Battery health prediction method and device, equipment and medium

Abstract

The embodiment of the invention provides a battery health prediction method and device, equipment and a medium. The method comprises the steps of firstly obtaining a graph structure model of a battery pack; then, obtaining operation data of each single cell node in a plurality of historical cycle periods, extracting time sequence features of the single cell nodes based on the operation data, and generating corresponding node initial feature representation; then, on the basis of a graph structure model, a graph attention network is adopted to carry out spatial information fusion on the initial features of all the nodes, and updated node representation is generated; and finally, based on the updated node representation, outputting a health state prediction value of the single battery cell and a consistency index of the battery pack through a multi-task prediction head. By means of the method, accurate prediction of the health state of the battery cell and consistency evaluation of the interior of the battery pack are achieved, and therefore a comprehensive and accurate decision basis is provided for safety management and maintenance of a battery system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery health prediction, and in particular to a battery health prediction method, device, equipment and medium. BACKGROUND

[0002] Lithium-ion batteries, as key energy storage elements, have been widely used in consumer electronics, electric vehicles (EV) and grid energy storage systems. In these application scenarios, the performance of the battery will gradually decline with the cycle use and time, and its state of health (SOH) and remaining useful life (RUL) are directly related to the safety, energy efficiency management and maintenance cost of the equipment. Therefore, real-time and accurate prediction of the SOH and RUL of the battery is crucial for ensuring the safe operation of the equipment, optimizing the energy management strategy, and realizing condition-based maintenance (CBM). Currently, the battery management system (BMS) collects real-time multi-dimensional streaming data such as voltage, current and temperature through sensors, and these data contain key information about the aging state of the battery. However, the existing technology still has problems such as insufficient use of raw information, reliance on manual feature engineering, and low computational efficiency when processing these data, and there is an urgent need for an end-to-end solution that can directly use raw data and capture the dynamic interaction within the battery pack.

[0003] The existing technology usually predicts the battery health state based on traditional machine learning models. Specifically, health indicators such as equivalent time interval of voltage curve or incremental capacity analysis parameters are designed manually as input features to train machine learning models to predict the battery health state.

[0004] The existing technology relies heavily on the knowledge of experts in the field for feature extraction, and the feature design is highly subjective and lacks universality. Moreover, the model is difficult to adapt to different battery types or complex working conditions, and has poor generalization ability. SUMMARY

[0005] The battery health prediction method, device, equipment and medium provided by the embodiments of the present application solve the problem of subjective and non-universal feature design in the prior art, and the model is difficult to adapt to different battery types or complex working conditions, and has poor generalization ability.

[0006] In a first aspect, the embodiments of the present application provide a battery health prediction method, comprising:

[0007] Obtaining a graph structure model of a battery pack, wherein the graph structure model includes a plurality of single cell nodes, and if two single cell nodes have a physical connection relationship, the two single cell nodes are connected to form an edge;

[0008] obtain running data of each of the single battery cell nodes in a plurality of historical cycle periods, and extract time sequence features of the single battery cell nodes based on the running data to generate corresponding initial feature representations of the nodes;

[0009] based on the graph structure model, perform spatial information fusion on the initial features of each node by using a graph attention network to generate updated node representations;

[0010] based on the updated node representations, output a health state prediction value of the single battery cell and a consistency index of the battery pack through a multi-task prediction head.

[0011] In a possible implementation, the obtaining of the graph structure model of the battery pack comprises:

[0012] each single battery cell in the battery pack is defined as a node in the graph structure;

[0013] edges are established between corresponding nodes according to a physical connection relationship between the single battery cells, the physical connection relationship comprising at least one of a series connection relationship, a parallel connection relationship, and a thermal coupling relationship;

[0014] an adjacency matrix for representing a topology structure of the battery pack is constructed based on the nodes and the edges.

[0015] In a possible implementation, the obtaining of the running data of each of the single battery cell nodes in a plurality of historical cycle periods comprises:

[0016] sensor data of each single battery cell in a plurality of historical consecutive cycle periods is collected, the sensor data comprising at least one or more of a discharge capacity sequence, a charging voltage curve, and a discharge voltage curve;

[0017] wherein the charging voltage curve and / or the discharge voltage curve is processed by resampling to be uniform in a preset length to form a node input feature corresponding to each cycle period.

[0018] In a possible implementation, the extracting of the time sequence features of the single battery cell nodes based on the running data to generate corresponding initial feature representations of the nodes comprises:

[0019] the running data of each single battery cell in the plurality of historical cycle periods is input into a time sequence feature extraction network to learn a time-dependent relationship of the running data of the single battery cell with the cycle period;

[0020] a node initial feature vector of a fixed dimension is output as the initial feature representation of the corresponding single battery cell node;

[0021] The time sequence feature extraction network includes any one or a combination of a recurrent neural network, a long short-term memory network, a convolutional neural network, or a Transformer network.

[0022] In a possible implementation, the spatial information fusion of the initial features of each node by using the graph attention network to generate an updated node representation includes:

[0023] The initial features of the target monomer battery cell node and the initial features of the adjacent monomer battery cell nodes are linearly transformed to obtain corresponding transformed features.

[0024] An attention coefficient is calculated based on the transformed features of the target node and the transformed features of the adjacent nodes.

[0025] The attention coefficient is normalized to obtain an attention weight representing the importance of each neighbor node.

[0026] According to the attention weight, the linearly transformed features of each neighbor node are weighted and summed, and a nonlinear activation function is combined to generate an updated node representation of the target node.

[0027] In a possible implementation, the health state prediction value of the monomer battery cell and the consistency index of the battery pack are output by the multi-task prediction head, including:

[0028] The updated node representation of each monomer battery cell node is input into a first prediction subnetwork to output a health state prediction value of the corresponding monomer battery cell.

