the calculated battery capacity, Δt is the sampling interval, I is the charging current, t1 and t2 are the charging start and end times, respectively, SOC t1 It is the SOC data at the start of charging. t2 The SOC data for charging termination.

[0033] As a further description of the above technical solution: During the training process of the GCN-LSTM model, the model structure and hyperparameters are adjusted by a cross-validation method.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] The Graph Convolution Network (GCN) and Long Short-Term Memory (LSTM) networks are used to automatically capture the key local features and timing features of the battery degradation process, significantly improving the estimation accuracy of the battery health status.

[0036] A voltage segment extraction method suitable for the battery fast charging process is adopted to improve the quality of feature data and the generalization ability between different fast charging strategies;

[0037] Converting time series data into graph structure data can effectively preserve the intrinsic relationship between battery data and enrich the connection relationship and dependency between data;

[0038] A high-precision estimate of the battery health status can be achieved by using short-time (for example, set to 4 minutes) charging data, achieving a balance between high estimation accuracy and low data collection difficulty.

[0039] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the embodiments of the present invention are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0041] Figure 1 This is a flow chart of a method for estimating a battery health state based on a GCN-LSTM model in an embodiment;

[0042] Figure 2 It is the feature sequence extraction process under the 3.6C (30%)-6C (80%)-1C (100%) fast charging strategy in the embodiment, wherein (a) shows the voltage and current curves under different aging degrees, (b) shows the selection of voltage segments according to the current switching point, and (c) shows the extraction of voltage sequences using the interpolation method;

[0043] Figure 3 : This is the structure of the GCN-LSTM estimation model in the embodiment.

[0044] Figure 4 1 is a diagram of the SOH estimation result of the GCN+LSTM model in the embodiment, wherein (a) is the SOH estimation result and (b) is the error distribution. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0046] Please refer to Figure 1 This embodiment provides a method for estimating the health status of a lithium-ion battery under fast charging based on short-charge data, which is as follows:

[0047] 1. Collect short-time charging voltage data fragments containing current switching points of lithium-ion batteries, and process the short-time charging voltage data fragments into short-time charging voltage time series data based on the interpolation method.

[0048] The multi-stage constant current and constant voltage fast charging strategy can control the temperature rise and internal pressure of the battery while accelerating charging by setting charging currents of different rates, effectively reducing battery aging and loss. The strategy is expressed as: C1(Q1)-C2(Q2)-…-Cj(100%). Among them, C and Q represent the charging current rate and charging capacity of the battery (i.e., the current switching point), respectively, and j represents the charging level. The charging information of the lithium-ion battery is collected using the charging equipment, which only requires time, voltage, current and SOC data, and the above charging data is recorded and stored;

[0049] The battery capacity is used as the label value for health status estimation. When obtaining the label value of the sample for training the GCN-LSTM model, the battery capacity data is calculated using the following modified ampere-hour integral formula:

[0050]

[0051] Among them, C ais the calculated battery capacity, Δt is the sampling interval, I is the charging current (charging is a negative number), t1 and t2 are the start and end times of charging, respectively, SOC t1 SOC data for the start of charging, SOC t2 The SOC data for charging termination.

[0052] In this embodiment, the short-time charging voltage data segment is a four-minute voltage segment, that is, only the four-minute voltage segment needs to be extracted according to the current switching point. Figure 2 shown.

[0053] The voltage and current curves under different aging degrees are as follows Figure 2 As shown in Figure (a), near the current switching point, the voltage curve changes significantly. This is because the change in current rate exacerbates the polarization effect and electrode reaction rate, directly affecting the concentration of active substances and lithium ions inside the battery. Therefore, a short-term charging voltage segment of four minutes is selected near the current switching point, as shown in Figure 1. Figure 2 Compared with the traditional method of extracting data based on voltage interval, this method not only effectively reduces the difficulty of data acquisition, but also can be applied to various multi-level constant current and constant voltage fast charging strategies.

[0054] Subsequently, short-time charging voltage time series data is generated by interpolation based on the extracted short-time charging voltage data segments, such as Figure 2 As shown in Figure (c).

