A battery pack intelligent health management method based on a dynamic spatiotemporal graph neural network

Abstract

This invention relates to the fields of artificial intelligence and new energy management technology, and in particular to a method, device, equipment, and computer storage medium for intelligent health management of battery packs based on dynamic spatiotemporal graph neural networks. The intelligent health management method for battery packs based on dynamic spatiotemporal graph neural networks of this invention fully integrates multi-source, multi-modal data during battery pack operation, constructs a dynamically evolving graph structure to characterize the complex coupling relationships and degradation propagation paths between cells, introduces a GRU mechanism to model the state evolution characteristics in the time dimension, adapts to different battery operating conditions through a "pre-training-fine-tuning" transfer strategy, and is finally deployed in a battery management system (BMS) to achieve online health monitoring and intelligent fault early warning. This method effectively improves the accuracy, stability, and generalization ability of battery pack health assessment, overcoming the technical bottlenecks of traditional methods that cannot dynamically perceive cell state evolution and are difficult to adapt to complex operating environments.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and new energy management technology, and in particular to a method, device, equipment and computer storage medium for intelligent health management of battery packs based on dynamic spatiotemporal graph neural networks. Background Technology

[0002] With the widespread application of new energy vehicles and distributed energy storage systems, the health status of battery packs, as core energy units, has a decisive impact on the safety, reliability, and lifespan of the system. Battery packs are typically composed of multiple individual cells connected in series and parallel. During long-term charging and discharging, due to manufacturing differences, operating environment, and thermoelectric coupling effects, uneven aging of individual cells can easily occur. This not only affects overall performance but may also lead to safety hazards such as localized overheating and sudden capacity drops. Therefore, how to achieve accurate monitoring and lifespan prediction of the individual cells in the battery pack has become a key issue in the research of intelligent battery management systems (BMS).

[0003] Currently, traditional methods for predicting State of Health (SOH) / Relative Life (RUL) mostly rely on statistical modeling or empirical formula derivation, which is ill-suited to the highly dynamic and complexly coupled battery degradation process. Some studies have attempted to model battery aging using data-driven methods such as support vector machines and neural networks, but these generally suffer from the following shortcomings: First, the models are often based on individual cell historical data, neglecting the spatial interconnections and degradation coupling effects between cells; second, the graph structure is usually fixed and cannot dynamically respond to changes in temperature, current, internal resistance, and other state conditions during actual operation; third, they lack the ability to transfer data across battery models or usage scenarios, resulting in poor generalization; and fourth, in multi-task scenarios, such as simultaneously performing health prediction and anomaly diagnosis, the collaborative modeling capability is insufficient, leading to low overall prediction accuracy.

[0004] To overcome the above problems, there is an urgent need for an intelligent modeling method that can integrate multi-source data, dynamically characterize the internal structural evolution of the battery pack, and has cross-domain adaptability, so as to achieve high-precision prediction of SOH and RUL and abnormal cell location, thereby supporting the intelligent upgrade and safety assurance of the next generation of BMS system. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the problems of existing technologies that cannot accurately model the degradation coupling relationship between multiple cells in the battery pack, have limited health prediction accuracy, weak model generalization ability, and cannot achieve collaborative prediction of SOH and RUL and accurate location of abnormal cells.

[0006] To address the aforementioned technical problems, this invention provides a battery pack intelligent health management method based on a dynamic spatiotemporal graph neural network, comprising:

[0007] Acquire historical operating data of the source domain battery pack and corresponding battery pack structure information;

[0008] Preprocessing and feature extraction are performed on the multi-source operation history data of the source domain battery pack to generate a time-series cell state feature representation;

[0009] Based on the battery pack structure information data, an initial static graph structure is constructed. Based on the time-series cell state feature representation, the state difference degree between cell nodes in the initial static graph structure within each time step is calculated. An edge weight update function is introduced to adaptively adjust the static edge weights to generate a dynamic graph structure sequence.

[0010] A source domain dataset is constructed based on the time-series cell state feature representation and dynamic graph structure sequence. A pre-constructed dynamic graph neural network model is pre-trained based on the source domain dataset, and the trained model is transferred to the target domain for fine-tuning and optimization.

[0011] The optimized model is loaded into the battery management system, which outputs the battery health status and remaining lifespan prediction results in real time based on the target battery pack's operating data.

[0012] Preferably, the multi-source operation history data of the source domain battery pack includes key state parameters of cell operation under different operating conditions from different source battery packs, and the battery pack structure information data includes a diagram of the series and parallel connection relationship of cells within the battery pack and the cell structure arrangement numbering order.

