Battery health state evolution path prediction method based on subgraph representation learning

Abstract

The invention discloses a battery health state evolution path prediction method based on subgraph representation learning, and aims to overcome the defects in the prior art, obtain the conversion relation between different fault key safety states and support battery fault early warning. The method comprises the following steps: firstly, collecting battery characteristic data, cleaning, serializing and segmenting, and performing efficient compression by using an auto-encoder; secondly, extracting a key state by adopting a data flow clustering technology, regarding segments as small micro-clusters, and integrating charging sequences to form large micro-clusters which are used as key state nodes of an evolution process; then, a state transition diagram is constructed based on the time sequence transition relation of the battery between the micro-clusters, nodes represent key states, and edges represent state transition; then, for any to-be-predicted node pair, dynamically extracting a closed sub-graph, and designing a structure identification vector containing four-dimensional topological characteristics for node marking; then, constructing an enhanced sub-graph by injecting a negative sample edge, and carrying out representation learning by adopting a multi-head graph attention network; and finally, performing link prediction by using the trained model, screening high-probability connecting edges, and splicing the high-probability connecting edges according to a time sequence to form a directed evolution path. According to the method, through subgraph extraction and composite topology marking, negative sample enhanced representation learning and an evolution path splicing mechanism, the prediction precision is remarkably improved, accurate description of the evolution trajectory of the full life cycle health state of the battery is realized, and a visual and quantifiable technical support is provided for fault early warning.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery anomaly detection, and more specifically, relates to a method for predicting the evolution path of battery health status based on subgraph representation learning. Background Art

[0002] As the global energy structure transitions to a clean and low-carbon one, batteries, as an important basic unit supporting the grid connection of new energy, peak and frequency regulation, and electric transportation, play an increasingly critical role in the modern energy system. However, batteries are affected by a variety of complex factors throughout their life cycle, such as charge and discharge rates, temperature changes, and fluctuations in operating conditions. This results in the evolution of their health status being highly nonlinear, highly time-correlated, and difficult to accurately model and predict. Existing battery health status prediction methods mostly rely on physical modeling, statistical analysis, or traditional machine learning techniques. They often have difficulty capturing the potential spatiotemporal dependency structure in battery operating data, unable to depict a series of key states of the battery from inception to failure, and unable to reveal fault characteristics and their dynamic evolution laws. Consequently, they are unable to adapt to the precise prediction needs of multiple types of batteries and complex application scenarios.

[0003] Graph neural networks, a deep learning model that has emerged in recent years, are capable of modeling complex relationships between nodes and integrating structural information with dynamic features. They demonstrate significant advantages in processing multi-dimensional, heterogeneous, and dynamically evolving data. Therefore, developing a method for predicting the evolutionary path of battery health status based on graph neural networks is expected to break through the bottleneck of traditional battery health status modeling, enabling accurate state perception and intelligent prediction of the evolutionary path of battery failure throughout its lifecycle. This will provide more forward-looking decision support for battery management systems and enhance the safety and reliability of battery systems. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a battery health state evolution path prediction method based on subgraph representation learning to obtain the temporal coupling relationship between different fault-critical safety states, thereby supporting early warning of battery failures.

[0005] To achieve the above-mentioned object, the present invention provides a method for predicting the evolution path of a full-cycle battery health state based on subgraph representation learning, which is characterized by comprising the following steps:

[0006] (1) Collect the characteristics of each battery;

[0007] (2) Clean, serialize and segment the battery data according to the battery characteristics, train an autoencoder to achieve efficient data compression;

[0008] (3) Extraction of critical states based on data stream microclusters: Data stream clustering is used to extract the critical states of the battery. Each segment is considered as a small microcluster, and the small microclusters of different segments of the same charging sequence are integrated into a larger microcluster. The battery charging sequence of each battery is clustered one by one. After completing the clustering of all charging sequences of a battery, the same operation is performed on the sequence of the next battery. All the microclusters obtained in the end are regarded as the critical states in the battery fault evolution process.

[0009] (4) State transition graph construction: The identified key states are converted into nodes in a graph structure. The microclusters in which each battery has been clustered are tracked to map the transition path of each battery in different states. The states of each battery transitioning at adjacent time points are connected to form edges in the graph. After deduplication of the transition relationships of all batteries, a complete set of edges is formed, resulting in a graph structure with nodes as key states and edges as state transitions.

