Online SOH Estimation Method for Lithium-ion Batteries Based on Graph Neural Networks

Abstract

This invention relates to the field of deep neural network technology, specifically to an online SOH estimation method for lithium batteries based on graph neural networks. The method extracts charging, discharging, and resting recovery segments from the online operating sequence of a lithium battery, extracting segment operating features to generate segment nodes. Based on temporal adjacency, overlapping state-of-charge intervals, similar operating conditions, and consistency with historical SOH estimates, it constructs multiple types of associated edges to form an online degradation dynamic graph. Segment quality gating weights are generated based on the completeness of segment information and the degree of operating condition disturbance. The segment nodes, multiple types of associated edges, and segment quality gating weights are input into a temporal graph neural network for message propagation to obtain the health state embedding. Finally, the deep neural network decoder outputs the current SOH estimate and its confidence level, and updates the online degradation dynamic graph according to the confidence level. This invention can reduce the erroneous influence of short-term operating condition disturbances on health state estimation.

Description

Technical Field

[0001] This invention relates to the field of deep neural network technology, and more specifically to an online estimation method for SOH of lithium batteries based on graph neural networks. Background Technology

[0002] Online estimation of lithium battery health typically relies on voltage, current, temperature, state of charge (SOC), and time-series data collected by the battery management system (BMS). Existing conventional methods often use capacity calibration results as training labels, obtaining complete charge-discharge cycles under experimental or maintenance conditions, calculating the retention ratio of releaseable capacity relative to the initial capacity, and then using the charging curve, discharging curve, or incremental capacity curve as the basis for health estimation. In online applications, a common practice is to select features from the operational data, such as voltage changes, constant-current charging time, temperature rise, and resting recovery voltage within a fixed SOC range. These features are then input into empirical models, filtered models, support vector regression models, or deep neural network models to output the current health estimate. To reduce the impact of curve fluctuations, some methods smooth the input data, remove outliers, correct the rate, and compensate for temperature. They also perform mean processing or trend correction on estimates from adjacent time points to ensure the output closely approximates historical degradation curves. The basic assumption of this type of scheme is that the segment to be estimated and the training segment are comparable in terms of charge state range, sampling completeness and operating conditions. It usually requires a relatively complete data window to support feature calculation. Therefore, it can produce usable results under the conditions of stable data source, complete charge and discharge range and small temperature and rate changes.

[0003] With the application of graph neural networks (GNNs) for battery data modeling, existing technologies have begun to represent individual cells, cycle segments, feature intervals, or historical samples as graph nodes. Graph edges are established based on temporal adjacency, sample similarity, or operating condition similarity. Health status embeddings are then obtained through graph convolution, graph attention, or time-series graph propagation. A more recent approach treats each charge-discharge cycle or fixed-length segment as a node. Node features include voltage slope, temperature statistics, capacity increment, and cycle number. Edge weights are calculated using Euclidean distance, correlation coefficients, or time intervals. The embedding output from the graph neural network is then fed into a deep neural network decoder to obtain a health status estimate. To adapt to online data, some solutions employ rolling windows to truncate segments, using historical estimates as reference points in the graph structure, or setting confidence thresholds at the output to filter estimates that significantly deviate from historical trends. In these solutions, the graph structure is often pre-determined before sample input, edge weight updates depend on surface similarity between segments, and the propagation identity of segments after entering the graph is relatively fixed. Although the above scheme introduces graph structure association, fragment quality, working condition disturbance and degradation association are usually mixed in the same node feature or the same weight. When new fragments enter the graph structure, they are mostly connected to historical fragments according to fixed similarity rules. There is a lack of detailed control over the propagation range of low-quality fragments and the conditions for writing historical healthy memories.

[0004] In actual vehicle or energy storage operation, lithium battery data rarely presents complete, continuous, and stable standard charge-discharge cycles. Online data often consists of segments of short-term charging, shallow discharging, intermittent rest, and operating condition switching. The voltage response under the same health condition can vary depending on the state of charge range, rate, temperature, and sampling gap. Existing estimation methods based on fixed interval characteristics or fixed graph edges struggle to determine whether a segment is sufficient to update the health state when it is incomplete. They also struggle to distinguish between short-term curve shifts caused by changes in operating conditions and long-term degradation changes caused by capacity decay. This leads to low-quality and disturbed segments being assigned similar influence weights to high-quality segments in graph propagation or model regression. When these segments participate in updating historical health memory, subsequent estimates continue to propagate along the contaminated degradation trajectory, causing jumps in health state estimates between adjacent state of charge windows or adjacent operating segments that do not conform to the battery degradation pattern. Since online estimation usually requires continuous results, models find it difficult to avoid these data deficiencies by waiting for a complete cycle. The main technical problem thus arises is: under the conditions of lacking complete charge-discharge cycles and inconsistent quality of online segments, how to establish a graph neural network estimation mechanism that can simultaneously express segment availability, operating condition comparability, and historical degradation correlation, so that the online health status estimation is not erroneously driven by short-term operating condition disturbances. Summary of the Invention

[0005] The purpose of this invention is to provide an online SOH estimation method for lithium batteries based on graph neural networks, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The online SOH estimation method for lithium batteries based on graph neural networks includes: extracting charging segments, discharging segments, and resting recovery segments from the online operation sequence of the lithium battery, extracting segment operation features, and generating segment nodes; constructing multi-type association edges based on the temporal adjacency relationship, state of charge interval overlap relationship, operating condition similarity relationship, and historical SOH estimation consistency relationship between segments to form an online degradation dynamic graph; generating segment quality gating weights based on the completeness of segment information and the degree of operating condition disturbance; inputting the segment nodes, the multi-type association edges, and the segment quality gating weights into a time-series graph neural network for message propagation to obtain a health state embedding; inputting the health state embedding into a deep neural network decoder to output the current SOH estimate and confidence level, and updating the online degradation dynamic graph based on the confidence level.

[0007] Preferably, extracting charging, discharging, and resting recovery segments from the online operation sequence of a lithium battery includes: initially segmenting the online operation sequence according to current direction, state of charge continuity, and resting voltage recovery trend; performing sampling interval consistency verification, state of charge coverage verification, and voltage monotonic change verification on the data after initial segmentation; retaining segments that meet the verification conditions as candidate segments, and recording the start and end times, start and end states of charge, average rate, temperature distribution, and sampling gap location for each candidate segment; for candidate segments with sampling gaps, segment splicing is performed only when the voltage trend, state of charge trend, and current direction on both sides of the gap are consistent, to obtain the charging segment, the discharging segment, and the resting recovery segment.

[0008] Preferably, extracting segment operation features and generating segment nodes includes: establishing a basic feature group, a morphological feature group, and a compensation feature group for each segment node; the basic feature group includes segment duration, state-of-charge interval width, average current, average temperature, and sampling stability; the morphological feature group includes voltage change slope, local curve curvature, normalized capacity increment per unit current, and static recovery slope; the compensation feature group includes voltage response after temperature compensation, capacity increment response after rate normalization, and segment position encoding after state-of-charge interval standardization; concatenating the basic feature group, the morphological feature group, and the compensation feature group into a node feature vector and writing it into the corresponding segment node.