[0013] Preferably, the preprocessing and feature extraction of the multi-source operating history data of the source domain battery pack to generate a time-series cell feature representation includes:

[0014] Preprocessing operations, including missing value imputation, outlier removal, normalization, and standardization, are performed on the multi-source operation history data of the source domain battery pack to obtain continuous time series data;

[0015] The continuous time series data is divided into multiple consecutive time segments based on a sliding window mechanism;

[0016] A position encoding mechanism is introduced for the cell state feature vector matrix in each time segment, which maps the time position to a vector representation and concatenates it with the corresponding cell state feature vector matrix to generate a time-series cell state feature representation.

[0017] Preferably, constructing the initial static diagram structure based on the battery pack structure information data includes:

[0018] An electrical connection adjacency matrix is ​​constructed based on the series-parallel connection relationship diagram of the cells in the battery pack;

[0019] Calculate the structural coupling edge weights based on the cell structure layout numbering order, and construct the structural coupling adjacency matrix;

[0020] Based on the electrical connection adjacency matrix and the structural coupling adjacency matrix, an initial static graph adjacency matrix is ​​generated using a weighted fusion strategy, and a static graph structure is constructed.

[0021] Preferably, the step of calculating the state difference degree between cell nodes in the initial static graph structure within each time step based on the time-series cell state feature representation, and introducing an edge weight update function to adaptively adjust the static edge weights to generate a dynamic graph structure sequence includes:

[0022] Based on the time-series cell state characteristic representation, Euclidean distance is used to calculate the state difference between cell nodes in the initial static graph structure within each time step, and an exponential decay function is introduced to adaptively adjust the static edge weights to generate a dynamic graph adjacency matrix under time evolution.

[0023] The adjacency matrix of the dynamic graph under the time evolution is symmetric and normalized to construct a dynamic graph structure sequence.

[0024] Preferably, constructing the dynamic graph neural network includes:

[0025] A dynamic graph neural network is constructed, and a gated recurrent unit is used to model the temporal dependencies of the cell nodes. For each time step, the neighbor state information of each cell node is aggregated based on the dynamic graph structure, the hidden state of the cell node is updated, and graph convolution operation is performed on the updated hidden state of the cell node to extract the embedded features of the cell node. The embedded features of the cell node are then mapped to the predicted value.

[0026] Preferably, the step of constructing a source domain dataset based on time-series cell state feature representation and dynamic graph structure sequence, pre-training a pre-built dynamic graph neural network model based on the source domain dataset, and transferring the trained model to the target domain for fine-tuning and optimization includes:

[0027] Based on the source domain dataset, the spatiotemporal degradation patterns and structural evolution characteristics of battery packs from different sources are learned according to the multi-task loss joint optimization strategy, and the pre-constructed dynamic graph neural network model is pre-trained.

[0028] Collect historical operating data of the target battery pack and corresponding battery pack structural information data, generate corresponding time-series cell state feature representations and dynamic graph structure sequences, and construct the target domain dataset;

[0029] Based on the target domain dataset, the shared parameters for structural modeling are kept frozen, and only the task parameters are fine-tuned and optimized.

[0030] The present invention also provides a smart health management device for a battery pack, comprising:

[0031] The source domain data acquisition module is used to acquire multi-source operation history data of the source domain battery pack and the corresponding battery pack structure information data;

[0032] The time-series feature extraction module is used to preprocess and extract features from the multi-source operation history data of the source domain battery pack to generate a time-series cell state feature representation.

[0033] The dynamic graph structure construction module is used to construct an initial static graph structure based on the battery pack structure information data, calculate the state difference degree between cell nodes in the initial static graph structure within each time step based on the time-series cell state feature representation, and introduce an edge weight update function to adaptively adjust the static edge weights to generate a dynamic graph structure sequence.

[0034] The model training module is used to construct a source domain dataset based on the time-series cell state feature representation and dynamic graph structure sequence, pre-train the pre-constructed dynamic graph neural network model based on the source domain dataset, and transfer the trained model to the target domain for fine-tuning and optimization.

[0035] The online prediction module is used to load the optimized model into the battery management system and output the battery health status and remaining service life prediction results in real time based on the target battery pack's operating data.

[0036] The present invention also provides a smart health management device for a battery pack, comprising:

[0037] Memory, used to store computer programs;

[0038] A processor is used to implement the steps of the above-described intelligent health management method for a battery pack when executing the computer program.

