A random-constrained graph liquid neural network lithium battery life prediction method

Abstract

This invention relates to the field of lithium battery health management technology and discloses a stochastically constrained graph-liquid neural network method for predicting the lifetime of lithium batteries. The method includes: collecting a degradation observation sequence of the lithium battery, performing normalization and segmented reconstruction to obtain a windowed degradation subsequence; constructing a degradation graph composed of temporal nodes, step-connecting edges, and global degradation nodes; inputting the degradation graph into a graph-liquid neural network, performing neighborhood message propagation and liquid-gated state updates, aggregating and reading out the updated hidden states of nodes, extracting window-level degradation representations, predicting future degradation increments, and recursively reconstructing the future health state trajectory; introducing a drift-diffusion physical prior based on Wiener stochastic processes, and constructing a joint optimization objective by combining supervised loss and monotonicity constraints; and determining the remaining lifetime of the lithium battery based on the future health state trajectory. This invention exhibits higher prediction accuracy, robustness, and interpretability in both short-lifetime, highly volatile scenarios and long-lifetime, stable scenarios.

Description

Technical Field

[0001] This invention relates to the field of lithium battery health management technology, specifically to a stochastically constrained graph-liquid neural network method for predicting lithium battery life. Background Technology

[0002] Lithium-ion batteries are widely used in electric vehicles, aerospace, grid energy storage, and portable devices. Their service safety and economic efficiency depend heavily on the accurate assessment of degradation status and remaining life. As the number of cycles increases, battery capacity typically shows a long-term downward trend. However, this macro-degradation trend is superimposed with short-term fluctuations, stage transitions, and random disturbances caused by changes in operating conditions. This makes remaining life prediction essentially a degradation modeling problem with structural dependence, continuous-time evolution, and random uncertainty. Current data-driven methods, such as support vector machines, Gaussian process regression, bidirectional long short-term memory networks (Bi-LSTM), and transformer models, are able to learn degradation patterns from historical data. In recent years, graph neural networks have shown structural awareness advantages in relational modeling, and liquid neural networks have good adaptability in continuous-time representation, providing new technical paths for refined modeling of degradation processes.

[0003] However, in practical battery health management scenarios facing strong fluctuations, time-varying operating conditions, or long-life degradation, the above methods still face several constraints in engineering applications. First, due to the difficulty in explicitly incorporating knowledge of degradation mechanisms, the predicted trajectories are prone to local rebounds inconsistent with physical laws, affecting the physical rationality of the prediction results. Second, since most existing methods treat degradation as a one-dimensional sequence, it is difficult to simultaneously characterize the relationship between local cyclic dependence and global degradation stages, resulting in insufficient utilization of degradation structure information. For scenarios with strong fluctuations, time-varying operating conditions, or long-life degradation, the prediction stability and interpretability are still insufficient. In addition, existing methods are still mainly based on deterministic modeling and have not yet unified the modeling of random drift-diffusion laws and structured continuous time in battery degradation, which restricts their engineering deployment capability under actual operating conditions. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for predicting the lifetime of lithium batteries using a stochastically constrained graph-liquid neural network, comprising:

[0006] The degradation observation sequence of lithium battery is collected, and the degradation observation sequence is normalized and reconstructed in segments to obtain a windowed degradation subsequence.

[0007] For the windowed degenerate subsequence, a degenerate graph consisting of temporal nodes, step-connecting edges, and global degenerate nodes is constructed;

[0008] The degraded graph is input into a graph-liquid neural network. Neighborhood message propagation and liquid gating state updates are performed on the node information in the graph structure. The updated hidden states of the nodes are aggregated and read out to extract window-level degradation representations. Based on the window-level degradation representations, future degradation increments are predicted, and the future health state trajectory is recursively reconstructed.

[0009] During the training process of the graph liquid neural network, a drift-diffusion physical prior based on Wiener stochastic processes is introduced, and a joint optimization objective is constructed by combining supervised loss and monotonicity constraints to constrain the future health state trajectory; the remaining lifespan of the lithium battery is determined based on the predicted step size of the future health state trajectory.

[0010] As a preferred embodiment of the stochastically constrained graph-liquid neural network lithium battery lifetime prediction method described in this invention, the following steps are included: collecting a degradation observation sequence of the lithium battery, and normalizing and reconstructing the degradation observation sequence by means of collecting the capacity value, state health index value or equivalent degradation observation value of the lithium-ion battery during continuous cycling to form a degradation observation sequence.

[0011] The degraded observation sequence is normalized to obtain a standardized degraded sequence;

[0012] The degradation increment between two adjacent cycles is determined based on the adjacent cycle observations in the standardized degradation sequence, and a degradation increment sequence is obtained based on the degradation increment.

[0013] The degradation increment sequence is segmented and reconstructed using a sliding window with a preset length and a preset step size to obtain a windowed degradation subsequence; wherein each windowed degradation subsequence corresponds to a set of time nodes, and the time nodes in the set of time nodes are used to characterize the local degradation state at the corresponding position within the window.

[0014] As a preferred embodiment of the stochastically constrained graph-liquid neural network lithium battery lifetime prediction method of the present invention, the construction of the degradation graph composed of temporal nodes, cross-connection edges and global degradation nodes includes: using the degradation increment between two adjacent cycles as the feature of the temporal node, and establishing sequential edges and cross-connection edges between the temporal nodes; the sequential edges are used to characterize short-range temporal dependencies, and the cross-connection edges are used to characterize medium-range degradation correlations.

[0015] Based on the set of time-series nodes, a global degradation node is introduced, and global connection edges are established between the global degradation node and all time-series nodes in the current window to obtain a window-level degradation graph. The feature vector of the global degradation node is composed of window statistical descriptors, which include window mean, window standard deviation, window net decay magnitude, first-order difference mean, and second-order difference mean, used to characterize the degradation stage, fluctuation intensity, and trend curvature.

[0016] As a preferred embodiment of the graph-liquid neural network lithium battery lifetime prediction method with random constraints described in this invention, the step of performing neighborhood message propagation and liquid gating state update on node information in the graph structure includes, for the node in the degenerate graph, performing a linear transformation on the current hidden state vector of the neighboring node, and performing weighted aggregation according to the connection weight between the neighboring node and the node to obtain the message vector of the node.

[0017] Based on the current hidden state vector of the node and the message vector, a candidate hidden state vector of the node is determined by a nonlinear activation function; based on the current hidden state vector of the node and the message vector, a liquid update gate vector is determined by a Sigmoid function after vector concatenation; the liquid update gate vector is used to control the fusion ratio between the current hidden state vector and the candidate hidden state vector.

