A method and system for collaborative estimation of state of health and remaining life of lithium-ion batteries

Abstract

This application provides a method and system for co-estimating the health status and remaining lifetime of a lithium-ion battery, relating to the field of battery state estimation technology. The method includes: acquiring charging operation data of a target lithium-ion battery in multiple charging cycles, constructing multi-cycle input samples corresponding to the charging cycle to be estimated; encoding intra-cycle degradation features in the cyclic data corresponding to each charging cycle to obtain cyclic embedding vectors; using the cyclic embedding vectors as graph nodes, modeling inter-cycle degradation dependencies based on temporal relationships to obtain inter-cycle node embedding vectors corresponding to each charging cycle; generating a local health status sequence based on the inter-cycle node embedding vectors, and determining the estimated health status and predicted remaining lifetime corresponding to the charging cycle to be estimated. The method provided in this application can jointly utilize intra-cycle degradation features and long-term inter-cycle degradation dependencies, improving the accuracy, consistency, and reliability of the co-estimation of health status and remaining lifetime.

Description

Technical Field

[0001] This application belongs to the field of battery testing technology, and in particular relates to a method and system for co-estimating the health status and remaining life of lithium-ion batteries. Background Technology

[0002] Lithium-ion batteries, characterized by high energy density, long cycle life, and good output stability, are widely used in new energy vehicles, energy storage systems, portable electronic devices, and other electrical equipment. During long-term use, lithium-ion batteries are affected by factors such as charge / discharge rate, ambient temperature, operating conditions, and material aging, gradually leading to phenomena such as capacity decay, increased internal resistance, and decreased performance consistency. To ensure the safety, reliability, and economy of the battery system, battery management systems typically need to estimate the battery's health status and remaining lifespan to provide a basis for charge / discharge control, maintenance, replacement, and safety warnings.

[0003] Existing methods for estimating the state of lithium-ion batteries mainly include model-based methods and data-driven methods. Model-based methods typically describe the battery's operating state using equivalent circuit models, electrochemical models, or empirical degradation models, and then adjust the model parameters based on collected data. However, battery aging is influenced by various factors, making it difficult to accurately determine model parameters, and some models have high online computational complexity, limiting their engineering applications. Data-driven methods can establish a mapping relationship between input data and battery state using battery operating data, reducing reliance on complex mechanistic models. However, in practical applications, battery operating data suffers from significant variations in operating conditions, individual differences, incomplete available data, and strong nonlinearity in the degradation process, which can easily lead to insufficient stability of the estimation results. Furthermore, although both health state estimation and remaining service life prediction are related to the degree of battery degradation, existing methods often treat them separately, making it difficult to guarantee the consistency between the two types of estimation results and the reliability of long-term predictions.

[0004] Therefore, in the estimation of the health status and remaining service life of lithium-ion batteries, the battery degradation process is affected by multiple coupled factors, the operating data is difficult to fully reflect the long-term aging trend, and the consistency between the health status estimation results and the remaining service life prediction results is insufficient, which have become problems that urgently need to be solved. Summary of the Invention

[0005] This application provides a method and system for co-estimating the health status and remaining life of lithium-ion batteries, aiming to solve the technical problems in the existing lithium-ion battery state estimation process, such as the battery degradation process being affected by multiple factors coupled together, the operational data being difficult to fully reflect the long-term aging trend, and the insufficient consistency between the health status estimation results and the remaining life prediction results, which lead to insufficient accuracy and reliability of the co-estimation of health status and remaining life.

[0006] In a first aspect, this application provides a method for co-estimating the health status and remaining life of a lithium-ion battery. The method includes: acquiring charging operation data of a target lithium-ion battery in multiple charging cycles, wherein the multiple charging cycles include the charging cycle to be estimated and multiple historical charging cycles prior to the charging cycle to be estimated. Based on the charging operation data, a multi-cycle input sample corresponding to the charging cycle to be estimated is constructed. The multi-cycle input sample includes multiple cycle data arranged in chronological order. The multiple cycle data correspond one-to-one with the multiple charging cycles and are obtained from the charging operation data of the corresponding charging cycle. The cyclic data corresponding to each charging cycle in the multi-cycle input samples are encoded with intra-cycle degradation features to obtain the cyclic embedding vector corresponding to each charging cycle. Using the cycle embedding vectors corresponding to each charging cycle as graph nodes, inter-cycle degradation dependency modeling is performed based on the temporal relationship between each charging cycle to obtain the inter-cycle node embedding vectors corresponding to each charging cycle. Based on the inter-cycle node embedding vectors corresponding to each charging cycle, the estimated health status value corresponding to each charging cycle is determined, and the local health status sequence corresponding to the multi-cycle input samples is generated in chronological order. Determine the health state estimate corresponding to the charging cycle to be estimated from the local health state sequence; Based on the inter-cycle node embedding vector corresponding to the charging cycle to be estimated and the local health state sequence, the remaining lifetime prediction value corresponding to the charging cycle to be estimated is determined; Output the estimated health status and predicted remaining lifetime corresponding to the charging cycle to be estimated.

[0007] In one possible design, the charging operation data includes current data, voltage data, and charging capacity data; The construction of the multi-cycle input sample corresponding to the charging cycle to be estimated based on the charging operation data includes: The charging segments within the set voltage range are extracted from the charging operation data of each charging cycle to obtain the charging segment data corresponding to each charging cycle. The charging segment data corresponding to each charging cycle is subjected to fixed-length resampling and normalization to obtain the cycle data corresponding to each charging cycle. Based on the charging cycle to be estimated, a preset number of historical cycle data are selected from the cycle data corresponding to each historical charging cycle before the charging cycle to be estimated, according to a preset sampling interval. Arrange the preset number of historical cycle data and the cycle data corresponding to the charging cycle to be estimated in chronological order to construct a multi-cycle input sample corresponding to the charging cycle to be estimated.

[0008] In one possible design, the step of performing intra-cycle degradation feature encoding on the cyclic data corresponding to each charging cycle in the multi-cycle input samples to obtain the cyclic embedding vector corresponding to each charging cycle includes: The cyclic data corresponding to each charging cycle are processed by feature mapping to obtain the initial latent space features corresponding to each charging cycle. The cyclic position codes corresponding to each charging cycle are fused with the initial latent space features to obtain the position enhancement features corresponding to each charging cycle. The position enhancement features corresponding to each charging cycle are encoded by intracyclic degradation features to obtain the cyclic embedding vector corresponding to each charging cycle.

[0009] In one possible design, the step of encoding the position enhancement features corresponding to each charging cycle with intra-cycle degradation features to obtain the cyclic embedding vector corresponding to each charging cycle includes: Intra-channel temporal encoding is performed on the temporal features of different data channels in the position enhancement features corresponding to each charging cycle to obtain the intra-channel temporal features corresponding to each charging cycle. Cross-channel feature fusion is performed on the time-series features within each channel corresponding to each charging cycle to obtain the cross-channel fused features corresponding to each charging cycle. Channel aggregation processing is performed on the cross-channel fusion features corresponding to each charging cycle to obtain the cyclic embedding vector corresponding to each charging cycle.

[0010] In one possible design, the step of using the cycle embedding vectors corresponding to each charging cycle as graph nodes, and performing inter-cycle degradation dependency modeling based on the temporal relationship between each charging cycle, yields the inter-cycle node embedding vectors corresponding to each charging cycle, including: Using the cycle embedding vectors corresponding to each charging cycle as graph nodes, and establishing directed edges based on the temporal sequence between each charging cycle, a directed graph between cycles is obtained; wherein, in the directed graph between cycles, the graph node corresponding to the subsequent charging cycle is connected to the graph node corresponding to the preceding charging cycle located before the subsequent charging cycle, so that the graph node corresponding to the subsequent charging cycle can aggregate the node features of the graph node corresponding to the preceding charging cycle. Graph attention processing is performed on the inter-cycle directed graph to obtain the inter-cycle node embedding vectors corresponding to each charging cycle.

[0011] In one possible design, the graph attention processing includes multi-head graph attention processing; The step of performing graph attention processing on the inter-cycle directed graph to obtain the inter-cycle node embedding vectors corresponding to each charging cycle includes: For any graph node in the intercyclic directed graph, under each attention head, determine the attention weight corresponding to the preceding graph node connected to the graph node; Under each attention head, the node features corresponding to the preceding graph node are weighted and aggregated based on the attention weight to obtain the aggregated node features of the graph node under that attention head; The aggregated node features of the graph node under multiple attention heads are fused to obtain the inter-cycle node embedding vector corresponding to the graph node.

[0012] In one possible design, determining the predicted remaining lifetime value corresponding to the charging cycle to be estimated based on the inter-cycle node embedding vector corresponding to the charging cycle to be estimated and the local health state sequence includes: Based on the inter-cycle node embedding vector corresponding to the charging cycle to be estimated, the remaining lifetime is predicted to obtain the feature-driven lifetime prediction result. Based on the local health state sequence, the remaining lifespan is predicted to obtain the health state-guided lifespan prediction result. The feature-driven lifetime prediction result and the health-state-guided lifetime prediction result are fused to obtain the remaining lifetime prediction value corresponding to the charging cycle to be estimated.

[0013] In one possible design, the method further includes: Acquire training samples, which include sample charging operation data of sample batteries in multiple sample charging cycles, health status labels corresponding to each sample charging cycle, and remaining life labels corresponding to the sample charging cycle to be predicted. The multiple sample charging cycles include the sample charging cycle to be predicted and multiple historical sample charging cycles located before the sample charging cycle to be predicted. The training samples are input into the collaborative estimation model to obtain the health status estimate corresponding to each sample charging cycle and the remaining lifetime prediction value corresponding to the charging cycle of the sample to be predicted; wherein, the collaborative estimation model is used to perform the intra-cycle degradation feature encoding, the inter-cycle degradation dependency modeling, the local health status sequence determination, and the remaining lifetime prediction value determination. The estimated health status loss is determined based on the estimated health status value and health status label corresponding to each sample charging cycle, and the predicted remaining lifetime loss is determined based on the predicted remaining lifetime value and remaining lifetime label corresponding to the sample charging cycle to be predicted. A joint loss is constructed based on the health status estimation loss and the remaining lifespan prediction loss, and the model parameters of the collaborative estimation model are updated based on the joint loss to obtain the trained collaborative estimation model.

[0014] In one possible design, the method further includes: When adapting the trained collaborative estimation model to the battery type to be adapted, the target domain samples corresponding to the battery type to be adapted are obtained, wherein the battery type to be adapted is a lithium-ion battery type not covered by the training samples. Freeze the model parameters used to perform in-loop degenerate feature encoding in the trained collaborative estimation model; In the trained collaborative estimation model, a low-rank adaptation branch is set in the model part used to perform inter-cycle degenerate dependency modeling. The low-rank adaptation branch is set in parallel with the main branch of the model part and is used to adapt and correct the output results of the inter-cycle degenerate dependency modeling. Based on the target domain samples, the parameters of the low-rank adaptation branch are updated, and the parameters of the output part of the trained collaborative estimation model used to determine the health status estimate and the remaining life prediction are adjusted to obtain a collaborative estimation model adapted to the battery type to be adapted.

[0015] Secondly, this application provides a system for co-estimating the health status and remaining life of a lithium-ion battery, the system comprising: The data acquisition module is used to acquire charging operation data of the target lithium-ion battery in multiple charging cycles, wherein the multiple charging cycles include the charging cycle to be estimated and multiple historical charging cycles located before the charging cycle to be estimated. The sample construction module is used to construct a multi-cycle input sample corresponding to the charging cycle to be estimated based on the charging operation data. The multi-cycle input sample includes multiple cycle data arranged in chronological order. The multiple cycle data correspond one-to-one with the multiple charging cycles and are obtained from the charging operation data of the corresponding charging cycle. The intra-cycle encoding module is used to encode the intra-cycle degradation features of the cyclic data corresponding to each charging cycle in the multi-cycle input samples to obtain the cyclic embedding vector corresponding to each charging cycle. The inter-cycle modeling module is used to model inter-cycle degradation dependencies based on the temporal relationship between charging cycles, using the cycle embedding vectors corresponding to each charging cycle as graph nodes, and to obtain the inter-cycle node embedding vectors corresponding to each charging cycle. The health state sequence determination module is used to determine the health state estimate corresponding to each charging cycle based on the inter-cycle node embedding vector corresponding to each charging cycle, and generate the local health state sequence corresponding to the multi-cycle input sample in chronological order. A health status determination module is used to determine the estimated health status value corresponding to the charging cycle to be estimated from the local health status sequence; The remaining lifetime prediction module is used to determine the remaining lifetime prediction value corresponding to the charging cycle to be estimated based on the inter-cycle node embedding vector corresponding to the charging cycle to be estimated and the local health state sequence. The output module is used to output the estimated health status and predicted remaining life corresponding to the charging cycle to be estimated.

[0016] Thirdly, this application provides an electronic device, including: a memory and at least one processor; The memory stores computer-executed instructions; The at least one processor executes computer execution instructions stored in the memory, causing the at least one processor to perform the method described in the first aspect or various possible designs of the first aspect.

[0017] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed, implement the method described in the first aspect or various possible designs of the first aspect.

[0018] Fifthly, this application provides a computer program product, which includes computer program code that, when run on a computer, causes the computer to implement the method described in the first aspect or various possible designs of the first aspect.

