A method, device, and medium for detecting a state of health of a battery

Abstract

The application discloses a kind of battery health state detection method, equipment and medium, it is related to battery health management technical field, including, the historical operation data of battery and full-band EIS scanning data are collected, battery dataset is constructed, and state parameter is separately collected;Deep reinforcement learning DRL agent is trained by battery dataset, the state parameter of battery is scanned decision using the DRL agent that training is completed, and sparse frequency band decision vector is output;Sparse frequency band decision vector is used to control EIS equipment to carry out sparse scanning to target battery, and sparse frequency band EIS data is obtained;A variety of model batteries are used to construct meta-knowledge base, and the meta-knowledge base is trained by meta-learning algorithm, and reference transfer model is obtained;Sparse frequency band EIS data is handled using reference transfer model with few samples adaptation.The application carries out sparse EIS scanning by dynamically selecting frequency point, and realizes efficient real-time diagnosis.

Description

Technical Field

[0001] This invention relates to the field of battery health management technology, and in particular to a method, device and medium for detecting the health status of a battery. Background Technology

[0002] Battery health status monitoring is an essential technology in battery management, widely used in electric vehicles and energy storage systems. Currently, the industry primarily uses monitoring of operating data such as voltage, current, and temperature, combined with methods like capacity testing, internal resistance measurement, and electrochemical impedance spectroscopy to assess battery aging. These methods, by acquiring and analyzing battery operating parameters in real time, can monitor battery performance degradation and provide predictions of aging trends.

[0003] Currently, the industry generally relies on full-band electrochemical impedance spectroscopy scanning, which results in long scanning time and high computational resource consumption. This makes it difficult to efficiently meet the needs of real-time detection, especially in complex battery systems where it is difficult to achieve rapid diagnosis, and it is impossible to fully trace the aging propagation path and predict the system-level health status. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method for detecting battery health status to solve the problem of difficulty in achieving rapid diagnosis in complex battery systems.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for detecting the health status of a battery, which includes collecting historical operating data and full-band EIS scan data of the battery, constructing a battery dataset, and collecting status parameters separately.

[0008] A deep reinforcement learning (DRL) agent is trained using a battery dataset. The trained DRL agent is then used to scan and make decisions on the battery's state parameters, outputting a sparse frequency band decision vector.

[0009] The sparse frequency band decision vector control EIS device is used to perform sparse scanning of the target battery to obtain sparse frequency band EIS data.

[0010] A meta-knowledge base is constructed by using various battery models, and a benchmark transfer model is obtained by training the meta-knowledge base using a meta-learning algorithm.

[0011] The baseline migration model is used to perform few-sample adaptation processing on sparse frequency band EIS data to obtain the virtual reference DRT baseline and SOH estimate of the target battery;

[0012] The electrical connection topology of the battery is obtained by a multi-channel scanner, and the SOH estimate and state parameters of the target battery are mapped to node and edge features to construct a heterogeneous battery connection graph.

[0013] Based on the virtual reference DRT benchmark, message passing and aggregation processing are performed on the battery connection heterogeneous graph to obtain the aging coupling relationship representation, and the aging propagation path and health status diagnosis results are calculated.

[0014] As a preferred embodiment of the battery health status detection method of the present invention, the method involves: collecting historical operating data and full-band EIS scan data of the battery to construct a battery dataset, and separately collecting status parameters. The specific steps are as follows: collecting historical electrical data and historical thermal data during battery operation.

[0015] Perform a full-band EIS scan to acquire the impedance spectrum data of the battery;

[0016] A battery dataset was constructed by integrating historical electrical data, historical thermal data, and impedance spectral data.

[0017] Status parameters are collected and stored individually using sensors.

[0018] As a preferred embodiment of the battery health status detection method of the present invention, the following steps are taken: a DRL agent is trained by a battery dataset, the trained DRL agent is used to scan and make decisions on the battery's state parameters, and a sparse frequency band decision vector is output.

[0019] The DRL agent is trained using the battery dataset, and the policy network of the DRL agent is optimized to obtain the trained DRL agent.

[0020] The real-time state parameters of the target battery are input into the trained DRL agent for scanning decision-making, generating sparse frequency band decision vectors.

[0021] As a preferred embodiment of the battery health status detection method of the present invention, the method involves: using sparse frequency band decision vectors to control an EIS device to perform sparse scanning on the target battery and obtain sparse frequency band EIS data. The specific steps are as follows: parsing the sparse frequency band decision vectors to obtain a sequence of specific frequency points to be scanned.