[0055] In this embodiment, a linear interpolation method is used to generate short-time charging voltage time series data. i ,v i ),i∈[1,n]}, and use linear interpolation to generate an equally spaced voltage sequence with a time interval of △t:

[0056]

[0057] Where j is the voltage point number; v k and v k+1 They are t k and t k+1 The voltage value at the moment; △t is the time interval, which is 5 seconds in this embodiment; t' j is the jth interpolation time point, v' j is the jth interpolation voltage.

[0058] 2. Convert the short-time charging voltage time series data into graph structure data, wherein the voltage values ​​and corresponding time points in the short-time charging voltage time series data are represented as nodes and edges of the graph, respectively.

[0059] In order to construct graph-structured data, the characteristic matrix and adjacency matrix of the graph must be determined.

[0060] First, the voltage sequence obtained by interpolation (short-time charging voltage time series data) is evenly divided into five segments, each of which represents the characteristic vector of a node, such as Figure 2 (c) is shown. Then the nodes in the graph are interconnected according to the similarity between the nodes, so that the graph preserves the mutual information between the nodes. The interconnected edges between the nodes represent the correlation between the nodes, which is calculated by equation (3):

[0061]

[0062] Among them, S pq represents the similarity between the pth and qth nodes; x p and x q are the feature vectors of the pth and qth nodes respectively; n is the total number of nodes.

[0063] In order to solve the scale difference of edge weights between nodes, this embodiment uses standardization to standardize the similarity values ​​of all nodes to a unified scale. This makes them comparable and reduces the impact of scale differences to enhance the stability and convergence of the algorithm. The adjacency matrix A of the graph is expressed as:

[0064]

[0065] Among them, A pq Denotes the element in the pth row and qth column of A. The degree matrix of a graph is defined as the degree of each node (the number of connected nodes). D is used to represent the connectivity of each node and provide basic statistical information about the topology and properties of the graph.

[0066] 3. Input the graph structure data into the GCN-LSTM model built and trained based on the graph convolution network and the long short-term memory network, and use the graph convolution network to capture the local features and topological relationships in the battery aging process from the graph structure data; use the long short-term memory network to extract and learn the temporal evolution of the battery aging characteristics from the local features and topological relationships in the battery aging process, and obtain the key local and temporal features of the battery in the entire aging cycle; the LSTM layer predicts the features of each node (the key local and temporal features of the battery in the entire aging cycle) through linear fusion through the fully connected layer (FCL) to generate a global label for the graph structure data, which corresponds to the health status of the battery. This global label can be used to evaluate the current aging state of the battery and predict its remaining effective service life.

[0067] In this embodiment, GCN aggregates the information between connected nodes and fuses the relationship between edges to update the node status. This method effectively expands the receptive field of the node, thereby improving the expressiveness and learning efficiency of the model. The information aggregation mechanism allows each node to capture not only its own features, but also the features of its adjacent nodes, thereby enriching the connection relationships and dependencies between nodes in the graph structure data.

[0068] The structure of the GCN-LSTM estimation model is as follows Figure 3 As shown in the figure, the GCN-LSTM model is trained using historical data to optimize model parameters and improve prediction capabilities. The model structure and hyperparameters are adjusted through cross-validation methods to prevent overfitting and ensure that the model has good generalization performance.

[0069] GCN is a component module of the model. In the kth convolutional layer, an aggregation function is first used to aggregate information from adjacent nodes of the graph structure data, and then the aggregated information is integrated and updated with the current state of the central node. This process aims to update the representation of the node to automatically capture the key local features between nodes, as shown in equations (5) and (6):

[0070]

[0071] Among them, f a and f c are the aggregation function and the update function respectively, N(i) is the node adjacent to i, σ is the activation function, is the aggregate information obtained by node i from its neighboring nodes, is the feature vector of the jth node in the k-1th layer, is the feature vector of the i-th node in the k-th layer. Combining these operations, the propagation rule of the k-th graph convolution layer is explained as:

[0072]

[0073] Among them, Z k is the node layer output vector, ω k is the weight matrix, A is the adjacency matrix, X is the feature matrix, is the normalized adjacency matrix. At the same time, Z k As the input of the LSTM model, this setting uses the memory capacity of LSTM to effectively capture the temporal characteristics of node features and enhance the accuracy of the model SOH (health status) estimation. When the time step is t, the forget gate f of the LSTM model t , input gate i t , output gate o t and input unit status C t The definition is as follows:

[0074]

[0075] Among them, ω is the weight matrix of each gate; U is the weight matrix of the hidden state; h t-1 represents the hidden state at time t-1, and is the four bias vectors, b is the threshold, Z i is the feature vector of the input i-th node, tanh is the activation function; σ is the activation function. t It is the node feature under step size t (i.e., local and temporal aging features). In order to estimate SOH, a fully connected layer is used to linearly transform the hidden state of each node into the global label of the graph structure data (i.e., the health status of the battery) output.