[0019] This application provides a method and system for co-estimating the health status and remaining life of a lithium-ion battery. The method constructs a multi-cycle input sample by acquiring charging operation data from the charging cycle to be estimated and several previous historical charging cycles, enabling the estimation process to utilize the current operating state of the charging cycle to be estimated and the degradation evolution information reflected by historical charging cycles. By encoding intra-cycle degradation features into the cycle data corresponding to each charging cycle, a cycle embedding vector is obtained for each charging cycle, converting the operating data in each charging cycle into a feature representation that characterizes the degradation state of that cycle. Furthermore, based on the temporal relationship between charging cycles, inter-cycle degradation dependency modeling is performed to obtain inter-cycle node embedding vectors for each charging cycle, allowing the degradation correlation and long-term aging trend between different charging cycles to participate in subsequent estimation. On this basis, a local health status sequence is generated according to the inter-cycle node embedding vectors, and the remaining life prediction value is determined by combining the inter-cycle node embedding vectors corresponding to the charging cycle to be estimated and the local health status sequence, thus establishing a correlation between the health status estimation result and the remaining life prediction process. Therefore, this application can improve the consistency, accuracy and reliability of the co-estimation of the health status and remaining life of lithium-ion batteries, even when the battery degradation process is affected by multiple coupled factors and the operating data is difficult to fully reflect the long-term aging trend. Attached Figure Description

[0020] Figure 1 A flowchart illustrating a method for co-estimating the health status and remaining life of a lithium-ion battery, provided in an embodiment of this application; Figure 2 A schematic diagram of the architecture of a lithium-ion battery health status and remaining life co-estimation model provided in an embodiment of this application; Figure 3 A schematic diagram of the architecture of an intra-cycle encoder provided in an embodiment of this application; Figure 4 A schematic diagram illustrating the relationship between the predicted and actual values ​​of each battery in the test set provided in this application embodiment; Figure 5 This is a schematic diagram of the structure of a lithium-ion battery health status and remaining life co-estimation system provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims and drawings of this application are intended to cover non-exclusive inclusion.

[0023] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of the phrase "embodiment" in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0024] In this article, the term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists, A and B can exist simultaneously, and B exists. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0025] Furthermore, the terms "first," "second," etc., in the specification and claims of this application or in the aforementioned drawings are used to distinguish different objects rather than to describe a specific order, and may explicitly or implicitly include one or more of the features.

[0026] In the description of this application, unless otherwise stated, "multiple" and "at least two" mean two or more (including two), and similarly, "multiple groups" and "at least two groups" mean two or more (including two groups).

[0027] In the description of this application, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, "connected" or "linked" can refer not only to a physical connection, but also to an electrical connection or a signal connection. For instance, it can be a direct connection, i.e., a physical connection, or an indirect connection through at least one intermediate component, as long as the circuit is connected. It can also refer to the internal connection between two components. A signal connection can refer not only to a signal connection through a circuit, but also to a signal connection through a medium, such as radio waves. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0028] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, different technical features in this application can be combined with each other.

[0029] Among related technologies, lithium-ion batteries are widely used in electric vehicles, portable electronic devices, grid energy storage, and aerospace due to their high energy density and long cycle life. During long-term charge and discharge cycles, irreversible electrochemical and mechanical aging may occur within the battery, leading to a gradual decrease in usable capacity, a gradual increase in internal resistance, and potentially safety risks such as thermal runaway. Therefore, battery management systems (BMS) typically need to estimate the battery's State of Health (SOH) and Remaining Useful Life (RUL) to provide a basis for energy management, safety warnings, and predictive maintenance.

[0030] Currently, methods for estimating the health status and predicting the remaining life of lithium-ion batteries mainly include model-based methods and data-driven methods. Model-based methods typically use equivalent circuit models (ECMs) or electrochemical models (EMs) to describe the battery's operating behavior and adjust the model parameters based on measurement data to achieve battery state estimation. ECMs characterize the battery's external dynamic response through equivalent electrical components such as resistance and capacitance, offering high computational efficiency and ease of engineering implementation, but their ability to depict complex aging mechanisms is limited. Electrochemical models describe the internal electrochemical reaction processes of the battery through differential equations, providing better mechanistic explanation and simulation accuracy, but their complex structure and numerous parameters result in high computational costs for model construction, parameter identification, and online prediction. Therefore, model-based methods often struggle to balance modeling accuracy, computational efficiency, and engineering applicability in practical applications.

[0031] Data-driven approaches typically establish a mapping between degradation characteristics and capacity decay based on large amounts of battery operating data, without requiring the construction of complex physical models. These methods can be further divided into traditional machine learning methods and deep learning methods. Traditional machine learning methods usually require manually extracting health features from raw battery data, and then using regression models such as Gaussian Process Regression (GPR) and Support Vector Machine (SVM) to estimate health status or predict remaining lifespan. However, manual feature extraction is highly dependent on domain experience, the feature construction process is time-consuming, and the applicability of features varies across different battery types and operating conditions, limiting its widespread application in practical industrial scenarios. Deep learning methods can automatically learn the latent features of battery degradation behavior from raw measurement data or charging curves. Commonly used models include Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and their combinations. In recent years, graph neural networks (GNNs) have been used for battery degradation modeling because they can process graph-structured data and characterize the dependencies between nodes. Examples include graph convolutional networks (GCNs), attention-enhanced graph neural networks, and graph convolutional networks combined with bidirectional long short-term memory networks (GCN-BiLSTM).

[0032] However, in actual battery state estimation, the degradation process of lithium-ion batteries is affected by multiple factors such as battery material system, charge / discharge rate, ambient temperature, cycle number, and operating conditions. The degradation behavior is nonlinear, cumulative, and exhibits individual variability. Some existing data-driven methods primarily focus on the temporal characteristics of raw measurement data over time, failing to adequately utilize the interrelationships between variables such as current, voltage, and charging capacity during charging. While some methods introduce graph structure models, they often emphasize dependencies within a single cycle or between artificial features, insufficiently characterizing long-term degradation correlations across multiple charging cycles. Furthermore, although both battery health state estimation and remaining lifespan prediction are related to the degree of battery degradation, in some existing schemes they are often treated as independent tasks, or the health state degradation trend is estimated first, and then the remaining lifespan is inferred based on the failure threshold. This sequential processing approach is prone to error propagation, affecting the accuracy and stability of long-term predictions.

[0033] Therefore, in estimating the health status and remaining lifespan of lithium-ion batteries, how to more fully characterize the battery degradation state based on battery charging operation data, improve the ability to depict long-term degradation trends, and enhance the consistency between the health status estimation results and the remaining lifespan prediction results have become problems that need to be further solved.

[0034] To address the aforementioned issues, this application proposes a method and system for co-estimating the health status and remaining life of lithium-ion batteries. Instead of relying solely on data from a single charging cycle for isolated battery state estimation, this method uses the charging cycle to be estimated and multiple preceding historical charging cycles as the basis for state estimation. First, degradation characteristics are characterized for the cycle data corresponding to each charging cycle. Then, the temporal relationship between charging cycles is combined to model cross-cycle degradation correlations, allowing current operating state and historical degradation evolution information to jointly participate in the health status and remaining life estimation. Furthermore, after obtaining the health status estimation results for each charging cycle, this application generates a health status sequence reflecting local degradation trends. This health status sequence is used to assist in predicting the remaining life of the charging cycle to be estimated, establishing a correlation between the health status estimation results and the remaining life prediction process. This improves the accuracy, consistency, and reliability of the co-estimation of lithium-ion battery health status and remaining life.

[0035] Figure 1 This is a flowchart illustrating a method for co-estimating the health status and remaining life of a lithium-ion battery, as provided in an embodiment of this application. Figure 1 As shown, the method provided in this application embodiment specifically includes S101 to S108, and S101 to S108 will be described in detail below.

[0036] It should be noted that the method provided in this application embodiment can be executed by a lithium-ion battery health status and remaining life collaborative estimation system, or by a battery management system, vehicle controller, edge computing device, collaborative estimation system or other electronic devices with data processing capabilities. This application does not limit the execution of such methods.

[0037] Figure 2 This is a schematic diagram illustrating the architecture of a lithium-ion battery health status and remaining lifespan co-estimation model provided in an embodiment of this application. Figure 2 As shown, the model may include an input layer, an intra-loop dependency modeling module, an inter-loop dependency modeling module, and an output header. The input layer is used to receive loop data corresponding to the charging cycle to be estimated and multiple previous historical charging cycles.

[0038] The intra-cycle dependency modeling module is used to encode intra-cycle degradation features for each charging cycle in the multi-cycle input samples. Specifically, the cyclic data corresponding to each charging cycle is first mapped to the latent space by linear projection and then fused with the cyclic position encoding to preserve the temporal order information of each charging cycle in the input window. Subsequently, the intra-cycle encoder models the intra-variable temporal dependencies and inter-variable relationships within a single charging cycle to obtain the cyclic embedding vector corresponding to each charging cycle. Figure 2 Multiple parallel intra-cycle encoding processing units in the middle indicate that different charging cycles can be processed by intra-cycle dependency modeling with the same structure to obtain the corresponding cyclic embedding.

[0039] The inter-cycle dependency modeling module is used to take the cycle embedding vectors corresponding to each charging cycle as graph nodes and construct an inter-cycle directed graph based on the temporal relationship between each charging cycle. Figure 2 The graph attention layer in the module is used to perform graph attention processing on the directed graph between cycles, enabling graph nodes corresponding to subsequent charging cycles to aggregate the node features of graph nodes corresponding to previous charging cycles, and dynamically distinguishing the differences in the contributions of different historical charging cycles to the current degradation state. After processing by the inter-cycle dependency modeling module, the updated node embeddings corresponding to each charging cycle can be obtained, that is, the inter-cycle node embedding vectors.

[0040] The output head is used for health status estimation and remaining lifetime prediction based on the inter-cycle node embedding vectors. Specifically, the health status estimation head can output the health status estimate for each charging cycle based on the inter-cycle node embedding vectors corresponding to each charging cycle within the window, and construct a local health status sequence in chronological order. The dual-branch remaining lifetime prediction head includes a feature-driven lifetime prediction branch and a health status-guided lifetime prediction branch. The feature-driven lifetime prediction branch predicts the remaining lifetime based on the inter-cycle node embedding vectors corresponding to the charging cycle to be estimated, while the health status-guided lifetime prediction branch predicts the remaining lifetime based on the local health status sequence. The prediction results of the two branches are fused to obtain the remaining lifetime prediction value corresponding to the charging cycle to be estimated. Thus, Figure 2 The model shown can achieve short-term degradation feature extraction within a cycle, long-term degradation dependency modeling between cycles, and co-estimation of health status and remaining lifespan within the same framework.

[0041] S101. Obtain charging operation data of the target lithium-ion battery in multiple charging cycles.

[0042] The plurality of charging cycles includes the charging cycle to be estimated and a plurality of historical charging cycles preceding the charging cycle to be estimated.

[0043] It should be noted that the target lithium-ion battery is a single battery cell that requires health status estimation and remaining life prediction.

[0044] A charging cycle refers to the operational process corresponding to the completion of one charging process of a target lithium-ion battery. In practical applications, the charging stage that meets the preset acquisition conditions during a single charging process can also be used as the charging cycle data source for state estimation.

[0045] The charging cycle to be estimated is the charging cycle for which the current health status estimate and remaining lifetime prediction are required.

[0046] Historical charging cycles are charging cycles that occur before the charging cycle to be estimated and can reflect the historical usage and degradation evolution of the target lithium-ion battery.

[0047] Charging operation data refers to the operational measurement data collected during the charging process, which is used to reflect the changes in the electrochemical state and degradation-related performance of the target lithium-ion battery in the corresponding charging cycle.

[0048] In this embodiment, by simultaneously acquiring charging operation data from the charging cycle to be estimated and multiple historical charging cycles, the subsequent estimation process no longer relies solely on the local information of a single charging cycle. Instead, it can use the current charging cycle's operating state and the degradation evolution information accumulated in historical charging cycles as the estimation basis, providing a temporally continuous original data foundation for subsequent health status estimation and remaining life prediction.

[0049] S102. Construct multi-cycle input samples corresponding to the charging cycle to be estimated based on the charging operation data.

[0050] The multi-cycle input sample includes multiple cycle data arranged in chronological order. Each of the multiple cycle data corresponds to one of the multiple charging cycles and is obtained from the charging operation data of the corresponding charging cycle.

[0051] It should be noted that the cycle data is a single-cycle data representation obtained by processing the charging operation data of the corresponding charging cycle.

[0052] For any one of the multiple charging cycles, a cycle data can be formed based on the charging operation data corresponding to that charging cycle; multiple cycle data are arranged in chronological order of the corresponding charging cycles to form a multi-cycle input sample corresponding to the charging cycle to be estimated.

[0053] The multi-cycle input sample is not an isolated single-cycle data, but a dataset containing multiple time-related cycle data. It includes both the cycle data corresponding to the charging cycle to be estimated and the cycle data corresponding to multiple historical charging cycles that precede the charging cycle to be estimated.

[0054] In one implementation, the multi-cycle input samples can be constructed using a sliding window approach. Using the charging cycle to be estimated as the end of the window, several charging cycles are selected from the charging cycle to be estimated and several historical charging cycles preceding it. The cyclic data corresponding to these charging cycles are then arranged in chronological order to obtain the multi-cycle input samples corresponding to the charging cycle to be estimated. This method preserves the local degradation evolution process within a certain historical range preceding the charging cycle to be estimated in a single input sample, enabling subsequent models to handle single-cycle operation characteristics and cross-cycle degradation trends within a unified input structure.

[0055] S103. Perform intracyclic degradation feature encoding on the cyclic data corresponding to each charging cycle in the multi-cycle input samples to obtain the cyclic embedding vector corresponding to each charging cycle.

[0056] Among them, the intra-cycle degradation feature encoding is used to extract features from the running data within a single charging cycle, so that the cyclic data corresponding to each charging cycle is converted into a cyclic embedding vector that can characterize the degradation state of the charging cycle.

[0057] For any charging cycle in the multi-cycle input sample, the corresponding cycle data can be encoded to extract operational change features related to the battery's health status within that charging cycle. These operational change features can reflect the local dynamic behavior of the target lithium-ion battery within that charging cycle, such as data change trends, data format differences, and implicit features related to capacity decay during the charging process.