[0018] Based on the liquid update gate vector, the current hidden state vector and the candidate hidden state vector of the node are fused by element-wise multiplication to obtain the updated hidden state vector of the node; wherein, the liquid gate state update makes the node state update rate adaptively change with input features and neighborhood information, thereby adapting to the slow decay stage and the rapid decay stage at the end of the lifetime.

[0019] As a preferred embodiment of the graph-liquid neural network lithium battery lifetime prediction method with random constraints described in this invention, the introduction of the drift-diffusion physical prior based on Wiener stochastic process includes using Wiener stochastic process to model the degradation increment with drift-diffusion prior, and using the drift term and diffusion intensity of Wiener stochastic process to characterize the expected drift rate and random fluctuation degree of the degradation increment, respectively.

[0020] Based on the future degradation increment output of the graph liquid neural network, a predicted degradation increment sequence is constructed. The sample mean and sample variance of the predicted degradation increment sequence are calculated. The drift term and the diffusion intensity are combined to construct a stochastic physical constraint loss to constrain the mean and variance of the predicted degradation increment sequence, thereby achieving a physically consistent probabilistic lifetime prediction.

[0021] The drift term and the diffusion intensity are updated online using an exponential moving average method based on the sample mean and sample variance of the predicted degradation increment sequence of the current cycle.

[0022] As a preferred embodiment of the graph-liquid neural network lithium battery lifetime prediction method with random constraints described in this invention, the method of constructing a joint optimization objective by combining supervised loss and monotonic constraints includes, in order to ensure that the reconstructed degradation trajectory satisfies the monotonic decay law, constructing monotonic constraints by utilizing the portion of the future health state trajectory where adjacent prediction step lengths show positive growth.

[0023] A joint optimization objective is constructed based on the stochastic physical constraint loss, the supervised loss, and the monotonicity constraint; the joint optimization objective is used to suppress positive growth predictions that do not conform to the battery degradation law and improve long-term extrapolation stability.

[0024] As a preferred embodiment of the graph-liquid neural network lithium battery lifetime prediction method with random constraints described in this invention, the method for determining the remaining lifetime of the lithium battery based on the prediction step size of the future health state trajectory includes determining a failure threshold based on the initial health state value of the lithium battery and a failure ratio threshold.

[0025] In the future health status trajectory, when the predicted health status value first reaches the failure threshold, the remaining lifespan of the lithium battery at the current moment is determined according to the corresponding minimum prediction step size.

[0026] A stochastically constrained graph-based neural network system for predicting the lifespan of lithium batteries, wherein:

[0027] The degradation observation sequence module collects the degradation observation sequence of lithium batteries, performs normalization processing and segmented reconstruction on the degradation observation sequence to obtain a windowed degradation subsequence;

[0028] The degradation graph construction module constructs a degradation graph consisting of temporal nodes, step-connecting edges, and global degradation nodes for the windowed degradation subsequence.

[0029] The future health status module inputs the degradation graph into a graph-liquid neural network, performs neighborhood message propagation and liquid gating state updates on the node information in the graph structure, aggregates and reads out the updated hidden states of the nodes, extracts window-level degradation representations, predicts future degradation increments based on the window-level degradation representations, and recursively reconstructs the future health status trajectory.

[0030] The constraint and remaining lifetime module introduces a drift-diffusion physical prior based on Wiener stochastic processes during the training of the graph liquid neural network, and constructs a joint optimization objective by combining supervised loss and monotonicity constraints to constrain the future health state trajectory; the remaining lifetime of the lithium battery is determined based on the predicted step size of the future health state trajectory.

[0031] A computer device includes: a memory and a processor; the memory stores a computer program, wherein: when the processor executes the computer program, it implements the steps of the method described in any one of the present invention.

[0032] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described in any one of the present invention.

[0033] The beneficial effects of this invention are as follows: By elevating the lithium battery degradation sequence from a simple time series to a degradation graph representation including time-series nodes, step-connecting edges, and global degradation nodes, this invention enables unified modeling of local cyclic dependencies and global degradation stages within the same framework, effectively improving the model's ability to perceive and utilize degradation structure information. A liquid-gated continuous update mechanism is introduced during graph structure message passing, allowing the node state update rate to adaptively change with input features and neighborhood information, thus effectively adapting to both slow degradation stages and rapid degradation stages at the end of the lifespan. The drift term and diffusion intensity of the Wiener process are used as stochastic physics prior embedding loss functions, while simultaneously constraining the mean and variance of the predicted degradation increment sequence, achieving physically consistent probabilistic lifetime prediction. Furthermore, a joint optimization constructed with supervised loss and monotonicity loss suppresses positive growth predictions that do not conform to battery degradation patterns, improving the stability of long-term extrapolation. This invention combines degradation structure modeling capabilities, dynamic state expression capabilities, and stochastic mechanism constraint capabilities, exhibiting higher prediction accuracy, robustness, and interpretability in both short-lifespan, highly volatile scenarios and long-lifespan, stable scenarios. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 The flowchart illustrates a random-constrained graph-liquid neural network method for predicting the lifespan of lithium batteries, as provided in this embodiment of the invention.

[0036] Figure 2 A computer device diagram for a random-constrained graph-liquid neural network method for predicting the lifespan of lithium batteries, provided in an embodiment of the present invention.

[0037] Figure 3 The image shows the tracking capability of the XJU dataset provided in this embodiment of the invention for real degraded trajectories.

[0038] Figure 4 The image shows the trajectory prediction results for the HUST dataset and long-term stationary degradation scenarios provided in this embodiment of the invention.

[0039] Figure 5 Error distribution diagram under the XJU dataset provided in the embodiments of the present invention.

[0040] Figure 6 This is an error distribution diagram of the HUST dataset provided in an embodiment of the present invention.

[0041] Figure 7 A comparison chart of root mean square errors of different models provided in the embodiments of the present invention. Detailed Implementation

[0042] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0043] Example 1, referring to Figure 1 and Figure 2 As an embodiment of the present invention, a method for predicting the lifetime of a lithium battery using a stochastically constrained graph neural network is provided, comprising:

[0044] S1: Collect the degradation observation sequence of lithium battery, normalize and reconstruct the degradation observation sequence in segments to obtain windowed degradation subsequence.