[0010] (5) Subgraph extraction and node labeling: For any pair of nodes to be predicted in the graph structure, subgraph extraction and node labeling are performed. A dynamic range subgraph extraction method is adopted: first, bidirectional expansion is performed with the two target nodes as the starting point, and all reachable neighbor nodes within the preset hop count range are collected to form an initial joint neighborhood set. When the number of nodes exceeds the preset upper limit, the topological importance screening mechanism is activated. By calculating the weighted sum of the inverse of the distance between each node and the two central nodes as the importance score, the preset number of nodes with the highest score are retained to form a closed subgraph. A composite structure labeling mechanism is designed for subgraph nodes: by constructing a structure identification vector containing four-dimensional topological features, it comprehensively reflects the shortest path distance from the node to the two central nodes, the local structure label generated based on the neighborhood topology iteration, and the network centrality value that represents the global importance of the node. It provides extremely rich and highly discriminative initial node features for subsequent representation learning, and explicitly and effectively encodes the topological position, local structure role and global importance of the node relative to the target pair to be predicted.

[0011] (8) Representation learning based on negative sample enhanced subgraph: After completing subgraph extraction and node labeling, subgraph enhancement is performed, and the enhanced subgraph is learned using an attention graph neural network. When constructing training samples, in order to prevent the graph neural network from simply relying on the existence of positive sample edges for memorization, the present invention introduces a negative sample injection strategy. First, a number of positive sample edges (i.e., edges that actually exist in the original graph) are sampled in the closed subgraph, and an equal number of negative sample edges (i.e., edges that do not exist in the original graph) are collected at the same time. These negative sample edges are then temporarily added to the graph structure to construct an enhanced subgraph. Node embedding calculations are performed on the enhanced subgraph, so that the model can see both the real positive sample edges and the existence of negative sample edges during training, thereby forcing the model to pay attention to richer topological and attribute information, rather than simply judging links by the presence or absence of edges. On the enhanced subgraph after negative sample injection, a multi-head graph attention network is introduced as the backbone architecture to aggregate and represent the contextual information of each node.

[0012] (9) Prediction of potential links in critical states: After training, the resulting model can be used to predict links for any pair of nodes and output the corresponding probability of existence. For a set of known critical fault states in the battery system, all combinations between these nodes or specific combinations of user interest can be scored, and node pairs with probabilities higher than a preset threshold are predicted as potential fault evolution edges. Subsequently, based on the time sequence and direction information of these predicted edges, they are spliced ​​into a series of directed paths, which intuitively show the evolution trajectory of the fault state in the battery system from early to late stages, providing visual and quantifiable technical support for fault warning and diagnosis.

[0013] The object of the present invention is achieved in this way.

[0014] The present invention first collects and preprocesses battery operation data; performs streaming clustering based on dimensionality reduction using an automatic encoder; regards the obtained microclusters as critical fault states, and splices the cluster centers of the sub-microclusters within each microcluster into node attributes; then constructs a state transition graph based on the temporal migration relationship of battery sequences between microclusters; for any pair of states to be predicted, extracts a closed subgraph containing two central nodes and their neighbors with a certain number of hops from the graph, and obtains the topological relative position identifier by constructing a structural identification vector of four-dimensional topological features; then, after fusing the attribute features of the nodes with the label encoding, a multi-headed graph attention network with a negative sample injection strategy is used to represent and learn the subgraph, and finally, the representation of the two central nodes is output as the existence probability through a link classifier; after scoring all key state pairs in the battery system, the high-confidence prediction edges are spliced ​​into directed paths in chronological order to obtain the fault evolution path. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1This is a flowchart of a specific implementation method of the battery health state evolution path prediction method based on subgraph representation learning of the present invention;

[0016] Figure 2 is a schematic diagram of the segmentation of the charging sequence in the present invention, wherein the curve represents the change of the voltage of a charging sequence over time and the dotted line represents the segmentation;

[0017] Figure 3 is a schematic diagram of the autoencoder in the present invention, wherein, is the original input of the encoder, represents the reconstructed output of the decoder, d k represents implicit representation;

[0018] Figure 4 It is a flow chart of a specific implementation of a specific implementation of the path construction method of the present invention.