[0058] The cyclic embedding vector corresponding to the charging cycle is the vectorized representation of the charging cycle in the feature space, which has a higher level of degenerate expression capability compared to the original cyclic data.

[0059] By encoding the degradation features within each charging cycle, each charging cycle can form an independent degradation state representation, providing a node feature basis for subsequent inter-cycle degradation dependency modeling using each charging cycle as a graph node.

[0060] S104. Using the cycle embedding vectors corresponding to each charging cycle as graph nodes, perform cycle degradation dependency modeling based on the temporal relationship between each charging cycle to obtain the cycle node embedding vectors corresponding to each charging cycle.

[0061] Specifically, each charging cycle in the multi-cycle input samples can be abstracted as a graph node, and the cycle embedding vector corresponding to the charging cycle can be used as the initial node feature of the graph node.

[0062] Since lithium-ion battery degradation is typically cumulative and time-dependent, the degradation state of a later charging cycle is usually influenced by multiple previous historical charging cycles. Therefore, inter-cycle dependencies can be established based on the temporal relationship between charging cycles, allowing graph nodes corresponding to subsequent charging cycles to utilize the historical degradation information contained in graph nodes corresponding to previous charging cycles.

[0063] Intercycle degradation dependency modeling is used to transfer and fuse degradation-related information between multiple charging cycles, so that the node features of subsequent graph nodes not only include the intracycle degradation features of the corresponding charging cycle itself, but also the cross-cycle degradation dependency information related to the preceding charging cycle.

[0064] The inter-cycle node embedding vector, obtained after inter-cycle degradation dependency modeling, can characterize the comprehensive degradation state of the corresponding charging cycle after incorporating historical degradation evolution information. Compared to using only the cycle embedding vector, the inter-cycle node embedding vector can more fully reflect the long-term aging trend of the target lithium-ion battery across multiple charging cycles.

[0065] S105. Based on the inter-cycle node embedding vector corresponding to each charging cycle, determine the health state estimate corresponding to each charging cycle, and generate the local health state sequence corresponding to the multi-cycle input sample in chronological order.

[0066] Among them, the health status estimate is used to characterize the health level of the target lithium-ion battery under the corresponding charging cycle.

[0067] For each charging cycle in the multi-cycle input sample, the corresponding inter-cycle node embedding vector can be input into the health state estimation unit, which then outputs the health state estimate for each charging cycle. Since the inter-cycle node embedding vector can characterize the intra-cycle degradation features and inter-cycle degradation dependency information of the corresponding charging cycle, determining the health state estimate based on this inter-cycle node embedding vector allows the health state estimation process to utilize single-cycle local states and cross-cycle historical degradation information.

[0068] After obtaining the health state estimates for each charging cycle, the health state estimates can be arranged according to the time order of each charging cycle in the multi-cycle input samples to generate a local health state sequence corresponding to the multi-cycle input samples.

[0069] Local health state sequences represent the changes in health state over multiple charging cycles near the estimated charging cycle, reflecting the degradation trend of the target lithium-ion battery within a local timeframe. Compared to outputting only a single health state estimate, local health state sequences provide more continuous information on health state changes, offering explicit evidence of health evolution for subsequent remaining lifetime prediction.

[0070] S106. Determine the estimated health state value corresponding to the charging cycle to be estimated from the local health state sequence.

[0071] Since the local health state sequence includes the health state estimates corresponding to each charging cycle within the multi-cycle input sample, and the health state estimates are arranged in chronological order of the corresponding charging cycles, the health state estimate corresponding to the charging cycle to be estimated can be determined from the local health state sequence based on the time position of the charging cycle to be estimated in the multi-cycle input sample.

[0072] In general, the charging cycle to be estimated is the last charging cycle in the time sequence of the multi-cycle input samples. In this case, the last health state estimate in the local health state sequence can be determined as the health state estimate corresponding to the charging cycle to be estimated.

[0073] In this embodiment, the local health state sequence obtained from multiple charging cycles in the multi-cycle input samples can be converted into the health state output result of the charging cycle to be estimated.

[0074] S107. Determine the predicted remaining lifetime value corresponding to the charging cycle to be estimated based on the inter-cycle node embedding vector corresponding to the charging cycle to be estimated and the local health state sequence.

[0075] Among them, the inter-cycle node embedding vector corresponding to the charging cycle to be estimated represents the implicit degradation state of the charging cycle to be estimated after combining the degradation dependency of the historical charging cycle; the local health state sequence represents the trend of health state change between multiple adjacent or spaced historical charging cycles and the charging cycle to be estimated.

[0076] The remaining lifetime prediction process utilizes both the inter-cycle node embedding vectors corresponding to the charging cycles to be estimated and the local health state sequence, which allows the prediction results to be constrained by the current overall degradation state and guided by the evolution trend of the health state.

[0077] In a practical implementation, the inter-cycle node embedding vector corresponding to the charging cycle to be estimated can be used as the input of the implicit degradation feature, and the local health state sequence can be used as the input of the explicit health evolution information. Based on the two, the remaining lifetime prediction value corresponding to the charging cycle to be estimated can be determined.

[0078] The remaining lifetime prediction can be used to represent the number of cycles that a target lithium-ion battery is expected to continue to experience from the estimated charging cycle until it reaches the preset failure condition, or it can be used to represent other lifetime indicators related to remaining lifetime.

[0079] In this embodiment, since the determination process of the remaining lifetime prediction value utilizes the local health state sequence, the health state estimation result is no longer separated from the remaining lifetime prediction result, which helps to reduce the error propagation caused by first estimating the health state degradation trend and then separately inferring the remaining lifetime.

[0080] S108. Output the estimated health status and predicted remaining life corresponding to the charging cycle to be estimated.

[0081] The estimated health status and predicted remaining life can be output as a collaborative estimation result of the target lithium-ion battery under the estimated charging cycle. The output can be stored or displayed locally, or sent to a battery management system, vehicle controller, energy storage management platform, collaborative estimation system, or maintenance terminal. The estimated health status is used to determine the current health level of the target lithium-ion battery, while the predicted remaining life is used to assist in formulating charge / discharge control strategies, maintenance and replacement plans, safety warning strategies, or operation scheduling strategies.

[0082] Since both the health status estimate and the remaining life expectancy prediction are based on the same multi-cycle input sample and utilize degradation representations obtained from in-cycle degradation feature encoding and inter-cycle degradation dependency modeling, the two types of output results are consistent in terms of data source, feature basis, and degradation representation. This avoids the inconsistency issues that arise from separate modeling of health status estimation and remaining life expectancy prediction.

[0083] This application first acquires charging operation data of the charging cycle to be estimated and several previous historical charging cycles, and constructs multi-cycle input samples arranged in chronological order, so that the current operating state and historical degradation evolution information can be jointly included in the estimation process. Then, the cycle data corresponding to each charging cycle are encoded with intra-cycle degradation features to obtain a cycle embedding vector that can characterize the local degradation state of a single cycle. Subsequently, using the cycle embedding vector as graph nodes, and combining the temporal relationship between each charging cycle, inter-cycle degradation dependency modeling is performed, so that the long-term degradation correlation across cycles is integrated into the inter-cycle node embedding vector. Further, a local health state sequence is generated based on the inter-cycle node embedding vector, and the remaining lifetime prediction value is jointly determined using the local health state sequence and the inter-cycle node embedding vector corresponding to the charging cycle to be estimated, so that the health state estimation result can participate in the remaining lifetime prediction process. Therefore, this application can improve the accuracy, consistency, and reliability of the co-estimation of the health state and remaining lifetime of lithium-ion batteries, especially when the degradation process of lithium-ion batteries is affected by multiple coupled factors, existing methods cannot fully utilize operating data to characterize long-term aging trends, and the consistency between the health state estimation result and the remaining lifetime prediction result is insufficient.

[0084] In one possible embodiment, the charging operation data includes current data, voltage data, and charging capacity data.

[0085] Among them, current data is used to characterize the current change of the target lithium-ion battery during the charging process, voltage data is used to characterize the terminal voltage change of the target lithium-ion battery during the charging process, and charging capacity data is used to characterize the capacity accumulation or capacity change of the target lithium-ion battery during the charging process.

[0086] By simultaneously using current data, voltage data, and charging capacity data, the operating state of the target lithium-ion battery during the charging cycle can be described from different dimensions, providing a multivariate input basis for subsequent construction of cycle data.

[0087] In one possible embodiment, the method steps shown in S102 can be implemented by S1021 to S1024, which are described in detail below.

[0088] S1021. Extract the charging segments within the set voltage range from the charging operation data of each charging cycle to obtain the charging segment data corresponding to each charging cycle.

[0089] The voltage range can be set according to the target lithium-ion battery type, rated voltage range, or actual data acquisition conditions.

[0090] For example, in one implementation, only the charging segment with a voltage range of 3.8V to 4.2V in each charging cycle can be selected. Since the changes in current, voltage, and charging capacity within this voltage range can reflect degradation-related characteristics during battery charging, effective input data for state estimation can be obtained without relying on complete charge-discharge data by selecting charging segments within a set voltage range.

[0091] S1022. Perform fixed-length resampling and normalization on the charging segment data corresponding to each charging cycle to obtain the cycle data corresponding to each charging cycle.

[0092] The length of charging segments within a set voltage range may vary in different charging cycles. To ensure a unified data structure for the data across different charging cycles, each charging segment can be resampled into data with a fixed length of T time steps. Resampling can employ interpolation, equal-interval sampling, or other methods capable of converting variable-length sequences into fixed-length sequences. After fixed-length resampling, each charging cycle forms a single-cycle data representation with the same number of time steps, facilitating subsequent batch input and feature encoding.

[0093] During the normalization process, the current data, voltage data, and charging capacity data can be scaled separately. Taking the t-th time step as an example, the normalization process can be expressed as: , , ;in, This represents the normalized current data. This represents the normalized charging capacity data. This represents the normalized voltage data. This represents the current data before normalization. This represents the charging capacity data before normalization. This represents the voltage data before normalization. The rated capacity of the target lithium-ion battery. To extract the maximum voltage in the sequence, t∈{1,2,…,T}, where T is the number of time steps after fixed-length resampling.

[0094] The above normalization process can reduce the impact of different battery rated capacities, different charging segment voltage ranges, and differences in the dimensions of different variables on the model input, so that the cyclic data corresponding to each charging cycle has a more consistent numerical scale.

[0095] S1023. Based on the charging cycle to be estimated, select a preset number of historical cycle data from the cycle data corresponding to each historical charging cycle before the charging cycle to be estimated, according to a preset sampling interval.

[0096] The charging cycle to be estimated is the charging cycle for which health status estimation and remaining lifetime prediction are currently required. Using the charging cycle to be estimated as a reference means using the time position of the charging cycle to be estimated among multiple charging cycles as a reference position, and backtracking to select historical charging cycles before the charging cycle to be estimated.

[0097] The preset sampling interval is used to limit the cycle interval between adjacent selected historical charging cycles, and the preset quantity is used to limit the number of historical cycle data that participate in the construction of multi-cycle input samples.

[0098] For example, if the charging cycle to be estimated is the k-th charging cycle, and the preset sampling interval is s, then from the historical charging cycles preceding the k-th charging cycle, the ks-th charging cycle, the (k-2s-th)-th charging cycle can be selected sequentially at intervals s, until a historical charging cycle meeting the preset quantity requirement is selected. Accordingly, multiple historical cycle data can be obtained. By selecting historical cycle data according to the preset sampling interval, it is possible to avoid insufficient coverage of historical degradation information due to using only a few charging cycles adjacent to the charging cycle to be estimated, and also to avoid data redundancy caused by including too many consecutive historical cycles in the input sample. Therefore, the selected historical cycle data can reflect the degradation evolution within a certain historical range before the charging cycle to be estimated, while also controlling the data scale of multiple cycle input samples.

[0099] In practical applications, the preset sampling interval and preset quantity can be set according to the cycle life range of the target lithium-ion battery, sampling density, model input size, and computational resource conditions. For example, while ensuring coverage of historical degradation trends, the input redundancy can be reduced by increasing the preset sampling interval, and the time coverage of historical degradation information can be expanded by increasing the preset quantity.

[0100] S1024. Arrange the preset number of historical cycle data and the cycle data corresponding to the charging cycle to be estimated in chronological order to construct a multi-cycle input sample corresponding to the charging cycle to be estimated.

[0101] Specifically, after obtaining a preset number of historical cyclic data, the historical cyclic data can be arranged from earliest to latest according to the occurrence time of their corresponding historical charging cycles, and the cyclic data corresponding to the charging cycle to be estimated can be arranged last, forming a multi-cycle input sample corresponding to the charging cycle to be estimated. Therefore, the cyclic data in the multi-cycle input sample has a clear temporal order, enabling subsequent intra-cycle degradation feature encoding and inter-cycle degradation dependency modeling to identify the sequential relationship between different cyclic data.

[0102] In one implementation, a sliding window strategy can be used to construct multi-cycle input samples. If the charging cycle to be estimated is the k-th charging cycle, and the multi-cycle input samples include N cycle data points with a preset sampling interval of s, then the multi-cycle input samples can be represented as: ;in, For the multi-cycle input sample corresponding to the k-th charging cycle, The data corresponds to the charging cycle to be estimated. to The data consists of historical cyclic data selected according to a preset sampling interval s. N represents the total number of cyclic data included in the multi-cycle input samples, 3 indicates that each cyclic data includes three data channels: current data, voltage data, and charging capacity data, and T represents the number of time steps corresponding to each data channel. Represents the real number field. Indicates multiple cyclic input samples It is a real tensor with dimensions N×3×T.