[0045] In this embodiment, the capacity of lithium batteries typically exhibits a long-term downward trend during the cyclic aging process. However, this macro-degradation trend is superimposed with short-term fluctuations and random disturbances caused by stage transitions. If the original degradation observation sequence is directly modeled, on the one hand, the difference in dimensions and the inconsistency in magnitude will affect the model convergence; on the other hand, the absolute health state value is difficult to reflect the local degradation between adjacent cycles. At the same time, if the degradation process is treated as a one-dimensional continuous sequence, it is difficult to simultaneously characterize the correlation between local cyclic dependence and global degradation stages within a unified framework. Therefore, before calling the graph liquid neural network for degradation modeling, the original degradation observation sequence needs to be normalized and incrementally transformed. Then, through sliding window segmented reconstruction, the continuous degradation sequence is transformed into a windowed degradation subsequence that takes into account both local degradation and window-level structural information, providing a structured original input for the subsequent degradation graph construction.

[0046] Specifically, the process of collecting the degradation observation sequence of lithium batteries and normalizing and reconstructing the degradation observation sequence includes the following steps: collecting capacity values, state health index values, or equivalent degradation observation values ​​of lithium-ion batteries during continuous cycling to form the original degradation observation sequence. Here, the cycle refers to one charge-discharge process completed by the battery according to a preset charge-discharge regime, or the smallest periodic unit corresponding to obtaining one health status observation value. Indicates the cycle number. This represents the total number of loops.

[0047] The original degraded observation sequence is normalized to obtain a standardized degraded sequence. The specific formula is as follows:

[0048] ;

[0049] in, Represents the normalized i-th One cyclic observation; Indicates the first One cyclic observation; and These represent the minimum and maximum values ​​in the current battery sample observation sequence, respectively.

[0050] The degradation increment between two adjacent cycles is determined based on the adjacent cycle observations in the standardized degradation sequence, and a degradation increment sequence is obtained based on the degradation increment, wherein the degradation increment is defined as:

[0051] ;

[0052] in, This represents the degradation increment between two adjacent loops; Represents the normalized i-th One cyclic observation; Represents the normalized i-th A cyclical observation.

[0053] It should be noted that this invention uses degradation increment prediction instead of direct prediction of absolute health state values; the degradation increment sequence has better local stationarity than the original degradation observation sequence, which is conducive to the model focusing on local degradation between adjacent cycles rather than long-term drift trends; at the same time, when recursively reconstructing the future health state trajectory, the health state prediction value is obtained by incremental accumulation, which makes the error propagation process more controllable and provides a direct object for subsequently introducing the drift-diffusion physical prior of Wiener stochastic process to impose constraints on the statistical laws of the predicted degradation increment sequence.

[0054] Furthermore, a preset length is adopted. and preset step size The sliding window is used to segment and reconstruct the degenerate increment sequence to obtain the first... A windowed degenerate subsequence:

[0055] ;

[0056] in, Indicates the first A windowed degenerate subsequence; Represents the normalized i-th A cyclical observation.

[0057] Each of the aforementioned windowed degenerate subsequences corresponds to a set of time-series nodes, the first... The set of time-series nodes within a window is represented as follows:

[0058] ;

[0059] in, Indicates the first The set of time-series nodes corresponding to each sliding window; Indicates the sliding window number; Indicates the length of the sliding window; Indicates the first Within the first sliding window There are 1 time-series nodes, among which , is used to characterize the local degradation state at the corresponding position within the window.

[0060] It should be noted that, to address the limitation of treating the degradation process as a one-dimensional continuous sequence, which makes it difficult to simultaneously characterize the relationships between local cyclic dependencies and global degradation stages within a unified framework, a sliding window segmentation reconstruction is used to transform the incremental degradation sequence into a structured windowed degradation subsequence. Each window corresponds to a set of temporal nodes, explicitly dividing the originally uniformly laid-out one-dimensional degradation sequence into window units with local degradation structures. Based on this, subsequent steps can simultaneously perceive local cyclic dependencies and window-level overall degradation trends within each window, laying a structured data foundation for constructing a degradation graph containing temporal nodes, step-connecting edges, and global degradation nodes. This, in turn, supports graph liquid neural networks in performing structure-aware modeling of the degradation process that combines local correlations and global trends, effectively improving the ability to utilize degradation structure information.

[0061] S2: For the windowed degenerate subsequence, construct a degenerate graph consisting of temporal nodes, step-connecting edges, and global degenerate nodes.

[0062] After completing the windowed degradation subsequence construction in step S1, a graph model that can explicitly express the degradation structure relationship needs to be established for each windowed degradation subsequence. However, if only adjacent connections are established between each temporal node within the window, only the sequential dependency relationship between adjacent cycles can be captured. The cross-cycle association between non-adjacent temporal nodes cannot be explicitly established. Moreover, there is a lack of dedicated nodes that can aggregate the overall degradation trend at the window level in the same graph structure. This results in the inability to uniformly express local temporal information and global degradation stage features in the same structure, and the utilization of degradation structure information is still insufficient. Therefore, it is necessary to explicitly construct the windowed degradation subsequence as a degradation graph containing temporal nodes, cross-step connection edges, and global degradation nodes. This will represent the battery degradation process as an information propagation structure that combines local association and global trend, providing a complete graph structure input for the subsequent neighborhood message propagation and liquid gating state update of the graph liquid neural network.

[0063] Specifically, for the windowed degenerate subsequence, constructing a degenerate graph consisting of temporal nodes, step-connecting edges, and global degenerate nodes includes the following steps: using the degenerate increment between two adjacent loops as the feature of the temporal node, establishing sequential edges between the temporal nodes. and step-connecting edge The sequential edge is used to connect the first... The timing nodes corresponding to adjacent loop positions within a sliding window represent short-range temporal dependencies; the step-connecting edge is used to connect nodes with a preset step size. The two time-series nodes represent the mid-range degradation correlation.

[0064] Introduce global degradation nodes based on the aforementioned time-series node set. Establish global connection edges between the global degenerate node and all temporal nodes within the current window. The global degradation node is connected to all temporal nodes within the current window to aggregate and represent the overall window-level degradation information, resulting in a window-level degradation graph:

[0065] ;

[0066] in, Indicates the first Window-level degradation graph corresponding to each sliding window; Indicates the first The global degradation node corresponding to each sliding window. and All belong to the first A sliding window, superscript Indicates the window number to which the node belongs. For the first The index of the time sequence node within each window Indicates the position of the node in the window. For the first The global degradation node corresponding to each window, index Indicates the global node identifier; Represents a set of edges, including ordered edges. Step-connection edge and global connection edges Globally connected edges Used to connect the global degradation node with all time-series nodes within the window, so as to realize the interactive propagation between local degradation information and global degradation context; This represents the node feature matrix.