[0022] Configure the scanning parameters of the EIS device according to the specific frequency point sequence to be scanned, control the EIS device to perform a rapid EIS scan on the target battery according to the scanning parameters, collect and record the impedance data of the target battery, and form sparse frequency band EIS data.

[0023] As a preferred embodiment of the battery health status detection method of the present invention, the method involves: constructing a meta-knowledge base using multiple battery models, and training the meta-knowledge base using a meta-learning algorithm to obtain a baseline transfer model. The specific steps are as follows: collecting full-band EIS data and corresponding DRT spectra of multiple different battery models.

[0024] Aging characteristic peak parameters are extracted from full-band EIS data and corresponding DRT spectra, and a meta-knowledge base is constructed using aging characteristic peak parameters.

[0025] A meta-learning algorithm is used to train the battery aging state mapping relationship on the meta-knowledge base to obtain the baseline transfer model.

[0026] As a preferred embodiment of the battery health status detection method of the present invention, the following steps are taken: using a benchmark transfer model to perform few-sample adaptation processing on sparse frequency band EIS data to obtain the virtual reference DRT benchmark and SOH estimate of the target battery. The specific steps are as follows: using prior knowledge in the meta-knowledge base to perform few-sample fast adaptation on the benchmark transfer model to obtain the adapted benchmark transfer model.

[0027] Based on the adapted benchmark migration model, the sparse frequency band EIS data of the target battery is fitted and reconstructed to generate a virtual reference DRT benchmark for the target battery.

[0028] By comparing the aging characteristic peak parameters in the virtual reference DRT benchmark of the target battery with the mapping relationship of battery aging state, the SOH estimate of the target battery is obtained.

[0029] As a preferred embodiment of the battery health status detection method of the present invention, the method involves: acquiring the electrical connection topology of the battery through a multi-channel scanner, mapping the SOH estimate and state parameters of the target battery to node and edge features, and constructing a battery connection heterogeneous graph. The specific steps are as follows: acquiring the electrical connection types of all batteries through a multi-channel scanner and generating the electrical connection topology.

[0030] The SOH estimate and state parameters of each battery are mapped to node features in a graph neural network;

[0031] The edges in the graph neural network are defined based on the electrical connection topology, and the electrical connection type and state parameters are mapped to edge features.

[0032] Based on node and edge features, a heterogeneous graph of battery connections is constructed.

[0033] As a preferred embodiment of the battery health status detection method of the present invention, the method is as follows: based on the virtual reference DRT benchmark, message passing and aggregation processing are performed on the battery connection heterogeneous graph to obtain the aging coupling relationship representation, and the aging propagation path and health status diagnosis result are calculated. The specific steps are as follows: the aging characteristic peak parameters in the virtual reference DRT benchmark are injected into the battery connection heterogeneous graph as prior knowledge.

[0034] Perform multi-round message passing on the heterogeneous graph of battery connections, aggregate the aging state information of adjacent battery nodes along the topological edges, and generate an overall aging coupling relationship representation of the battery through the readout function;

[0035] Based on the aging coupling relationship representation, the aging propagation path is traced and the health status diagnosis results of each battery node are calculated.

[0036] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein the computer program, when executed by the processor, implements any step of the battery health status detection method as described in the first aspect of the present invention.

[0037] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the battery health status detection method as described in the first aspect of the present invention.

[0038] The beneficial effects of this invention are as follows: by dynamically selecting frequency points for sparse EIS scanning, the scanning time and computational resource consumption are significantly reduced, achieving efficient and real-time diagnosis, and solving the problem of low efficiency in traditional full-band EIS scanning; in addition, by constructing a battery connection heterogeneous graph and combining it with graph neural networks to analyze the aging coupling relationship between batteries, the aging propagation path can be accurately traced, and the battery pack-level health status can be accurately predicted, overcoming the limitation of not being able to fully analyze the mutual influence of aging. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 A flowchart for a method to detect battery health status.

[0041] Figure 2 This is a schematic diagram of data acquisition and processing.

[0042] Figure 3This is a flowchart of DRL decision-making and sparse scanning.

[0043] Figure 4 This is a flowchart for meta-learning and health diagnosis. Detailed Implementation

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0045] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0046] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0047] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a method for detecting battery health status, including the following steps:

[0048] S1. Collect historical operating data and full-band EIS scan data of the battery to construct a battery dataset, and collect status parameters separately, including the following steps:

[0049] Historical electrical and thermal data are collected during battery operation. Specifically, voltage and current sensors are used to monitor the battery's operating status and record voltage and current data for each charge-discharge cycle. At the same time, the number of charge-discharge cycles and timestamps are recorded by counting to obtain historical electrical data. Temperature sensors are installed on the battery surface and inside to record temperature changes during operation, forming historical thermal data.