[0103] Through the above arrangement, the multi-cycle input sample can simultaneously contain the current operating status information of the charging cycle to be estimated and the degradation evolution information of multiple historical charging cycles. Since the cycle data in the multi-cycle input sample are arranged in chronological order, the subsequent model can extract the degradation features of individual charging cycles within a cycle and utilize the temporal evolution relationship between historical charging cycles and the charging cycle to be estimated between cycles when processing this multi-cycle input sample, thereby providing a more complete input basis for health status estimation and remaining lifetime prediction.

[0104] In one possible embodiment, the method steps shown in S103 can be implemented by S1031 to S1033, which are described in detail below.

[0105] S1031. Perform feature mapping processing on the cyclic data corresponding to each charging cycle to obtain the initial latent space features corresponding to each charging cycle.

[0106] Among them, feature mapping processing is used to map the cyclic data corresponding to each charging cycle from the original data space to the latent space, so as to improve the feature representation ability of the cyclic data.

[0107] Specifically, a learnable linear projection matrix can be used to linearly project each cyclic data in the multi-cyclic input samples, so that each cyclic data is mapped from the original time step dimension to a preset latent dimension space.

[0108] In one implementation, if the multi-cycle input sample corresponding to the charging cycle to be estimated is... ,and Then, a learnable linear projection matrix E can be used to... Perform feature mapping.

[0109] in, , Let be the dimension of the latent space. After feature mapping, the cyclic data corresponding to each charging cycle is transformed into the corresponding initial latent space features. The initial latent space features can provide a unified latent space input for subsequent cyclic position encoding fusion and intra-cycle degenerate feature encoding.

[0110] S1032. The cyclic position codes corresponding to each charging cycle are fused with the initial latent space features to obtain the position enhancement features corresponding to each charging cycle.

[0111] Among them, the cyclic position encoding is used to characterize the temporal order position of each charging cycle in the multi-cycle input sample.

[0112] Since the multiple cyclic data in the multi-cycle input sample are arranged in chronological order, the relative positions of different cyclic data in the sample can reflect the temporal relationship between the corresponding charging cycle and the charging cycle to be estimated. Therefore, fusing cyclic position encoding into the initial latent space features can enable the subsequent model to perceive the sequential information of each charging cycle when extracting degradation features.

[0113] In one implementation, the cyclic position code can be a learnable cyclic position code. , Where 1 indicates that the cyclic position code is shared across different data channels. Let be the dimension of the hidden dimension space. By sharing the cyclic position encoding across different data channels, it can be ensured that different data channels have consistent cyclic position information within the same charging cycle.

[0114] In one implementation, the location-enhancing feature can be represented as: ;in, For location enhancement features, .

[0115] Through the above fusion process, the location enhancement features not only include the latent space features obtained by mapping the cyclic data, but also the temporal order information of each charging cycle in the multi-cycle input samples.

[0116] S1033. Encode the position enhancement features corresponding to each charging cycle with intracyclic degradation features to obtain the cyclic embedding vector corresponding to each charging cycle.

[0117] Among them, the position enhancement features can be input into the intracyclic encoder, which encodes the degradation-related features within each charging cycle to obtain the cyclic embedding vector corresponding to each charging cycle.

[0118] Cyclic embedding vectors are used to characterize the intra-cycle degradation state of the corresponding charging cycle and serve as graph node features in subsequent inter-cycle degradation dependency modeling.

[0119] In this embodiment, the cyclic data corresponding to each charging cycle is first mapped to the latent space, then the cyclic position encoding is fused to preserve the temporal order information of each charging cycle in the multi-cycle input samples, and finally the cyclic embedding vector corresponding to each charging cycle is obtained based on the position enhancement features. This provides a node feature foundation that simultaneously includes intra-cycle degradation representation and cycle order information for subsequent inter-cycle degradation dependency modeling.

[0120] Figure 3 This is a schematic diagram of the architecture of an intra-cycle encoder provided in an embodiment of this application. Figure 3 As shown, the intracyclic encoder may include a depthwise convolution processing section, a residual connection section, a pointwise convolution processing section, and an output section. The input of the intracyclic encoder can be the position enhancement features corresponding to each charging cycle. These position enhancement features include latent space features obtained by mapping the cyclic data and temporal order information provided by the cyclic position encoding.

[0121] The depthwise convolution processing section is used to perform intra-channel temporal encoding on different data channels. Specifically, different data channels, such as current data, voltage data, and charging capacity data, can be subjected to one-dimensional temporal convolution to extract the temporal dynamic features of the corresponding variables within each data channel. Since depthwise convolution is performed independently within each data channel, it can avoid excessive mixing of features of different variables in the early stages, thus fully preserving the temporal variation patterns of each variable. Figure 3 The residual connection in the input is used to add the position enhancement features of the input depthwise convolution processing part to the features after depthwise convolution processing, so as to retain the original feature information while extracting the temporal features within the channel, and reduce feature degradation or information loss.

[0122] The pointwise convolution processing section is used to perform cross-channel feature fusion and channel aggregation on the intra-channel temporal features obtained from depthwise convolution. Specifically, the first pointwise convolution can interactively fuse features from different data channels along the channel dimension to extract correlation features between current data, voltage data, and charging capacity data; the second pointwise convolution can further aggregate the multi-channel fused features into single-channel cyclic embedding vectors, so that each charging cycle forms a vector representation that can be used as a graph node feature. Thus, Figure 3 The intra-cycle encoder shown uses a processing method of "intra-channel temporal coding - cross-channel feature fusion - channel aggregation" to simultaneously model the temporal dependencies within a single variable and the correlations between multiple variables, thereby improving the ability of the cyclic embedding vector to represent the degradation state of a single charging cycle.

[0123] In one possible embodiment, the method steps shown in S1033 can be implemented by S10331 to S10333, and S10331 to S10333 are described in detail below.

[0124] S10331. Perform intra-channel timing encoding on the timing features of different data channels in the position enhancement features corresponding to each charging cycle to obtain the intra-channel timing features corresponding to each charging cycle.

[0125] The location enhancement features include features corresponding to multiple data channels, such as features corresponding to the current data channel, voltage data channel, and charging capacity data channel.

[0126] For the same charging cycle, different data channels have their own timing variation patterns. If different data channels are directly mixed in the early stage of timing encoding, it may weaken the timing dynamic expression of each data channel. Therefore, in this embodiment, the timing characteristics of different data channels can be encoded within each channel first to extract the internal timing dynamic characteristics of each data channel.

[0127] In one implementation, depthwise convolution can be used to perform intra-channel temporal encoding of the positional enhancement features corresponding to each charging cycle. Depthwise convolution can perform one-dimensional temporal convolution independently on each data channel, allowing the current data channel, voltage data channel, and charging capacity data channel to complete temporal feature extraction within their own channels. This avoids premature mixing of different data channels in the early encoding stage and ensures that the local change trends and dynamic features within each data channel are fully extracted.

[0128] Location-enhanced features For example, the timing coding within a channel can be represented as: ;in, Indicates the timing characteristics within the channel. This represents a convolution operation with both C input and output channels, where C is the number of data channels. The kernel size is [size]. This represents a non-linear activation function, which can be the GELU activation function. BN represents the batch normalization operation, and the last term... This represents the positional enhancement features directly passed to the output. These positional enhancement features are added to the features obtained through convolution to form a residual connection. Through residual connections, the original feature information in the positional enhancement features can be preserved while extracting temporal features within a channel, reducing feature degradation or information loss.

[0129] S10332. Perform cross-channel feature fusion on the timing features within the channel corresponding to each charging cycle to obtain the cross-channel fused features corresponding to each charging cycle.

[0130] Among them, the intra-channel timing features mainly reflect the dynamic changes in time within each data channel. However, the degradation performance of lithium-ion batteries during charging is not only reflected in a single data channel; there are also interrelationships between current data, voltage data, and charging capacity data. To further extract the correlations between different data channels, cross-channel feature fusion can be performed on the intra-channel timing features.

[0131] In one implementation, pointwise convolution can be used to perform cross-channel feature fusion of temporal features within a channel. Pointwise convolution can be a convolution operation with a kernel size of 1, which can perform feature interaction in the data channel dimension without changing the time step dimension, thereby extracting spatial correlation features between different data channels.

[0132] Cross-channel feature fusion can be represented as: ;in, This indicates cross-channel fusion features, and kernel=1 indicates that the convolution kernel size is 1.

[0133] Through the above processing, based on the completion of intra-channel timing coding in each data channel, the correlation between current, voltage and charging capacity can be fused and modeled to obtain cross-channel fusion features that better characterize the internal degradation state of the charging cycle.

[0134] S10333. Perform channel aggregation processing on the cross-channel fusion features corresponding to each charging cycle to obtain the cyclic embedding vector corresponding to each charging cycle.

[0135] The cross-channel fusion features still retain multiple data channel dimensions. In order for each charging cycle to participate as a graph node in the degradation dependency modeling between subsequent cycles, the multi-channel features corresponding to the charging cycle need to be further aggregated into a single cyclic embedding vector, which is used as the node feature representation of the corresponding charging cycle.

[0136] In one implementation, pointwise convolution can be used for channel aggregation, aggregating multi-channel features into a single-channel cyclic embedding vector. Channel aggregation can be represented as: ;in, This represents the cyclic embedding vector after channel aggregation. This represents a pointwise convolution operation with C input channels and 1 output channel. Through this channel aggregation process, the fused features of different data channels in each charging cycle can be compressed and integrated into a cyclic embedding vector for the corresponding charging cycle.

[0137] It should be noted that the method steps shown in S10331 to S10333 can be executed by an intra-loop encoder. This intra-loop encoder first uses depthwise convolution to perform independent intra-channel temporal encoding on different data channels, then uses pointwise convolution to perform cross-channel feature fusion, and finally obtains the cyclic embedding vector through channel aggregation. This separation-then-fusion encoding method can, on the one hand, preserve the temporal dynamic features of current data, voltage data, and charging capacity data, and on the other hand, further extract the correlation between different data channels, forming a single-loop feature representation suitable for modeling inter-loop degradation dependencies.

[0138] This embodiment enables hierarchical modeling of intra-variable temporal dependencies and inter-variable relationships within a single charging cycle. The resulting cyclic embedding vector contains both the temporal degradation features of each data channel and the interactive degradation features between different data channels. This improves the ability of the cyclic embedding vector to represent the degradation state of a single charging cycle, providing a more accurate node feature foundation for subsequent inter-cycle degradation dependency modeling.

[0139] In one possible embodiment, the method steps shown in S104 can be implemented by S1041 to S1042, which are described in detail below.

[0140] S1041. Using the cycle embedding vectors corresponding to each charging cycle as graph nodes, and establishing directed edges based on the temporal relationship between each charging cycle, a directed graph between cycles is obtained.

[0141] In the directed graph between cycles, the graph node corresponding to the subsequent charging cycle is connected to the graph node corresponding to the preceding charging cycle, so that the graph node corresponding to the subsequent charging cycle can aggregate the node features of the graph node corresponding to the preceding charging cycle.

[0142] It should be noted that the multi-cycle input samples include multiple cycles of data arranged in chronological order. After encoding the degradation features within each cycle, each charging cycle has a corresponding cycle embedding vector. To further characterize the long-term degradation correlation between different charging cycles, the cycle embedding vectors corresponding to each charging cycle can be used as graph node features in the inter-cycle directed graph. That is, each charging cycle corresponds to a graph node, and the initial node feature of each graph node is the cycle embedding vector corresponding to that charging cycle.

[0143] Because the degradation process of lithium-ion batteries is cumulative over time, the degradation state of a subsequent charging cycle is usually related to the degradation state of one or more preceding charging cycles. Therefore, when constructing a directed graph between cycles, directed edges can be established according to the temporal relationship between each charging cycle. This allows graph nodes corresponding to subsequent charging cycles to access and aggregate the node features of graph nodes corresponding to preceding charging cycles. Through this directed edge configuration, the directed graph between cycles can reflect the directional relationship of historical degradation information transmission to subsequent charging cycles.

[0144] In one implementation, if the multi-cycle input samples include N cycle data, the corresponding graph nodes can be recorded in chronological order from earliest to latest. , … ,in, Corresponding to the charging cycle to be estimated, to The corresponding historical charging cycle is located before the charging cycle to be estimated.

[0145] For any node in the post-order graph It can make the nodes of the post-sequence graph With the preceding nodes of the sequence graph that satisfy the time order Connections. Accordingly, the adjacency relationships in a directed graph with cycles can be represented as: ;in, The adjacency matrix of a directed graph between cycles represents the i-th... The graph node and the first The connection relationships between nodes in the graph. Indicates the first The first graph node can access the second graph node during feature aggregation. Each graph node Indicates the first The first graph node is not accessed during feature aggregation. Each graph node and All are graph node numbers.

[0146] Through the above adjacency relationships, a lower triangular adjacency matrix can be formed, enabling each subsequent graph node to access all its preceding graph nodes as well as its own graph nodes, thereby perceiving complete historical information during feature aggregation.

[0147] If self-connection is not required in actual implementation, the above connection relationship can also be set as a subsequent graph node accessing its preceding graph node. This application does not limit this.

[0148] In this embodiment, by constructing a directed graph between cycles, the multiple charging cycles originally arranged sequentially in the multi-cycle input samples are transformed into a graph structure with time-direction constraints, allowing the degradation relationships between charging cycles to be expressed through the graph structure. Compared with processing multiple cycle data directly in a sequential manner, this graph structure can more intuitively limit the access relationship of subsequent charging cycles to the historical information of previous charging cycles, which is beneficial for subsequent modeling of long-term degradation dependencies across cycles.

[0149] S1042. Perform graph attention processing on the inter-cycle directed graph to obtain the inter-cycle node embedding vector corresponding to each charging cycle.