[0076]

[0077] Among them, ω y and b y is the weight and threshold, h t is the feature vector of the node.

[0078] Accurate estimation of SOH is crucial to improving the safety and reliability of batteries. This embodiment uses the GCN-LSTM model to automatically capture the key local and temporal features of lithium ions during fast charging. Only 4 minutes of charging data is needed to achieve high-precision battery health status estimation, and it can be applied to various charging strategies. Compared with existing estimation methods, it further expands the estimation means of battery health status in high-rate fast charging scenarios, which can effectively guide battery health management and performance maintenance; it proposes an efficient charging fragment extraction method to avoid subjectivity in the data acquisition process, and greatly reduces the demand for data volume, achieving a balance between high estimation accuracy and low data collection difficulty; it innovatively combines two deep learning methods, graph convolutional neural networks and long short-term memory neural networks, to achieve automatic extraction of key local features and timing features in the battery degradation process, and has significant advantages in estimation accuracy; it can maintain extremely high estimation accuracy under various fast charging strategies, has excellent generalization ability, and can adapt to various complex battery usage scenarios.

[0079] The battery health status estimation results of this embodiment are compared with other existing methods. Figure 4 As shown in Table 1. Table 1 Statistical comparison of estimation accuracy of various methods

[0080]

[0081] The present invention extracts charging segments that are highly correlated with the battery health status by supervising the current switching points, and can carry out data analysis on sensitive areas of complex electrochemical changes inside the battery. The battery charging mode transitions at the current switching point, accompanied by significant battery polarization effects and rapidly changing electrode reaction rates. In this process, the utilization rate of active materials and lithium ion concentrations inside the battery will be directly affected, and therefore contain key battery degradation information. In addition, the current switching point is the most sensitive moment for changes in battery characteristics, which can not only reflect the battery's immediate health status information, but also characterize the trend of battery performance evolution. Compared with traditional voltage threshold-based extraction methods, this method can better capture mutations in electrochemical reactions inside the battery, providing deeper insights into electrochemical processes for subsequent health status estimation; this method can effectively avoid the problems caused by the subjectivity of voltage threshold setting under different working conditions, and can maintain good generalization capabilities in a variety of different data environments.

[0082] The present invention innovatively transforms traditional time series data into graph structured data based on data correlation, which not only preserves the intrinsic characteristics of the data, but also enhances the expression of complex relationships between data. Time series data often contains complex dependencies and dynamic characteristics, but existing methods often only focus on the linear patterns and surface statistical characteristics of time series data. Battery degradation is a highly coupled electrochemical process, and its evolution process shows significant nonlinearity. Therefore, it is necessary to enrich the multi-level relationships between data and dependencies on different time scales by constructing a graph structure. Graph structured data is very flexible and can achieve the expansion of small-scale data to large-scale data sets by expanding nodes and edges, thereby meeting the needs of data analysis tasks in a big data environment. In addition, graph structured data can intuitively represent the dependencies between nodes and achieve more accurate descriptions by setting the weights and types of edges. Graph structured data can use a variety of advanced graph analysis techniques to reveal deep patterns and relationships in the data, and allow the model to learn and generalize the relationship between nodes, so that the model maintains the accuracy of estimation when facing new data.

[0083] The present invention combines graph convolutional neural networks and long short-term memory networks to carry out battery health state estimation tasks, and can integrate the local feature extraction capabilities of GCN and the time series analysis capabilities of LSTM to solve the analysis problem of battery degradation, a complex, highly coupled and nonlinear electrochemical process. GCN can capture the local mutual influence in the electrochemical process inside the battery by applying convolution operations on the nodes of the graph. The specific process includes domain aggregation and feature fusion, which realizes the update of node information and the expansion of the receptive field, thereby reflecting the deep local characteristics of the data. LSTM effectively captures the hidden evolutionary trends in time series data through its gate structure, and avoids the gradient disappearance and explosion problems faced by traditional recurrent neural networks. The degradation phenomenon of the battery takes a long time to appear. LSTM can effectively utilize the rich historical data of the battery to capture the aging characteristics and decay rates of the battery at different stages of use, thereby realizing dynamic adjustment of the model and adapting to different usage and charging conditions. Combining the local feature capture capability of GCN and the time series analysis capability of LSTM, it can provide a more comprehensive monitoring of the changes in the battery health state, effectively increasing the estimation accuracy of the model. Compared with other traditional estimation methods, this method can more deeply learn the complex electrochemical reaction process of the battery, thereby better adapting to the needs of battery health management and maintenance.