[0009] Preferably, multiple types of associated edges are constructed based on the temporal adjacency relationship, state of charge interval overlap relationship, operating condition similarity relationship, and historical SOH estimation consistency relationship between segments. This includes: establishing temporal adjacency edges between adjacent segment nodes according to the acquisition sequence; establishing interval overlap edges between cross-cycle segment nodes according to the overlap ratio of state of charge intervals; establishing operating condition similarity edges according to the joint distance of average magnification, temperature distribution, and segment duration; establishing health consistency edges according to the consistency of historical SOH estimates, confidence level, and degradation order; and setting edge attribute vectors for different types of associated edges, wherein the edge attribute vectors include edge type identifier, segment interval time, interval overlap information, operating condition difference information, and health residual information.

[0010] Preferably, the segment quality gating weights are obtained by fusing the segment integrity component, the operating condition disturbance component, and the degradation comparability component; the segment integrity component is determined based on the sampling gap ratio, the state of charge coverage range, and the continuity of the voltage curve; the operating condition disturbance component is determined based on the number of current abrupt changes, the rate of temperature change, and the degree of abnormality in static recovery; the degradation comparability component is determined based on the comparability between the candidate segment and the historical healthy anchor segment in the state of charge range, the multiplier range, and the temperature range; the segment integrity component, the operating condition disturbance component, and the degradation comparability component are mapped to gating weights, and the gating weights are bound to the corresponding segment nodes.

[0011] Preferably, the node feature vectors undergo intra-segment consistency encoding before being input into the time-series graph neural network. The intra-segment consistency encoding includes: generating sub-vectors for the basic feature group, the morphological feature group, and the compensation feature group in chronological order; generating consistency markers based on the matching relationships between the voltage change direction, current direction, and state of charge change direction within the same segment; generating perturbation markers based on the morphological offsets before and after temperature compensation; embedding the consistency markers and the perturbation markers into the node feature vectors, and using the embedded node feature vectors as inputs for the segment nodes to participate in message propagation.

[0012] Preferably, the online degradation dynamic graph is updated according to a rolling time window. The update process includes: after receiving a new fragment node within the current rolling time window, firstly calculating the multi-type association edges between the new fragment node and existing fragment nodes; then pruning the neighborhood range of the new fragment node according to the edge attribute vector, retaining adjacent nodes that simultaneously satisfy time continuity, charge state comparability, and operating condition similarity; subsequently, adding historical healthy anchor nodes to the pruned neighborhood range, and setting anchor point identifiers for historical healthy anchor nodes that cannot be covered by the current low-confidence fragments; and generating a dynamic graph subgraph corresponding to the current rolling time window based on the neighborhood range and the anchor point identifiers.

[0013] Preferably, when the segment quality gating weights participate in message propagation, a hierarchical gating process is adopted. The hierarchical gating process includes: using segment nodes with gating weights in the high-quality range as state update nodes; using segment nodes with gating weights in the middle-quality range as neighborhood reference nodes; and using segment nodes with gating weights in the low-quality range as restricted propagation nodes. The state update nodes participate in the joint update of health state embedding and historical SOH graph nodes, the neighborhood reference nodes only participate in message aggregation of adjacent segment nodes, and the restricted propagation nodes only retain unidirectional message inputs related to time adjacency edges.

[0014] Preferably, when the temporal graph neural network performs message propagation on the dynamic graph subgraph, it generates edge type attention coefficients for temporally adjacent edges, interval overlapping edges, operating condition similar edges, and health consistent edges, respectively; multiplies the edge type attention coefficients with the segment quality gating weights to obtain joint propagation weights; aggregates the node feature vectors of adjacent nodes, the health embeddings of historical health anchor nodes, and the segment position encodings of newly added segment nodes according to the joint propagation weights; applies degradation order constraints and short-term reversible fluctuation constraints to the aggregated embeddings to form the health state embedding.

[0015] Preferably, updating the online degradation dynamic graph based on the credibility includes: performing a degradation consistency comparison between the current SOH estimate output by the deep neural network decoder and the SOH sequence corresponding to the historical healthy anchor node; when the current SOH estimate satisfies degradation consistency, the corresponding fragment node belongs to the state update node, and the credibility meets the writing condition, writing the current SOH estimate to a new healthy anchor node; when the corresponding fragment node belongs to the neighborhood reference node or the restricted propagation node, only retaining the temporary association edge between the current SOH estimate and the fragment node; and updating the online degradation dynamic graph according to the new healthy anchor node and the temporary association edge.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention transforms charging, discharging, and resting recovery segments in online operational data into segment nodes. Before node formation, it limits the current direction, charge state continuity, resting recovery trend, sampling interval, voltage monotonicity, and gap splicing conditions, ensuring that data segments entering the graph structure have clear temporal and state boundaries. Segment nodes further carry basic feature groups, morphological feature groups, and compensation feature groups. Multiple types of associated edges describe temporal adjacency, charge state interval overlap, operating condition similarity, and historical health consistency relationships. Segment quality gating weights distinguish the way segments participate in message propagation based on completeness, perturbation degree, and degradation comparability. Therefore, when aggregating neighborhood information, the temporal graph neural network no longer treats all online segments as equally reliable data sources, but determines the transmission strength of health information according to segment quality and edge type relationships. The health state embedding received by the deep neural network decoder simultaneously includes the joint constraints of current segment observation, comparable historical segments, and health anchors. Perturbed segments can be used as reference information in the current estimation but do not directly rewrite the long-term degradation graph; only high-quality segments consistent with the historical degradation order are written into new health anchors. By combining fragment gating, dynamic graph association, and confidence writing, the online estimation process can recover continuous degradation trajectories using fragmented fragments even in the absence of complete charge-discharge cycles. This reduces estimation jumps caused by changes in the state of charge window, rate changes, and temperature disturbances, making the health status output more consistent with the temporal continuity of lithium battery capacity degradation. Because the graph propagation process is simultaneously constrained by edge type attention coefficients and quality gating weights, short-term curve shifts will not spread to all neighboring nodes along a single path, and historical health anchors will not be covered by low-confidence fragments. This creates a stable data loop between the current estimate and long-term health memory, preventing a single anomalous fragment from continuously influencing the estimation results of subsequent time windows.