[0029] The updated node representations of all monomer battery cell nodes are subjected to a graph-level readout operation to generate a global graph representation vector, and the graph-level readout operation includes at least one of summation, averaging, or maximum.

[0030] The global graph representation vector is input into a second prediction subnetwork to output the consistency index of the battery pack, and the consistency index includes any one or more of a variance, a range, or a dispersion index of the health state within the battery pack.

[0031] In a possible implementation, the method further includes model training.

[0032] The model training process includes:

[0033] A training sample set is obtained, and the training sample set includes operation data of a sample battery pack, real labeled values of monomer battery cell health states, and real labeled values of battery pack consistency indexes.

[0034] construct a joint loss function, the joint loss function is a first loss term composed of mean square error based on all single cell state of health prediction value and its actual value, and a second loss term composed of mean square error based on consistency index prediction value of the battery pack and its actual value;

[0035] Based on the joint loss function, the parameters of the time series feature extraction network, the graph attention network and the multi-task prediction head are iteratively updated by back propagation until the preset convergence condition is met.

[0036] In a second aspect, the embodiments of the present application provide a battery health prediction device, comprising:

[0037] The acquisition module is configured to acquire a graph structure model of the battery pack, the graph structure model comprising a plurality of single cell nodes, and if two single cell nodes have a physical connection relationship, the two single cell nodes are connected to form an edge.

[0038] The first processing module is configured to acquire running data of each single cell node in a plurality of historical cycle periods, and extract time series features of the single cell nodes based on the running data to generate corresponding node initial feature representations.

[0039] The fusion module is configured to perform spatial information fusion on the initial features of each node based on the graph structure model by using a graph attention network to generate updated node representations.

[0040] The second processing module is configured to output the state of health prediction value of the single cell and the consistency index of the battery pack based on the updated node representations through a multi-task prediction head.

[0041] In a possible implementation, the first acquisition module is specifically configured to:

[0042] Each single cell in the battery pack is defined as a node in the graph structure.

[0043] According to the physical connection relationship between the single cells, edges are established between the corresponding nodes, and the physical connection relationship includes at least one of a series connection relationship, a parallel connection relationship and a thermal coupling relationship.

[0044] An adjacency matrix for representing the topology of the battery pack is constructed based on the nodes and the edges.

[0045] In a possible implementation, the first processing module is specifically configured to:

[0046] The sensor data of each single cell in a plurality of historical consecutive cycle periods is collected, and the sensor data includes at least one or more of a discharge capacity sequence, a charging voltage curve and a discharge voltage curve.

[0047] The charging voltage curve and / or the discharging voltage curve are resampled to be uniform in length to form a node input feature corresponding to each cycle period.

[0048] In a possible implementation, the first processing module is specifically configured to:

[0049] input the operation data of each single cell in the plurality of historical cycle periods into a time sequence feature extraction network to learn the time dependence of the single cell operation data with respect to the cycle period;

[0050] output a fixed-dimension node initial feature vector as an initial feature representation of the corresponding single cell node;

[0051] The time sequence feature extraction network includes any one or a combination of a recurrent neural network, a long short-term memory network, a convolutional neural network, or a Transformer network.

[0052] In a possible implementation, the fusion module is specifically configured to:

[0053] perform linear transformation on the initial feature of the target single cell node and the initial features of adjacent single cell nodes to obtain corresponding transformed features;

[0054] calculate an attention coefficient based on the transformed feature of the target node and the transformed features of the adjacent nodes;

[0055] perform normalization processing on the attention coefficient to obtain an attention weight representing the importance of each adjacent node;

[0056] According to the attention weight, perform weighted summation on the features of each adjacent node after linear transformation, and combine a nonlinear activation function to generate an updated node representation of the target node.

[0057] In a possible implementation, the second processing module is specifically configured to:

[0058] input the updated node representation of each single cell node into a first prediction subnetwork to output a health state prediction value of the corresponding single cell;

[0059] perform a graph-level readout operation on the updated node representations of all single cell nodes to generate a global graph representation vector, and the graph-level readout operation includes at least one of summation, averaging, or maximum value calculation;

[0060] input the global graph representation vector into a second prediction subnetwork to output a consistency index of the battery pack, and the consistency index includes any one or more of a variance, a range, or a dispersion index of the health states within the battery pack.

[0061] In a possible implementation, the second processing module is further configured to model training.

[0062] The model training process comprises:

[0063] obtaining a training sample set, the training sample set comprising operation data of a sample battery pack, real label values of single battery cell health states, and real label values of battery pack consistency indicators;

[0064] constructing a joint loss function, the joint loss function comprising a first loss term based on mean square errors between all single battery cell health state prediction values and real values thereof, and a second loss term based on a mean square error between the battery pack consistency indicator prediction value and a real value thereof;

[0065] iteratively updating parameters of the time series feature extraction network, the graph attention network, and the multi-task prediction head based on the joint loss function using back propagation until a preset convergence condition is met.

[0066] In a third aspect, an electronic device is provided, comprising: a memory, a processor;

[0067] The memory stores computer execution instructions.

[0068] The processor executes the computer execution instructions stored in the memory, so that the processor executes the first aspect and / or various possible implementation manners of the first aspect.

[0069] In a fourth aspect, a computer readable storage medium is provided, the computer readable storage medium stores computer execution instructions, and the computer execution instructions are executed by a processor to implement the first aspect and / or various possible implementation manners of the first aspect.