[0019] Figure 5 It is a system framework diagram of a specific implementation method of the battery health state evolution path prediction method based on subgraph representation learning of the present invention. DETAILED DESCRIPTION

[0020] The following describes the specific embodiments of the present invention in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, such descriptions will be omitted here.

[0021] Figure 1 This is a flowchart of a specific implementation of the battery health status evolution path prediction method based on subgraph representation learning of the present invention.

[0022] In this embodiment, if Figure 1 As shown, the battery health state evolution path prediction method based on subgraph representation learning of the present invention includes the following steps:

[0023] S1: Data collection and preprocessing

[0024] During battery operation, a variety of key parameters are generated, including voltage, current, temperature, operating status, and battery capacity. Taking the lithium-ion battery in an electric vehicle as an example, to better characterize the battery's health throughout its lifecycle, we will use the characteristics collected during the charging process as an example, including operating voltage, current, and battery module temperature.

[0025] To represent the state of the battery at a certain moment, we can use a feature vector to describe multiple feature values ​​at that moment. is the jth feature of the battery at the i-th time point. For example, the vector [52,3.5,345] indicates that the current voltage of the battery pack is 345 volts (the first feature), the operating current is 3.5 amps (the second feature), and a temperature in the battery module is 45 degrees Celsius (the third feature).

[0026] The data preprocessing stage mainly includes steps such as missing value filling, feature enhancement, serialization and segmentation, post-processing and data aggregation:

[0027] 1.1) Missing value filling: For the remaining missing value problems in the data set, a forward filling strategy is adopted to backfill the current blank with the first subsequent valid data point. in is the first valid data point subsequently.

[0028] 1.2) Feature Enhancement: Calculate the mean and variance of the temperature and voltage data of each sensor at each timestamp, and append these values ​​as additional features to the original features to comprehensively observe the data. For the i-th time point, the enhanced temperature feature is:

[0029]

[0030]

[0031] where n T is the number of temperature sensors. Similarly, the mean and variance of the voltage are The enhanced feature vector is:

[0032]

[0033] 1.4) Serialization and segmentation: Figure 2 As shown, all the data of a single battery charge is regarded as an independent charging sequence. Each charge is divided into four stages based on the voltage fluctuation during the charging process. Each stage represents a specific voltage range in the battery charging process. Assuming the voltage range is V1, V2, V3, V4, V5, the data sequence for each stage is:

[0034]

[0035] Where k = 0, 1, 2, 3, 4, Represents the mean voltage value at the i-th moment.

[0036] 1.5) Post-processing: Before data analysis, for each segmented data, those segments with less than 10 record entries were removed to enhance the robustness of the statistical analysis.

[0037] 1.6) Data aggregation: Take the average value of the corresponding fields of all data in each segment to obtain the feature vector of this segment, thereby extracting a representative feature vector. k , we need to take the average value of each dimension feature j over all timestamps i. The specific formula is:

[0038]

[0039] in represents the average value of the jth feature in the kth segment; |S k | represents the number of timestamps contained in the kth segment. For each segment S k , we perform the above aggregation operation on all features j and finally obtain the feature vector of the segment:

[0040]

[0041] Where m is the total number of features.

[0042] S2: Low-dimensional representation of data based on autoencoders

[0043] After data preprocessing, for example, electric vehicle batteries typically contain dozens of temperature and voltage sensors. Therefore, each segment may maintain a data structure with hundreds or even more dimensions, which greatly hinders subsequent data stream clustering to extract key states. This paper employs autoencoder technology to achieve efficient data compression and reconstruction.

[0044] like Figure 3 As shown in Figure 2, the autoencoder consists of two parts: the encoder and the decoder. The encoder maps the high-dimensional charging segment data to the low-dimensional latent space. Its mathematical representation is as follows:

[0045]

[0046] in, is the kth segment original feature vector of the input, d k The decoder maps the low-dimensional latent representation back to the original high-dimensional data space.

[0047] The decoder maps the low-dimensional latent representation back to the original high-dimensional data space.

[0048]

[0049] here, is the reconstructed output, which is as close as possible to the original input

[0050] Using mean square error as the loss function, calculate the original input and reconstructed output The differences between:

[0051]

[0052] In the actual use stage, only the encoder part is needed to map the input data into a low-dimensional latent space to obtain the compressed representation of the data d k This compressed representation contains the main features of the input data and can be used for subsequent data stream clustering.