[0150] In a directed graph with cyclic connections, each graph node has a corresponding cyclic embedding vector, and directed edges define the direction of information transfer between the nodes. Graph attention processing on a directed graph with cyclic connections involves distinguishing the contribution of different preceding graph nodes to subsequent graph nodes within the node connection relationships defined by the graph, and updating node features based on the degenerate associations between graph nodes.

[0151] In actual degradation processes, the degree of influence of different historical charging cycles on the degradation state of the current or subsequent charging cycles may vary. For example, some historical charging cycles may contain more significant capacity decay or state change information, contributing more to the state estimation of subsequent charging cycles; while others may have a relatively weaker impact on subsequent charging cycles. Therefore, when processing directed graphs between cycles, graph attention processing can be used to dynamically determine the differences in contribution of different preceding graph nodes to subsequent graph nodes, and the node features of preceding graph nodes can be aggregated based on these contribution differences to update the feature representation of subsequent graph nodes.

[0152] The inter-cycle node embedding vector obtained after graph attention processing not only contains the intra-cycle degradation features of the corresponding charging cycle itself, but also incorporates the degradation information of the preceding charging cycles related to that charging cycle. For the charging cycle to be estimated, its corresponding inter-cycle node embedding vector can characterize the comprehensive degradation state of the charging cycle after combining the degradation evolution information of multiple historical charging cycles. This inter-cycle node embedding vector can serve as the feature basis for subsequent health status estimation and remaining lifetime prediction.

[0153] In this embodiment, the cycle embedding vectors corresponding to each charging cycle are organized into a directed inter-cycle graph. Directed edges are used to define the access relationships between subsequent charging cycles and the historical information of preceding charging cycles. Graph attention processing is then applied to the directed inter-cycle graph to dynamically represent the differences in contributions of different historical charging cycles to subsequent charging cycles. This allows for a more comprehensive modeling of the long-term degradation dependencies between multiple charging cycles, improves the expressive power of inter-cycle node embedding vectors for cross-cycle degradation trends, and provides a more reliable feature basis for subsequent health status estimation and remaining lifetime prediction.

[0154] In one possible embodiment, the graph attention processing includes multi-head graph attention processing.

[0155] Multi-head graph attention processing refers to calculating the degradation dependencies between graph nodes using multiple attention heads, and then fusing the results from the multiple attention heads to represent the degradation relationships between different charging cycles from multiple feature subspaces.

[0156] In one possible embodiment, the method steps shown in S1042 may include implementation in S10421 to S10423, which are described in detail below.

[0157] S10421. For any graph node in the intercyclic directed graph, under each attention head, determine the attention weight corresponding to the preceding graph node connected to the graph node.

[0158] In this model, any graph node can represent any charging cycle in the multi-cycle input sample, and a preceding graph node can represent a graph node that precedes the corresponding charging cycle in time and is connected to the current graph node. Since different historical charging cycles may have different degrees of influence on the degradation state of subsequent charging cycles, attention weights can be used to characterize the contribution of different preceding graph nodes to the current graph node when performing inter-cycle degradation dependency modeling.

[0159] In one implementation, for the first... The first graph node and the first node connected to it. The nth node of the preceding graph can be determined based on the nth node of the preceding graph. The node characteristics of the first graph node and the first The relevance score is calculated based on the node features of each node in the preceding graph. The relevance score can be expressed as: ;in, Indicates the first The node of the preceding graph is paired with the node of the first... The relevance score of each graph node. Indicates the first Node characteristics of a graph node Indicates the first Node characteristics of a preorder graph node. This indicates a feature concatenation operation. This represents the learnable feature transformation matrix. This represents a learnable attention vector. Let denote the transpose of the attention vector a, and LeakyReLU denotes the linear unit activation function with leakage correction.

[0160] Furthermore, the relevance scores can be normalized to obtain the attention weights. The attention weights can be expressed as: ;in, Indicates the first The node of the preceding graph is paired with the node of the first... Attention weights for each graph node. Indicates the relationship with the first Each graph node connects to the set of preceding graph nodes. This indicates that the set of nodes in the preceding graph and the first... Let k be the normalized set of nodes formed by the nodes themselves in the graph. Any node in the graph, This indicates that the k-th graph node is paired with the k-th graph node. The relevance score of each graph node.

[0161] By calculating the attention weights as described above, different preceding graph nodes can participate in subsequent feature aggregation according to their contribution to the current graph node, thereby avoiding the situation where simple average aggregation cannot distinguish the importance of different historical charging cycles.

[0162] S10422. Under each attention head, the node features corresponding to the preceding graph node are weighted and aggregated based on the attention weight to obtain the aggregated node features of the graph node under that attention head.

[0163] After determining the attention weights corresponding to the preceding graph nodes, the node features of the preceding graph nodes can be weighted and summed based on these attention weights to obtain the aggregated node features of the current graph node under this attention head. These aggregated node features can characterize the node representation of the current graph node after fusing historical degradation information within a feature subspace.

[0164] In one implementation, the first The graph node at the th ... The aggregated node features under each attention head can be represented as: ;in, Indicates the first The graph node at the th ... Features of aggregated nodes under attention heads Indicates the first The first thing to pay attention to. The node of the preceding graph is paired with the node of the first... Attention weights for each graph node. Indicates the first The learnable feature transformation matrix corresponding to each attention head. Indicates the first The node characteristics of each node in the preceding graph.

[0165] Through this weighted aggregation process, preceding graph nodes with higher attention weights contribute more to the aggregated node features, while preceding graph nodes with lower attention weights contribute less. Therefore, the aggregated node features can reflect the differences in the degree of influence of different historical charging cycles on the degradation state of the current charging cycle.

[0166] S10423. The aggregated node features of the graph node under multiple attention heads are fused to obtain the inter-cycle node embedding vector corresponding to the graph node.

[0167] After processing by S10422, the i-th graph node can obtain corresponding aggregated node features under multiple attention heads. Different attention heads can represent the degradation contribution relationship of the preceding graph node to the current graph node from different feature subspaces. To enhance the expressive power and stability of inter-cycle degradation dependency modeling, the aggregated node features obtained from multiple attention heads can be fused to obtain the inter-cycle node embedding vector corresponding to the i-th graph node.

[0168] In one implementation, the aggregated node features obtained from multiple attention heads can be concatenated to obtain the inter-cycle node embedding vector corresponding to the i-th graph node. That is, if the i-th graph node obtains an aggregated node feature under each attention head, the aggregated node features output by each attention head can be concatenated according to their feature dimensions to form the final node embedding representation. This final node embedding representation is the inter-cycle node embedding vector corresponding to the i-th graph node.

[0169] In this embodiment, through the multi-head fusion processing described above, the inter-cycle node embedding vector corresponding to the i-th graph node can integrate the modeling results of historical degradation information from multiple attention heads, representing the influence of previous charging cycles on the degradation state of the current charging cycle from multiple feature subspaces. Therefore, the ability of the inter-cycle node embedding vector to express long-term degradation dependencies across cycles can be improved, and the impact of the instability of single attention head modeling results on subsequent health status estimation and remaining lifetime prediction can be reduced.

[0170] In one implementation, a health state estimation head can be used to estimate the health state of the inter-cycle node embedding vectors corresponding to each charging cycle.

[0171] For any selected charging cycle within the window, its health status estimate can be expressed as: ;in, Indicates the first Health status estimate for each charging cycle This represents a health status estimation head composed of fully connected layers.

[0172] By arranging the health state estimates corresponding to each charging cycle in chronological order, a local health state sequence can be constructed. : ;in, This represents the sequence of local health states corresponding to the charging cycle to be estimated. This represents the estimated health state value corresponding to the charging cycle to be estimated. to This represents the estimated health status value corresponding to the historical charging cycles selected according to the preset sampling interval.

[0173] In one possible embodiment, the method steps shown in S107 can be implemented by S1071 to S1073, which are described in detail below.

[0174] S1071. Based on the inter-cycle node embedding vector corresponding to the charging cycle to be estimated, the remaining lifetime is predicted to obtain the feature-driven lifetime prediction result.

[0175] Among them, the inter-cycle node embedding vector corresponding to the charging cycle to be estimated can be the current cycle node embedding obtained after modeling the inter-cycle degradation dependency. The inter-cycle node embedding vector corresponding to the charging cycle to be estimated contains the intra-cycle degradation features of the charging cycle to be estimated itself and the inter-cycle degradation dependency information passed from the previous charging cycle.

[0176] Feature-driven lifetime prediction results are used to characterize the remaining lifetime results directly predicted based on the implicit degradation representation in the inter-cycle node embedding vector corresponding to the charging cycle to be estimated.

[0177] In one implementation, the inter-cycle nodes corresponding to the charging cycle to be estimated can be embedded into the vector input feature-driven branch, and the remaining lifetime can be directly predicted using a multilayer perceptron. The feature-driven lifetime prediction result can be expressed as: ;in, This represents the predicted feature-driven lifetime for the charging cycle to be estimated. Indicates feature-driven branches, represents the inter-cycle node embedding vector corresponding to the charging cycle to be estimated, and k represents the cycle number of the charging cycle to be estimated.

[0178] By using feature-driven branches, the remaining lifetime corresponding to the charging cycle to be estimated can be predicted by utilizing the degradation representation information implicit in the inter-cycle node embedding vector.

[0179] S1072. Based on the local health state sequence, predict the remaining lifespan to obtain the health state-guided lifespan prediction result.

[0180] The local health state sequence comprises health state estimates corresponding to each charging cycle in the multi-cycle input samples, arranged in chronological order. This local health state sequence reflects the health state change trend from multiple historical charging cycles prior to the charging cycle to be estimated. Since there is a strong correlation between the health state change trend and remaining lifetime, the local health state sequence can be used as explicit health state evolution information input into the health state guidance branch to assist in remaining lifetime prediction.

[0181] In one implementation, a sequence of local health states can be input into a health state-guided branch, and a multilayer perceptron can be used to predict remaining lifetime using the degradation trend of the health states. The health state-guided lifetime prediction result can be expressed as: ;in, This represents the health state-guided lifetime prediction result corresponding to the charging cycle to be estimated. The health status guides the branch. This represents a sequence of local health states.

[0182] By using a health status-guided branch, the remaining life prediction process can utilize the explicit degenerative trends reflected in the health status estimates, thereby enhancing the correlation between health status estimates and remaining life prediction.

[0183] S1073. The feature-driven lifetime prediction result and the health-state-guided lifetime prediction result are fused to obtain the remaining lifetime prediction value corresponding to the charging cycle to be estimated.

[0184] The feature-driven lifetime prediction result is obtained based on the inter-cycle node embedding vector corresponding to the charging cycle to be estimated, which can reflect the implicit degradation representation in the node embedding; the health state-guided lifetime prediction result is obtained based on the local health state sequence, which can reflect the explicit degradation evolution trend represented by the health state sequence. Since the two predict the remaining lifetime from the perspectives of implicit degradation features and explicit health state trends respectively, they can be fused to obtain the final remaining lifetime prediction value corresponding to the charging cycle to be estimated.

[0185] In one implementation, the prediction results of the two branches can be added together to obtain the predicted remaining lifetime value corresponding to the charging cycle to be estimated: ;in, This represents the predicted remaining lifetime value corresponding to the charging cycle to be estimated. By adding the prediction results of the two branches, the implicit degradation information in the inter-cycle node embedding vector and the explicit health state evolution information in the local health state sequence can jointly participate in the remaining lifetime prediction.

[0186] In this embodiment, a feature-driven branch is used to predict the remaining lifetime based on the inter-cycle node embedding vector corresponding to the charging cycle to be estimated, and a health state-guided branch is used to assist in predicting the remaining lifetime based on the local health state sequence. The prediction results of the two branches are then fused to obtain the final remaining lifetime prediction value. Therefore, within a unified framework, the degradation representation implicit in the cyclic embedding and the explicit degradation evolution trend reflected by the health state sequence can be utilized simultaneously, reducing the error propagation caused by the separate processing of health state estimation and remaining lifetime prediction, and improving the accuracy and robustness of remaining lifetime prediction.

[0187] In one possible embodiment, the method further includes training a collaborative estimation model. This training process includes steps S801 to S804, which are described in detail below.

[0188] S801. Obtain training samples, the training samples include sample charging operation data of sample batteries in multiple sample charging cycles, health status labels corresponding to each sample charging cycle, and remaining life labels corresponding to the sample charging cycle to be predicted. The multiple sample charging cycles include the sample charging cycle to be predicted and multiple historical sample charging cycles located before the sample charging cycle to be predicted. Here, the sample battery is a single lithium-ion battery cell used for model training. The sample charging cycle is the charging cycle formed during the charging process of the sample battery. The sample charging cycle to be predicted is the current sample cycle for which the remaining lifetime needs to be predicted during the training phase, and the historical sample charging cycles are the sample charging cycles that precede the sample charging cycle to be predicted.

[0189] The training samples are constructed in the same way as the multi-cycle input samples corresponding to the charging cycle to be estimated. That is, multi-cycle input samples corresponding to the charging cycle to be predicted can be constructed based on the sample charging operation data of the sample battery in multiple sample charging cycles. These multi-cycle input samples can include the cycle data corresponding to the charging cycle to be predicted, as well as the cycle data corresponding to multiple historical sample charging cycles, arranged in chronological order.

[0190] The health status label corresponding to each sample charging cycle is used to characterize the actual health status of the sample battery under the corresponding sample charging cycle. The remaining lifetime label corresponding to the sample charging cycle to be predicted is used to characterize the actual remaining lifetime of the sample battery from the start of the sample charging cycle to be predicted until the preset failure condition is reached.

[0191] By simultaneously setting the health status label for each sample's charging cycle and the remaining lifetime label for the sample to be predicted's charging cycle, the collaborative estimation model can learn both the health status estimation task and the remaining lifetime prediction task during training.