[0070] The battery health prediction method, device, equipment and medium provided by the embodiments of the application first model the battery pack as a graph structure to depict the physical connection and interaction relationship between single cells; on this basis, the operation data of each single cell in multiple historical cycle periods is obtained, and time-related features such as capacity attenuation and voltage evolution are automatically modeled through a time sequence feature extraction mechanism, so as to form a node initial feature representation that can accurately represent the aging state of the single cell; subsequently, a graph attention network is introduced based on the graph structure model, the initial features of each single cell node are weighted and aggregated, the influence weight of different adjacent cells on the target cell is adaptively learned, the effective fusion of spatial correlation information such as electrical coupling and thermal coupling is realized, and an updated node representation containing both single cell historical degradation information and intra-pack interaction influence information is obtained; finally, the updated node representation is input into a multi-task prediction head, on the one hand, the fine-grained prediction of the health state of each single cell is realized, and on the other hand, the overall index representing the consistency of the battery pack is output through the graph-level feature aggregation, so that the single cell health evaluation accuracy and the battery pack consistency description ability are simultaneously improved in a single model framework, and more comprehensive and reliable health state evaluation and operation decision basis is provided for the battery management system. BRIEF DESCRIPTION OF DRAWINGS

[0071] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate embodiments consistent with the present application and, together with the description, further serve to explain the principles of the application.

[0072] Figure 1 Flowchart of the battery health prediction method provided by the embodiments of the application Figure 1 ;

[0073] Figure 2 Flowchart of the battery health prediction method provided by the embodiments of the application Figure 2 ;

[0074] Figure 3 Structure diagram of the battery health prediction device provided by the embodiments of the application;

[0075] Figure 4 Structure diagram of the electronic device provided by the embodiments of the application.

[0076] Through the above drawings, the specific embodiments of the application have been shown, and will be described in more detail hereinafter. These drawings and textual descriptions are not intended to limit the scope of the concept of the application by any means, but to illustrate the concept of the application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0077] The exemplary embodiments will be described in detail herein with reference to the attached drawings. The following description is made with reference to the accompanying drawings in which like reference numerals refer to like elements, unless the context dictates otherwise. The following description of exemplary embodiments is not representative of all embodiments consistent with the present application. Rather, it is merely an example of apparatus and methods consistent with some aspects of the present application as detailed in the appended claims.

[0078] Lithium-ion batteries have been widely used in consumer electronics, electric vehicles, and grid energy storage systems as key energy storage components. In these application scenarios, the performance of the battery gradually deteriorates with cycling and time, and its SOH and RUL are directly related to the safety, energy efficiency management, and maintenance cost of the equipment. Therefore, real-time and accurate prediction of the SOH and RUL of the battery is crucial for ensuring the safe operation of the equipment, optimizing energy management strategies, and achieving state-based maintenance. Currently, the BMS collects real-time multi-dimensional streaming data such as voltage, current, and temperature through sensors, and these data contain key information about the aging state of the battery. However, existing technologies still have problems such as insufficient use of raw information, reliance on manual feature engineering, and low computational efficiency when processing these data, and there is an urgent need for an end-to-end solution that can directly use raw data and capture the dynamic interaction within the battery pack.

[0079] The existing technology usually predicts the battery health state based on traditional machine learning models, specifically: by manually designing health indicators such as equivalent time interval of voltage curve or incremental capacity analysis parameters as input features, training machine learning models to predict the battery health state.

[0080] The existing technology relies heavily on the knowledge of domain experts for feature extraction, and the feature design is highly subjective and lacks universality. In addition, the process of compressing high-dimensional sensor data into scalar features leads to information loss, and the model is difficult to adapt to different battery types or complex working conditions (such as dynamic load of electric vehicles), and has poor generalization ability.

[0081] Based on this, this application proposes a battery health prediction method. Addressing the problems of insufficient utilization of raw information, reliance on manual feature engineering, and low computational efficiency in existing technologies, the inventors propose a battery health prediction method based on the physical mechanism that individual cells within a battery pack exhibit independent temporal aging behavior and form a spatially interconnected network through electrical and thermal coupling. This method abstracts the battery pack into a graph structure model, using individual cells as nodes and electrical or thermal connections as edges. It introduces battery pack topology information without relying on manual rules and directly utilizes raw high-dimensional sensor data such as multi-cycle capacity and voltage curves. A time-series feature extraction network automatically learns the aging evolution characteristics of individual cells, avoiding information loss caused by artificial compression. Furthermore, a graph attention network is used to adaptively weight and fuse the spatial influence between individual cells, enabling the model to dynamically characterize the changes in the degree of influence between cells under different operating conditions and aging stages. Finally, a multi-task prediction mechanism simultaneously achieves individual cell health status prediction and battery pack consistency assessment, significantly improving the accuracy, robustness, and generalization ability across battery types and complex operating conditions of health prediction.

[0082] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0083] Figure 1 A flowchart illustrating the battery health prediction method provided in this application embodiment. Figure 1 ;like Figure 1 As shown, the method includes:

[0084] S101. Obtain the graph structure model of the battery pack.

[0085] In one possible approach, each individual cell in the battery pack is first defined as a node in a graph structure; then, edges are established between the corresponding nodes based on the physical connections between the individual cells; finally, an adjacency matrix is ​​constructed based on the nodes and edges to characterize the topology of the battery pack.

[0086] The graph structure model includes multiple individual cell nodes, and if two individual cell nodes have a physical connection, the two individual cell nodes are connected to form an edge; the physical connection relationship includes at least one of the following: series connection relationship, parallel connection relationship, and thermal coupling relationship.