[0017] 2. This invention also limits the data organization and graph update boundaries in the online estimation process. In the segment truncation stage, constraints are imposed on sampling gaps, state of charge coverage, and voltage trends to prevent incorrectly spliced ​​discontinuous segments from entering subsequent health state inferences. In the node feature stage, the voltage change slope, unit current normalized capacity increment, resting recovery slope, temperature compensation response, and rate normalization response are jointly encoded, and intra-segment consistency and perturbation markers are introduced, enabling the time-series graph neural network to identify matching relationships between voltage, current, and state of charge change directions within the same segment. In the graph update stage, a rolling time window is used to generate dynamic graph subgraphs, and the neighborhood range is pruned using edge attribute vectors, ensuring that new segments only propagate messages to historical segments that are time-continuous, have comparable states of charge, and are similar in operating conditions. Historical health anchors are marked as not to be covered by low-confidence segments, and, in conjunction with hierarchical gating of state update nodes, neighborhood reference nodes, and restricted propagation nodes, the writing range of low- and medium-quality segments to historical health memories is limited. The aforementioned constraints ensure that the graph structure maintains clear data source boundaries and health memory boundaries during online updates, preventing fragments caused by sampling gaps, short-term static anomalies, or current surges from being misused as evidence of degradation. Separating temporary associated edges and new health anchors also isolates the current estimate from long-term degradation graph updates, allowing subsequent fragments to reference current fragment information without inheriting its uncertainty. Edge attributes, node labels, and write conditions jointly constrain the incremental updates of the dynamic graph, ensuring that the online degradation graph maintains the correspondence between node sources, propagation paths, and health status records as the running data expands. This improves the stability and traceability of the continuous online estimation process and reduces the cumulative offset caused by repeated participation of low-confidence fragments in neighborhood propagation, facilitating subsequent backtracking of the fragment nodes and associated edges upon which each health status estimate is based. Attached Figure Description

[0018] Figure 1 This is a flowchart of the overall process for online SOH estimation of lithium batteries based on graph neural networks according to the present invention. Figure 2 This is a flowchart illustrating the online degradation dynamic graph scrolling update and health anchor point writing process of the present invention. Detailed Implementation

[0019] refer to Figure 1In one embodiment, a method for online SOH estimation of lithium batteries based on graph neural networks is provided, comprising: extracting charging segments, discharging segments, and resting recovery segments from the online operation sequence of the lithium battery, extracting segment operation features, and generating segment nodes; constructing multi-type association edges based on the temporal adjacency relationship, state of charge interval overlap relationship, operating condition similarity relationship, and historical SOH estimation consistency relationship between segments to form an online degradation dynamic graph; generating segment quality gating weights based on the completeness of segment information and the degree of operating condition disturbance; inputting the segment nodes, the multi-type association edges, and the segment quality gating weights into a temporal graph neural network for message propagation to obtain a health state embedding; inputting the health state embedding into a deep neural network decoder to output the current SOH estimate and confidence level, and updating the online degradation dynamic graph based on the confidence level.

[0020] In this embodiment, the online operating sequence is the voltage sequence, current sequence, temperature sequence, state of charge sequence, and sampling time sequence already formed during lithium battery management. During processing, the above sequences are aligned using the same time index, isolated data points that cannot correspond to any time index are deleted, and sequence segments that can describe the continuous changes in the battery's operating state are retained. The charging segment is determined by the current direction and the upward trend of the state of charge, the discharging segment is determined by the current direction and the downward trend of the state of charge, and the resting recovery segment is determined by the current approaching zero value range and the terminal voltage recovery trend. The segment operating characteristics are not estimated based on a single instantaneous sampling point, but rather by forming node representations based on the voltage trajectory, state of charge range, temperature change, and current change within the segment. The online degradation dynamic graph uses segment nodes as graph nodes and the temporal continuity relationship and state comparability relationship between segments as graph edges. The temporal graph neural network propagates segment embeddings along the edges in the dynamic graph and forms healthy state embeddings. The deep neural network decoder converts the healthy state embeddings into healthy state estimates and confidence levels through nonlinear mapping. The advantage of this embodiment is that the online estimation does not rely on a complete charge-discharge cycle, but instead uses the available segment data to establish a continuous degradation relationship.

[0021] Specifically, fragment truncation is not simply divided according to a fixed duration, but rather searches for local intervals in the operating sequence that can express changes in the electrochemical response. The voltage sequence is used to express the shape of the external terminal voltage changes with charging and discharging, the current sequence is used to identify the operating modes of charging, discharging, and resting recovery, the temperature sequence is used to characterize the impact of thermal disturbances on the voltage response, the state of charge sequence is used to determine whether the fragments are in comparable capacity ranges, and the sampling time sequence is used to determine whether there are excessively long sampling intervals within the fragments. When generating fragment nodes, a unique fragment identifier, start and end time identifier, start and end state of charge identifier, fragment type identifier, and feature vector identifier are established for each fragment, so that subsequent graph construction can distinguish fragments from different sources. The working principle is to move the fragment quality judgment forward to before graph propagation, represent the differences in operating conditions in the edge attributes, and use the historical health state as anchor information in the graph structure, so that low-quality fragments will not affect the degradation state memory in the same way as high-quality fragments. The advantage of this embodiment is that there is a clear correspondence between the fragment source, fragment state, and graph propagation path.

[0022] Furthermore, the online degradation dynamic graph expands gradually with the addition of new segments. Whenever a new candidate segment is formed in the online running sequence, the new segment is compared with the existing segments in terms of temporal adjacency, state of charge interval, operating condition similarity, and historical health consistency. Temporal adjacency is used to connect adjacent running segments, state of charge interval is used to connect cross-time segments in the same or similar state of charge range, operating condition similarity is used to limit segments with large differences in temperature, magnification, and duration from directly transmitting health information, and historical health consistency is used to establish a degradation relationship between the current segment and a trusted health anchor. The time-series graph neural network encodes different edge types separately during message propagation and then introduces segment quality gating weights into the propagation weight calculation to avoid short-term abnormal segments from having too wide a range of neighborhood influences in the graph. After the decoder outputs the current health state estimate, it also uses the credibility as a condition for whether to write it into the dynamic graph. The advantage of this embodiment is that a closed loop is formed between the current estimate and the long-term degradation graph update.

[0023] In one embodiment, the fragment quality gating weight is determined by fragment integrity, degree of operational disturbance, and degradation comparability, and can be calculated using the following formula:

[0024] in, This represents the segment quality gating weight for the i-th segment node. The closer the value is to 1, the more suitable the segment is for participating in health status propagation. Indicating fragment completeness, its physical meaning is the degree of usability formed by the combined factors of sampling continuity, state of charge coverage, and voltage curve continuity. This indicates the degree of disturbance to the operating conditions, and its physical meaning is the intensity of the disturbance to the segment observation caused by sudden changes in current, temperature changes, and abnormal resting conditions. Indicating degradation comparability, its physical meaning is the degree to which the current segment and the historical healthy anchor segment are comparable across the states of charge, rate range, and temperature range. , and These represent the fusion weights for completeness, degree of perturbation, and comparability, respectively. This represents exponential operations with the natural constant as the base. For example, when , , , , , When the value inside the parentheses is 0.40×0.80-0.35×0.30+0.25×0.70=0.39, substituting it in, we get... The calculated result is approximately 0.596. The advantage of this embodiment is that the gating weight expresses the available information and the perturbation information on the same scale.

[0025] Table 1 is a table of fragment node features. Table 1 is used to explain the main data items written when generating node feature vectors for different fragment types and their usage positions in subsequent graph propagation.