[0053] S3: Automatic extraction of battery key status based on data stream micro-clusters

[0054] 3.1) Establishment of micro-clusters

[0055] After compressing and characterizing the battery data, the next key step is to apply a data stream clustering algorithm to extract key battery status characteristics. To effectively mine key status information from the continuous battery data stream, this paper introduces an innovative segmented data stream clustering strategy. This strategy divides the continuously influx of data into multiple segments and then performs cluster analysis on each segment, thereby enhancing the control of the dynamic characteristics of the data stream.

[0056] When processing streaming data, especially in scenarios related to analyzing battery performance, a significant challenge lies in the strong time correlation and dynamic evolution characteristics of such data. In order to accurately capture these complex characteristics that change over time and carefully explore the unique behaviors within each specific time period, this paper proposes a novel data processing method, namely the sequence segmentation method. Its core is to regard each segment as a small microcluster, and then integrate these small microclusters into a larger microcluster to facilitate comprehensive analysis. Specifically, the form of describing the i-th time series microcluster in the model is as follows:

[0057] MC i ={mc1,mc2,mc3,mc4},

[0058] Among them mc j The small micro-cluster representing the j-th segment is as follows:

[0059] mc j =(LS,N,C,I),j∈{1,2,3,4},

[0060] Where LS represents the sum of the eigenvectors of each segment, and the calculation formula is The feature vector representing the jth segment in sequence i is obtained from the data representation of step S2; N represents the number of sequences clustered in this small microcluster; C represents the cluster center, calculated as: C = LS / N; I is a set representing the number of batteries clustered in this segment. It should be noted that in actual applications, the battery charging process may be incomplete, so some sequences may not contain the complete four segments. To solve this problem, the present invention allows the formula MC i Small and medium microclusters mc j The number of is less than 4.

[0061] 3.2) Model input and initialization

[0062] This invention uses a personalized data stream clustering method to update the model. Specifically, the invention clusters each battery's charging sequence one by one. After completing the clustering of all charging sequences for a battery, the invention switches to the next battery's sequence and performs the same operation.

[0063] In this setting, the initialization process of the model is greatly simplified. For the first charging sequence of the first battery, there is no need for complex preprocessing or parameter adjustment. Instead, it is directly used as the first microcluster of the model. Assuming that this sequence comes from battery c, it can be recorded as The set of segments owned by The first micro-cluster initialized is expressed as:

[0064]

[0065] Among them, d j The characteristic vector representing the jth segment of this charging sequence is obtained from the data representation of S2. Here, N=1 means there is only one segment of data, and the initialization center C=d j / 1, I is initialized to contain only the battery number c.

[0066] 3.3) Model update and post-processing

[0067] After the model is initialized, new sequence inputs need to be evaluated to determine whether to create a new microcluster or merge it into an existing microcluster. The method of the present invention compares the new sequence with the microclusters in the model one by one and makes a decision based on a series of preset constraints.

[0068] First, a preset distance threshold, λ, is introduced as a key parameter for evaluating the similarity between a new sequence and existing microclusters. By calculating the distance between a new sequence and each microcluster in the model and comparing it with λ, the new sequence's affiliation is determined: if the distance between the input sequence and a microcluster is less than λ, it is merged into that microcluster; conversely, if the distance exceeds λ, a new microcluster is created to accommodate the sequence.

[0069] In addition, the distance formula designed by the present invention specifically solves the problem of non-overlapping segments between microclusters after sequence segmentation. is divided into multiple segments, and the set of these segments is recorded as In the model, each microcluster MC i It is composed of several small microclusters, and the segment set covered by these small microclusters is recorded as s i When calculating the battery sequence With micro cluster MC i The distance between the two is calculated using the following formula:

[0070]

[0071] Here, d j Represents the battery sequence The eigenvector of the jth segment in C j Microcluster MC i The distance calculation of the present invention only considers those segments that exist in both the sequence and the microcluster, that is, and s i The intersection of .

[0072] To maintain the sequential nature of the model, the model maintains a clustering index, begin_index, for each battery. This index records the microcluster number in the model from which the current sequence should begin distance calculations. This is because the present invention clusters the data of each battery one by one. To ensure that different sequences for the same battery are not reversed in time, the begin_index is set to limit the starting point for comparison between the current sequence and the microclusters in the model. Specifically, begin_index records the microcluster number of the most recent sequence for the current battery; subsequent sequences must start from this number to compare with the microclusters in the model.