[0087] It should be understood that in the present embodiment, in order to accurately depict the structural relationship and mutual influence mechanism between each single cell in the battery pack, the battery pack is first abstracted into a graph structure model. Specifically, first, a battery pack containing M single cells is mapped to a graph G=(V, E), where the node set V={v1, v2, …, vM}, and each node vi uniquely represents a single cell. The edge set E is defined according to the actual physical connection relationship between the single cells: if two single cells are directly connected in series or parallel in electricity, an edge is established between their corresponding nodes; in addition, based on the thermal design model of the battery module or pack, if two single cells are adjacent in position and there is a significant heat conduction path, an edge should also be established between the corresponding nodes to model the thermal coupling effect. Based on this graph definition, an MxM adjacency matrix A can be generated, where Aij=1 if there is an edge between nodes i and j, and 0 otherwise.

[0088] For example: when two single cells are in the same series branch or parallel branch, an electrical connection edge is established between their corresponding nodes; when two single cells are adjacent in spatial position and there is a significant heat conduction path, a thermal coupling edge is established between their corresponding nodes. Finally, based on the node set and the edge set, an adjacency matrix is constructed to represent the overall topology of the battery pack and serve as the structural input for subsequent graph neural network processing.

[0089] It can be understood that by the above method, the electrical structure and thermal coupling relationship of the battery pack are explicitly modeled as a graph structure, enabling the model to directly perceive the objective physical association between single cells and providing a structural basis for subsequent spatial information fusion, thereby avoiding the modeling bias caused by simply regarding the battery pack as independent units.

[0090] S102, obtain the running data of each single cell node in multiple historical cycle periods, and extract the time sequence features of the single cell node based on the running data to generate the corresponding node initial feature representation.

[0091] It can be understood that by the above method, the internal laws of single cell capacity decay and voltage evolution changing with cycle period can be automatically learned without relying on artificial experience feature design, so that the node features completely retain the key information contained in the original high-dimensional sensor data; the node initial features formed thereby can more accurately and stably represent the true aging state and running trend of the single cells, providing high-quality input for subsequent spatial information fusion and health state prediction, thereby effectively improving the prediction accuracy, robustness and generalization ability of the overall model under different battery types and complex working conditions.

[0092] S103, based on the graph structure model, using a graph attention network to perform spatial information fusion on the initial features of each node to generate updated node representations.

[0093] In an implementable manner, first, the initial features of a target monomer battery cell node and the initial features of its adjacent monomer battery cell nodes are respectively subjected to linear transformation to obtain corresponding transformed features; then, attention coefficients are calculated based on the transformed features of the target node and the transformed features of the adjacent nodes; then, the attention coefficients are subjected to normalization processing to obtain attention weights representing the importance of each adjacent node; finally, the features of each adjacent node after linear transformation are weighted and summed according to the attention weights, and an updated node representation of the target node is generated in combination with a nonlinear activation function.

[0094] It should be understood that the embodiments of the present application simulate the dynamic mutual influence of the states of the battery cells in the battery pack by using a graph attention network (GAT). Unlike simply treating all neighbor information equally, the GAT dynamically learns and quantifies the degree to which a battery cell is affected by its neighbors at a certain moment (or aging stage) through an attention mechanism, thereby achieving accurate and adaptive information fusion. This step is implemented by stacking K graph attention layers. In the kth GAT layer, the state vector of node i is updated according to its state in the previous layer and the states of all adjacent nodes (where j is a neighbor of i). Specifically, for node i and each of its neighbor nodes j, the model calculates the correlation strength between them through a shared attention network to obtain an attention coefficient:

[0095]

[0096] where is the linear transformation weight matrix of the kth layer. This attention mechanism is a single-layer feedforward network that maps the concatenated two node vectors to a scalar representing the importance of node j to node i.

[0097] Further, in order to facilitate comparison and weighting, the attention coefficients of all neighbors of a node are normalized using a Softmax function:

[0098]

[0099] where, i.e., the importance score of neighbor j learned automatically by the model when updating the state of node i.

[0100] Finally, the transformed features of the neighbor nodes are weighted and summed using the calculated attention weights, and an updated state vector of node i at the kth layer is obtained through a nonlinear activation function (such as ReLU):

[0101]

[0102] wherein, is a nonlinear activation function. Up to now, by stacking K layers of GAT, each node can aggregate the information of its K-hop neighbors, thus capturing more global intra-battery interactions.

[0103] It can be understood that by introducing the graph attention mechanism, the model can automatically identify and quantify the complex and nonlinear interactions within the battery pack, thus dynamically depicting the changes of the electrical and thermal coupling relationships within the battery pack at different aging stages and operating conditions, and improving the accuracy and robustness of spatial information modeling.

[0104] S104, based on the updated node representation, outputting the health state prediction value of the single battery cell and the consistency index of the battery pack through the multi-task prediction head.

[0105] It should be understood that through the multi-task prediction mechanism, the single battery cell health state evaluation and the battery pack consistency analysis are simultaneously realized in the same model framework, which not only improves the comprehensive description ability of the model for local and global states, but also enhances the prediction value of the running risk and performance degradation trend of the battery pack.

[0106] The embodiments of the present application provide a battery health prediction method, which first models the battery pack as a graph structure to depict the physical connection and interaction relationship between the single battery cells; on this basis, the running data of each single battery cell in multiple historical cycle periods is obtained, and the time-related features such as capacity decay and voltage evolution are automatically modeled through a time series feature extraction mechanism, thereby forming a node initial feature representation that can accurately represent the aging state of the single battery cell; then, based on the graph structure model, a graph attention network is introduced to weight and aggregate the initial features of each single battery cell node, and through adaptive learning of the influence weight of different adjacent battery cells on the target battery cell, the spatial correlation information such as electrical coupling and thermal coupling is effectively fused, and an updated node representation containing both the historical degradation information of the single battery cell and the interaction influence information within the group is obtained; finally, the updated node representation is input into the multi-task prediction head, on the one hand, the fine prediction of the health state of each single battery cell is realized, and on the other hand, the overall index representing the consistency of the battery pack is output through the graph-level feature aggregation, thereby simultaneously improving the single health evaluation accuracy and the battery pack consistency description ability in a single model framework, and providing a more comprehensive and reliable health state evaluation and operation decision basis for the battery management system.