[0050] It should be noted that the historical electrical and thermal data are collected in real time by sensors to ensure the continuity and accuracy of the data; high-precision sensors are used during the collection process and are calibrated regularly to reduce measurement errors; the historical electrical and thermal data are recorded at fixed time intervals, which is applicable to different types of batteries.

[0051] A full-band EIS scan was performed to acquire the battery's impedance spectrum data. Specifically, an electrochemical workstation was connected to the battery, and a wide-band AC signal was applied to the battery to measure its impedance response at different frequencies. The full-band EIS scan was performed point by point from high frequency to low frequency, and the real and imaginary parts of the impedance at each frequency point were recorded to form the battery's impedance spectrum data represented by a Nyquist plot.

[0052] It should be noted that the full-band EIS scan is performed under controlled temperature and state of charge conditions to ensure data comparability; the scanning process requires the use of automated script control to avoid manual intervention and improve efficiency.

[0053] A battery dataset is constructed by integrating historical electrical data, historical thermal data, and impedance spectral data. Specifically, the collected voltage data, current data, cycle count, historical thermal data, and battery impedance spectral data are imported into a unified data processing platform. Median filtering is used to remove noise and outliers. Timestamps are used for data alignment to ensure the temporal consistency of historical electrical data, historical thermal data, and battery impedance spectral data. Each record corresponds to a time point and contains all data fields, stored in tabular form to obtain the battery dataset.

[0054] State parameters are collected and stored individually using sensors. Specifically, voltage, current, and temperature sensors are used to collect real-time battery voltage, current, and temperature data. The cycle count is directly recorded using a counting device. The state of charge (SOC) is obtained based on the voltage and current data using the ampere-hour integration method, expressed as:

[0055] ;

[0056] in, For the current moment The state of charge, To measure the battery's open-circuit voltage (voltage data) after the battery has reached electrochemical equilibrium, the initial value of the state of charge (SOC) is obtained by querying a preset open-circuit voltage and SOC key table. For rated capacity, For current data, For the current moment, From initial time 0 to current time The time between.

[0057] It should be noted that the acquisition of state parameters is independent of historical electrical data, historical thermal data, and battery impedance spectrum data, ensuring modularity of data acquisition; state parameters are sampled at high speed to ensure real-time performance, and a compressed format is used for storage to save space; voltage data is used to determine the initial voltage required for the ampere-hour integration method. And perform periodic error correction.

[0058] S2. Train a deep reinforcement learning (DRL) agent using the battery dataset. Utilize the trained DRL agent to scan and make decisions regarding the battery's state parameters, outputting a sparse frequency band decision vector. This includes the following steps:

[0059] The state space, action space, and reward function of the DRL agent are defined based on state parameters. Specifically, the collected state parameters are organized into feature vectors, which serve as the state space of the DRL agent, representing the current operating state of the battery. The frequency bands scanned by the full-band EIS are pre-divided into N candidate frequency points. Each action is an N-dimensional binary vector, defined as a combination of frequency points that the agent can choose, forming a multi-dimensional action space. The comprehensive calculation expression of the reward function is as follows:

[0060] ;

[0061] in, As a reward value, The complete EIS spectrum is generated by reconstructing a model from impedance data at frequency points selected by the agent through pre-training. This is the true spectrum of the full scan in the dataset. This is the normalized root mean square error, used to calculate the reconstruction error between the two. The total number of frequency points selected for this action. For each selected frequency point Pre-calibrated measurement time, , , and As positive coefficients used to balance the various weights, in subsequent training processes, , , and The specific values ​​will be dynamically optimized based on the statistical characteristics such as the reconstruction error distribution and scan time range reflected in the battery dataset.

[0062] Furthermore, when <5 or >15:00: The contribution of the reconstruction accuracy reward is too weak, and the DRL agent tends to select very few frequency points (average). <3), leading to >15%, which fails to meet basic accuracy requirements; when >20 or <5: The contribution of the reconstruction accuracy reward is too strong, and the DRL agent tends to select too many frequency points. >80% of candidate points), sparse scanning time is extended to almost the time of a full scan, resulting in insufficient efficiency improvement; therefore ∈[8,12], ∈[8,12], a balance between accuracy and efficiency can be achieved; similarly, when <0.01 or When <0.001: the efficiency penalty is too weak, and the DRL agent lacks sparsity constraints, making... >15, limited improvement in time efficiency; when >0.2 or When the value is greater than 0.02: the efficiency penalty is too strong, and the DRL agent excessively pursues sparsity, resulting in... <5, resulting in NRMSE exceeding 10%; therefore ∈[0.05,0.15], ∈[0.005,0.015], the average number of scan points can be controlled between 5 and 10 while ensuring NRMSE < 10%; using prior knowledge of typical battery datasets, the initial value is set to =10, =10, =0.1, =0.01, and the value is optimized through sliding window statistics within a fixed training period.