[0067] Furthermore, the feature vector of the globally degenerate node It consists of window statistical descriptors, including the window mean, window standard deviation, net window decay magnitude, first-order difference mean, and second-order difference mean, used to characterize the degradation stage, fluctuation intensity, and trend curvature. The specific formula is as follows:

[0068] ;

[0069] in, The mean of the window; The standard deviation of the window; This represents the net attenuation amplitude of the window. It is the first-order difference mean; It is the second-order difference mean.

[0070] It should be noted that ordinary sequence modeling can only establish sequential dependencies between adjacent nodes, and cross-period associations between non-adjacent cycles cannot be explicitly captured. Furthermore, it is difficult to uniformly express local temporal information and window-level overall degradation trends within the same structure. To address this limitation, by introducing cross-stepping edges and global degradation nodes, the windowed degenerate subsequences are explicitly constructed as a ternary degradation graph structure. Cross-stepping edges break through the constraints of adjacent nodes, enabling direct cross-period associations between non-adjacent temporal nodes. Global degradation nodes, connected to all temporal nodes within the window, contain information including the window mean, window standard deviation, and window net decay. The window statistical descriptors of the reduction value, the first-order difference mean, and the second-order difference mean are used as feature vectors to comprehensively characterize the overall degradation state at the window level from three dimensions: degradation stage, fluctuation intensity, and trend curvature. Through the collaborative modeling of temporal nodes, cross-step connection edges, and global degradation nodes, the local temporal dependencies, cross-period correlations, and global degradation trends of the degradation process are simultaneously represented in the same degradation graph structure. This provides a structured input with both local correlations and global trends for the subsequent neighborhood message propagation and liquid gating state update of the graph liquid neural network, effectively improving the model's ability to perceive and utilize degradation structure information.

[0071] S3: Input the degraded graph into the graph liquid neural network, perform neighborhood message propagation and liquid gating state update on the node information in the graph structure, aggregate and read out the updated node hidden state, extract window-level degradation representation, predict future degradation increment based on the window-level degradation representation, and recursively reconstruct the future health state trajectory.

[0072] In this embodiment, it is necessary to perform evolutionary modeling on the degradation information carried by all nodes in the obtained degradation graph to extract window-level degradation representations that can reflect the current window degradation and predict future degradation increments accordingly. However, the gating update mechanism adopts a fixed state update method for all nodes, which cannot adjust the state update step size according to the rate of change of each node's degradation stage. Especially when there are differences in dynamic characteristics between the slow degradation stage and the rapid degradation stage at the end of the life of lithium batteries, the fixed gating mechanism is difficult to effectively adapt to the degradation at different stages of the entire life cycle. Therefore, this invention constructs a graph liquid neural network, which performs evolutionary modeling on the hidden state of nodes through neighborhood message propagation and liquid gating state update, so that the node state update rate adapts to the node input features and neighborhood information, thereby adapting to the slow degradation stage and the rapid degradation stage at the end of the life. On this basis, the updated hidden state of nodes is aggregated and read out to recursively reconstruct the future health state trajectory.

[0073] Specifically, the degenerate graph is input into a graph-liquid neural network, and the node information in the graph structure is updated through neighborhood message propagation and liquid gating as follows: For the nodes in the degenerate graph... The current hidden state vector of the neighboring nodes is linearly transformed by the message mapping weight matrix, and then weighted and aggregated according to the connection weights between the neighboring nodes to obtain the message vector of the node. The corresponding neighborhood message aggregation result is as follows:

[0074] ;

[0075] in, Indicates the first In a layered graph liquid neural network, nodes The message vector is obtained by weighted aggregation of information from its neighboring nodes; Indicates the identifier of the target node to be updated; Represents a node The neighborhood node index; Represents a node The set of neighboring nodes; Represents a node To the node Connection weights; This represents the message mapping weight matrix; Represents a node In the The hidden state vector of the layer; Indicates the neighboring nodes The message representation obtained after linear transformation of the current hidden state; This represents the weighted message contribution after considering connection strength.

[0076] Based on the node's current hidden state vector and the message vector, a candidate hidden state vector for the node is determined using a non-linear activation function, expressed as:

[0077] ;

[0078] in, Represents a node In the Candidate hidden state vectors of the layer; Represents a nonlinear activation function; The transformation weight matrix represents the current node's own state; Represents a node In the The current hidden state vector of the layer; The transformation weight matrix representing the aggregated message; Represents a node In the A layer is a message vector aggregated from its neighboring nodes; This represents the bias vector.

[0079] Further, based on the node's current hidden state vector and the message vector, a liquid update gate vector is determined by concatenating the vectors and then using the Sigmoid function. Based on the liquid update gate vector, the node's current hidden state vector and the candidate hidden state vector are fused using element-wise multiplication to obtain the updated hidden state vector of the node, thus completing the liquid-gated state update.

[0080] ;

[0081] ;

[0082] in, Represents a node In the The liquid update gate vector of the layer is used to control the fusion ratio between the current hidden state vector and the candidate hidden state vector; This represents the weight matrix of the updated gate. , , as well as All are trainable parameter matrices; Represents a node In the The current hidden state vector of the layer; Represents a node In the A layer is a message vector aggregated from its neighboring nodes; Represents a node In the The hidden state of the layer; Indicates the bias term; Represents the Sigmoid function; This indicates vector concatenation; This represents element-wise multiplication; Represents a node In the Candidate hidden state vectors of the layer.

[0083] It should be noted that, to address the shortcomings of the gated update mechanism, which uses a fixed-state update method for all nodes and is difficult to adapt to the differences in dynamic characteristics at different degradation stages throughout the lithium battery's life cycle, a liquid update gate vector is introduced. This gate vector is calculated by concatenating the current node's hidden state vector and the neighborhood message vector and then applying the Sigmoid function. It adaptively adjusts according to the node's input features and neighborhood information. When degradation enters the slow decay stage, the liquid update gate vector tends to retain more of the current hidden states to maintain a stable state evolution. When degradation enters the rapid decay stage at the end of the life cycle, the liquid update gate vector allows for a greater fusion of candidate states to quickly respond to changes in accelerated capacity decay. This allows the node state update rate to match the dynamic change rate of the current degradation stage, effectively improving the model's adaptability to the entire life cycle degradation process.

[0084] Furthermore, the hidden states of nodes updated via liquid gating are aggregated and read out to extract window-level degradation representations indicating current window degradation; then, future degradation increments are predicted based on these window-level degradation representations, and the future health state trajectory is reconstructed using a recursive approach.

[0085] ;

[0086] in, Indicates the future number Predicted health status values ​​for each cycle; Indicates the future number Predicted degradation increment for each cycle; This indicates the maximum prediction step size.