[0107] In an implementable manner, the specific process of obtaining the running data of each single battery cell node in multiple historical cycle periods includes:

[0108] Collecting the sensor data of each single battery cell in a plurality of consecutive historical cycle periods.

[0109] wherein the sensor data at least includes one or more of a discharge capacity sequence, a charging voltage curve and a discharging voltage curve; the charging voltage curve and / or the discharging voltage curve is resampled to be uniform in a preset length to form a node input feature corresponding to each cycle period.

[0110] It can be understood that, in the embodiment, in order to comprehensively and accurately characterize the aging evolution characteristics of the single battery cell in the use process, the running data of each single battery cell node in multiple historical cycle periods is first acquired. Specifically, first, for each single battery cell in the battery pack, the time sequence data of the single battery cell in the continuous N historical charge and discharge cycle periods is collected through the BMS connected to the single battery cell. The core sensor data includes a scalar sequence and a high-dimensional curve sequence. The scalar sequence is the discharge capacity corresponding to each cycle period, forming a capacity degradation sequence with a length of N , which is one of the most direct correlation indicators of SOH; the high-dimensional curve sequence is the charging voltage curve and the discharging voltage curve in each period. These curves record the complete morphology of the change of voltage with time (or cumulative capacity) at a fixed sampling frequency, and contain rich information related to the internal electrochemical state (such as lithium ion concentration, internal resistance, phase change). Since the duration of charge and discharge and the number of sampling points of the battery cell in different cycle periods may be different, resulting in inconsistent lengths of voltage curves, in the embodiment, the collected charging voltage curve and / or discharging voltage curve is resampled to be uniform in a preset length (for example, 300 points), thereby forming a node input feature corresponding to each cycle period. Through the above processing, the running data from different cycle periods and different single battery cells is kept consistent in the feature dimension, providing a standardized and directly input data basis for subsequent time series feature extraction and model training.

[0111] It can be understood that, by collecting and normalizing the capacity and voltage of the single battery cell and other raw sensor data in multiple historical cycle periods, the key information contained in the battery cell aging process can be completely retained, avoiding information loss caused by non-uniform data format or manual compression, thereby providing high-quality and structure-consistent input data for subsequent time series feature extraction, and effectively improving the accuracy and stability of the characterization of the battery cell aging state.

[0112] In an implementable manner, generating the corresponding node initial feature representation includes:

[0113] First, the running data of each single battery cell in multiple historical cycle periods is input into a time series feature extraction network to learn the time-dependent relationship of the running data of the single battery cell with the cycle period; then, a node initial feature vector with a fixed dimension is output as the initial feature representation of the corresponding single battery cell node.

[0114] The time sequence feature extraction network includes any one of a recurrent neural network, a long short-term memory network, a convolutional neural network, or a Transformer network, or a combination thereof.

[0115] It should be understood that, in the present embodiment, in order to extract a feature representation capable of representing the aging state and evolution trend of the single battery cell from the operation data of the single battery cell in multiple historical cycle periods, the operation data is input into the time sequence feature extraction network for processing. Specifically, for each single battery cell node, the multi-dimensional data sequence of each battery cell after preprocessing in N consecutive periods is mapped into a fixed-dimensional embedding vector. As input, the time sequence feature extraction network is fed in to learn the time dependence and evolution law of the operation data of the single battery cell with the cycle period. Through the time sequence modeling mechanism or attention mechanism inside the network, the operation features in different cycle periods are associated and modeled. Subsequently, the time sequence feature extraction network outputs a fixed-dimensional feature vector as the initial feature representation of the corresponding single battery cell node, which is used for subsequent spatial information fusion processing based on the graph structure.

[0116] It should be noted that, if a recurrent neural network or its variant long short-term memory network is used, the network will process the data of each period in time step order, update and transfer information through its internal hidden state, thereby modeling the short-term and long-term dependence between periods, and finally taking the hidden state of the last time step or the integration of all hidden states as the initial feature vector of the node. If a convolutional neural network is used, the network automatically extracts local pattern features (such as specific fluctuation patterns of voltage curves) and their combinations in the data sequence through convolution and pooling operations in the time dimension, and maps them into a fixed-dimensional feature vector through a fully connected layer. If a Transformer encoder is used, the global association weight between all historical period data is directly calculated using the self-attention mechanism, dynamically focusing on the historical time point that is most important for predicting the current state (such as the inflection point period of capacity rapid decline), thereby more accurately capturing the long-range nonlinear decline trajectory.

[0117] It should be understood that, through the above-mentioned manner, the time evolution features of the single battery cell can be automatically learned from the multi-period operation data without manual design of features, so that the initial feature of the node not only reflects the operation state at a single time point, but also comprehensively represents the overall trend of the battery cell with cycle aging, thereby providing a node input with sufficient information and stronger expression ability for subsequent spatial feature fusion of the graph attention network, effectively improving the accuracy of battery health prediction and the generalization ability of the model.

[0118] Figure 2 Flowchart of the battery health prediction method provided by the present embodiment​Figure 2 As shown in Figure 1 the embodiment, on the basis of the embodiment, the process of acquiring the health state prediction value of the single battery cell and the battery pack consistency index is described in detail, and the method comprises the following steps: Figure 2

[0119] S201, input the updated node representation of each single battery cell node into the first prediction sub-network to output the health state prediction value of the corresponding single battery cell.