[0063] It should be noted that normalizing the state space by normalizing the root mean square error ensures that state parameters of different dimensions have a consistent impact on the DRL agent, accelerating the convergence of the DRL agent and improving training stability. The action space design using binary vectors makes the policy output intuitive and easy to decode, allowing direct control of the hardware for scanning. The reward function transforms the abstract data quality and scanning efficiency objectives into calculable numerical signals. Since different battery types, aging states, and usage environments lead to differences in the statistical characteristics of the dataset, the initial parameters are usually not optimal and must be adapted through dynamic tuning.

[0064] The DRL agent is trained using a battery dataset to optimize its policy network, resulting in a fully trained DRL agent. Specifically, the battery dataset is divided into training and validation sets based on timestamps. The DRL agent is initialized using the Proximal Policy Optimization (PPO) algorithm. Under strictly controlled ideal conditions in the laboratory, the batteries are scanned across the entire frequency range using EIS to obtain ground truth values ​​across the entire frequency band. For each data sample in the training set, the policy network of the DRL agent receives the data and then propagates it forward to obtain the selection probability distribution for each candidate frequency point. An action vector is generated from the selection probability distribution using multinomial sampling, and the corresponding frequency point data is extracted from the ground truth values ​​across the entire frequency band based on the action vector, thus obtaining a reward value. The agent updates the relevant adjustment parameters of the policy network using interaction experience. This process is repeated until the reward value on the validation set no longer increases. The relevant adjustment parameters of the policy network at this point are saved, resulting in a fully trained agent.

[0065] It should be noted that the training fully utilizes historical data, avoiding the time-consuming and risky process of repeated interaction with real battery systems. The training process is safe, efficient, and repeatable. The early stopping strategy based on the validation set effectively prevents overfitting and ensures that the trained DRL agent has good generalization ability and can be applied to unseen battery states.

[0066] The real-time state parameters of the target battery are input into the trained DRL agent for scanning and decision-making. After the decision, a sparse frequency band decision vector is generated. Specifically, the real-time state parameters are preprocessed and normalized in the same way as in the training phase and combined into a real-time state vector. The real-time state vector is then input into the deployed and trained DRL agent for forward computation, outputting a probability distribution. Finally, the action with the highest probability is sampled or selected to generate a binary decision vector, which is the sparse frequency band decision vector.

[0067] It should be noted that the decision-making process involves minimal computation, requiring only one forward propagation of the neural network, and can be completed in milliseconds, meeting the computational power and timing requirements for online real-time decision-making. The output decision vector has a clear meaning and compact structure, which can be directly converted into hardware control instructions to drive the measurement equipment to scan specific frequency points marked in the sparse frequency band decision vector, thereby achieving fast and accurate adaptive EIS measurement. This represents a leap from "fixed full-frequency scanning" to "intelligent adaptive scanning," enhancing the practicality of EIS in online monitoring.

[0068] S3. Use sparse frequency band decision vector control to perform sparse scanning of the target battery using the EIS device to obtain sparse frequency band EIS data, including the following steps:

[0069] The sparse frequency band decision vector is parsed to obtain the specific frequency point sequence to be scanned. Specifically, the N-dimensional binary decision vector output by the DRL agent is mapped to a predefined global frequency point list. The decision vector is traversed to extract the index position corresponding to all elements with a value of "1". The specific frequency value is mapped from the global list according to the index position to generate the specific frequency point sequence to be scanned.

[0070] It should be noted that the parsing process is simple and direct with extremely low computational complexity, ensuring a rapid response from decision to execution; mapping through a predefined global frequency list ensures the accuracy and reliability of the correspondence between the decision vector and the actual physical frequency, avoiding instruction ambiguity.