[0073] Based on the above discussion, the model update rules are as follows:

[0074] If the distance between the new sequence and an existing micro-cluster is less than or equal to λ, the new sequence has a high similarity with the micro-cluster and should be included in the micro-cluster. The micro-cluster should be updated and begin_index should be set to the position of the current micro-cluster.

[0075] If the distance between the new sequence and all existing microclusters is greater than λ, the new sequence is considered to be sufficiently different, and a new microcluster should be created at the next position of the microcluster position begin_index of the previous sequence to accommodate the sequence. The begin_index should be updated to the position of the next microcluster.

[0076] After all microclusters are constructed, post-processing operations are required. In order to effectively identify abnormal microclusters, the present invention creates a microcluster MC for each microcluster. i Perform a specific operation: The total number of sequences of the micro-cluster is obtained by summing the sum of the sequences of all small micro-clusters in the micro-cluster. If this total number of sequences is lower than the preset deletion threshold del_num, the micro-cluster will be deleted.

[0077] 3.4) Micro-cluster update and creation

[0078] This paper proposes a segmented independent update strategy for battery charging sequences and microclusters. The main advantage of this strategy is its flexibility, which can effectively integrate new sequence data while ensuring the stability of existing information in the microclusters. In the specific implementation process, consider the nth charging sequence of battery c. Integration into micro cluster MC i When the micro cluster MC is i Adjust and update the small micro clusters in:

[0079]

[0080] The detailed explanation of this formula is as follows:

[0081] For the segments shared by the current sequence and the micro-cluster (i.e. ), add the features of these segments of the current sequence to the micro-clusters: the corresponding segments x of the current sequence j The eigenvector of is added to the existing eigenvector LS to form the sum of the new eigenvectors LS+x j , to reflect the newly added features; the number of sequences is increased by one due to the addition of new sequences, and updated to N+1; the new cluster center is the sum of the updated feature vectors divided by the new number of sequences, that is, To maintain the central characteristics of the small micro-cluster; add the current battery number c to the existing battery number set I and update it to I∪{c}.

[0082] For segments unique only to the sequence Create small microclusters of these segments of the current sequence in the microcluster: the sum of the feature vectors LS = x j ; Only contains the current sequence, so N = 1; cluster center C = x j / 1 is the feature vector; the battery number set only contains the current battery c.

[0083] For segments unique to microclusters only (i.e. ), the state of the small micro-cluster remains unchanged.

[0084] In this way, not only the data structure in the micro-cluster is updated, but also the micro-cluster can effectively reflect the latest battery charging sequence information.

[0085] Creating a micro cluster is relatively simple. For the current charging sequence The number of segments owned is The newly created micro cluster MC new It can be expressed as:

[0086]

[0087] Here, each micro-cluster in each segment only contains the segment data corresponding to the current sequence of battery c when initialized.

[0088] S4: Prediction of key state evolution paths based on subgraph representation learning

[0089] 4.1) Construct key state transition diagram

[0090] In the previous paper, we successfully extracted many microclusters through data stream clustering methods. These microclusters are considered to be the key states in the evolution of battery failure. In order to further explore the dynamic transition relationship between these states, this paper first converts the identified microclusters into nodes in a graph structure. Each node not only represents a unique state, but also contains detailed information about the state through its feature vector. For each node in the graph, its corresponding microcluster MC i The eigenvector z i It can be expressed as:

[0091] z i =C1‖C2‖C3‖C4,

[0092] The symbol ‖ represents the concatenation operation of a vector sequence. Specifically, C j This characterization refers to microclusters MC i Internal j-th microcluster mc j The cluster center vector of , which not only condenses the essential attributes of the micro-cluster, but also gives detailed and descriptive features to the nodes in the graph. In practice, if the micro-cluster MC i If the jth subordinate micro-cluster is not included, then C j Set it to a one-dimensional matching all-zero vector to maintain the consistency of the feature vector dimension and the convenience of the processing flow.