[0087] It should be noted that, in response to the problem that deterministic modeling leads to the occurrence of local rebound phenomena in the predicted trajectory that contradict the physical law of monotonic battery degradation, the prediction process is improved by replacing the direct prediction of the absolute value of the health state with the prediction of degradation increment, and reconstructing the future health state trajectory through recursive accumulation. This gives the generation process of the predicted trajectory a natural basis of incremental constraints.

[0088] S4: During the training process of the liquid neural network in the figure, a drift-diffusion physical prior based on Wiener stochastic process is introduced, and a joint optimization objective is constructed by combining supervised loss and monotonicity constraint to constrain the future health state trajectory; the remaining lifespan of the lithium battery is determined according to the predicted step size of the future health state trajectory.

[0089] Furthermore, whether the future health state trajectory reconstructed in step S3 conforms to the physical laws of battery degradation directly determines the reliability and interpretability of the remaining life prediction results. However, methods based on deterministic modeling and pure supervised learning ignore the objectively existing random drift and diffusion characteristics in the degradation increment sequence. The predicted trajectory trained solely on supervised loss lacks explicit constraints of the physical mechanism and is prone to local rebound phenomena that do not conform to the monotonic degradation law of the battery. In scenarios with strong fluctuations or long life degradation, the prediction stability and interpretability are insufficient. Therefore, by designing a joint optimization mechanism that combines drift-diffusion random prior, monotonicity constraints, and supervised learning, the statistical prior of the Wiener stochastic process is embedded into the loss function. At the same time, a penalty is imposed on the positive growth in the reconstructed degradation trajectory to achieve consistency constraints on the degradation mechanism and perception of random fluctuations. Based on this, the remaining life of the lithium battery is determined according to the future health state trajectory.

[0090] Specifically, in the training process of the graph liquid neural network, the introduction of a drift-diffusion physical prior based on the Wiener stochastic process includes the following steps: using the Wiener stochastic process to model the degradation increment for drift-diffusion prior, with the drift term and diffusion intensity of the Wiener stochastic process representing the expected drift rate and the degree of random fluctuation of the degradation increment, respectively:

[0091] ;

[0092] in, This represents the degradation increment between two adjacent loops; Indicates the average drift term; Indicates diffusion intensity; This represents a random disturbance term that follows a standard normal distribution.

[0093] Based on the future degradation increment output of the graph-liquid neural network, a predicted degradation increment sequence is constructed. The sample mean and sample variance of the predicted degradation increment sequence are calculated. A stochastic physical constraint loss is constructed by combining the drift term and the diffusion intensity to constrain the mean and variance of the predicted degradation increment sequence, achieving physically consistent probabilistic lifetime prediction. The specific formula is as follows:

[0094] ;

[0095] in, This represents the sample mean of the predicted degradation increment sequence; This represents the sample variance of the predicted degradation increment sequence; Indicates the average drift term; Indicates diffusion intensity; This represents the variance constraint weighting coefficient.

[0096] Furthermore, during the training process of the graph-liquid neural network, an exponential moving average method is used to update the drift term and the diffusion intensity online based on the sample mean and sample variance of the predicted degradation increment sequence in the current cycle.

[0097] ;

[0098] ;

[0099] in, Represents the smoothing coefficient, and ; Indicates the first A cycle of drift terms; Indicates the first The diffusion intensity of each cycle; Indicates the first The sample mean of the predicted degradation increment sequence for each cycle; Indicates the first The sample variance of the predicted degradation increment sequence for each cycle.

[0100] It should be noted that, to address the problem of neglecting the objectively existing random drift and diffusion characteristics in the degradation increment sequence, which leads to a lack of physical prior constraints in the prediction results, a Wiener stochastic process is introduced to model the degradation increment a priori. The drift term and diffusion intensity are used to characterize the expected drift rate and random fluctuation degree of the degradation increment, respectively. This incorporates the random uncertainty in the battery degradation process into the modeling framework, providing a statistical prior basis for the subsequent construction of the stochastic physical constraint loss, and achieving a physically consistent probabilistic lifetime prediction. Secondly, an exponential moving average method is used to update the drift term and diffusion intensity online, so that the physical prior parameters can be adjusted as the degradation process progresses, thereby continuously tracking the non-stationary characteristics of the degradation process and ensuring that the stochastic physical constraint loss can effectively reflect the statistical law of the current degradation at different degradation stages.

[0101] Furthermore, the construction of the joint optimization objective by combining supervised loss and monotonicity constraints includes: to ensure that the reconstructed degenerate trajectory satisfies the monotonic decay law, constructing monotonicity constraints using the portion of the future healthy state trajectory where adjacent predicted step sizes show positive growth. :

[0102] ;

[0103] in, Indicates the future number Predicted health status values ​​for each cycle; This indicates the maximum prediction step size.

[0104] A joint optimization objective is constructed based on the stochastic physical constraint loss, the supervised loss, and the monotonicity constraint. :

[0105] ;

[0106] in, Indicates the monitoring error loss; Represents the loss due to random physical constraints; Indicates monotonicity constraints; and The weighting coefficient is indicated; the joint optimization objective is used to suppress positive growth predictions that do not conform to the battery degradation law and improve the long-term extrapolation stability.

[0107] Finally, the remaining lifespan of the lithium battery is determined based on the predicted step size of the future health state trajectory. Specifically, a failure threshold is determined based on the initial health state value and the failure ratio threshold of the lithium battery; in the future health state trajectory, the battery is progressively searched in ascending order of predicted step size, and the minimum predicted step size corresponding to when the predicted health state value first reaches the preset failure threshold is determined as the remaining lifespan of the lithium battery at the current moment.

[0108] ;

[0109] in, Indicates remaining lifespan; Indicates the future number Predicted health status values ​​for each cycle; Indicates the failure rate threshold; This represents the initial health status value.

[0110] It should be noted that, to address the issue that relying solely on supervised loss for training leads to a lack of physical mechanism constraints in the predicted trajectory, resulting in local rebounds that are inconsistent with the monotonic degradation pattern of batteries, this paper proposes a method that utilizes supervised loss, stochastic physical constraint loss, and monotonicity constraint to jointly construct an optimization objective. At the training level, this method imposes constraints on the prediction results from both statistical and physical dimensions: the stochastic physical constraint loss ensures that the mean and variance of the predicted degradation increment sequence are consistent with the Wiener process drift-diffusion prior, achieving awareness of random fluctuations; the monotonicity constraint penalizes positive growth in the reconstructed degradation trajectory, ensuring that the predicted trajectory conforms to the physical law of battery monotonic degradation. These two types of constraints, along with the supervised loss, are synergistically optimized, effectively improving the stability and interpretability of long-term extrapolation. Furthermore, by retrieving the minimum prediction step size corresponding to the first time the predicted health state value reaches the failure threshold in the future health state trajectory and mapping it to the remaining lifetime estimate at the current moment, this method directly corresponds to engineering practices in battery failure determination and is applicable to online health management and predictive maintenance scenarios for lithium batteries.