[0192] In one implementation, the health status label corresponding to the sample charging cycle can be determined based on the capacity calibration data of the sample lithium-ion battery in the corresponding cycle.

[0193] Specifically, the health status label corresponding to the kth cycle can be determined based on the ratio between the current discharge capacity and the rated capacity of the sample lithium-ion battery in the kth charge-discharge cycle.

[0194] The health state corresponding to the kth charge-discharge cycle can be represented as: .in, Indicates health status; This represents the current discharge capacity of the sample lithium-ion battery in the kth charge-discharge cycle; This indicates the rated capacity of the sample lithium-ion battery.

[0195] In one implementation, when the remaining capacity of a sample lithium-ion battery decays to a preset capacity threshold, this state can be defined as the End of Life (EOL). The preset capacity threshold is used to determine the battery's end-of-life cycle. For example, the preset capacity threshold can be 80% of the rated capacity.

[0196] In one implementation, the remaining lifetime tag corresponding to a sample charging cycle can be determined based on the difference between the cycle number when the sample lithium-ion battery reaches the end of its lifespan and the current cycle number. The remaining lifetime corresponding to the current cycle can be expressed as: Where β represents the cycle number corresponding to when the sample lithium-ion battery reaches the end of its lifespan; α represents the current cycle number. The unit of RUL can be expressed as the number of cycles.

[0197] It should be noted that the above methods for calculating health status and remaining lifetime are mainly used to determine health status labels and remaining lifetime labels during the training or evaluation phase. During the model inference phase, the multi-cycle input samples corresponding to the charging cycle to be estimated can be input into the trained co-estimation model, and the co-estimation model can output the estimated health status value and predicted remaining lifetime value corresponding to the charging cycle to be estimated.

[0198] In one implementation, sample batteries can be divided at the individual battery cell level to form training, validation, and test sets. This avoids data from the same battery cell appearing in both the training and test sets, thus more accurately reflecting the generalization ability of the co-estimation model for unknown battery cells. Before training, the health status label and remaining lifetime label can be standardized to reduce the impact of the difference in numerical dimensions between health status and remaining lifetime on joint training. During the evaluation phase, the model output can be transformed back to its original scale before calculating the error.

[0199] S802. Input the training samples into the collaborative estimation model to obtain the health status estimate corresponding to each sample charging cycle and the remaining lifetime prediction value corresponding to the charging cycle of the sample to be predicted.

[0200] The collaborative estimation model is used to perform the intra-cycle degradation feature encoding, the inter-cycle degradation dependency modeling, the local health state sequence determination, and the remaining lifetime prediction.

[0201] The collaborative estimation model includes a model part for performing intra-cycle degradation feature encoding, a model part for performing inter-cycle degradation dependency modeling, a model part for determining the local health state sequence, and a model part for determining the remaining lifetime prediction.

[0202] After inputting the training samples into the collaborative estimation model, the collaborative estimation model can, according to the processing flow in the aforementioned embodiments, encode the intra-cycle degradation features of each charging cycle in the multi-cycle input samples to obtain the cycle embedding vector corresponding to each charging cycle; then, based on the temporal relationship between the charging cycles of each sample, perform inter-cycle degradation dependency modeling to obtain the inter-cycle node embedding vector corresponding to each charging cycle of each sample; subsequently, output the health status estimate corresponding to each charging cycle of each sample based on the inter-cycle node embedding vector corresponding to each charging cycle of each sample, and form a sample local health status sequence; further, based on the inter-cycle node embedding vector corresponding to the charging cycle of the sample to be predicted and the sample local health status sequence, output the remaining lifetime prediction value corresponding to the charging cycle of the sample to be predicted.

[0203] Therefore, the collaborative estimation model can obtain the health status estimate and the remaining lifetime prediction of each sample charging cycle within the window in a single forward calculation, enabling the health status estimation task and the remaining lifetime prediction task to be executed together in a unified model framework.

[0204] S803. Determine the estimated health status loss based on the estimated health status value and health status label corresponding to each sample charging cycle, and determine the predicted remaining lifetime loss based on the predicted remaining lifetime value and remaining lifetime label corresponding to the sample charging cycle to be predicted.

[0205] The health status estimation loss measures the deviation between the estimated health status value and the health status label for each sample charging cycle. Since the collaborative estimation model can output the estimated health status value for each sample charging cycle in the multi-cycle input samples, health status label supervision can be applied to the estimated health status values ​​for all selected sample charging cycles within the window during the training phase, enabling the model to learn feature representations with degradation awareness.

[0206] The remaining lifetime prediction loss measures the deviation between the predicted remaining lifetime value and the remaining lifetime label for the charging cycle of the sample to be predicted. In other words, the health status estimation loss is oriented towards multiple charging cycles of the multi-cycle input sample, while the remaining lifetime prediction loss is oriented towards the charging cycle of the sample to be predicted. This allows the model to learn both local health status changes and the remaining lifetime prediction relationship corresponding to the current cycle.

[0207] S804. Construct a joint loss based on the health status estimation loss and the remaining lifespan prediction loss, and update the model parameters of the collaborative estimation model based on the joint loss to obtain the trained collaborative estimation model.

[0208] In one implementation, a joint loss function can be used to simultaneously optimize the health status estimation objective and the remaining lifespan prediction objective. The overall loss can be defined as the weighted sum of the health status estimation loss and the remaining lifespan prediction loss: Where L represents the joint loss, M is the number of training samples, and N is the number of charging cycles per training sample. Indicates the first In the training samples, the th The health status label corresponding to each sample charging cycle Indicates the first In the training samples, the th Health status estimate for each sample charging cycle Indicates the first The remaining lifetime label corresponding to the charging cycle of the sample to be predicted in each training sample. Indicates the first The remaining lifetime prediction value corresponding to the charging cycle of the sample to be predicted in each training sample, where λ is the loss tradeoff coefficient used to balance the loss of health status estimation and the loss of remaining lifetime prediction.

[0209] The first term in the joint loss above is used to supervise the health status estimate of each sample charging cycle within the window, and the second term is used to supervise the remaining lifetime prediction of the sample charging cycle to be predicted. By jointly optimizing these two objectives, the collaborative estimation model can learn both health status estimation and remaining lifetime prediction simultaneously based on shared degradation representations, avoiding the separation between health status estimation and remaining lifetime prediction.

[0210] During parameter updates, a gradient-based optimizer can be used to iteratively update the model parameters of the co-estimation model based on the joint loss. In one implementation, the AdamW optimizer can be used to update the model parameters, and an early stopping strategy can be implemented based on validation set performance to stop training early when validation performance no longer improves, reducing the risk of overfitting. After parameter updates, a trained co-estimation model is obtained. This trained co-estimation model is used in subsequent online estimation processes to output the estimated health state and remaining lifetime of the target lithium-ion battery under the estimated charging cycles.

[0211] In this embodiment, health status estimation and remaining lifetime prediction are incorporated into the same collaborative estimation model for end-to-end joint optimization during the training phase. A joint loss is used to ensure that the health status estimation results of each sample charging cycle within the window and the remaining lifetime prediction results of the sample to be predicted jointly constrain the model parameter updates. This allows the model to learn a shared degradation representation that simultaneously serves health status estimation and remaining lifetime prediction, reducing the accumulated error caused by estimating the health status degradation trend first and then inferring the remaining lifetime in traditional sequential estimation methods, thus improving the predictive consistency and generalization ability of the collaborative estimation model.

[0212] In one possible embodiment, the method further includes a transfer adaptation process for adapting the trained collaborative estimation model to the battery type to be adapted. This transfer adaptation process includes steps S901 to S904, which are described in detail below.

[0213] S901. When adapting the trained collaborative estimation model to the battery type to be adapted, obtain the target domain samples corresponding to the battery type to be adapted, wherein the battery type to be adapted is a lithium-ion battery type not covered by the training samples.

[0214] The battery type to be adapted can be a lithium-ion battery type different from the sample batteries in the training samples. For example, the training samples can come from one or more existing battery chemistry systems, while the battery type to be adapted can be another battery chemistry system not covered in the training phase. The target domain samples can include a small number of labeled samples corresponding to the battery type to be adapted. The labeled samples can include charging operation data, health status labels, and remaining life labels of the battery of the battery type to be adapted in multiple charging cycles.

[0215] It should be noted that the trained co-estimation model has already learned the ability to extract intra-cycle degradation features, model inter-cycle degradation dependencies, and co-estimate health status and remaining lifetime for the source domain battery type. However, different battery types may differ in capacity range, cycle lifetime, degradation rate, and degradation mode. If the trained co-estimation model is directly applied to the battery type to be adapted, the estimation accuracy may decrease due to the offset in degradation mode distribution between the source and target domains. Therefore, the parameters of the trained co-estimation model can be efficiently adapted based on samples from the target domain to better match the degradation characteristics of the battery type to be adapted.

[0216] S902. Freeze the model parameters used to perform cyclic degenerate feature encoding in the trained collaborative estimation model.

[0217] The model component used to perform intra-cycle degradation feature encoding may include the aforementioned intra-cycle encoder. The intra-cycle encoder is used to extract local degradation features within a single charging cycle. These features typically reflect local dynamic changes in data such as current, voltage, and charging capacity within the charging cycle. Since the local electrochemical dynamics within a single charging cycle share certain commonalities across different lithium-ion battery types, the local feature representations learned by the trained intra-cycle encoder have good cross-type transfer value.

[0218] During the transfer adaptation phase, the model parameters used for performing in-loop degenerate feature encoding can be frozen, preventing these parameters from participating in gradient updates on the target domain samples. For example, the model parameters of the in-loop encoder can be denoted as... And fix during the migration and adaptation phase By freezing these parameters, the general local degradation feature representations learned by the trained co-estimation model from the training samples can be preserved. This avoids destroying existing feature representations due to over-updating the in-loop encoder when the number of samples in the target domain is small, thereby reducing the risk of overfitting and catastrophic forgetting.

[0219] S903. In the trained collaborative estimation model, a low-rank adaptation branch is set in the model part used to perform inter-cycle degenerate dependency modeling. The low-rank adaptation branch is set in parallel with the main branch of the model part and is used to adapt and correct the output results of the inter-cycle degenerate dependency modeling.

[0220] The model portion used for performing inter-cycle degradation dependency modeling may include a Graph Attention Network version 2 (GATv2) structure, which is used to model long-term degradation dependencies between multiple charging cycles. Compared to local dynamic features within a cycle, the differences in long-term degradation patterns between different battery types are usually more pronounced. Therefore, during the transfer adaptation stage, a Low-Rank Adaptation (LoRA) branch can be set for the model portion that models inter-cycle degradation dependencies to adapt the inter-cycle degradation dependency modeling results to the target domain.

[0221] Specifically, a low-rank adaptation branch parallel to the main branch can be set up alongside the main branch for inter-loop degenerate dependency modeling. This low-rank adaptation branch can employ the same or corresponding multi-head aggregation method as the main branch, but its feature transformation matrix is ​​parameterized using low-rank matrix factorization. In one implementation, the adaptation output corresponding to the low-rank adaptation branch can be expressed as: ;in, Indicates the first The low-rank adaptation branch outputs corresponding to each graph node, where H represents the number of attention heads. This means concatenating the outputs of H attention heads. Indicates the first The first thing to pay attention to. The graph node pairs with the first Attention weights for each graph node. and This represents the low-rank matrix under the h-th attention head. This represents the node characteristics of the j-th graph node.

[0222] in, , , Represents a low-rank dimension. Indicates the input feature dimension. This indicates the output feature dimension, and By replacing the originally large eigenvalue transformation matrix with a low-rank matrix. and The product of these parameters can significantly reduce the number of parameters that need to be updated during the migration and adaptation phase.

[0223] In one implementation, the output of the low-rank adaptation branch can be added to the output of the main branch to obtain the adapted node representation: ;in, Indicates the first The adapted node representation corresponding to each graph node. This represents the inter-loop node embedding vector output by the main branch of the inter-loop degradation dependency modeling in the trained collaborative estimation model. This represents the adaptation correction amount of the low-rank adaptation branch output.

[0224] In this way, while retaining the source domain degradation knowledge already learned by the main branch, a low-rank adaptation branch can be introduced to correct the long-term degradation dependency of the target domain.

[0225] S904. Based on the target domain samples, update the parameters of the low-rank adaptation branch, and adjust the parameters of the output part of the trained collaborative estimation model used to determine the health status estimate and the remaining life prediction, to obtain a collaborative estimation model adapted to the battery type to be adapted.

[0226] In the transfer adaptation phase, the weights of the main branch for modeling inter-cycle degradation dependencies can be frozen, and only the parameters in the low-rank adaptation branch can be updated. This allows the model to efficiently adjust parameters for the inter-cycle degradation patterns of the battery type to be adapted without retraining the entire collaborative estimation model.

[0227] Furthermore, due to differences in capacity range, lifespan, and degradation characteristics among different battery types, the output portions of the health state estimate and remaining lifespan prediction also need to be recalibrated based on the target domain samples. Therefore, the parameters of the output portions used to determine the health state estimate and remaining lifespan prediction can be adjusted based on the target domain samples. The output portions may include a health state estimation header, a feature-driven lifespan prediction branch, and a health state-guided lifespan prediction branch.

[0228] In one implementation, the parameters of the output section can be updated jointly during the migration adaptation phase. ,in, Parameters representing the health status estimation head, The parameters representing the feature-driven lifetime prediction branch. The parameters that indicate the health status-guided lifespan prediction branch.

[0229] When updating parameters based on samples from the target domain, the same or similar health status estimation loss, remaining lifetime prediction loss, and joint loss as those used in the aforementioned training process can be employed to fine-tune the low-rank adaptation branch and output. After the update, a collaborative estimation model adapted to the battery type to be adapted can be obtained. This adapted collaborative estimation model can be used for health status estimation and remaining lifetime prediction of lithium-ion batteries of the battery type to be adapted.