[0039] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described intelligent health management method for a battery pack.

[0040] The technical solution of the present invention has the following advantages compared with the prior art:

[0041] The intelligent health management method for battery packs based on dynamic spatiotemporal graph neural networks described in this invention fully integrates multi-source, multi-modal data during battery pack operation, constructs a dynamically evolving graph structure to characterize the complex coupling relationships and degradation propagation paths between cells, introduces a GRU mechanism to model the state evolution characteristics in the time dimension, adapts to different battery operating conditions through a "pre-training-fine-tuning" transfer strategy, and is finally deployed in the battery management system (BMS) to achieve online health monitoring and intelligent fault early warning. This method effectively improves the accuracy, stability, and generalization ability of battery pack health assessment, overcoming the technical bottlenecks of traditional methods that cannot dynamically perceive cell state evolution and are difficult to adapt to complex operating environments. Attached Figure Description

[0042] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0043] Figure 1 This is a flowchart illustrating the implementation of a smart health management method for battery packs provided by the present invention.

[0044] Figure 2 This is a schematic diagram of a GRU for a battery pack intelligent health management method based on a dynamic spatiotemporal graph neural network according to the present invention.

[0045] Figure 3 This is a schematic diagram of the GCN of the intelligent health management method for battery packs based on dynamic spatiotemporal graph neural networks according to the present invention. Detailed Implementation

[0046] The core of this invention is to provide a method, device, equipment, and computer storage medium for intelligent health management of battery packs, which effectively suppresses [damage / injury].

[0047] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Please refer to Figure 1. Figure 1 The flowchart illustrates the implementation of a smart health management method for a battery pack provided by this invention; the specific operation steps are as follows:

[0049] S101: Obtain historical operating data of the source domain battery pack and the corresponding battery pack structure information data;

[0050] S102: Preprocess and extract features from the multi-source operation history data of the source domain battery pack to generate a time-series cell state feature representation;

[0051] S103: Based on the battery pack structure information data, an initial static graph structure is constructed. Based on the time-series cell state feature representation, the state difference degree between cell nodes in the initial static graph structure within each time step is calculated. An edge weight update function is introduced to adaptively adjust the static edge weights to generate a dynamic graph structure sequence.

[0052] S104: Construct a source domain dataset based on the time-series cell state feature representation and dynamic graph structure sequence; pre-train a pre-constructed dynamic graph neural network model based on the source domain dataset; and transfer the trained model to the target domain for fine-tuning and optimization.

[0053] S105: Load the optimized model into the battery management system and output the battery health status and remaining service life prediction results in real time based on the target battery pack operation data.

[0054] Based on the above embodiments, this embodiment will provide a detailed description of step S101:

[0055] Among them, the multi-source operating history data of the source domain battery pack includes key state parameters of cell operation under different operating conditions from battery packs from different sources. These key state parameters include, but are not limited to, key state parameters such as temperature, voltage, current, and internal resistance of each cell. Specifically:

[0056] Cell temperature sequence: (Unit: °C), Cell voltage sequence: (Unit: V), Cell current sequence: (Unit: A), Cell AC or equivalent series internal resistance sequence: (Unit: mΩ)

[0057] Among them, the battery pack structure information data includes a diagram of the series and parallel connection relationship of the cells in the battery pack and the cell structure arrangement numbering order, which is used to build a data foundation that comprehensively reflects the battery state evolution.

[0058] Based on the above embodiments, this embodiment will provide a detailed description of step S102:

[0059] In the following embodiments, for the collected structured numerical data, missing value filling, outlier removal and normalization operations are performed, and the data is divided into continuous time segments based on the sliding window mechanism. Each segment introduces a sine and cosine position encoding mechanism to map the time position into a vector representation, which is then concatenated with the state features to form a time-aware node representation.