[0111] On the other hand, this embodiment also provides a stochastically constrained graph-liquid neural network lithium battery lifetime prediction system, which includes:

[0112] The degradation observation sequence module collects degradation observation sequences of lithium batteries, performs normalization processing and segmented reconstruction on the degradation observation sequences, and obtains windowed degradation subsequences.

[0113] The degradation graph construction module constructs a degradation graph consisting of temporal nodes, step-connecting edges, and global degradation nodes for the windowed degradation subsequence.

[0114] The future health status module inputs the degradation graph into a graph-liquid neural network, performs neighborhood message propagation and liquid gating state updates on the node information in the graph structure, aggregates and reads out the updated hidden states of the nodes, extracts window-level degradation representations, predicts future degradation increments based on the window-level degradation representations, and recursively reconstructs the future health status trajectory.

[0115] The constraint and remaining lifetime module introduces a drift-diffusion physical prior based on Wiener stochastic processes during the training of the graph liquid neural network, and constructs a joint optimization objective by combining supervised loss and monotonicity constraints to constrain the future health state trajectory; the remaining lifetime of the lithium battery is determined based on the predicted step size of the future health state trajectory.

[0116] like Figure 2 As shown, if the above functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0117] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0118] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0119] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0120] Example 2, refer to Figures 3-7 As an embodiment of the present invention, a random-constrained graph liquid neural network method for predicting the lifespan of lithium batteries is provided. To verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculations and simulation experiments.

[0121] To verify the effectiveness, stability, and generalization ability of the lithium battery remaining life prediction method proposed in this invention under different degradation scenarios, this embodiment selects the XJU dataset and the HUST dataset for verification. The XJU dataset contains 55 cylindrical lithium-ion cells, each with a nominal capacity of 2000mAh, a rated voltage of 3.6V, a charging cutoff voltage of 4.2V, and a discharging cutoff voltage of 2.5V. All experiments were conducted at room temperature and continuously cycled until the battery life ended. The XJU dataset covers various operating conditions, including constant current-constant voltage charging / discharging conditions, conditions with randomly varying discharge currents in different cycles, and conditions with equivalent charging / discharging strategies. Therefore, the degradation trajectory in the dataset is significantly affected by operating condition switching and random disturbances, and the capacity decay process is usually accompanied by strong local fluctuations and non-stationary changes. This dataset is suitable for verifying the invention's ability to jointly model degradation structure relationships, dynamic evolution laws, and remaining life under complex operating conditions, strong random disturbances, and short-term rapid degradation scenarios.

[0122] The HUST dataset was collected from a lithium iron phosphate / graphite battery system, comprising 77 batteries, each with a nominal capacity of 1100mAh and a rated voltage of 3.3V. All batteries underwent cycle testing at a constant temperature of 30℃, using the same charging protocol and corresponding to 77 different multi-stage discharge strategies until the batteries reached the end of their lifespan. Compared to the XJU dataset, the capacity decay trajectory in the HUST dataset is smoother overall, the degradation process is more gradual, and there are relatively fewer fluctuations between cycles, demonstrating strong stable operating conditions. This dataset is suitable for verifying the invention's ability to continuously track the overall degradation trend and its long-term extrapolation stability under long-term, stable, and gradual capacity decay scenarios. Furthermore, the XJU dataset represents non-stationary degradation scenarios under multi-condition perturbation, while the HUST dataset represents progressive degradation scenarios driven by multi-stage discharge strategies under isothermal conditions. The two datasets differ in battery system, nominal parameters, discharge strategies, and degradation fluctuation characteristics, which can verify the applicability and robustness of the present invention in random degradation scenarios and stationary progressive degradation scenarios from different perspectives.

[0123] To ensure consistency in model training, testing, and comparative verification, this invention employs a unified data preprocessing workflow for the XJU and HUST datasets. This includes: extracting capacity sequences, health status index sequences, or equivalent degradation observation sequences for each battery during cycling; normalizing the original sequences; and then reconstructing them into window-level samples using a sliding window of length L. For each window, a degradation graph is constructed, consisting of temporal nodes, step-connecting edges, and global degradation nodes, mapping the original one-dimensional degradation sequence into a graph representation that combines local structure and global contextual information. During model training, the constructed degradation graph is input into a graph-liquid neural network. Window-level degradation representations are obtained through neighborhood message aggregation and liquid-gated state updates to predict future degradation increment sequences. Training is performed using a joint optimization strategy combining supervised loss, stochastic physical constraint loss, and monotonicity loss. In the real-time prediction phase, an exponential moving average method is used to recursively update the drift term and diffusion intensity of the Wiener process to adapt to changes in degradation statistical characteristics under time-varying conditions.

[0124] like Figure 3 The figure shows the tracking capability of this invention for the real degradation trajectory on the XJU dataset. The horizontal axis represents the number of cycles (Cycle), and the vertical axis represents the battery capacity (Capacity). On representative battery samples in the XJU dataset, the predicted capacity trajectory output by this invention maintains a high degree of consistency with the real capacity trajectory: In the early stage of degradation, the predicted curve can track the changing trend of the slow capacity decay stage well, without obvious initial deviation; in the middle stage of degradation, when the real capacity curve shows local fluctuations and short-term fluctuations, the predicted curve can still maintain a basically synchronous change with the real trajectory, indicating that this invention has a good adaptability to random disturbances; in the late stage of degradation, facing the rapid capacity decay stage, the predicted curve can still accurately reflect the change of the degradation slope, without obvious lag or excessive oscillation.

[0125] Furthermore, from Figure 3 It can also be seen that the present invention can not only maintain an accurate representation of the overall downward trend, but also take into account the local fluctuation response and the stability of the global trend. It does not exhibit common problems such as error accumulation amplification, local overfitting, or increased end-of-life shift. This indicates that the present invention, through degradation graph structure representation and liquid dynamic update mechanism, can effectively extract degradation evolution features under strong random disturbance and non-stationary degradation conditions, and achieve stable tracking of the real degradation trajectory.