[0230] In this embodiment, given a limited number of target domain samples for the battery types to be adapted, the model parameters used for intra-cycle degradation feature encoding are frozen to preserve common local degradation feature representations across battery types. A low-rank adaptation branch is set in the model portion used for inter-cycle degradation dependency modeling to adapt and correct long-term degradation dependencies in the target domain with a small number of trainable parameters. Furthermore, the outputs of health state estimation and remaining lifetime prediction are parameter-adjusted to match the capacity range and lifetime characteristics of the battery types to be adapted. This reduces the risks of overfitting and catastrophic forgetting associated with full parameter fine-tuning, achieving efficient parameter adaptation for lithium-ion battery types not covered by the training samples without retraining the entire co-estimation model.

[0231] In one specific embodiment, a publicly available lithium-ion battery dataset can be used to train, validate, and test the lithium-ion battery health status and remaining life co-estimation method provided in this application. This publicly available lithium-ion battery dataset may include three different types of lithium-ion batteries: NCA, NCM, and NCM+NCA, totaling 99 individual battery cells. Specifically, NCA represents lithium nickel cobalt aluminum oxide batteries, NCM represents lithium nickel cobalt manganese oxide batteries, and NCM+NCA represents hybrid batteries of lithium nickel cobalt manganese oxide and lithium nickel cobalt aluminum oxide.

[0232] During data partitioning, the three types of batteries can be merged at the individual battery level and then divided into training, validation, and test sets according to a preset ratio. For example, 99 individual battery cells can be divided into training, validation, and test sets in a 6:2:2 ratio. Since the data partitioning is performed at the individual battery level, the data of individual battery cells in the test set will not appear in the training set, thus enabling a more realistic evaluation of the co-estimation model's generalization ability to unseen battery cells.

[0233] During model training, the total number of cyclic data N in the multi-cycle input samples can be set to 10, the number of time steps T after fixed-length resampling to 128, and the preset sampling interval s to 3. During intra-cycle degradation feature encoding, the depthwise convolution kernel size can be set to 7; during inter-cycle degradation dependency modeling, the Graph Attention Network (GATv2) version 2 can be configured with 3 attention heads, i.e., H = 3; the latent dimension can be set to 16, i.e., D = 16. The loss tradeoff coefficient λ in the joint loss can be set to 0.6 to balance the impact of the health status estimation task and the remaining lifespan prediction task on model training.

[0234] During the optimization training process, the AdamW optimizer can be used to update the model parameters of the co-estimation model, with a learning rate set to 1×10^-3 and a total training epochs of 100. To reduce the risk of overfitting, an early stopping strategy can be adopted, terminating training prematurely when the validation set performance fails to improve for five consecutive epochs. In one implementation, the co-estimation model can be implemented based on the PyTorch framework and can be trained and evaluated on an NVIDIA RTX 4090 24GB GPU.

[0235] During the testing phase, mean absolute error, mean absolute percentage error, and root mean square error can be used to evaluate the health status estimation results and remaining life expectancy prediction results. The mean absolute error is denoted as MAE (Mean Absolute Error); the mean absolute percentage error is denoted as MAPE (Mean Absolute Percentage Error); and the root mean square error is denoted as RMSE (Root Mean Square Error).

[0236] The test results are shown in Table 1.

[0237] Table 1 shows the test set results of the proposed model. As shown in Table 1, on the test set, the proposed model has a mean absolute error of 0.31% and a mean absolute percentage error of 0.36% for State of Health (SOH), and a root mean square error of 0.42% for Remaining Life (RUL); the mean absolute error is 19.94 cycles, the mean absolute percentage error is 5.97%, and the root mean square error is 28.46 cycles. These results demonstrate that the method provided in this application can achieve high-precision state of health estimation and remaining life prediction on different types of lithium-ion battery data, and has good generalization ability.

[0238] Figure 4 This diagram illustrates the relationship between the predicted and actual values ​​of each battery in the test set provided in this application embodiment. Figure 4 As shown, this figure includes several subplots to illustrate the correspondence between the estimated health status and predicted remaining life for different battery types and the actual values. The horizontal axis represents the actual value, and the vertical axis represents the model's predicted value. The red diagonal line represents the ideal reference line when the predicted value perfectly matches the actual value. The closer the scatter plot or curve is to this reference line, the closer the model's prediction is to the actual value, and the smaller the prediction error.

[0239] Figure 4 The subplot showing the health status estimation results illustrates the relationship between the predicted and actual health status of different test batteries. The plot shows that the overall health status prediction results are close to the ideal reference line, indicating that the model can accurately estimate the health status of different batteries in the test set. Figure 4 The remaining lifetime prediction subplot in the figure is used to illustrate the relationship between the predicted remaining lifetime and the actual remaining lifetime for different test batteries. Although remaining lifetime prediction is usually more affected by long-term degradation trends and individual battery differences than health state estimation, the prediction results in the figure still generally follow the trend of actual remaining lifetime changes, indicating that the model can effectively utilize inter-cycle node embedding vectors and local health state sequences for remaining lifetime prediction.

[0240] also, Figure 4 The small error distribution plot in the table is used to show the concentration of prediction errors. The more concentrated the error distribution is near zero, the smaller the overall deviation of the model prediction results. Combined with the mean absolute error, mean absolute percentage error, and root mean square error results in Table 1, it can be further demonstrated that the lithium-ion battery health status and remaining life co-estimation method provided in this application embodiment has good estimation accuracy and generalization ability on the test set, and can simultaneously achieve high-precision health status estimation and remaining life prediction.

[0241] In one possible embodiment, the trained collaborative estimation model can be embedded as a software module in the Battery Management System (BMS) for online health status estimation and remaining life prediction of in-service lithium-ion batteries.

[0242] During online operation, the battery management system can collect current, voltage, and charging capacity data in real time during the charging process of the target lithium-ion battery, and preprocess each charging cycle according to the same processing method as in the model training phase. Specifically, charging segments within a set voltage range can be extracted, and these segments can be resampled and normalized to a fixed length. Then, a sliding window approach is used to construct multi-cycle input samples based on the charging cycle to be estimated. Since this online operation process only needs to utilize charging segment data from the charging cycle to be estimated and several previous historical charging cycles, without relying on complete charge and discharge data, it is more suitable for "charge-as-you-go" application scenarios in actual vehicles or energy storage systems.

[0243] Furthermore, after inputting the constructed multi-cycle input samples into the trained collaborative estimation model, it can sequentially undergo intra-cycle degradation feature encoding, inter-cycle degradation dependency modeling, health state estimation, and remaining lifetime prediction. A single inference can simultaneously output the estimated health state value and the predicted remaining lifetime value corresponding to the charging cycle to be estimated. Thus, the battery management system can perform battery health assessment, maintenance warning, charge / discharge control, or operation scheduling based on the output results.

[0244] In one specific embodiment, the trained collaborative estimation model can be deployed on an NVIDIA Jetson Orin Nano Super 8GB platform to simulate a real automotive or embedded operating environment. Testing showed that the model's floating-point computation is approximately 1.443 MFLOPs, the number of parameters is approximately 0.030 M, and the average inference time per sample is approximately 9.985 ms. These test results demonstrate that the collaborative estimation model has low computational complexity and parameter size, and its inference latency is significantly shorter than the duration of a single charging cycle, thus meeting the requirements of battery management systems for real-time online estimation and embedded deployment.

[0245] In another possible embodiment, when it is necessary to apply the trained collaborative estimation model to battery types not covered during the training phase, a transfer adaptation approach can be used to improve the model's applicability to the new battery types. Specifically, the model parameters used to perform intra-cycle degradation feature encoding in the trained collaborative estimation model can be frozen, and a low-rank adaptation branch can be set in the model part used to perform inter-cycle degradation dependency modeling. Simultaneously, the parameters of the output part used to determine the health state estimate and remaining lifetime prediction can be adjusted. Through this process, while retaining the local degradation feature representations already learned in the source domain battery types, the cross-cycle long-term degradation dependencies and output mapping relationships can be adapted and corrected using a small number of target domain samples corresponding to the battery types to be adapted.

[0246] In one specific embodiment, leave-one-out cross-domain migration experiments can be performed between three battery chemistry systems: NCA, NCM, and NCM+NCA. Specifically, a co-estimation model can be trained on data from two of the battery chemistry systems and evaluated on data from the remaining battery chemistry system. Migration settings can include: NCA&NCM→NCM+NCA, NCA&NCM+NCA→NCM, and NCM&NCM+NCA→NCA.

[0247] In the comparative experiments, two methods can be set up: direct transfer and low-rank adaptation-based transfer. Direct transfer involves directly applying the pre-trained model to the battery type to be adapted, without target domain adaptation. Low-rank adaptation-based transfer employs the aforementioned transfer adaptation strategy, selecting 10% of the total number of batteries from each operating condition in the target domain as fine-tuning data, and adjusting the parameters of the low-rank adaptation branch and the output part. Specifically, the low-rank decomposition parameter in the low-rank adaptation branch is set to 8, i.e., r = 8. The results of direct transfer are shown in Table 2, and the results of low-rank adaptation-based transfer are shown in Table 3.

[0248] Table 2 Direct migration results Table 3. LoRA-based migration results Where MAE stands for Mean Absolute Error; MAPE stands for Mean Absolute Percentage Error; and RMSE stands for Root Mean Square Error.

[0249] As shown in Tables 2 and 3, the direct migration method achieves acceptable state of health (SOH) estimation accuracy for some battery types, such as a RMSE of 0.66% under the NCA&NCM+NCA→NCM setting. However, the remaining lifetime (RUL) prediction error of the direct migration method is relatively high, indicating that long-term degradation modes are more sensitive to differences in battery chemistry. By adopting a migration method based on low-rank adaptation, both the SOH estimation performance and RUL prediction performance under the three migration settings are improved. For example, under the NCA&NCM→NCM+NCA setting, the RMSE of SOH decreased from 3.03% to 0.72%, and the RMSE of RUL decreased from 58.90 cycles to 15.86 cycles; under the NCA&NCM+NCA→NCM setting, the RMSE of SOH decreased from 0.66% to 0.39%, and the RMSE of RUL decreased from 62.73 cycles to 38.24 cycles; under the NCM&NCM+NCA→NCA setting, the RMSE of SOH decreased from 2.08% to 0.82%, and the RMSE of RUL decreased from 102.71 cycles to 60.99 cycles.

[0250] The above results show that the method provided in this application can not only be deployed in resource-constrained vehicle or embedded battery management systems for online health status estimation and remaining life prediction, but also achieve efficient parameter transfer across battery types through low-rank adaptation when there are few target domain samples of the battery types to be adapted, thereby improving the adaptability of the collaborative estimation model to different lithium-ion battery types and its engineering application value.

[0251] This application, by encoding intra-cycle degradation features in the cyclic data corresponding to each charging cycle, can more fully extract degradation-related features within the charging cycle. Specifically, the intra-cycle degradation feature encoding process can not only extract the time-series variation features of current data, voltage data, and charging capacity data, but also further integrate the correlation between different data channels. This overcomes the problem of focusing only on the time-series variation of a single variable while ignoring the correlation between variables, enabling the obtained cyclic embedding vector to more completely represent the short-term degradation behavior within a single charging cycle.

[0252] This application uses the cycle embedding vectors corresponding to each charging cycle as graph nodes and performs inter-cycle degradation dependency modeling based on the temporal relationship between charging cycles, enabling a unified representation of intra-cycle degradation features and inter-cycle degradation trends. Specifically, intra-cycle degradation feature encoding can characterize the local dynamic features of a single charging cycle, while inter-cycle degradation dependency modeling can characterize the long-term cumulative degradation relationship between multiple charging cycles. Combining the two can simultaneously capture short-term local degradation information and long-term aging evolution patterns, thereby improving the feature base quality of health status estimation and remaining lifespan prediction.

[0253] This application generates a local health state sequence based on the inter-cycle node embedding vectors corresponding to each charging cycle, and determines the remaining lifetime prediction value by combining the inter-cycle node embedding vectors corresponding to the charging cycle to be estimated and the local health state sequence. This enables collaborative processing of health state estimation and remaining lifetime prediction. Specifically, the local health state sequence can explicitly reflect the trend of health state changes within a certain historical range before the charging cycle to be estimated. The remaining lifetime prediction process simultaneously utilizes the implicit degradation representation in the inter-cycle node embedding vectors and the explicit degradation trend in the local health state sequence, thereby avoiding the separation between health state estimation and remaining lifetime prediction, reducing the accumulated errors easily introduced by serial inference methods, and improving the accuracy and reliability of long-term prediction.

[0254] This application reduces reliance on complete charge / discharge data and artificial health feature extraction by extracting charging segments from a set voltage range and performing fixed-length resampling and normalization on these segments. Since this method can directly construct multi-cycle input samples based on partial charging segments, it better suits application scenarios in real-world vehicles or energy storage systems where the charging process is incomplete, operating conditions are complex, and data acquisition is limited, thus facilitating online estimation for on-demand charging and testing.

[0255] The collaborative estimation model adopted in this application has low computational complexity and parameter size, and can be deployed as a software module in battery management systems or resource-constrained automotive and embedded hardware platforms. Since the model can simultaneously output the health state estimate and remaining lifetime prediction value corresponding to the charging cycle to be estimated in a single inference, and the inference latency is less than the duration of a single charging cycle, it can meet the requirements of battery management systems for real-time online estimation and engineering deployment feasibility.

[0256] This application can also improve the model's applicability to different lithium-ion battery types through transfer adaptation. When the trained co-estimation model needs to be adapted to a battery type not covered by the training samples, the intra-cycle degradation feature encoding part can be frozen, a low-rank adaptation branch can be set in the inter-cycle degradation dependency modeling part, and the parameters of the output part can be adjusted. Thus, efficient parameter adaptation across battery types can be achieved using only a small number of target domain samples, without retraining the entire model, thereby reducing the difficulty of model deployment when data acquisition costs for new battery types are high.