[0060] According to some embodiments, preprocessing and feature extraction are performed on the multi-source operation history data of the source domain battery pack to generate a time-series cell feature representation, including:

[0061] Preprocessing operations, including missing value imputation, outlier removal, normalization, and standardization, are performed on the multi-source operation history data of the source domain battery pack to obtain continuous time series data;

[0062] The continuous time series data is divided into multiple consecutive time segments based on a sliding window mechanism. ,in The length of the divided graph sequence is specified, and each graph snapshot contains cell status information within a fixed-length time window;

[0063] A positional encoding mechanism is introduced for the cell state feature vector matrix in each time segment, mapping the time position to a vector representation, and concatenating it with the corresponding cell state feature vector matrix to generate a time-series cell state feature representation. Specifically:

[0064] Time indexing is achieved using sine and cosine time coding methods. Mapped to a vector with the same dimension as the node embedding. The encoding formula is:

[0065]

[0066] in, This indicates the index position of the current snapshot or sample in the time series. Indicates the feature embedding dimension index. This represents the dimension of the node feature vector. After normalizing the state feature vectors of individual battery cells (such as temperature, voltage, current, and internal resistance), a multimodal node feature representation with time-aware capabilities is obtained by concatenating the position encoding vector. ,in, This indicates the result after preprocessing. Indicates the first The temperature characteristic vector of each cell within a time segment reflects the thermal state evolution of the cell during that time segment; This represents the voltage observation value of each cell within the same time period, used to measure the potential change trend of the cell; The current input characteristics of the battery cell are used to characterize the current carrying behavior during the charging and discharging process; This represents the AC internal resistance or equivalent series internal resistance of the battery cell during that time period, reflecting its degree of health degradation and changes in conductivity. Encode the location information at that moment.

[0067] Based on the above embodiments, this embodiment will provide a detailed description of step S103:

[0068] In the following embodiments, an electrical connection adjacency matrix is ​​constructed based on the electrical connection relationships between battery cells. Simultaneously, structural coupling edge weights are calculated based on the arrangement order of the battery cell numbers to generate a structural coupling adjacency matrix, using fusion coefficients. Weighted synthesis of static initial adjacency matrix Establish a graph structure .

[0069] According to some embodiments, constructing an initial static diagram structure based on the battery pack structure information data includes:

[0070] An electrical connection adjacency matrix is ​​constructed based on the series-parallel connection relationship diagram of the cells in the battery pack, specifically:

[0071] Assume the battery pack is made of A battery cell assembly is composed of individual cells, denoted as a set of cells. With each cell Let's consider a node in the graph. Based on the collected structural information, we first construct an electrical connection adjacency matrix. If the battery cell and If there is a series or parallel electrical connection, then let Otherwise, it is 0.

[0072] Based on the cell structure layout numbering order, the structural coupling edge weights are calculated, and a structural coupling adjacency matrix is ​​constructed. Specifically:

[0073] To further reflect the interrelationship of battery cells in structural arrangement, a structural coupling adjacency matrix is ​​constructed. The adjacent arrangement index distance is defined based on the arrangement order of cell numbers in the battery pack design. And use this to construct the edge weights:

[0074]

[0075] This form, without introducing actual spatial coordinates or thermal resistance modeling, uses the numbering order to approximately express the proximity of the arrangement and the potential for structural coupling, serving as an effective guide for the initialization of the diagram structure as a structural a priori heuristic.

[0076] Based on the electrical connection adjacency matrix and the structural coupling adjacency matrix, an initial static graph adjacency matrix is ​​generated using a weighted fusion strategy, and a static graph structure is constructed. Specifically:

[0077] A weighted fusion strategy is used to define the static initial graph adjacency matrix. for:

[0078]

[0079] in, The structural fusion weights are used to adjust the combined contribution of electrical connection structures and adjacent arrangement priors to the graph topology; the final constructed static graph structure is represented as follows: It is used to initialize the topology of the dynamic graph neural network model and serve as the basis for the subsequent graph snapshot sequence that evolves over time.

[0080] In the following embodiments, the state difference degree (Euclidean distance) between nodes is calculated by combining the node state characteristics within each time step, and an exponential decay function is introduced to adaptively adjust the static edge weights, forming a dynamic graph adjacency matrix under time evolution. Then, normalization is performed to obtain the dynamic graph structure sequence.

[0081] According to some embodiments, based on the time-series cell state feature representation, the state difference degree between cell nodes in the initial static graph structure within each time step is calculated, and an edge weight update function is introduced to adaptively adjust the static edge weights to generate a dynamic graph structure sequence, including:

[0082] Based on the time-series cell state characteristic representation, Euclidean distance is used to calculate the state difference between cell nodes in the initial static graph structure within each time step, and an exponential decay function is introduced to adaptively adjust the static edge weights to generate a dynamic graph adjacency matrix under time evolution. Specifically:

[0083] Based on the aforementioned initial static graph structure Combining each time segment Multi-source state data collected internally is used to adaptively and dynamically adjust the connection weights of edges in the graph to capture the time-varying characteristics of state similarity and degradation propagation relationships between battery cells. Let the set of nodes in the battery pack be... Each node corresponds to one battery cell, where Indicates the cell number. Time step Next, node The state characteristics are represented as follows:

[0084]

[0085] in , , , They represent battery cells respectively. At time step The following are the pre-processed temperature, voltage, current, and internal resistance characteristics. For the first The temporal location encoding vector corresponding to each snapshot of the graph. Any pair of nodes and The feature difference metric is defined as:

[0086]

[0087] Euclidean distance is used to measure the degree of state difference between nodes at a given time step; to reflect the dynamic impact of this difference on the structure, an edge weight update function is introduced. Dynamic adjacency matrix The elements are given by the following formula:

[0088]

[0089] in, This represents the original weight of the edge in the static topology. The edge weight attenuation coefficient controls the degree of influence of state differences on edge strength; this update method ensures that when two cells are in highly similar states... Its connection weights remain close When the states are significantly differentiated, the edge weights will decay exponentially, reflecting the physical meaning of the weakening of their potential coupling relationship.

[0090] The adjacency matrix of the dynamic graph under the aforementioned time evolution is symmetric and normalized to construct a sequence of dynamic graph structures, specifically:

[0091] To ensure the structural stability and propagation rationality of the dynamic graph, the adjacency matrix was further refined. Symmetry and normalization are performed using the following operations:

[0092]

[0093] in, for The degree matrix, It refers to the adjacency matrix. The result after symmetric normalization has diagonal elements defined as This processing can effectively eliminate the impact of unbalanced node degrees on the information propagation process; in extended scenarios with many edge types, edge weight update functions can be constructed and generated independently for electrically connected edges and structurally coupled edges. and Ultimately, through adjustable parameters The fusion forms a dynamic graph adjacency matrix. :

[0094]

[0095] Ultimately, based on continuity Inputting a time segment yields a sequence of dynamic graphs:

[0096]

[0097] This graph sequence serves as the input structural information stream, along with the aforementioned node state feature matrix sequence. The data are input into a dynamic graph neural network model to model the spatiotemporal interaction process and degradation propagation mechanism between cell states.

[0098] Based on the above embodiments, this embodiment will provide a detailed description of step S104:

[0099] like Figure 2 and Figure 3 In the following embodiments, a gated recurrent unit (GRU) is used to model the temporal dependencies of nodes, and a dynamic graph adjacency matrix is ​​used to aggregate neighbor states, thereby realizing the dynamic update of node states over time. Furthermore, a graph convolutional layer is used to extract the node embedding features at each time step. .

[0100] According to some embodiments, constructing a dynamic graph neural network includes:

[0101] A dynamic graph neural network is constructed, employing gated recurrent units to model the temporal dependencies of cell nodes. For each time step, neighbor state information of each cell node is aggregated based on the dynamic graph structure to update the hidden state of the cell node. Graph convolution is then performed on the updated hidden state to extract the embedded features of the cell nodes. These embedded features are then mapped to predicted values. Specifically:

[0102] Dynamic graph convolution includes: introducing a mechanism for the evolution of node states over time, modeling the dependencies of node representations in the time dimension, thereby capturing long-term dependencies and evolutionary trends during the degradation process. Specifically, let the node set be... The time series is The node state at each time step is represented by the input feature matrix. The corresponding diagram structure is a snapshot of the dynamic diagram after integrating the electrical connection relationship and the structural coupling relationship. .

[0103] To model the temporal evolution of node states, a graph neural network evolution mechanism based on gated recurrent units (GRUs) is adopted. At each time step... For each node Its previous hidden state Current neighbor aggregation information The information is input into the GRU unit for updating. This includes neighbor aggregation information. In the graph structure Next, node For all its neighboring nodes Hidden state Perform a weighted summation, with the weights being the normalized edge weights. It is used to perceive the state characteristics of neighboring nodes in the previous moment, reflecting the spatiotemporal propagation information received by the current node:

[0104]

[0105]

[0106] in, It is a dynamic graph adjacency matrix that has been symmetrically normalized. Represents a node In the figure The set of neighboring nodes in the middle, This represents a standard gated loop unit, which contains an interaction mechanism between update gates and reset gates, and can effectively model the state change trend of nodes at different time steps.

[0107] All nodes in time By concatenating the hidden states, we obtain the hidden state matrix of the nodes. Based on this, perform the convolution operation shown in the figure below to generate output features:

[0108]

[0109] in, The weight matrix to be trained. For bias terms, It is a non-linear activation function. For the final time step The next node is embedded.