[0084] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for estimating the health status of a lithium-ion battery under fast charging, characterized in that: include: Collecting short-time charging voltage data segments of the lithium-ion battery including current switching points, and processing the short-time charging voltage data segments into short-time charging voltage time series data based on an interpolation method; Convert the short-time charging voltage time series data into graph structure data, wherein the voltage values ​​and corresponding time points in the short-time charging voltage time series data are represented as nodes and edges of the graph respectively; Input the graph structure data into the GCN-LSTM model built and trained based on the graph convolution network and the long short-term memory network, use the graph convolution network to capture the local features and topological relationships in the battery aging process from the graph structure data, and use the long short-term memory network to extract and learn the temporal evolution law of the battery aging characteristics from the local features and topological relationships in the battery aging process, and obtain the key local and temporal features of the battery in the entire aging cycle. Then, estimate the health status of the battery based on the key local and temporal features of the battery in the entire aging cycle; The method of converting short-time charging voltage time series data into graph structure data includes: The short-time charging voltage time series data is evenly divided into five segments, each of which represents the feature vector of a node in the graph; The nodes in the graph are interconnected according to the similarity between the nodes. The interconnected edges between the nodes represent the correlation between the nodes. The calculation formula is: Among them, S pq represents the similarity between the pth and qth nodes; x p and x q are the feature vectors of the pth and qth nodes respectively; n is the total number of nodes; The similarities of all nodes are normalized to a uniform scale; the adjacency matrix A of the graph is expressed as: Among them, A pq represents the element in the p-th row and q-th column of A.

2. The method for estimating the health status of a lithium-ion battery under rapid charging according to claim 1, characterized in that: The short-time charging voltage data segment is a voltage segment of four minutes.

3. The method for estimating the health status of a lithium-ion battery under rapid charging according to claim 1, characterized in that: Based on the linear interpolation method, the short-time charging voltage data fragments are processed into short-time charging voltage time series data, and the formula is: Where j is the voltage point number; v k and v k+1 They are t k and t k+1 The voltage value at the moment; △t is the time interval; t′ j is the jth interpolation time point, v′ j is the jth interpolation voltage.

4. The method for estimating the health status of a lithium-ion battery under rapid charging according to claim 1, characterized in that: In the kth convolutional layer of the graph convolutional network, an aggregation function is first used to aggregate information from adjacent nodes of the graph structure data, and then the aggregated information is integrated and updated with the current state of the central node. The formula is as follows: Among them, f a and f c are the aggregation function and the update function respectively, N(i) is the node adjacent to i, σ is the activation function, is the aggregate information obtained by node i from its neighboring nodes, is the feature vector of the jth node in the k-1th layer, is the feature vector of the i-th node in the k-th layer; The propagation rule of the kth convolutional layer is explained as: Among them, Z k is the node layer output vector, ω k is the weight matrix, A is the adjacency matrix, X is the feature matrix, is the normalized adjacency matrix; Long short-term memory network is based on Z k As input, the temporal characteristics of node features are captured.

5. The method for estimating the health status of a lithium-ion battery under rapid charging according to claim 1, characterized in that: The battery capacity is used as the label value for health status estimation. When obtaining the label value of the sample for training the GCN-LSTM model, the battery capacity data is calculated using the following modified ampere-hour integral formula: Among them, C a is the calculated battery capacity, Δt is the sampling interval, I is the charging current, t1 and t2 are the charging start and end times, respectively, SOC t1 It is the SOC data at the start of charging. t2 The SOC data for charging termination.

6. The method for estimating the health status of a lithium-ion battery under rapid charging according to claim 1, characterized in that: During the training process of the GCN-LSTM model, the model structure and hyperparameters are adjusted through the cross-validation method.

Images