[0093] In the process of building a graph model, the edges between nodes reflect the state transition, which is a crucial step. The present invention uses the battery sequence information recorded in the micro-cluster creation phase to construct these edges. Specifically, first traverse all micro-clusters in order, and for each micro-cluster, all sub-micro-clusters m j Traverse and record the battery sequence I contained in each jThis step can track the microclusters in which each battery has been clustered, thereby mapping the transition path of each battery in different states. Since each microcluster has a time sequence, the changes in the aggregation of batteries in the microcluster naturally reveal the key state transition process of battery failure evolution. Based on this, the present invention connects the states of each battery transferred at adjacent time points to form edges in the graph.

[0094] In addition, the process of creating edges may cause multiple edges to be repeated because there are multiple batteries transferred between two microclusters, so it is necessary to remove the duplicate edges. Finally, the key state graph G = (V, E, Z) is obtained, where V is the key state vertex cluster. For any node v i ∈V, its eigenvector z i ∈Z corresponds to micro-cluster MC i The characteristic vector of , E is the transfer relationship obtained from the existing data.

[0095] Due to the scarcity of data, some transition relationships (i.e., state evolution paths) are not obtained. Next, link prediction technology is used to predict the transitions between these hidden key states, such as Figure 4 shown.

[0096] 4.2) Subgraph extraction

[0097] For a graph G = (V, E), for any pair of nodes to be predicted in the graph structure, denoted as nodes u and v, the dynamic range subgraph extraction method is used to perform bidirectional K-hop neighborhood expansion starting from u and v respectively to obtain the joint neighborhood node set like (preset node upper limit), then filter by topological importance: for each node Calculate the importance score:

[0098]

[0099] Keep the N with the highest score max nodes, forming a closed subgraph where d(·) is the shortest path distance and α is a smoothing constant.

[0100] 4.3) Node Marking

[0101] For subgraph For all assigned composite structure labels in , for any node w, first generate a four-dimensional structural feature vector:

[0102] r(w)=[d(u,w),d(v,w),WL k (w),PR(w)] T

[0103] WLk (w) is the topological label generated by the Weisfeiler-Lehman algorithm after k iterations, and PR(w) is the PageRank centrality value.

[0104] 4.4) Feature Learning

[0105] After obtaining the structural features of the subgraph, it is passed through a linear layer to obtain an embedding vector, combined with the original feature vector Z sub , connect the structural feature embedding of each node to Z sub In the corresponding rows in [ ], the model can learn from both features simultaneously. Based on the above discussion, we have obtained the subgraph surrounding each link and the feature vectors of each node within it. The next step is to input the subgraph into the graph neural network for feature learning. To this end, this paper selects the graph attention network as the main network architecture for subgraph processing.

[0106] 4.5) Training strategy

[0107] Imagine that we have obtained a subgraph of a network This includes a set of sampled positive training links and a set of negative training links E n , and with E p There is no intersection. If you directly Generate node embeddings on the training link, which will record the connection information that exists in the training link (because ). Graph neural networks can quickly identify the existence of such links and optimize only by fitting this part of the information. This approach will lead to poor generalization performance. The strategy adopted by this invention is to temporarily n Add to E sub In the new network This operation ensures that both positive and negative training links have the same link existence information in the embedding, thus preventing the graph neural network from classifying links by fitting only this part of information. This strategy is called negative injection.

[0108] 4.6) Backbone network

[0109] Using a graph attention network as the backbone network, after processing the subgraph, the backbone network obtains updated representations of each connected node in the subgraph. To perform link prediction, the new representations of the two nodes u and v on the edge are taken and the Hadamard product is performed, which is the element-wise multiplication of corresponding positions. This result is then input into a fully connected layer, where a ReLU activation function introduces nonlinearity, and then another fully connected layer is used to obtain the final link probability.

[0110] 4.7) Loss Function

[0111] The loss function of the link prediction module uses binary cross entropy loss, which is formulated as:

[0112] BCE(y,p)=-[y·log(p)+(1-y)·log(1-p)],

[0113] Among them, y is the true label, that is, whether there is a link, and p is the probability that the model predicts it to be the positive class.

[0114] Figure 5 It is a system framework diagram for constructing a time series evolution path of a specific implementation method of the battery health state evolution path prediction method based on subgraph representation learning of the present invention.

[0115] In this embodiment, if Figure 5 As shown in FIG, when the evolutionary path prediction system is running, the following steps are included: (1) preprocessing the charging sequence and using the trained encoder to represent the data; (2) clustering the data using the data stream micro-clusters proposed in the present invention to obtain the key states; (3) using the link prediction algorithm based on subgraph representation learning to construct the evolutionary path with missing key states.