[0026] Table 1. Fragment Node Feature Composition Table

[0027] In one embodiment, extracting charging, discharging, and resting recovery segments from the online operation sequence of a lithium battery includes: initially segmenting the online operation sequence according to current direction, state of charge continuity, and resting voltage recovery trend; performing sampling interval consistency verification, state of charge coverage verification, and voltage monotonicity change verification on the data after initial segmentation; retaining segments that meet the verification conditions as candidate segments; and recording the start and end times, start and end states of charge, average rate, temperature distribution, and sampling gap location for each candidate segment; for candidate segments with sampling gaps, only the voltage trend and state of charge on both sides of the gap are checked. When the state trend and current direction are consistent, the segments are spliced ​​to obtain the charging segment, the discharging segment, and the resting recovery segment. In specific processing, the interval between continuous sampling times is compared with the median sampling interval within the segment. If the current signs on both sides of a gap are the same and the direction of change of state of charge is the same, the gap mark is allowed to be retained and spliced. If the operating mode on both sides of the gap changes, it is split into two candidate segments. The resting recovery segment also needs to satisfy that the voltage change direction within the resting interval has a recovery trend. The advantage of this embodiment is that the interference of non-continuous operating intervals on subsequent estimation can be blocked during the segment truncation stage.

[0028] In a preferred embodiment, the voltage monotonicity check does not require all sampling points within a segment to be strictly monotonic. Instead, it performs a majority consistency judgment on the voltage change direction within the segment and identifies short-term noise by combining local swing amplitude. Specifically, a symbol sequence is established for adjacent voltage differences within the segment. If the change opposite to the segment type in the symbol sequence only appears in a few local locations and the adjacent charge states still maintain the change direction corresponding to the segment type, then the opposite change is recorded as a disturbance marker instead of directly discarding the segment. The charge state coverage range check is judged by the overlap between the current segment's actual coverage range and the historical anchor point range. The sampling interval consistency check is judged by the dispersion of the interval between adjacent sampling times within the segment. All check results are written into the basic feature group or gating weight calculation input of the segment node. The spliced ​​segment retains the gap position index, enabling the time-series graph neural network to identify incomplete intervals within the segment during message propagation. The advantage of this embodiment is that short-term noise will not cause excessive segment discarding, while incomplete information is still explicitly recorded.

[0029] In one embodiment, extracting segment operation features and generating segment nodes includes establishing a basic feature group, a morphological feature group, and a compensation feature group for each segment node. The basic feature group includes segment duration, state-of-charge interval width, average current, average temperature, and sampling stability. The morphological feature group includes voltage change slope, local curve curvature, normalized capacity increment per unit current, and resting recovery slope. The compensation feature group includes temperature-compensated voltage response, rate-normalized capacity increment response, and segment position encoding after state-of-charge interval standardization. The basic feature group, the morphological feature group, and the compensation feature group are then combined... The state feature group and the compensation feature group are concatenated into a node feature vector and written into the corresponding segment node. In specific implementation, the segment duration is obtained from the start and end time difference, the width of the state of charge interval is obtained from the start and end state of charge difference, the sampling stability is obtained from the sampling interval dispersion, the voltage change slope is obtained from the rate of change of terminal voltage relative to time or state of charge, the curvature of the local curve is obtained from the difference between adjacent slopes, and the normalized capacity increment per unit current is obtained from the state of charge change and current integral relationship within the segment. The advantage of this embodiment is that the node features simultaneously express the basic conditions, degradation morphology and operating condition compensation state of the segment.

[0030] Furthermore, before entering the temporal graph neural network, the node feature vectors can be used to form extended node representations through intra-segment consistency encoding, which can be expressed by the following formula:

[0031] in, This represents the extended node feature vector of the i-th fragment node. Represents the basic feature set vector. Represents a vector of morphological feature groups. Represents the compensation feature group vector. Indicates the consistency marker within the fragment. This indicates an indicator operation; the expression returns 1 if the condition within the parentheses is true and 0 if it is false. This represents the overall change in voltage across the segment. Indicates the encoding of the current direction in the segment. This represents the total change in the state of charge of a segment. Indicates a disturbance marker. This represents the average voltage response before compensation. This represents the average voltage response after temperature compensation, for example, during a charging segment. , , When, the product is greater than 0 and ,when , hour, The advantage of this embodiment is that the node vector can preserve the consistency of the curve direction and the traces of temperature disturbance within the segment.

[0032] In one embodiment, multiple types of associated edges are constructed based on the temporal adjacency, state of charge (SCO) interval overlap, operating condition similarity, and historical SOH estimation consistency among segments. This includes establishing temporal adjacency edges between adjacent segment nodes according to the acquisition sequence; establishing interval overlap edges between cross-cycle segment nodes according to the SCO interval overlap ratio; establishing operating condition similarity edges based on the joint distance of average scaling factor, temperature distribution, and segment duration; and establishing health consistency edges based on the consistency of historical SOH estimates, confidence levels, and degradation order. Edge attribute vectors are set for different types of associated edges. These attribute vectors include edge type identifiers, segment interval time, interval overlap information, operating condition difference information, and health residual information. Specifically, temporal adjacency edges preserve the operating order between adjacent segments; interval overlap edges connect segments within the same SCO range; operating condition similarity edges connect segments with similar scaling factors and temperature states; and health consistency edges connect the current segment with historical health anchors. The edge attribute vectors serve as input for the attention calculation of the graph neural network. The advantage of this embodiment is that the multidimensional relationships between segments are decomposed into different edge types.

[0033] In a preferred embodiment, multi-type associated edges are not fixedly connected using a single similarity threshold. Instead, candidate neighborhoods are established according to edge type, and pruning is performed within the candidate neighborhoods. Temporal adjacency edges retain directional information between adjacent sampled segments, enabling message propagation to distinguish between the transmission of historical segments to the current segment and the writing back of the current segment to historical segments. Interval overlap edges form part of the edge attributes based on the ratio of the overlap length of the charge state interval to the union length of the two segments. Operating condition similarity edges constitute the joint distance based on the average ratio difference, temperature distribution difference, and duration difference. Health consistency edges are jointly established based on the estimated value, confidence level, and formation time of historical health anchors, ensuring that historical segments with low confidence are not used as high-level anchors. The edge type identifier uses a learnable embedding vector written into the edge encoding input of the temporal graph neural network. Segment interval time, interval overlap information, operating condition difference information, and health residual information jointly participate in the subsequent joint propagation weight calculation. The advantage of this embodiment is that similar segments and comparable segments are processed separately.