[0071] The scanning parameters of the EIS device are configured according to the specific frequency point sequence to be scanned. The EIS device is then controlled to perform a rapid EIS scan on the target battery based on the scanning parameters, collecting and recording the impedance data of the target battery to form sparse frequency band EIS data. Specifically, the specific frequency point sequence to be scanned is sent to the electrochemical workstation or embedded EIS measurement device through the communication interface. On the device side, the scanning mode is set to fixed-point frequency scan or list scan, and the specific frequency point sequence to be scanned is written into the scanning parameter list. After setting the amplitude of the excitation current (or voltage) corresponding to each frequency point, the scanning process is started. The device will apply AC excitation signals only sequentially at the frequency points specified in the specific frequency point sequence to be scanned and measure the impedance response of the battery, accurately collecting the real and imaginary impedance data at each frequency point. After the scan is completed, the real and imaginary impedance data collected at each frequency point are organized and stored in frequency order to form sparse frequency band EIS data.

[0072] It should be noted that by utilizing the decision-making results of the DRL agent, the traditional full-band continuous scanning mode was abandoned, and "on-demand scanning" was achieved. Only a few specific frequency points were scanned, which reduced the total scanning time by orders of magnitude compared to full-band scanning, greatly reducing the time cost and energy consumption of online monitoring, and reducing the risk of changes in battery operating status during scanning.

[0073] S4. Construct a meta-knowledge base based on multiple battery models, and train the meta-knowledge base using a meta-learning algorithm to obtain a baseline transfer model, including the following steps:

[0074] Full-band EIS data and corresponding DRT spectra of various battery models were collected. Specifically, battery samples from different manufacturers, chemical systems, capacities, and specifications were collected. The battery samples were periodically subjected to full-band EIS scanning under controlled SOC and temperature conditions during their respective aging experiments to obtain full-band EIS data of various battery models. The relaxation time distribution of the full-band EIS data of various battery models was analyzed using inversion algorithms such as regularization to generate their corresponding DRT spectra.

[0075] It should be noted that collecting diverse battery data ensures the broad representativeness of the meta-knowledge base, which is the basis for the model's ability to transfer across models; collecting EIS and calculating DRT under unified conditions ensures the comparability between data from different sources.

[0076] Aging characteristic peak parameters are extracted based on full-band EIS data and corresponding DRT spectra. A meta-knowledge base is then constructed using these aging characteristic peak parameters. Specifically, peak detection is performed on each set of DRT spectra to identify characteristic peaks related to the aging mechanism. For each characteristic peak related to the aging mechanism, the peak position, peak height, peak area, and peak width are extracted to obtain characteristic peak parameters. The characteristic peak parameters from different battery models are associated and labeled with the corresponding battery states and stored in a feature matrix structure to construct the meta-knowledge base.

[0077] It should be noted that the DRT spectrum can separate the overlapping relaxation processes in the full-band EIS data, thereby clearly and physically characterizing the different aging modes inside the battery; by extracting peak parameters, the complex spectral information is transformed into a set of quantitative, low-dimensional and physically meaningful feature vectors, which greatly compresses the amount of data.

[0078] A meta-learning algorithm is used to train the battery aging state mapping relationship of the meta-knowledge base to obtain the baseline transfer model. Specifically, the data of each battery model in the meta-knowledge base is regarded as an independent task; the MAML meta-learning algorithm is used for training; in the inner loop of the MAML meta-learning algorithm, it learns how to quickly adapt to a small amount of feature peak data of a certain battery model and establish a mapping relationship from feature peak parameters to battery aging state; in the outer loop of the MAML meta-learning algorithm, the initial parameters are optimized so that the meta-knowledge base can work on the new task after a small number of gradient updates; through multiple rounds of cross-task training, the baseline transfer model is obtained.

[0079] It should be noted that the obtained baseline transfer model is not a specific state estimator, but a "pre-trained model" or "model generator". When faced with a brand new battery model with very little data, it can be quickly adapted into a high-precision dedicated state estimation model with only a small amount of gradient updates or fine-tuning. This completely solves the problem of needing a large amount of data to retrain due to changes in battery models, and enables rapid transfer and generalization across models.

[0080] S5. Using a benchmark migration model, perform few-sample adaptation processing on the sparse frequency band EIS data to obtain the virtual reference DRT benchmark and SOH estimate of the target battery, including the following steps:

[0081] The baseline transfer model is quickly adapted with a small number of samples using prior knowledge from the meta-knowledge base to obtain the adapted baseline transfer model. Specifically, a limited number of sparse frequency band EIS data collected under different health conditions and the corresponding real SOH values ​​obtained through capacity testing are obtained from the recent history of the target battery to form a support set. The support set is used as fine-tuning samples to train the baseline transfer model with a small number of iterations of gradient descent to obtain the adapted baseline transfer model.