[0257] Figure 5 This is a schematic diagram of a lithium-ion battery health status and remaining life co-estimation system provided in an embodiment of this application. Figure 5As shown, the collaborative estimation system 500 includes a data acquisition module 501, a sample construction module 502, an intra-loop encoding module 503, an inter-loop modeling module 504, a health state sequence determination module 505, a health state determination module 506, a remaining lifespan prediction module 507, and an output module 508.

[0258] The data acquisition module 501 is used to acquire charging operation data of the target lithium-ion battery in multiple charging cycles, wherein the multiple charging cycles include the charging cycle to be estimated and multiple historical charging cycles located before the charging cycle to be estimated.

[0259] The sample construction module 502 is used to construct a multi-cycle input sample corresponding to the charging cycle to be estimated based on the charging operation data. The multi-cycle input sample includes multiple cycle data arranged in chronological order. The multiple cycle data corresponds one-to-one with the multiple charging cycles and is obtained from the charging operation data of the corresponding charging cycle.

[0260] The intra-cycle encoding module 503 is used to perform intra-cycle degradation feature encoding on the cyclic data corresponding to each charging cycle in the multi-cycle input samples to obtain the cyclic embedding vector corresponding to each charging cycle.

[0261] The inter-cycle modeling module 504 is used to perform inter-cycle degradation dependency modeling based on the temporal sequence relationship between each charging cycle, using the cycle embedding vectors corresponding to each charging cycle as graph nodes, to obtain the inter-cycle node embedding vectors corresponding to each charging cycle.

[0262] The health status sequence determination module 505 is used to determine the health status estimate corresponding to each charging cycle based on the inter-cycle node embedding vector corresponding to each charging cycle, and generate the local health status sequence corresponding to the multi-cycle input sample in chronological order.

[0263] The health status determination module 506 is used to determine the health status estimate value corresponding to the charging cycle to be estimated from the local health status sequence.

[0264] The remaining lifetime prediction module 507 is used to determine the remaining lifetime prediction value corresponding to the charging cycle to be estimated based on the inter-cycle node embedding vector corresponding to the charging cycle to be estimated and the local health state sequence.

[0265] The output module 508 is used to output the estimated health status and predicted remaining life corresponding to the charging cycle to be estimated.

[0266] It should be noted that the specific process of each module in the collaborative estimation system executing the above method has been described in detail in the above embodiments, and this embodiment does not make specific limitations on it.

[0267] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 6 As shown, the electronic device 600 provided in this embodiment includes a memory 601 and a processor 602.

[0268] The memory 601 can be a separate physical unit, connected to the processor 602 via a bus 603. Alternatively, the memory 601 and processor 602 can be integrated and implemented in hardware. The memory 601 stores program instructions, which the processor 602 calls to execute the operations performed by the cooperative estimation system in any of the above method embodiments.

[0269] Optionally, when some or all of the methods in the above embodiments are implemented by software, the electronic device 600 may also include only the processor 602. A memory 601 for storing programs is located outside the electronic device 600, and the processor 602 is connected to the memory via circuits / wires to read and execute the programs stored in the memory. The processor 602 may be a central processing unit (CPU), a network processor (NP), or a combination of a CPU and an NP. The processor 602 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The PLD may be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), or any combination thereof.

[0270] The memory 601 may include volatile memory, such as random-access memory (RAM); the memory may also include non-volatile memory, such as flash memory, hard disk drive (HDD) or solid-state drive (SSD); the memory may also include a combination of the above types of memory.

[0271] For example, this application provides a chip including: an interface circuit and a logic circuit. The interface circuit is used to receive signals from other chips outside the chip and transmit them to the logic circuit, or to send signals from the logic circuit to other chips outside the chip. The logic circuit is used to perform the operations performed by the cooperative estimation system in the above method embodiments.

[0272] For example, this application provides a computer-readable storage medium having computer program instructions stored thereon, which are executed by a processor of an electronic device to cause the electronic device to perform the operations performed by the cooperative estimation system in the above method embodiments.

[0273] For example, this application provides a computer program product that, when run on an electronic device, causes the electronic device to perform the operations performed by the cooperative estimation system in the above method embodiments.

[0274] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments described herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for co-estimating the health status and remaining life of a lithium-ion battery, characterized in that, The method includes: Acquire charging operation data of the target lithium-ion battery in multiple charging cycles, wherein the multiple charging cycles include the charging cycle to be estimated and multiple historical charging cycles preceding the charging cycle to be estimated. Based on the charging operation data, a multi-cycle input sample corresponding to the charging cycle to be estimated is constructed. The multi-cycle input sample includes multiple cycle data arranged in chronological order. The multiple cycle data correspond one-to-one with the multiple charging cycles and are obtained from the charging operation data of the corresponding charging cycle. The cyclic data corresponding to each charging cycle in the multi-cycle input samples are encoded with intra-cycle degradation features to obtain the cyclic embedding vector corresponding to each charging cycle. Using the cycle embedding vectors corresponding to each charging cycle as graph nodes, inter-cycle degradation dependency modeling is performed based on the temporal relationship between each charging cycle to obtain the inter-cycle node embedding vectors corresponding to each charging cycle. Based on the inter-cycle node embedding vectors corresponding to each charging cycle, the estimated health status value corresponding to each charging cycle is determined, and the local health status sequence corresponding to the multi-cycle input samples is generated in chronological order. Determine the health state estimate corresponding to the charging cycle to be estimated from the local health state sequence; Based on the inter-cycle node embedding vector corresponding to the charging cycle to be estimated and the local health state sequence, the remaining lifetime prediction value corresponding to the charging cycle to be estimated is determined; Output the estimated health status and predicted remaining lifetime corresponding to the charging cycle to be estimated.

2. The method according to claim 1, characterized in that, The charging operation data includes current data, voltage data, and charging capacity data; The construction of the multi-cycle input sample corresponding to the charging cycle to be estimated based on the charging operation data includes: The charging segments within the set voltage range are extracted from the charging operation data of each charging cycle to obtain the charging segment data corresponding to each charging cycle. The charging segment data corresponding to each charging cycle is subjected to fixed-length resampling and normalization to obtain the cycle data corresponding to each charging cycle. Based on the charging cycle to be estimated, a preset number of historical cycle data are selected from the cycle data corresponding to each historical charging cycle before the charging cycle to be estimated, according to a preset sampling interval. Arrange the preset number of historical cycle data and the cycle data corresponding to the charging cycle to be estimated in chronological order to construct a multi-cycle input sample corresponding to the charging cycle to be estimated.

3. The method according to claim 1, characterized in that, The step of encoding intra-cycle degradation features for the cyclic data corresponding to each charging cycle in the multi-cycle input samples to obtain the cyclic embedding vector corresponding to each charging cycle includes: The cyclic data corresponding to each charging cycle are processed by feature mapping to obtain the initial latent space features corresponding to each charging cycle. The cyclic position codes corresponding to each charging cycle are fused with the initial latent space features to obtain the position enhancement features corresponding to each charging cycle. The position enhancement features corresponding to each charging cycle are encoded by intracyclic degradation features to obtain the cyclic embedding vector corresponding to each charging cycle.

4. The method according to claim 3, characterized in that, The step of encoding intra-cycle degradation features for the position enhancement features corresponding to each charging cycle to obtain the cyclic embedding vector corresponding to each charging cycle includes: Intra-channel temporal encoding is performed on the temporal features of different data channels in the position enhancement features corresponding to each charging cycle to obtain the intra-channel temporal features corresponding to each charging cycle. Cross-channel feature fusion is performed on the time-series features within each channel corresponding to each charging cycle to obtain the cross-channel fused features corresponding to each charging cycle. Channel aggregation processing is performed on the cross-channel fusion features corresponding to each charging cycle to obtain the cyclic embedding vector corresponding to each charging cycle.

5. The method according to claim 1, characterized in that, The method involves using the cycle embedding vectors corresponding to each charging cycle as graph nodes, and performing inter-cycle degradation dependency modeling based on the temporal relationship between each charging cycle to obtain the inter-cycle node embedding vectors corresponding to each charging cycle, including: Using the cycle embedding vectors corresponding to each charging cycle as graph nodes, and establishing directed edges based on the temporal sequence between each charging cycle, a directed graph between cycles is obtained; wherein, in the directed graph between cycles, the graph node corresponding to the subsequent charging cycle is connected to the graph node corresponding to the preceding charging cycle located before the subsequent charging cycle, so that the graph node corresponding to the subsequent charging cycle can aggregate the node features of the graph node corresponding to the preceding charging cycle. Graph attention processing is performed on the inter-cycle directed graph to obtain the inter-cycle node embedding vectors corresponding to each charging cycle.

6. The method according to claim 5, characterized in that, The graph attention processing includes multi-head graph attention processing; The step of performing graph attention processing on the inter-cycle directed graph to obtain the inter-cycle node embedding vectors corresponding to each charging cycle includes: For any graph node in the intercyclic directed graph, under each attention head, determine the attention weight corresponding to the preceding graph node connected to the graph node; Under each attention head, the node features corresponding to the preceding graph node are weighted and aggregated based on the attention weight to obtain the aggregated node features of the graph node under that attention head; The aggregated node features of the graph node under multiple attention heads are fused to obtain the inter-cycle node embedding vector corresponding to the graph node.

7. The method according to claim 1, characterized in that, The step of determining the predicted remaining lifetime value corresponding to the charging cycle to be estimated based on the inter-cycle node embedding vector corresponding to the charging cycle to be estimated and the local health state sequence includes: Based on the inter-cycle node embedding vector corresponding to the charging cycle to be estimated, the remaining lifetime is predicted to obtain the feature-driven lifetime prediction result. Based on the local health state sequence, the remaining lifespan is predicted to obtain the health state-guided lifespan prediction result. The feature-driven lifetime prediction result and the health-state-guided lifetime prediction result are fused to obtain the remaining lifetime prediction value corresponding to the charging cycle to be estimated.

8. The method according to claim 1, characterized in that, The method further includes: Acquire training samples, which include sample charging operation data of sample batteries in multiple sample charging cycles, health status labels corresponding to each sample charging cycle, and remaining life labels corresponding to the sample charging cycle to be predicted. The multiple sample charging cycles include the sample charging cycle to be predicted and multiple historical sample charging cycles located before the sample charging cycle to be predicted. The training samples are input into the collaborative estimation model to obtain the health status estimate corresponding to each sample charging cycle and the remaining lifetime prediction value corresponding to the charging cycle of the sample to be predicted; wherein, the collaborative estimation model is used to perform the intra-cycle degradation feature encoding, the inter-cycle degradation dependency modeling, the local health status sequence determination, and the remaining lifetime prediction value determination. The estimated health status loss is determined based on the estimated health status value and health status label corresponding to each sample charging cycle, and the predicted remaining lifetime loss is determined based on the predicted remaining lifetime value and remaining lifetime label corresponding to the sample charging cycle to be predicted. A joint loss is constructed based on the health status estimation loss and the remaining lifespan prediction loss, and the model parameters of the collaborative estimation model are updated based on the joint loss to obtain the trained collaborative estimation model.

9. The method according to claim 8, characterized in that, The method further includes: When adapting the trained collaborative estimation model to the battery type to be adapted, the target domain samples corresponding to the battery type to be adapted are obtained, wherein the battery type to be adapted is a lithium-ion battery type not covered by the training samples. Freeze the model parameters used to perform in-loop degenerate feature encoding in the trained collaborative estimation model; In the trained collaborative estimation model, a low-rank adaptation branch is set in the model part used to perform inter-cycle degenerate dependency modeling. The low-rank adaptation branch is set in parallel with the main branch of the model part and is used to adapt and correct the output results of the inter-cycle degenerate dependency modeling. Based on the target domain samples, the parameters of the low-rank adaptation branch are updated, and the parameters of the output part of the trained collaborative estimation model used to determine the health status estimate and the remaining life prediction are adjusted to obtain a collaborative estimation model adapted to the battery type to be adapted.

10. A system for co-estimating the health status and remaining life of a lithium-ion battery, characterized in that, The system includes: The data acquisition module is used to acquire charging operation data of the target lithium-ion battery in multiple charging cycles, wherein the multiple charging cycles include the charging cycle to be estimated and multiple historical charging cycles located before the charging cycle to be estimated. The sample construction module is used to construct a multi-cycle input sample corresponding to the charging cycle to be estimated based on the charging operation data. The multi-cycle input sample includes multiple cycle data arranged in chronological order. The multiple cycle data correspond one-to-one with the multiple charging cycles and are obtained from the charging operation data of the corresponding charging cycle. The intra-cycle encoding module is used to encode the intra-cycle degradation features of the cyclic data corresponding to each charging cycle in the multi-cycle input samples to obtain the cyclic embedding vector corresponding to each charging cycle. The inter-cycle modeling module is used to model inter-cycle degradation dependencies based on the temporal relationship between charging cycles, using the cycle embedding vectors corresponding to each charging cycle as graph nodes, and to obtain the inter-cycle node embedding vectors corresponding to each charging cycle. The health state sequence determination module is used to determine the health state estimate corresponding to each charging cycle based on the inter-cycle node embedding vector corresponding to each charging cycle, and generate the local health state sequence corresponding to the multi-cycle input sample in chronological order. A health status determination module is used to determine the estimated health status value corresponding to the charging cycle to be estimated from the local health status sequence; The remaining lifetime prediction module is used to determine the remaining lifetime prediction value corresponding to the charging cycle to be estimated based on the inter-cycle node embedding vector corresponding to the charging cycle to be estimated and the local health state sequence. The output module is used to output the estimated health status and predicted remaining life corresponding to the charging cycle to be estimated.

Images