[0034] In one embodiment, the segment quality gating weight is obtained by fusing the segment integrity component, the operating condition disturbance component, and the degradation comparability component. The segment integrity component is determined based on the sampling gap ratio, the state of charge coverage range, and the continuity of the voltage curve. The operating condition disturbance component is determined based on the number of current mutations, the rate of temperature change, and the degree of abnormality in resting recovery. The degradation comparability component is determined based on the comparability between the candidate segment and the historical healthy anchor segment in the state of charge range, the rate range, and the temperature range. The segment integrity component, the operating condition disturbance component, and the degradation comparability component are mapped to gating weights, and the gating weights are bound to the corresponding segment nodes. Specifically, the higher the sampling gap ratio, the lower the segment integrity component; the closer the state of charge coverage range is to the historical anchor coverage range, the higher the degradation comparability component; and the higher the number of current mutations and the rate of temperature change, the higher the operating condition disturbance component. The mapping process uses a monotonic function to increase the gating weight when the integrity and comparability are improved, and to decrease the gating weight when the disturbance level is improved. The advantage of this embodiment is that whether a segment participates in the health memory update is controlled by calculable quality information.

[0035] Furthermore, the gating weights, after being bound to fragment nodes, do not change the original node feature vectors. Instead, they participate in neighborhood aggregation as external control variables for message propagation, allowing the propagation range of the same fragment across different edge types to be adjusted. Charging fragments, discharging fragments, and static recovery fragments all form comparable weights using the same gating calculation method. The operating condition disturbance component does not directly delete fragments but reduces the impact of fragments on neighborhood health embedding during message propagation. The degradation comparability component is used to determine whether a fragment can form a highly reliable connection with historical health anchors. The fragment integrity component is used to determine whether a fragment can participate in long-term health graph node updates. The gating weights also retain the input components at the time of formation, facilitating subsequent backtracking of why a fragment was used as a state update node, neighborhood reference node, or restricted propagation node. The advantage of this embodiment is that fragment quality control and fragment feature preservation are separated from each other.

[0036] In one embodiment, the node feature vectors undergo intra-segment consistency encoding before being input into the time-series graph neural network. This intra-segment consistency encoding includes generating sub-vectors for the basic feature group, the morphological feature group, and the compensation feature group in chronological order; generating consistency markers based on the matching relationships between the voltage change direction, current direction, and state of charge change direction within the same segment; generating perturbation markers based on the morphological offsets before and after temperature compensation; embedding the consistency markers and perturbation markers into the node feature vectors; and using the embedded node feature vectors as input for the segment nodes to participate in message propagation. Specifically, in a charging segment, when the overall voltage and state of charge increase and the current direction conforms to the charging definition, the consistency marker is considered a match; in a discharging segment, when the overall voltage and state of charge decrease and the current direction conforms to the discharging definition, the consistency marker is considered a match. If the morphological difference before and after temperature compensation is significant, the perturbation marker is increased and written into the extended node feature vector. The advantage of this embodiment is that graph propagation can identify the degree of self-consistency within a segment.

[0037] In this embodiment, after the basic feature group, morphological feature group, and compensation feature group generate sub-vectors, they are embedded in position according to the time order within the segment. Position embedding is used to express whether the feature is generated in the starting interval, middle interval, or ending interval of the segment. The voltage change direction is determined by the difference between the segment endpoints and the majority slope direction within the segment. The current direction is determined by the dominance of the current sign within the segment. The charge state change direction is determined by the difference between the starting and ending charge states. The disturbance label does not depend solely on the temperature value itself, but is determined by the offset of the morphological features before and after temperature compensation. Therefore, it can distinguish between high-temperature stable segments and rapidly changing temperature segments. When the embedded node feature vector enters the graph neural network, it is simultaneously input with the edge attribute vector, so that the consistency within the node and the comparability between nodes jointly determine the propagation path. The advantage of this embodiment is that the erroneous direction change within the segment will not be masked as ordinary feature fluctuation.

[0038] refer to Figure 2 In one embodiment, the online degradation dynamic graph is updated according to a rolling time window. The update process includes receiving a new fragment node within the current rolling time window, first calculating the multi-type association edges between the new fragment node and existing fragment nodes, then pruning the neighborhood range of the new fragment node according to the edge attribute vector, retaining adjacent nodes that simultaneously satisfy time continuity, charge state comparability, and operating condition similarity, then adding historical health anchor nodes to the pruned neighborhood range, and setting anchor identifiers for historical health anchor nodes that cannot be covered by the current low-confidence fragments, generating a dynamic graph subgraph corresponding to the current rolling time window based on the neighborhood range and the anchor identifiers. In specific implementation, the rolling time window is not required to contain a complete charge-discharge cycle, but rather to receive nodes that have passed fragment verification. After the new fragment node enters the graph structure, it only propagates with the pruned neighborhood nodes. The historical health anchor node retains the corresponding health state estimate, confidence level, and write time. The advantage of this embodiment is that the neighborhood range will not expand indefinitely when the dynamic graph is updated incrementally with online fragments.

[0039] Preferably, the dynamic graph subgraph is constructed with the current segment node as the center. The neighborhood range of the center node is jointly determined by temporal continuity, charge state comparability, and operating condition similarity. Temporal continuity is used to retain segments near the current operating stage. Charge state comparability is used to select segments with comparable value within similar capacity ranges. Operating condition similarity is used to exclude nodes with excessively large temperature or rate differences. When historical health anchor nodes are connected to the subgraph, they are not completely restricted by ordinary neighborhood pruning, but are connected to the center node according to health consistency edges. Anchor point identification ensures that low-confidence segments can only read degradation reference information from historical health anchors and cannot overwrite the existing health state records of historical health anchors. The dynamic graph subgraph can retain temporary associated edges after message propagation ends, and can also generate new health anchors according to confidence. The advantage of this embodiment is that online estimation has local computational boundaries and historical memory protection boundaries.

[0040] In one embodiment, when a temporal graph neural network performs message propagation on a dynamic graph subgraph, it generates edge type attention coefficients for temporally adjacent edges, interval overlapping edges, edge with similar operating conditions, and edge with consistent health, respectively. The joint propagation weight is obtained by multiplying the edge type attention coefficients by the segment quality gating weights, which can be calculated using the following formula:

[0041] in, Let $\mathbf{j}$ represent the edge type attention coefficient between the $j$-th adjacent segment node and the $i$-th center segment node in the $k$-th associated edge. and These represent the query embeddings of the central fragment node and its adjacent fragment nodes, respectively. This represents the learnable matrix corresponding to the k-th type of associated edge. This represents the edge attribute mapping vector corresponding to the k-th type of associated edge. This represents the edge attribute vector between the i-th node and the j-th node. Let i represent the set of neighborhoods of the i-th node under the k-th type of associated edge. Indicates the weight of joint communication. This represents the segment quality gating weight of adjacent segment nodes. For example, after normalizing the exponent terms of two adjacent nodes of the same edge type, we get... and And the corresponding gating weights are respectively and At that time, the joint propagation weights were 0.56 and 0.12, respectively. The advantage of this embodiment is that the edge relationship strength and fragment quality jointly constrain the message source.