[0082] It should be noted that only a very small number of targeted fine-tuning samples are needed for fine-tuning, which greatly reduces the amount of data required for new batteries and avoids the massive amounts of data and computing resources required to train a large model from scratch for each new battery model.

[0083] Based on the adapted benchmark transfer model, the sparse frequency band EIS data of the target battery is fitted and reconstructed to generate a virtual reference DRT benchmark for the target battery. Specifically, the newly acquired sparse frequency band EIS data of the target battery is input into the adapted benchmark transfer model; the sparse frequency band EIS data is validated, standardized in format, and correlated with battery state parameters (voltage, temperature, etc.) to obtain a standardized sparse EIS feature vector that can be directly read by the model; the benchmark transfer model uses prior knowledge learned from the meta-knowledge base (i.e., complete DRT spectra of multiple batteries) to connect the sparse EIS feature vectors into a smooth and continuous relaxation time distribution curve (DRT distribution curve) through a regularized constraint optimization algorithm; the time-domain relaxation intensity distribution is transformed using the integral equation in DRT theory to output the predicted complete relaxation time distribution DRT spectrum, i.e., the virtual reference DRT benchmark of the target battery.

[0084] It should be noted that even if the sparse frequency band EIS data of the target battery is highly sparse, the adapted baseline transfer model can still generate a high-fidelity, complete virtual DRT spectrum based on prior knowledge.

[0085] The SOH estimate of the target battery is obtained by comparing the aging characteristic peak parameters in the virtual reference DRT benchmark with the battery aging state mapping relationship. Specifically, peak analysis is performed on the generated virtual reference DRT benchmark to extract new aging-related characteristic peak parameters. The new aging-related characteristic peak parameters are input into the adapted benchmark transfer model to obtain the parameter state mapping relationship. Since the parameter state mapping relationship has been optimized during the meta-knowledge base training and few-sample adaptation process, after inputting the new characteristic peak parameters, the adapted benchmark transfer model can directly use the parameter state mapping relationship for forward propagation to output the SOH estimate of the target battery.

[0086] It should be noted that by analyzing the evolution of characteristic peaks in the virtual DRT benchmark to infer the SOH estimate of the target battery, the estimation result is not only a numerical output, but also has physical interpretability; it realizes the rapid and accurate assessment of the battery's health status using a very small amount of sparse scan data, completing a complete closed loop from "sparse data" to "rich information" and then to "accurate state estimation".

[0087] S6. Obtain the electrical connection topology of the battery using a multi-channel scanner, and map the SOH estimate and state parameters of the target battery into node and edge features to construct a heterogeneous battery connection graph, including the following steps:

[0088] The electrical connection types of all batteries are obtained by a multi-channel scanner, and an electrical connection topology is generated. Specifically, a multi-channel data acquisition unit or a dedicated topology scanner is used to automatically identify the physical connection relationships between all batteries, such as series, parallel, and series-to-parallel connections, by applying detection signals to the battery pack bus or measuring the potential difference and current path between each terminal. Each battery is abstracted as a node, and the physical connection lines between batteries are abstracted as edges, thereby generating a topology diagram describing the electrical connection relationships of the entire battery pack.

[0089] It should be noted that the automated and non-intrusive identification of the complex electrical connection structure of the battery pack avoids the problems of error and inefficiency in manual inspection; the generated topology map exists in the form of a computer-readable adjacency matrix or edge list, ensuring that the physical relationships between batteries are accurately modeled.

[0090] The SOH estimate and state parameters of each battery are mapped to node features in a graph neural network. Specifically, each battery is regarded as a node in the graph. The SOH estimate of the battery, as well as the real-time collected state parameters such as voltage, current, temperature, and SOC, are organized and normalized to form a multi-dimensional feature vector, which is then used as the node features of the corresponding node.

[0091] The edges in the graph neural network are defined based on the electrical connection topology, and the electrical connection type and state parameters are mapped to edge features. Specifically, based on the electrical connection topology generated in the previous step, undirected or directed edges are established between two battery nodes that have a physical connection. Furthermore, the physical connection type represented by each edge is encoded as a one-hot vector, and the measured or estimated parameters such as current and resistance on the connection line are optionally added as additional features to form edge features.

[0092] It should be noted that defining edges establishes channels for interaction between nodes, enabling graph neural networks to simulate current flow and state dependencies; assigning features to edges, especially distinguishing connection types, allows graph neural networks to learn key physical constraints such as consistent current and superimposed voltage in series connections, and consistent voltage and superimposed current in parallel connections. This greatly enhances the ability of graph neural networks to integrate prior physical knowledge and improves the accuracy of their predictions of the overall behavior of battery packs.