[0116] Although the above describes the illustrative specific embodiments of the present invention to facilitate understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concepts of the present invention are protected.

Claims

1. A method for predicting battery health state evolution path based on subgraph representation learning, characterized in that: The following steps are involved: (1) Collect the characteristics of each battery; (2) Clean, serialize and segment the battery data according to the battery characteristics, train an autoencoder to achieve efficient data compression; (3) Extraction of critical states based on data stream microclusters: Data stream clustering is used to extract the critical states of the battery. Each segment is considered as a small microcluster, and the small microclusters of different segments of the same charging sequence are integrated into a larger microcluster. The battery charging sequence of each battery is clustered one by one. After completing the clustering of all charging sequences of a battery, the same operation is performed on the sequence of the next battery. All the microclusters obtained in the end are regarded as the critical states in the battery fault evolution process. (4) State transition graph construction: The identified key states are converted into nodes in a graph structure. The microclusters in which each battery has been clustered are tracked to map the transition path of each battery in different states. The states of each battery transitioning at adjacent time points are connected to form edges in the graph. After deduplication of the transition relationships of all batteries, a complete set of edges is formed, resulting in a graph structure with nodes as key states and edges as state transitions. (5) Subgraph extraction and node labeling: For any pair of nodes to be predicted in the graph structure, subgraph extraction and node labeling are performed. A dynamic range subgraph extraction method is adopted: first, bidirectional expansion is performed with the two target nodes as the starting point, and all reachable neighbor nodes within the preset hop count range are collected to form an initial joint neighborhood set. When the number of nodes exceeds the preset upper limit, the topological importance screening mechanism is activated. By calculating the weighted sum of the inverse of the distance between each node and the two central nodes as the importance score, the preset number of nodes with the highest score are retained to form a closed subgraph. A composite structure labeling mechanism is designed for subgraph nodes: by constructing a structure identification vector containing four-dimensional topological features, it comprehensively reflects the shortest path distance from the node to the two central nodes, the local structure label generated based on the neighborhood topology iteration, and the network centrality value that represents the global importance of the node. It provides extremely rich and highly discriminative initial node features for subsequent representation learning, and explicitly and effectively encodes the topological position, local structure role and global importance of the node relative to the target pair to be predicted. (8) Representation learning based on negative sample enhanced subgraph: After completing subgraph extraction and node labeling, subgraph enhancement is performed, and the enhanced subgraph is learned using an attention graph neural network. When constructing training samples, in order to prevent the graph neural network from simply relying on the existence of positive sample edges for memorization, the present invention introduces a negative sample injection strategy. First, a number of positive sample edges (i.e., edges that actually exist in the original graph) are sampled in the closed subgraph, and an equal number of negative sample edges (i.e., edges that do not exist in the original graph) are collected at the same time. These negative sample edges are then temporarily added to the graph structure to construct an enhanced subgraph. Node embedding calculations are performed on the enhanced subgraph, so that the model can see both the real positive sample edges and the existence of negative sample edges during training, thereby forcing the model to pay attention to richer topological and attribute information, rather than simply judging links by the presence or absence of edges. On the enhanced subgraph after negative sample injection, a multi-head graph attention network is introduced as the backbone architecture to aggregate and represent the contextual information of each node. (9) Prediction of potential links in critical states: After training, the resulting model can be used to predict links for any pair of nodes and output the corresponding probability of existence. For a set of known critical fault states in the battery system, all combinations between these nodes or specific combinations of user interest can be scored, and node pairs with probabilities higher than a preset threshold are predicted as potential fault evolution edges. Subsequently, based on the time sequence and direction information of these predicted edges, they are spliced ​​into a series of directed paths, which intuitively show the evolution trajectory of the fault state in the battery system from early to late stages, providing visual and quantifiable technical support for fault warning and diagnosis.