[0042] Furthermore, temporal graph neural networks can use multiple rounds of propagation to form healthy state embeddings, and node updates can be performed using the following formula:

[0043] in, Indicates that the i-th fragment node is in the... Node embedding during round propagation This indicates the node embedding after the next round of propagation. Represents a nonlinear transformation function. Let K represent the self-maintaining mapping matrix of the central node, and let K represent the set of associated edge types. This represents the mapping matrix of adjacent nodes corresponding to the k-th type of associated edge. Indicates the weight of joint communication. This represents a deep neural network decoder. Let represent the estimated health status of the i-th segment node, and T represent the number of propagation rounds. For example, when the self-maintaining term of a certain dimension's central node is 0.20, and the contributions of two adjacent nodes after mapping are 0.50 and 0.30 respectively, with joint propagation weights of 0.56 and 0.12 respectively, the aggregated input is 0.20 + 0.56 × 0.50 + 0.12 × 0.30 = 0.516. Using the hyperbolic tangent function, the embedding of this dimension is approximately 0.474. The advantage of this embodiment is that graph propagation can incorporate information from comparable historical segments into the current segment according to controlled weights.

[0044] Table 2 is the fragment node propagation identity table. Table 2 is used to explain how the gating weight, credibility, and write conditions jointly determine the different processing methods of fragment nodes in the dynamic graph update process.

[0045] Table 2. Fragment Node Propagation Identity Table

[0046] In one embodiment, a hierarchical gating process is used when segment quality gating weights participate in message propagation. The hierarchical gating process includes using segment nodes with gating weights in the high-quality range as state update nodes, using segment nodes with gating weights in the intermediate-quality range as neighborhood reference nodes, and using segment nodes with gating weights in the low-quality range as restricted propagation nodes. The state update nodes participate in the joint update of health state embedding and historical SOH graph nodes. The neighborhood reference nodes only participate in message aggregation of adjacent segment nodes. The restricted propagation nodes only retain unidirectional message inputs related to temporally adjacent edges. In specific implementation, the high-quality range, intermediate-quality range, and low-quality range can be determined by the segment quality distribution in the training data. During propagation, the state update nodes are allowed to read information from multiple types of associated edges and generate health anchors when the credibility condition is met. The neighborhood reference nodes are used to supplement the local morphological information of adjacent segments. The restricted propagation nodes retain the minimum temporal connection to avoid graph structure breakage. The advantage of this embodiment is that different quality segments are given different propagation permissions.

[0047] In this embodiment, hierarchical gating processing and rolling time window updates are executed in conjunction. After a new segment enters the current dynamic graph subgraph, its propagation identity is first determined based on the segment quality gating weight. Then, the types of edges that can participate are set according to the propagation identity. State update nodes can participate in the propagation of temporally adjacent edges, interval overlapping edges, operating condition similar edges, and health-consistent edges. Neighborhood reference nodes can participate in the propagation of temporally adjacent edges and some operating condition similar edges. Restricted propagation nodes only receive unidirectional input from the previous adjacent segment or historical anchor point and do not actively propagate to other segments. If a restricted propagation node is subsequently confirmed to be in a continuous degradation trajectory as new segments arrive, its gating weight can be recalculated and its propagation identity updated. The propagation identity update does not change the original segment data, but only changes its propagation permission in the current subgraph. The advantage of this embodiment is that the disturbed segment can retain observation information without polluting long-term health memory.

[0048] In one embodiment, after receiving the health state embedding from the temporal graph neural network output, the deep neural network decoder uses a multi-layer nonlinear mapping to output the current health state estimate, and generates a confidence level based on segment quality gating weights, degradation order consistency, and historical anchor point bias. The result can be calculated and written into the judgment using the following formula:

[0049]

[0050] in, This indicates the confidence level of the health status estimate corresponding to the i-th segment node. , and These represent the fusion weights for gating weight, degradation order consistency, and historical anchor point bias, respectively. The score represents the consistency of the degradation order. This represents an estimate of the current health status. This indicates the health status value corresponding to the most recently written historical health anchor. Indicates the conditions for writing credibility. Indicates the conditions for writing the gating weight. This indicates the allowable short-term reversible fluctuation margin. This indicates that a new health anchor point is written when the result of the judgment is 1. For example, when... , , , , , When the calculated value inside the parentheses is 0.73, we get... It is approximately 0.675, if , and If it is established, then The advantage of this embodiment is that there is a computable constraint relationship between the decoding output and the graph update write.

[0051] In one embodiment, updating the online degradation dynamic graph based on the confidence level includes performing a degradation consistency comparison between the current SOH estimate output by the deep neural network decoder and the SOH sequence corresponding to the historical healthy anchor node. When the current SOH estimate satisfies degradation consistency, the corresponding fragment node belongs to the state update node, and the confidence level meets the writing condition, the current SOH estimate is written to a new healthy anchor node. When the corresponding fragment node belongs to the neighboring reference node or the restricted propagation node, only the temporary association edge between the current SOH estimate and the fragment node is retained. The online degradation dynamic graph is updated according to the new healthy anchor node and the temporary association edge. Specifically, the degradation consistency comparison does not require adjacent estimates to be absolutely monotonic, but rather determines whether the current estimate deviates from the historical degradation direction within the allowable short-term reversible fluctuation margin. After the state update node generates a new healthy anchor, the new healthy anchor records the current health state estimate, confidence level, fragment identifier, and writing time. The temporary association edge records the correspondence between the unwritten estimate and the fragment node. The advantage of this embodiment is that low-confidence estimates can be referenced but will not become a long-term degradation benchmark.

[0052] Furthermore, when updating the online degradation dynamic graph, two types of historical records are retained: one is the health anchor record, and the other is the temporary association record. The health anchor record is used for the construction of healthy consistent edges and the comparison of degradation order in the subsequent rolling time window. The temporary association record is used to retain fragment evidence in the current estimation process and participate in the neighborhood reference when a new fragment arrives. If multiple subsequent fragments form a stable relationship with a certain temporary association record in terms of time continuity, charge state comparability, and operating condition similarity, and the confidence after re-decoding meets the writing conditions, the corresponding temporary association record can be converted into a new health anchor record. If subsequent fragments show that the temporary association record is caused by current mutation, temperature change, or sampling gap, the temporary state is maintained and its propagation range is limited. The historical health anchor node is set with an anchor identifier that cannot be covered by low confidence fragments. When the anchor identifier is updated, only higher confidence state update nodes are allowed to establish successor anchors. Neighborhood reference nodes or restricted propagation nodes are not allowed to replace existing anchors. The advantage of this embodiment is that long-term health memory and short-term estimation evidence are stored in layers in the dynamic graph.

[0053] In a preferred embodiment, the temporal graph neural network and the deep neural network decoder can be jointly trained to obtain initial parameters. The training samples consist of historical running segments with health status labels. The historical segments are used to generate a training dynamic graph according to the same segment truncation, feature extraction, edge construction, and gating calculation methods as in the online stage. The training objectives include health status estimation error, degradation sequence consistency constraint, and perturbed segment propagation constraint. The health status estimation error is used to make the decoder output close to the labeled health status. The degradation sequence consistency constraint is used to limit the outputs that do not conform to the aging direction in the same battery historical anchor point sequence. The perturbed segment propagation constraint is used to reduce the contribution of low-gated weight segments to the neighborhood aggregation result. In the online running stage, the basic network parameters are fixed or only the local embeddings related to the target battery health anchor point are updated, so that the model does not need to be frequently retrained on the target battery. The advantage of this embodiment is that the degradation relationship formed by the offline calibration data can be transferred to the online segment estimation process through the dynamic graph structure.