[0093] Based on node features and edge features, a heterogeneous graph for battery connection is constructed. Specifically, all node sets, node feature sets, edge sets, and edge feature sets are integrated into a complete graph data structure. Since the edges in the graph data structure have different types, the graph data structure is defined as a heterogeneous graph.

[0094] It should be noted that the constructed battery connection heterogeneous graph is a precise mathematical abstraction of the entire battery pack system, which includes the individual states of all batteries, i.e., node characteristics, and the complex physical coupling relationships between batteries. The heterogeneous graph enables the model to perform state assessment, fault prediction, and anomaly detection at the battery pack level, rather than just at the level of a single battery.

[0095] S7. Based on the virtual reference DRT benchmark, perform message passing and aggregation processing on the battery connection heterogeneity diagram to obtain the aging coupling relationship representation, and calculate the aging propagation path and health status diagnosis results, including the following steps:

[0096] The aging characteristic peak parameters in the virtual reference DRT benchmark are injected as prior knowledge into the battery connection heterogeneous graph. Specifically, the virtual reference DRT benchmark generated for each battery is analyzed to extract the aging characteristic peak parameters. The aging characteristic peak parameters are then added as a new feature dimension to the node features of the corresponding battery node in the battery connection heterogeneous graph.

[0097] It should be noted that introducing the internal state depth information resolved from the electrochemical impedance level into the graph structure greatly enriches the representation dimensions of the nodes, enabling the graph neural network to see the external operating state of the battery. This provides crucial prior knowledge for subsequent accurate analysis of the mutual influence between batteries and the propagation of aging, laying the physical foundation for high-precision diagnosis.

[0098] Multi-round message passing is performed on the heterogeneous graph of battery connections. The aging state information of adjacent battery nodes is aggregated along the topological edges. The aging coupling relationship representation of the battery as a whole is generated through a readout function. Specifically, an RGCN graph neural network specially adapted for heterogeneous graphs is used, with enhanced node features and edge features as input for multi-round message passing. In each round, each node receives messages from its neighboring nodes. The messages are modulated by edge features and updated with its own state representation through neural network aggregation. After K rounds of iteration, the hidden state of each node is fused with the aging information of all batteries in the K-hop topological neighborhood. Using the readout function, the final hidden states of all nodes are aggregated into a single, fixed-size graph-level vector, which represents the aging coupling relationship representation of the battery as a whole.

[0099] It should be noted that the multi-round message passing process simulates the dynamic process of aging effects propagating and coupling in the battery pack through electrical connections; heterogeneous graph neural networks can distinguish the different impact patterns brought about by different connection types such as series and parallel connections, thereby modeling coupling relationships more accurately.

[0100] Based on the aging coupling relationship representation, the aging propagation path is traced and the health status diagnosis results of each battery node are calculated. Specifically, the graph-level aging coupling relationship representation obtained above is input into the diagnostic decoder. The diagnostic decoder contains two types of outputs: 1) Path tracing: The contribution of all nodes in the graph to the overall aging is obtained by using graph attention weights, and the battery node most likely to be the origin of aging and the main propagation path are traced back in reverse; 2) State diagnosis: The graph-level aging coupling relationship representation is fused with the hidden state of each node through a fully connected neural network, and the MLP regressor completes the final estimate of the health status of each node, outputting the SOH correction value and RUL prediction for each battery node.

[0101] This embodiment also provides a computer device applicable to a battery health status detection method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the battery health status detection method proposed in the above embodiment.

[0102] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0103] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the battery health status detection method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0104] In summary, this invention significantly reduces scanning time and computational resource consumption by dynamically selecting frequency points for sparse EIS scanning, achieving efficient and real-time diagnosis and solving the problem of low efficiency in traditional full-band EIS scanning. By constructing a heterogeneous battery connection graph and combining it with graph neural networks to analyze the aging coupling relationship between batteries, it accurately traces the aging propagation path and accurately predicts the system-level health status, overcoming the limitation of existing technologies that cannot comprehensively analyze the mutual influence of aging.