[0120] It should be understood that in the embodiment, the node representation vector updated by the graph attention network (GAT) has included the historical aging law of the battery cell itself and the correction information affected by the adjacent battery cells, and has sufficient features for final state regression. Therefore, through a lightweight, shared parameter fully connected neural network (i.e. the first prediction sub-network), these high-dimensional features can be efficiently mapped to specific SOH scalar prediction values. In specific implementation, the first prediction sub-network is usually composed of one or more fully connected layers, and finally reduces the dimension to 1 through a linear output layer (corresponding to the SOH value). The parameters are shared between different single battery cells. For any single battery cell node i, input the updated node representation into the first prediction sub-network to output the corresponding single battery cell health state prediction value , and the expression is:

[0121]

[0122] wherein, is the node representation after fusion of spatial information by the graph attention network, is used to learn the nonlinear mapping relationship between the node representation and the health state.

[0123] It can be understood that by predicting the health state of each single battery cell node respectively, the health difference of the single battery cells in the battery pack can be finely described, the local degradation risk can be avoided by relying only on the overall index of the battery pack, and the accuracy and interpretability of the single-level health evaluation can be improved.

[0124] S202, perform a graph-level readout operation on the updated node representation of all single battery cell nodes to generate a global graph representation vector.

[0125] wherein, the graph-level readout operation at least includes one of summation, averaging or maximum.

[0126] ​It should be understood that, since the overall characteristics (such as consistency) of the battery pack are not a simple list of individual cell properties, but a result of mutual comparison and joint action of all cell states. Therefore, after completing the single cell node level feature update, in order to represent the overall operation state of the battery pack, a graph level readout operation is performed on the updated node representation of all single cell nodes, so as to integrate the local state information distributed in all nodes into a global feature vector which can represent the overall state of the battery pack. In a specific implementation, the updated node representation of all single cell nodes in the battery pack is taken as input, and a global graph representation vector is generated by the graph level readout operation The expression is:

[0127]

[0128] Wherein, is used to map the node level features to a global feature vector which can represent the overall state of the battery pack.

[0129] It can be understood that, through the graph level readout operation, the local feature information distributed in multiple single cell nodes is effectively integrated into a unified global representation, so that the model can depict the operation state and internal difference of the battery pack from the overall level, and provide stable and compact feature input for subsequent consistency index prediction.

[0130] S203, input the global graph representation vector into the second prediction subnetwork to output the consistency index of the battery pack.

[0131] Wherein, the consistency index includes any one or more of the variance, range or dispersion index of the state of health within the battery pack.

[0132] After obtaining the global graph representation vector, the global graph representation vector is input into the second prediction subnetwork to output the consistency index of the battery pack. Specifically, the second prediction subnetwork also adopts a multi-layer perceptron structure, which is used to learn the mapping relationship between the global graph representation and the consistency of the battery pack, and its output can be one or more indexes for measuring the consistency degree of the battery pack, such as the variance, range or dispersion index of the state of health of the single cell within the battery pack, and its expression is:

[0133]

[0134] Wherein, represents the predicted consistency index of the battery pack.

[0135] It can be understood that, through the prediction of the consistency index of the battery pack, the dispersion degree of the health state of the internal single cell of the battery pack can be quantitatively represented, so as to assist in judging whether there is a significant short board cell or imbalance risk in the battery pack, and provide a basis for battery pack maintenance and safety management. ​

[0136] It should also be noted that before performing the embodiment Figure 3 Before performing the embodiment, model training needs to be continued, and the specific training process includes:

[0137] First, a training sample set is obtained, then a joint loss function is constructed, and finally, based on the joint loss function, the parameters of the time series feature extraction network, the graph attention network, and the multi-task prediction head are iteratively updated using backpropagation until the preset convergence condition is met.

[0138] The training sample set includes the operating data of the sample battery pack, the true labeled value of the single cell health state, and the true labeled value of the battery pack consistency index. The joint loss function is a first loss term based on the mean square error of the predicted value of all single cell health states and the true value, and a second loss term based on the mean square error of the predicted value of the battery pack consistency index and the true value.

[0139] It should be understood that in order for the entire end-to-end model (including the time series feature extraction network, the graph attention network, the first prediction subnetwork, and the second prediction subnetwork) to simultaneously optimize the single SOH and battery pack consistency two prediction targets, a multi-task joint training strategy must be used. The core principle is to construct a joint objective function that combines the loss of the two tasks, and simultaneously update all network parameters using the backpropagation algorithm, so that the underlying spatiotemporal feature representation learned by the model is beneficial to both tasks, thereby achieving positive transfer between tasks and enhancing shared representation. The specific implementation process is as follows: First, construct a training sample set, each sample must include: the original sensor data of the battery pack in N historical periods; the true value of the health state (SOH) of each single cell at the current time (or the N+1 period); and the true value of the battery pack consistency index (such as variance) calculated based on the true values of all cell SOH. Then, construct a joint loss function, and the specific joint loss function expression is:

[0140]

[0141]

[0142]

[0143] wherein, is the predicted value of the single cell health state, is the true value of the single cell health state, is the predicted value of the battery pack consistency index, is the true value of the battery pack consistency index, The preset weight coefficient is used for the preset weight coefficient. Finally, based on the joint loss function, the parameters of the time series feature extraction network, the graph attention network and the multi-task prediction head are iteratively updated by using the back propagation algorithm until the preset convergence condition (such as the loss no longer significantly decreases or the maximum training number of rounds is reached) is met.

[0144] It can be understood that, through the above end-to-end training mode based on the multi-task joint loss function, it is ensured that the node and graph representation vectors learned by the model are not only beneficial to the state regression of a single battery cell, but also beneficial to capturing the difference pattern between battery cells, thereby significantly improving the overall prediction accuracy, robustness and generalization ability of the model, so that a single model can perform a complex double health evaluation task.