[0054] In this embodiment, the training dynamic graph and the online degradation dynamic graph maintain the same data structure. The fragment nodes during training also include basic feature groups, morphological feature groups, compensation feature groups, consistency markers, and perturbation markers. The edge attribute vectors during training also include edge type identifiers, fragment interval time, interval overlap information, working condition difference information, and health residual information. The edge type mapping matrix and node mapping matrix after training are used for joint propagation weight calculation in the online stage. The health anchor writing rules formed in the training stage are used in the online stage to determine whether to update the long-term health memory. If the fragment quality distribution in the online data changes relative to the training data, the gating classification boundary can be adjusted only based on the recent fragment quality statistics without changing the learned graph neural network parameters and deep neural network decoder parameters. The advantage of this embodiment is that the input structure of the training process and the online inference process is consistent, reducing the estimation bias caused by changes in the data organization method.

[0055] In a preferred embodiment, the online estimation process employs a method of preserving evidence while restricting the propagation of anomalous segments. When a candidate segment has a sampling gap, a sudden current change, or a large temperature variation, the candidate segment is not directly deleted. Instead, the sampling gap location, disturbance marker, and segment integrity component are written into the segment node. If the gating weight is in the low-quality range, the segment node is used as a restricted propagation node and is only allowed to receive one-way message inputs from adjacent edges in the preceding time or historical health anchors. If subsequent adjacent segments maintain continuity with the restricted propagation node in terms of voltage trend, state of charge trend, and operating condition recovery state, the gating weight is recalculated and the restricted propagation node is adjusted to a neighborhood reference node. If subsequent segments show that the segment corresponding to the restricted propagation node still cannot form a comparable relationship with any historical anchor, the temporary association record is retained. The advantage of this embodiment is that the observation information of anomalous segments is not discarded indiscriminately, nor is it directly written into the long-term health graph structure.

[0056] In a preferred embodiment, the backtracking of health status estimation results is accomplished jointly by fragment nodes, edge attribute vectors, joint propagation weights, propagation identities, and write records. When it is necessary to determine the formation source of the health status estimate at a certain moment, the extended node feature vector of the corresponding central fragment node can be read, and the time-adjacent edges, interval overlapping edges, working condition similar edges, and health-consistent edges participating in the aggregation can be queried. The main adjacent fragment sources are obtained by sorting according to the joint propagation weights. Then, the participation mode of each adjacent fragment in the health status embedding is determined according to the fragment quality gating weights and propagation identities. If the health status estimate is written to a new health anchor, the credibility, degradation consistency comparison results, and original health anchor identifier at the time of writing are further read. If the health status estimate only forms a temporary associated edge, the quality reasons and disturbance reasons for not writing are recorded. The advantage of this embodiment is that each online estimation can be technically traced by the nodes, edges, and write records in the graph structure.

[0057] In a preferred embodiment, the deep neural network decoder can be composed of multiple fully connected nonlinear mappings. The input is the health state embedding output by the time-series graph neural network and the gating weights of the current segment node. The output is the health state estimate and confidence level. During decoder training, the health state estimate is constrained to a numerical range corresponding to the battery capacity retention rate. The confidence level output is jointly associated with segment integrity, degradation comparability, and historical anchor point deviation. In the online stage, when the deviation between the health state estimate output by the decoder and the most recent health anchor point exceeds the short-term reversible fluctuation margin, even if the segment node is a state update node, a new health anchor point is not directly written. Instead, it is converted into a temporary association record and waits for subsequent segment verification. If subsequent segments form continuous degradation evidence, the judgment result is recalculated and written. The advantage of this embodiment is that the decoder output does not independently determine the long-term health memory update.

[0058] In a preferred embodiment, different types of segments can be used collaboratively in the same online degradation dynamic graph. The charging segment mainly provides voltage rise pattern, normalized capacity increment per unit current, and temperature compensation response. The discharging segment mainly provides voltage drop pattern, state of charge range change, and rate normalization response. The resting recovery segment mainly provides recovery voltage slope and polarization release related pattern. During graph construction, different types of segments are not mixed into homogeneous nodes. Instead, segment type identifiers and segment position codes are written into node features. The connection relationship between different types of segments is distinguished by edge type identifiers. If the current rolling time window only contains short-term charging segments, the health anchor information in historical discharging segments or resting recovery segments can be referenced through interval overlapping edges and health consistency edges. If the current rolling time window only contains short-term discharging segments, the health state embedded by historical charging segments can be referenced through operating condition similarity edges. The advantage of this embodiment is that different operating segments can form controllable complementarity in a unified graph structure.

[0059] In a preferred embodiment, the storage structure of the online degradation dynamic graph may include a fragment node table, an edge attribute table, a health anchor table, and a temporary association table. The fragment node table records fragment identifier, fragment type, start and end times, start and end charge states, node feature vectors, gating weights, and propagation identity. The edge attribute table records the starting fragment identifier, ending fragment identifier, edge type identifier, fragment interval time, interval overlap information, operating condition difference information, and health residual information. The health anchor table records the written health state estimate, confidence level, forming fragment identifier, and writing time. The temporary association table records the estimation results, fragment identifier, reason for not writing, and subsequent verification status of fragments not written to long-term health memory. When updating the rolling time window, only local records related to the currently added fragment are read to generate a dynamic graph subgraph. After propagation and decoding are completed, the corresponding records are updated. The advantage of this embodiment is that graph computation data and health memory data are managed in a hierarchical manner.

[0060] In this embodiment, the complete online estimation process takes fragment verification as the entry point and dynamic graph update as the output. After candidate fragments undergo sampling interval consistency verification, charge state coverage verification, and voltage monotonic change verification, they form fragment nodes. After being encoded by basic feature group, morphological feature group, compensation feature group, consistency marker, and disturbance marker, the fragment nodes are connected to the online degradation dynamic graph. The dynamic graph generates multiple types of associated edges based on temporal adjacency, charge state interval overlap, operating condition similarity, and historical health consistency. The message propagation of the time-series graph neural network is controlled by gating weight, edge type attention coefficient, and hierarchical propagation identity. After the deep neural network decoder outputs the health state estimate and confidence level, it determines the generation of new health anchors or temporary associated edges according to degradation consistency, propagation identity, and write conditions. The advantage of this embodiment is that incomplete online fragments can form a health state estimation closed loop through quality constraints, relationship constraints, and write constraints.