[0126] Secondly, to verify the trajectory prediction performance of the method of the present invention in long-term stationary degradation scenarios, such as... Figure 4As shown in the representative battery samples of the HUST dataset, the predicted curve of this invention maintains a high degree of consistency with the actual capacity curve over a relatively long cycle interval: for the relatively flat capacity decay stage in the early stage of degradation, the prediction result can accurately reflect the slow downward trend of the actual curve; for the gradually deepening decay process in the middle and late stages of degradation, the predicted curve can continuously and smoothly track the changes in the actual trajectory, without obvious phase shift, local rebound, or long-cycle cumulative error. The figure also shows that as the number of cycles continues to increase, the predicted curve of this invention can still maintain good stability and continuity, indicating that this invention has strong long-term extrapolation capability in long-life, progressive degradation scenarios. Especially in the later stages of life, although the capacity decay rate changes to some extent, the prediction result can still reflect the trend of decay slope change well, indicating that this invention can not only describe the smooth degradation process, but also maintain good trend consistency and reliability in long-term prediction.

[0127] Furthermore, this embodiment also verifies the prediction error performance of the present invention, as shown in Table 1. On the XJU dataset, the present invention achieves low prediction errors for all four battery samples XJU1 to XJU4, with root mean square errors (RMSE) of 0.0034234, 0.0020234, 0.0020667, and 0.0026445, respectively, and a corresponding average of 0.0025395, along with a corresponding coefficient of determination. The values ​​are all close to 1, indicating that the present invention can accurately maintain the overall degraded shape, where Cell represents the battery sample number and MSE represents the mean square error.

[0128] Table 1. Prediction Error Performance Table for XJU Dataset

[0129] Cell MSE RMSE <![CDATA[R 2 ]]> XJU1 0.0000117 0.0034234 0.9987541 XJU2 0.0000041 0.0020234 0.9996153 XJU3 0.0000043 0.0020667 0.9996183 XJU4 0.0000070 0.0026445 0.9993795 Average 0.0000068 0.0025395 0.9993418

[0130] Compared to Table 1, as shown in Table 2, on the HUST dataset, the root mean square errors (RMSE) of this invention for the four battery samples HUST1 to HUST4 are 0.0012807, 0.0011094, 0.0009106, and 0.0017572, respectively, with a corresponding average of 0.0012645 and a coefficient of determination. The average value is 0.9997229; compared with the XJU dataset, the error distribution on the HUST dataset is smaller and the average value of the root mean square error (RMSE) is further reduced, indicating that the present invention can achieve higher prediction accuracy in stationary degradation scenarios.

[0131] Table 2. Prediction Error Performance Table for HUST Dataset

[0132] Cell MSE RMSE <![CDATA[R 2 ]]> HUST1 0.0000016 0.0012807 0.9997095 HUST2 0.0000012 0.0011094 0.9998004 HUST3 0.0000008 0.0009106 0.9998623 HUST4 0.0000031 0.0017572 0.9995192 Average 0.0000017 0.0012645 0.9997229

[0133] Furthermore, in order to verify the stability of the error distribution, from Figure 5 As shown in the error distribution plot of the XJU dataset, where the horizontal axis represents the battery sample number and the vertical axis represents the error range, the cyclic prediction error of this invention for each battery sample in the XJU dataset is generally concentrated near zero. Although there are a few long-tailed errors in XJU1, the overall violin plot center width is still relatively narrow, indicating that the large errors are only a few discrete points rather than persistent deviations. This shows that this invention still has good anti-disturbance capability under strong fluctuations and non-stationary degradation conditions. Secondly, from Figure 6 As shown in the error distribution diagram of the HUST dataset, the error distribution of each battery sample in the HUST dataset is more compact than that in the XJU dataset, with shorter tails and less overall dispersion. This further illustrates that the output results of this invention have better stability and consistency under long-life stable degradation scenarios.

[0134] Finally, to further verify the superiority of the present invention, this embodiment also selected the Bi-LSTM model and the Transformer model as comparison methods. "Proposed" indicates the present invention. Specific results are as follows: Figure 7 As shown in Table 3, it can be seen that the present invention achieves the lowest root mean square error (RMSE) on all eight test battery samples in both the XJU and HUST datasets. Specifically, on the XJU dataset, the average RMSE of the present invention is 0.002540, lower than 0.426897 for the Bi-LSTM model and 0.177887 for the Transformer model; on the HUST dataset, the average RMSE of the present invention is 0.001265, lower than 0.388658 for the Bi-LSTM model and 0.169331 for the Transformer model.

[0135] Table 3 Comparison of Methods

[0136] Cell Proposed Bi-lstm Transformer XJU1 0.003423 0.441970 0.372117 XJU2 0.002023 0.445234 0.202919 XJU3 0.002067 0.429428 0.068047 XJU4 0.002645 0.390955 0.068466 Average 0.002540 0.426897 0.177887 HUST1 0.001281 0.391089 0.108757 HUST2 0.001109 0.421444 0.167379 HUST3 0.000911 0.368368 0.108765 HUST4 0.001757 0.373731 0.292421 Average 0.001265 0.388658 0.169331

[0137] The above results show that traditional pure data-driven sequence models are easily affected by local disturbances and accumulate errors in strong fluctuation scenarios, and it is also difficult to guarantee the physical consistency of degradation evolution in long-term prediction. However, this invention, by introducing Wiener drift-diffusion stochastic prior and combining degradation graph representation and liquid dynamic update mechanism, can effectively suppress long-term prediction drift and improve the stability and reliability of remaining lifetime prediction.

[0138] Therefore, in summary Figures 3 to 7As shown in Tables 1 to 3, the present invention can achieve high-precision remaining lifetime prediction in both short-lifetime, highly volatile scenarios and long-lifetime, stable scenarios; it can maintain consistency between the predicted trajectory and the actual degradation trajectory, reducing non-physical deviations; it can maintain good generalization ability and robustness under different degradation modes and different lifetime scales; at the same time, compared with existing deep sequence models such as Bi-LSTM and Transformer models, the present invention shows advantages in prediction accuracy, long-term stability, and physical consistency. That is, through the synergistic effect of degradation graph structure representation, liquid dynamic update, and random physical constraints, the present invention can achieve high-precision remaining lifetime prediction in both highly volatile, short-lifetime scenarios and long-lifetime, stable scenarios, and has good robustness, stability, and engineering application value.