[0105] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method of detecting state of health of a battery, characterized by: comprising, Collecting historical operation data and full-band EIS scan data of the battery, constructing a battery data set, and separately collecting state parameters; Collecting historical electrical data and historical thermal data during the operation of the battery; Performing full-band EIS scanning to obtain impedance spectrum data of the battery; Integrating the historical electrical data, the historical thermal data, and the impedance spectrum data to construct the battery data set; Separately collecting state parameters through sensors and storing them; Training a deep reinforcement learning (DRL) agent through the battery data set, using the trained DRL agent to make scanning decisions on the state parameters of the battery, and outputting a sparse frequency band decision vector; Using the sparse frequency band decision vector to control the EIS equipment to perform sparse scanning on the target battery, and obtaining sparse frequency band EIS data; Constructing a meta-knowledge base through multiple types of batteries, and training the meta-knowledge base through a meta-learning algorithm to obtain a benchmark transfer model; Collecting full-band EIS data and corresponding DRT spectrum graphs of multiple different types of batteries; Extracting aging characteristic peak parameters from the full-band EIS data and the corresponding DRT spectrum graphs, and constructing a meta-knowledge base using the aging characteristic peak parameters; Training the meta-knowledge base using a meta-learning algorithm to obtain a benchmark transfer model; Using the benchmark transfer model to perform few-sample adaptation processing on the sparse frequency band EIS data, and obtaining a virtual reference DRT benchmark and an SOH estimate value of the target battery; Using prior knowledge in the meta-knowledge base to quickly adapt the benchmark transfer model to obtain an adapted benchmark transfer model; Based on the adapted benchmark transfer model, fitting and reconstructing the sparse frequency band EIS data of the target battery to generate a virtual reference DRT benchmark of the target battery; Comparing the aging characteristic peak parameters in the virtual reference DRT benchmark of the target battery with the battery aging state mapping relationship to obtain an SOH estimate value of the target battery; Obtaining the electrical connection topology of the battery through a multi-channel scanner, and mapping the SOH estimate value and the state parameters of the target battery into node and edge features to construct a battery connection heterogeneous graph; Based on the virtual reference DRT benchmark, performing message passing and aggregation processing on the battery connection heterogeneous graph to obtain an aging coupling relationship representation, and calculating an aging propagation path and a health status diagnosis result.

2. The method of detecting state of health of a battery of claim 1, wherein: Training a DRL agent through a battery data set, using the trained DRL agent to make scanning decisions on the state parameters of the battery, and outputting a sparse frequency band decision vector, the specific steps are as follows, Defining the state space, action space, and reward function of the DRL agent based on the state parameters; Training the DRL agent using the battery data set to optimize the policy network of the DRL agent, and obtaining a trained DRL agent; Inputting the real-time state parameters of the target battery into the trained DRL agent for scanning decision, and generating a sparse frequency band decision vector.

3. The method of detecting state of health of a battery of claim 2, wherein: Using the sparse frequency band decision vector to control the EIS equipment to perform sparse scanning on the target battery, and obtaining sparse frequency band EIS data, the specific steps are as follows, Analyzing the sparse frequency band decision vector to obtain a specific frequency point sequence to be scanned; According to the specific frequency point sequence to be scanned, the scanning parameters of the EIS device are configured, the EIS device is controlled to perform a rapid EIS scan on the target battery according to the scanning parameters, impedance data of the target battery is collected and recorded, and sparse frequency band EIS data is formed.

4. The method of detecting state of health of a battery of claim 1, wherein: The electrical connection topology of the battery is obtained through a multi-channel scanner, and the SOH estimation value and state parameters of the target battery are mapped into node and edge features to construct a battery connection heterogeneous graph, and the specific steps are as follows, The electrical connection topology of the battery is obtained through a multi-channel scanner, and the SOH estimation value and state parameters of the target battery are mapped into node and edge features to construct a battery connection heterogeneous graph, and the specific steps are as follows, The SOH estimation value and state parameters of each battery are mapped into node features in the graph neural network. The edges in the graph neural network are defined according to the electrical connection topology, and the electrical connection type and state parameters are mapped into edge features. Based on the node features and edge features, a battery connection heterogeneous graph is constructed.

5. The method of detecting state of health of a battery of claim 4, wherein: Based on the virtual reference DRT benchmark, message passing and aggregation processing are performed on the battery connection heterogeneous graph to obtain an aging coupling relationship representation, and aging propagation paths and health status diagnosis results are calculated, and the specific steps are as follows, The aging feature peak parameters in the virtual reference DRT benchmark are injected into the battery connection heterogeneous graph as prior knowledge. Multiple rounds of message passing are performed on the battery connection heterogeneous graph, the aging state information of adjacent battery nodes is aggregated along the topological edges, and the aging coupling relationship representation of the entire battery is generated through a readout function. Based on the aging coupling relationship representation, the aging propagation paths are traced and the health status diagnosis results of each battery node are calculated. 6.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to realize the steps of the battery health state detection method of any one of claims 1-5.

7. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the battery health state detection method of any one of claims 1-5.

Images