Claims

1. A method for online estimation of state of harmonics (SOH) of lithium batteries based on graph neural networks, characterized in that, include: The charging segment, discharging segment, and resting recovery segment are extracted from the online operation sequence of the lithium battery, and the segment operation features are extracted to generate segment nodes; Based on the temporal adjacency relationship between segments, the overlap relationship of charge state intervals, the similarity relationship of operating conditions, and the consistency relationship of historical SOH estimates, multiple types of associated edges are constructed to form an online degradation dynamic graph; Segment quality gating weights are generated based on the completeness of segment information and the degree of operational disturbance. The fragment nodes, the multi-type association edges, and the fragment quality gating weights are input into a temporal graph neural network for message propagation to obtain the health status embedding. The health status is embedded into the input deep neural network decoder, which outputs the current SOH estimate and confidence level, and updates the online degradation dynamic graph based on the confidence level.

2. The online SOH estimation method for lithium batteries based on graph neural networks according to claim 1, characterized in that, The charging segment, discharging segment, and resting recovery segment are extracted from the online operation sequence of the lithium battery, including: The online operation sequence is initially segmented according to the current direction, the continuity of the state of charge, and the recovery trend of the static voltage. Perform sampling interval consistency verification, state of charge coverage verification, and voltage monotonic change verification on the data after initial segmentation. Segments that meet the verification conditions are retained as candidate segments, and the start and end times, start and end states of charge, average magnification, temperature distribution and sampling gap location are recorded for each candidate segment. For candidate segments with sampling gaps, segments are spliced ​​only when the voltage trends, state of charge trends, and current directions on both sides of the gap are consistent, to obtain the charging segment, the discharging segment, and the static recovery segment.

3. The online SOH estimation method for lithium batteries based on graph neural networks according to claim 1, characterized in that, Extract fragment runtime features and generate fragment nodes, including: For each of the aforementioned fragment nodes, a basic feature group, a morphological feature group, and a compensation feature group are established; The basic feature set includes segment duration, state-of-charge interval width, average current, average temperature, and sampling stability. The morphological feature group includes voltage change slope, local curve curvature, normalized capacity increment per unit current, and static recovery slope. The compensation feature set includes the voltage response after temperature compensation, the capacity increment response after rate normalization, and the segment position encoding after standardization of the state of charge interval. The basic feature group, the morphological feature group, and the compensation feature group are concatenated into a node feature vector, and then written into the corresponding segment node.

4. The online SOH estimation method for lithium batteries based on graph neural networks according to claim 1, characterized in that, Based on the temporal adjacency, overlapping states of charge (SOC) intervals, similar operating conditions, and consistency of historical state of charge (SOH) estimates among the segments, multiple types of associated edges are constructed, including: Establish temporal adjacency edges between adjacent segment nodes according to the collection time sequence; Establish overlapping edges between nodes of cross-period segments according to the overlap ratio of charge state intervals; Establish working condition similarity edges based on the joint distance of average magnification, temperature distribution, and segment duration; Establish healthy, consistent edges based on historical SOH estimates, reliability, and consistency of degradation order; Set edge attribute vectors for different types of associated edges. The edge attribute vectors include edge type identifier, segment interval time, interval overlap information, working condition difference information, and health residual information.

5. The online SOH estimation method for lithium batteries based on graph neural networks according to claim 2, characterized in that, The segment quality gating weight is obtained by fusing the segment integrity component, the operating condition disturbance component, and the degradation comparability component; The fragment integrity component is determined based on the sampling gap ratio, the coverage range of the state of charge, and the continuity of the voltage curve; The operating condition disturbance components are determined based on the number of current abrupt changes, the rate of temperature change, and the degree of abnormality in the recovery after static conditions. The degradation comparability component is determined based on the comparability between candidate segments and historical healthy anchor segments in the state of charge range, rate range, and temperature range. The fragment integrity component, the operating condition disturbance component, and the degradation comparability component are mapped to gating weights, and the gating weights are bound to the corresponding fragment nodes.

6. The online SOH estimation method for lithium batteries based on graph neural networks according to claim 3, characterized in that, The node feature vectors undergo intra-segment consistency encoding before being input into the temporal graph neural network. This intra-segment consistency encoding includes: Sub-vectors are generated for the basic feature group, the morphological feature group, and the compensation feature group in chronological order. A consistency marker is generated based on the matching relationship between the voltage change direction, current direction, and state of charge change direction within the same segment; Disturbance markers are generated based on the morphological shifts before and after temperature compensation. The consistency tag and the perturbation tag are embedded into the node feature vector, and the embedded node feature vector is used as the input for the fragment node to participate in message propagation.

7. The online SOH estimation method for lithium batteries based on graph neural networks according to claim 4, characterized in that, The online degradation animation is updated according to a rolling time window. The update process includes: After receiving a new fragment node within the current scrolling time window, first calculate the multi-type association edges between the new fragment node and existing fragment nodes; Then, the neighborhood range of the newly added fragment nodes is pruned according to the edge attribute vector, retaining the adjacent nodes that simultaneously satisfy the time continuity, charge state comparability and operating condition similarity. Then, the historical health anchor node is added to the trimmed neighborhood range, and an anchor node identifier that cannot be covered by the current low-trust fragment is set for the historical health anchor node. A dynamic subgraph corresponding to the current scrolling time window is generated based on the neighborhood range and the anchor point identifier.

8. The online SOH estimation method for lithium batteries based on graph neural networks according to claim 5, characterized in that, When fragment quality gating weights participate in message propagation, hierarchical gating is used. Hierarchical gating includes: Use fragment nodes with gating weights in the high-quality range as state update nodes; Use fragment nodes with gating weights in the middle quality range as neighborhood reference nodes; Use fragment nodes with low-quality gating weights as restricted propagation nodes; The state update node participates in the joint update of health state embedding and historical SOH graph nodes, the neighborhood reference node only participates in the message aggregation of adjacent fragment nodes, and the restricted propagation node only retains unidirectional message input related to time adjacency edges.

9. The online SOH estimation method for lithium batteries based on graph neural networks according to claim 7, characterized in that, When the temporal graph neural network performs message propagation on the dynamic graph subgraph, it generates edge type attention coefficients for temporally adjacent edges, interval overlapping edges, working condition similar edges, and health consistent edges, respectively. The joint propagation weight is obtained by multiplying the edge type attention coefficient by the fragment quality gating weight; The node feature vectors of adjacent nodes, the health embeddings of historical health anchor nodes, and the fragment position encodings of newly added fragment nodes are aggregated based on the joint propagation weights. The healthy state embedding is formed by applying degradation order constraints and short-term reversible fluctuation constraints to the aggregated embedding.

10. The online SOH estimation method for lithium batteries based on graph neural networks according to claim 8, characterized in that, Updating the online degradation dynamic graph based on the stated confidence level includes: The current SOH estimate output by the deep neural network decoder is compared with the SOH sequence corresponding to the historical health anchor node for degradation consistency. When the current SOH estimate satisfies degradation consistency, the corresponding fragment node belongs to the state update node, and the confidence level meets the writing condition, the current SOH estimate is written to the new health anchor node. When the corresponding fragment node belongs to the neighboring reference node or the restricted propagation node, only the temporary association edge between the current SOH estimate and the fragment node is retained; The online degradation dynamic graph is updated according to the new health anchor node and the temporary associated edge.

Images