2. The device evolution path prediction method according to claim 1, characterized in that: In step (8), after completing the subgraph extraction and node labeling, the subgraph is enhanced, and the enhanced subgraph is learned using the attention graph neural network. The probability of the evolution path is calculated as: 4.1) Construct key state transition diagram In the previous paper, we successfully extracted many microclusters through data stream clustering methods. These microclusters are considered to be the key states in the evolution of battery failure. In order to further explore the dynamic transition relationship between these states, this paper first converts the identified microclusters into nodes in a graph structure. Each node not only represents a unique state, but also contains detailed information about the state through its feature vector. For each node in the graph, its corresponding microcluster MC i The eigenvector z i It can be expressed as: With i =C1‖C2‖C3‖C4, The symbol ‖ represents the concatenation operation of a vector sequence. Specifically, C j This characterization refers to microclusters MC i Internal j-th microcluster mc j The cluster center vector of , which not only condenses the essential attributes of the micro-cluster, but also gives detailed and descriptive features to the nodes in the graph. In practice, if the micro-cluster MC i If the jth subordinate micro-cluster is not included, then C j Set it to a one-dimensional matching all-zero vector to maintain the consistency of the feature vector dimension and the convenience of the processing flow. In the process of building a graph model, the edges between nodes reflect the state transition, which is a crucial step. The present invention uses the battery sequence information recorded in the micro-cluster creation phase to construct these edges. Specifically, first traverse all micro-clusters in order, and for each micro-cluster, all sub-micro-clusters m j Traverse and record the battery sequence I contained in each j This step can track the microclusters in which each battery has been clustered, thereby mapping the transition path of each battery in different states. Since each microcluster has a time sequence, the changes in the aggregation of batteries in the microcluster naturally reveal the key state transition process of battery failure evolution. Based on this, the present invention connects the states of each battery transferred at adjacent time points to form edges in the graph. In addition, the process of creating edges may cause multiple edges to be repeated because there are multiple batteries transferred between two microclusters, so it is necessary to remove the duplicate edges. Finally, the key state graph G = (V, E, Z) is obtained, where V is the key state vertex cluster. For any node v i ∈V, its eigenvector z i ∈Z corresponds to micro-cluster MC i The characteristic vector of , E is the transfer relationship obtained from the existing data. Due to the scarcity of data, some transition relationships (i.e., state evolution paths) have not been obtained. Next, link prediction technology is used to predict the transitions between these hidden key states, as shown in Figure 4. 4.2) Subgraph extraction For a graph G = (V, E), for any pair of nodes to be predicted in the graph structure, denoted as nodes u and v, the dynamic range subgraph extraction method is used to perform bidirectional K-hop neighborhood expansion starting from u and v respectively to obtain the joint neighborhood node set like (preset node upper limit), then filter by topological importance: for each node Calculate the importance score: Keep the N with the highest score max nodes, forming a closed subgraph where d(·) is the shortest path distance and α is a smoothing constant. 4.3) Node Marking For subgraph For all assigned composite structure labels in , for any node w, first generate a four-dimensional structural feature vector: r(w)=[d(u,w),d(v,w),WL k (w),PR(w)] T WL k (w) is the topological label generated by the Weisfeiler-Lehman algorithm after k iterations, and PR(w) is the PageRank centrality value. 4.4) Feature Learning After obtaining the structural features of the subgraph, it is passed through a linear layer to obtain an embedding vector, combined with the original feature vector Z sub , connect the structural feature embedding of each node to Z sub In the corresponding rows in [ ], the model can learn from both features simultaneously. Based on the above discussion, we have obtained the subgraph surrounding each link and the feature vectors of each node within it. The next step is to input the subgraph into the graph neural network for feature learning. To this end, this paper selects the graph attention network as the main network architecture for subgraph processing. 4.5) Training strategy Imagine that we have obtained a subgraph of a network This includes a set of sampled positive training links and a set of negative training links E n , and with E p There is no intersection. If you directly Generate node embeddings on the training link, which will record the connection information that exists in the training link (because ). Graph neural networks can quickly identify the existence of such links and optimize only by fitting this part of the information. This approach will lead to poor generalization performance. The strategy adopted by this invention is to temporarily n Add to E sub In the new network This operation ensures that both positive and negative training links have the same link existence information in the embedding, thus preventing the graph neural network from classifying links by fitting only this part of information. This strategy is called negative injection. 4.6) Backbone network Using a graph attention network as the backbone network, after processing the subgraph, the backbone network obtains updated representations of each connected node in the subgraph. To perform link prediction, the new representations of the two nodes u and v on the edge are taken and the Hadamard product is performed, which is the element-wise multiplication of corresponding positions. This result is then input into a fully connected layer, where a ReLU activation function introduces nonlinearity, and then another fully connected layer is used to obtain the final link probability.

Images