[0145] Figure 3 The structure schematic diagram of the battery health prediction device provided by the embodiment of the application is shown as follows: Figure 4 As shown in the figure, the device comprises:

[0146] The acquisition module 301 is configured to acquire a graph structure model of a battery pack, wherein the graph structure model comprises a plurality of single battery cell nodes, and if two single battery cell nodes have a physical connection relationship, the two single battery cell nodes are connected to form an edge.

[0147] The first processing module 302 is configured to acquire running data of each single battery cell node in a plurality of historical cycle periods, and extract time series features of the single battery cell node based on the running data to generate a corresponding node initial feature representation.

[0148] The fusion module 303 is configured to perform spatial information fusion on the initial features of each node based on the graph structure model by using a graph attention network to generate an updated node representation.

[0149] The second processing module 304 is configured to output a health state prediction value of the single battery cell and a consistency index of the battery pack based on the updated node representation through a multi-task prediction head.

[0150] In a possible implementation, the first acquisition module 301 is specifically configured to:

[0151] define each single battery cell in the battery pack as a node in the graph structure;

[0152] establish an edge between the corresponding nodes according to the physical connection relationship between the single battery cells, and the physical connection relationship at least includes one of a series connection relationship, a parallel connection relationship and a thermal coupling relationship;

[0153] construct an adjacency matrix for representing the topology structure of the battery pack based on the nodes and the edges.

[0154] In a possible implementation, the first processing module 302 is specifically configured to:

[0155] collecting sensor data of each single battery cell in a plurality of historical consecutive cycle periods, the sensor data at least including one or more of a discharge capacity sequence, a charging voltage curve and a discharging voltage curve;

[0156] wherein the charging voltage curve and / or the discharging voltage curve is resampled to be uniform in a preset length to form a node input feature corresponding to each cycle period.

[0157] In a possible implementation, the first processing module 302 is specifically configured to:

[0158] input the running data of each single battery cell in a plurality of historical cycle periods into a time sequence feature extraction network to learn a time-dependent relationship of the running data of the single battery cell with the cycle period;

[0159] output a node initial feature vector of a fixed dimension as an initial feature representation of a corresponding single battery cell node;

[0160] wherein the time sequence feature extraction network includes any one or a combination of a recurrent neural network, a long short-term memory network, a convolutional neural network or a Transformer network.

[0161] In a possible implementation, the fusion module 303 is specifically configured to:

[0162] perform linear transformation on the initial feature of a target single battery cell node and the initial features of its adjacent single battery cell nodes respectively to obtain corresponding transformed features;

[0163] calculate an attention coefficient based on the transformed feature of the target node and the transformed features of the adjacent nodes;

[0164] perform normalization processing on the attention coefficient to obtain an attention weight representing the importance of each adjacent node;

[0165] according to the attention weight, perform weighted summation on the features of each adjacent node after linear transformation, and combine a nonlinear activation function to generate an updated node representation of the target node.

[0166] In a possible implementation, the second processing module 304 is specifically configured to:

[0167] input the updated node representation of each single battery cell node into a first prediction sub-network to output a health state prediction value of the corresponding single battery cell;

[0168] perform a graph-level readout operation on the updated node representations of all single battery cell nodes to generate a global graph representation vector, the graph-level readout operation at least including one of summation, averaging or maximum value calculation;

[0169] The global graph representation vector is input to the second prediction sub-network to output a consistency index of the battery pack, the consistency index including any one or more of a variance, a range, or a dispersion index of the state of health within the battery pack.

[0170] In a possible implementation, the second processing module 304 is further configured to perform model training.

[0171] The process of model training includes:

[0172] Obtaining a training sample set, the training sample set including operation data of a sample battery pack, real labeled values of the state of health of the single battery cells, and real labeled values of the consistency index of the battery pack;

[0173] Constructing a joint loss function, the joint loss function including a first loss term based on a mean square error between predicted values of all the states of health of the single battery cells and real values thereof, and a second loss term based on a mean square error between predicted values of the consistency index of the battery pack and real values thereof;

[0174] Iteratively updating parameters of the time series feature extraction network, the graph attention network, and the multi-task prediction head based on the joint loss function by using back propagation until a preset convergence condition is met.

[0175] The battery health prediction apparatus provided in the embodiments of the present application can perform the method provided in the method embodiments, and has similar implementation principles and technical effects, which will not be described here again.

[0176] Figure 4 A structural schematic diagram of an electronic device provided in the embodiments of the present application is shown in FIG. 4. As shown in FIG. 4, the electronic device 40 provided in the embodiments of the present application includes at least one processor 401 and a memory 402. Optionally, the device 40 further includes a communication component 403. The processor 401, the memory 402, and the communication component 403 are connected through a bus 404. ​

[0177] In the specific implementation process, the at least one processor 401 executes computer execution instructions stored in the memory 402, so that the at least one processor 401 performs the method described above.

[0178] The specific implementation process of the processor 401 can refer to the method embodiments described above, and has similar implementation principles and technical effects, which will not be described here again.

[0179] ​In the above embodiments, it should be understood that the processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The steps of the method disclosed in combination with the application can be directly embodied as hardware processor execution, or executed by a combination of hardware and software modules in the processor.

[0180] The memory can include a random access memory (RAM), and can also include a non-volatile memory (NVM), such as at least one disk memory.

[0181] The bus can be an industry standard architecture (ISA) bus, a peripheral component (PCI) bus, an extended industry standard architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of the present application does not limit to only one bus or one type of bus.

[0182] The present application also provides a computer readable storage medium, and the computer readable storage medium stores computer execution instructions, when the processor executes the computer execution instructions, the above method is realized.

[0183] The above readable storage medium can be realized by any type of volatile or non-volatile storage device or their combination, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The readable storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0184] An example readable storage medium is coupled to the processor such that the processor can read information from the readable storage medium and can write information to the readable storage medium. Of course, the readable storage medium can also be a part of the processor. The processor and the readable storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the readable storage medium can also exist as discrete components in the device.