[0139] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A stochastically constrained graph-based neural network method for predicting the lifespan of lithium batteries, characterized in that, include: The degradation observation sequence of lithium battery is collected, and the degradation observation sequence is normalized and reconstructed in segments to obtain a windowed degradation subsequence. For the windowed degenerate subsequence, a degenerate graph consisting of temporal nodes, step-connecting edges, and global degenerate nodes is constructed; The degraded graph is input into a graph-liquid neural network. Neighborhood message propagation and liquid gating state updates are performed on the node information in the graph structure. The updated hidden states of the nodes are aggregated and read out to extract window-level degradation representations. Based on the window-level degradation representations, future degradation increments are predicted, and the future health state trajectory is recursively reconstructed. During the training process of the graph liquid neural network, a drift-diffusion physical prior based on Wiener stochastic processes is introduced, and a joint optimization objective is constructed by combining supervised loss and monotonicity constraints to constrain the future health state trajectory; the remaining lifespan of the lithium battery is determined based on the predicted step size of the future health state trajectory.

2. The stochastically constrained graph-liquid neural network method for predicting lithium battery lifetime as described in claim 1, characterized in that: Collecting a degradation observation sequence of lithium batteries, and normalizing and reconstructing the degradation observation sequence includes collecting capacity values, state health index values ​​or equivalent degradation observation values ​​of lithium-ion batteries during continuous cycling to form a degradation observation sequence. The degraded observation sequence is normalized to obtain a standardized degraded sequence; The degradation increment between two adjacent cycles is determined based on the adjacent cycle observations in the standardized degradation sequence, and a degradation increment sequence is obtained based on the degradation increment. The degradation increment sequence is segmented and reconstructed using a sliding window with a preset length and a preset step size to obtain a windowed degradation subsequence; wherein each windowed degradation subsequence corresponds to a set of time nodes, and the time nodes in the set of time nodes are used to characterize the local degradation state at the corresponding position within the window.

3. The stochastically constrained graph-liquid neural network method for predicting lithium battery lifetime as described in claim 2, characterized in that: The construction of the degradation graph consisting of temporal nodes, step-connecting edges, and global degradation nodes includes using the degradation increment between two adjacent cycles as the feature of the temporal nodes, and establishing sequential edges and step-connecting edges between the temporal nodes; the sequential edges are used to characterize short-range temporal dependencies, and the step-connecting edges are used to characterize medium-range degradation associations. Based on the set of time-series nodes, a global degradation node is introduced, and global connection edges are established between the global degradation node and all time-series nodes in the current window to obtain a window-level degradation graph. The feature vector of the global degradation node is composed of window statistical descriptors, which include window mean, window standard deviation, window net decay magnitude, first-order difference mean, and second-order difference mean, used to characterize the degradation stage, fluctuation intensity, and trend curvature.

4. The stochastically constrained graph-liquid neural network method for predicting lithium battery lifetime as described in claim 3, characterized in that: The process of performing neighborhood message propagation and liquid gating state update on node information in the graph structure includes, for a node in the degenerate graph, performing a linear transformation on the current hidden state vector of the neighboring node, and performing weighted aggregation according to the connection weight between the neighboring node and the node to obtain the node's message vector. Based on the current hidden state vector of the node and the message vector, a candidate hidden state vector of the node is determined by a nonlinear activation function; based on the current hidden state vector of the node and the message vector, a liquid update gate vector is determined by a Sigmoid function after vector concatenation; the liquid update gate vector is used to control the fusion ratio between the current hidden state vector and the candidate hidden state vector. Based on the liquid update gate vector, the current hidden state vector and the candidate hidden state vector of the node are fused by element-wise multiplication to obtain the updated hidden state vector of the node; wherein, the liquid gate state update makes the node state update rate adaptively change with input features and neighborhood information, thereby adapting to the slow decay stage and the rapid decay stage at the end of the lifetime.

5. The stochastically constrained graph-liquid neural network method for predicting lithium battery lifetime as described in claim 4, characterized in that: The introduction of the drift-diffusion physical prior based on Wiener stochastic processes includes using Wiener stochastic processes to model the degradation increment for drift-diffusion priors, with the drift term and diffusion intensity of the Wiener stochastic process representing the expected drift rate and the degree of random fluctuation of the degradation increment, respectively. Based on the future degradation increment output of the graph liquid neural network, a predicted degradation increment sequence is constructed. The sample mean and sample variance of the predicted degradation increment sequence are calculated. The drift term and the diffusion intensity are combined to construct a stochastic physical constraint loss to constrain the mean and variance of the predicted degradation increment sequence, thereby achieving a physically consistent probabilistic lifetime prediction. The drift term and the diffusion intensity are updated online using an exponential moving average method based on the sample mean and sample variance of the predicted degradation increment sequence of the current cycle.

6. The stochastically constrained graph-liquid neural network method for predicting lithium battery lifetime as described in claim 5, characterized in that: The method of constructing a joint optimization objective by combining supervised loss and monotonic constraints includes constructing monotonic constraints by utilizing the portion of the future healthy state trajectory where adjacent predicted step sizes show positive growth in order to ensure that the reconstructed degenerate trajectory satisfies the monotonic decay law. A joint optimization objective is constructed based on the stochastic physical constraint loss, the supervised loss, and the monotonicity constraint; the joint optimization objective is used to suppress positive growth predictions that do not conform to the battery degradation law and improve long-term extrapolation stability.

7. The stochastically constrained graph-liquid neural network method for predicting lithium battery lifetime as described in claim 6, characterized in that: Determining the remaining lifespan of a lithium battery based on the predicted step size of the future health status trajectory includes determining a failure threshold based on the initial health status value and the failure ratio threshold of the lithium battery. In the future health status trajectory, when the predicted health status value first reaches the failure threshold, the remaining lifespan of the lithium battery at the current moment is determined according to the corresponding minimum prediction step size.

8. A graph-liquid neural network-based lithium battery lifetime prediction system employing the method described in any one of claims 1-7, characterized in that: The degradation observation sequence module collects the degradation observation sequence of lithium batteries, performs normalization processing and segmented reconstruction on the degradation observation sequence to obtain a windowed degradation subsequence; The degradation graph construction module constructs a degradation graph consisting of temporal nodes, step-connecting edges, and global degradation nodes for the windowed degradation subsequence. The future health status module inputs the degradation graph into a graph-liquid neural network, performs neighborhood message propagation and liquid gating state updates on the node information in the graph structure, aggregates and reads out the updated hidden states of the nodes, extracts window-level degradation representations, predicts future degradation increments based on the window-level degradation representations, and recursively reconstructs the future health status trajectory. The constraint and remaining lifetime module introduces a drift-diffusion physical prior based on Wiener stochastic processes during the training of the graph liquid neural network, and constructs a joint optimization objective by combining supervised loss and monotonicity constraints to constrain the future health state trajectory; the remaining lifetime of the lithium battery is determined based on the predicted step size of the future health state trajectory.

9. A computer device, comprising: A memory and a processor; the memory stores a computer program, characterized in that: when the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images