[0185] The division of units is only a logical functional division, and in actual implementation, there can be another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0186] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, can be located in one place, or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0187] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit.

[0188] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the part of the present application that essentially contributes to the prior art or the part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the embodiments of the present application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0189] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware. The foregoing program can be stored in a computer readable storage medium. The program executes to perform the steps of the above-mentioned method embodiments; and the foregoing storage medium includes various media capable of storing program codes, such as ROM, RAM, magnetic disk, or optical disk.

[0190] Finally, it should be noted that other embodiments of the present application will readily occur to those skilled in the art upon consideration of the specification and practice of the present application disclosed herein. The present application is intended to include all such variations and modifications as fall within the scope of the present application, which is defined by the following claims, as well as the full scope of equivalents to which such claims are entitled. It is intended, therefore, that the present application be considered as including all possibilities falling within the scope of the application and their equivalents.

Claims

1. A battery health prediction method, characterized in that, include: Obtain a graph structure model of the battery pack, wherein the graph structure model includes multiple individual cell nodes, and if there is a physical connection between two individual cell nodes, then the two individual cell nodes are connected to form an edge; The operation data of each individual cell node in multiple historical cycles is obtained, and the time-series features of the individual cell nodes are extracted based on the operation data to generate the corresponding initial feature representation of the nodes. Based on the graph structure model, a graph attention network is used to fuse the initial features of each node spatially to generate an updated node representation. Based on the updated node representation, the health status prediction value of a single cell and the consistency index of the battery pack are output through the multi-task prediction head.

2. The method according to claim 1, characterized in that, The acquisition of the graph structure model of the battery pack includes: Each individual cell in the battery pack is defined as a node in the graph structure; Based on the physical connection relationship between individual battery cells, edges are established between the corresponding nodes. The physical connection relationship includes at least one of the following: series connection relationship, parallel connection relationship, and thermal coupling relationship. An adjacency matrix is ​​constructed based on nodes and edges to characterize the topology of the battery pack.

3. The method according to claim 1, characterized in that, The acquisition of operational data for each individual cell node over multiple historical cycles includes: Collect sensor data for each individual battery cell over multiple consecutive historical cycles. The sensor data includes at least one or more of the following: discharge capacity sequence, charging voltage curve, and discharge voltage curve. The charging voltage curve and / or the discharging voltage curve are resampled to be uniform to a preset length, so as to form the node input features corresponding to each cycle.

4. The method according to claim 1 or 3, characterized in that, The step of extracting the timing features of individual battery cell nodes based on the operational data and generating corresponding initial feature representations of the nodes includes: The operating data of each individual cell in the multiple historical cycle periods are input into the time-series feature extraction network to learn the time dependency of the individual cell operating data with the cycle period. Output a fixed-dimensional initial feature vector for the node, which serves as the initial feature representation for the corresponding individual cell node; The temporal feature extraction network includes any one or a combination of recurrent neural networks, long short-term memory networks, convolutional neural networks, or Transformer networks.

5. The method according to claim 1, characterized in that, The step of using a graph attention network to fuse the initial features of each node spatially to generate an updated node representation includes: The initial characteristics of the target single cell node and the initial characteristics of its adjacent single cell nodes are linearly transformed to obtain the corresponding transformed characteristics. The attention coefficient is calculated based on the transformation characteristics of the target node and the transformation characteristics of its neighboring nodes. The attention coefficients are normalized to obtain the attention weights that characterize the importance of each neighboring node; Based on the attention weights, the features of each neighboring node after linear transformation are summed in a weighted manner, and combined with a nonlinear activation function to generate the updated node representation of the target node.

6. The method according to claim 1, characterized in that, The method of outputting the predicted health status of individual cells and the consistency indicators of the battery pack through the multi-task prediction head includes: The updated node representation of each individual cell node is input into the first prediction sub-network to output the corresponding individual cell health status prediction value. Perform a graph-level readout operation on the updated node representations of all individual battery cell nodes to generate a global graph representation vector. The graph-level readout operation includes at least one of summation, averaging, or maximization. The global graph representation vector is input into the second prediction sub-network to output the consistency index of the battery pack. The consistency index includes any one or more of the variance, range, or dispersion index of the health status within the battery pack.

7. The method according to claim 1 or 4, characterized in that, The method also includes model training; The model training process includes: Obtain a training sample set, which includes the operating data of the sample battery pack, the true labeled values ​​of the health status of individual cells, and the true labeled values ​​of the battery pack consistency index. A joint loss function is constructed, which consists of a first loss term based on the mean square error between the predicted and actual values ​​of the health status of all individual cells, and a second loss term based on the mean square error between the predicted and actual values ​​of the consistency index of the battery pack. Based on the joint loss function, backpropagation is used to iteratively update the parameters of the temporal feature extraction network, graph attention network, and multi-task prediction head until the preset convergence condition is met.

8. A battery health prediction device, characterized in that, include: The acquisition module is used to acquire the graph structure model of the battery pack. The graph structure model includes multiple individual cell nodes. If there is a physical connection between two individual cell nodes, the two individual cell nodes are connected to form an edge. The first processing module is used to acquire the operating data of each individual cell node in multiple historical cycle periods, and extract the time-series features of the individual cell nodes based on the operating data to generate the corresponding initial feature representation of the nodes. The fusion module is used to perform spatial information fusion on the initial features of each node based on the graph structure model using a graph attention network to generate an updated node representation; The second processing module is used to output the predicted health status of individual cells and the consistency index of the battery pack based on the updated node representation through a multi-task prediction head.

9. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-7.

Images