[0110] This mechanism dynamically updates node states over time and fuses them with structure-aware convolution, enabling the model to capture fine-grained correlations in the state evolution between cells, providing high spatiotemporal resolution embedding support for subsequent health status assessment and lifetime prediction.

[0111] In the following embodiments, a dynamic graph neural network model is pre-trained based on source domain battery pack data to learn general degradation patterns and structural evolution mechanisms; then, task parameters are fine-tuned on target battery pack data while keeping structural parameters unchanged, thereby achieving accuracy improvement under small sample conditions.

[0112] According to some embodiments, a source domain dataset is constructed based on time-series cell state feature representation and dynamic graph structure sequence. A pre-built dynamic graph neural network model is pre-trained based on the source domain dataset, and the trained model is then transferred to the target domain for fine-tuning and optimization.

[0113] Based on the aforementioned source domain dataset, the spatiotemporal degradation patterns and structural evolution characteristics of battery packs from different sources are learned according to a multi-task loss joint optimization strategy. A pre-built dynamic graph neural network model is then pre-trained. Specifically:

[0114] Based on the pre-processed dynamic graph sequence, a "pre-training-fine-tuning" mechanism is proposed to address the differences in operating conditions and degradation modes of battery packs from different sources, thereby improving the model's predictive generalization ability and transfer adaptability on the target battery pack.

[0115] First, suppose the source domain training data comes from multiple battery pack models or usage environments, and the corresponding dynamic graph sequence is denoted as:

[0116]

[0117] in, This indicates that the graph sequence originates from a specific source region battery pack. Represents the source domain. A snapshot of the fused graph structure at each time step, showing the graph structure at each time step in the source domain. It is dynamically constructed based on the state characteristics and structural correlation information of the battery pack.

[0118] Combined with the node state features at each time step in the source domain On this source domain, for the defined dynamic graph neural network model Parameter training is performed. Among these steps... Represents the source domain Mid-time step The state feature matrix of all nodes. For each node's feature dimensions, This represents the total number of time steps in the source domain. The spatiotemporal degradation patterns and structural evolution characteristics of battery packs from different sources are learned to provide parameter initialization and model priors for subsequent transfer and adaptation to the target battery pack.

[0119]

[0120] in, Indicates the first in the source domain The node embedding representation of a graph snapshot. These represent shared parameters for structural modeling and dedicated parameters for task decoding, respectively.

[0121] The training objective employs a multi-task loss joint optimization strategy, defining the total loss function as:

[0122]

[0123] in, , , These are respectively: State of Health (SOH), Source Domain Battery Remaining Life (RUL), and Loss from Anomaly Detection Task. , , These are the weighting coefficients.

[0124] Collect historical operating data of the target battery pack and corresponding battery pack structural information data, generate corresponding time-series cell state feature representations and dynamic graph structure sequences, and construct the target domain dataset;

[0125] Based on the target domain dataset, the shared parameters for structural modeling are kept frozen, and only the task parameters are fine-tuned and optimized. Specifically:

[0126] After pre-training, the model will be transferred to the target battery pack. After collecting data from the target battery pack and completing graph construction and feature extraction, a dynamic graph sequence will be formed. and the corresponding node state feature sequence As input, maintain shared parameters in structural modeling. Freeze, only for task parameters Fine-tuning and optimization are performed, namely:

[0127]

[0128] The loss function for the fine-tuning stage is defined as:

[0129]

[0130] in, , , Consistent with the pre-training phase, this strategy effectively leverages source domain data to learn general structural degradation features and adapts to the specific operating conditions of the target battery pack, achieving high-performance health prediction with small sample sizes.

[0131] Based on the above embodiments, this embodiment provides a detailed description of step S105:

[0132] In the following embodiments, the final model is loaded into the battery management system (BMS) to receive cell operation data in real time. The graph neural network model outputs the SOH, RUL and abnormal cell identification results to achieve intelligent perception and efficient prediction of the battery pack's operating status.

[0133] According to some embodiments, the aforementioned dynamic graph neural network model that has undergone transfer learning optimization is deployed inside the battery management system (BMS). Through an embedded hardware and software collaborative mechanism, real-time inference of key health indicators and fault identification during battery pack operation are achieved. Specifically:

[0134] Utilizing a pre-trained, fine-tuned, and optimized dynamic graph neural network model loaded into BMS. Perform the online prediction process:

[0135]

[0136] in, Embed vectors for nodes. The final model parameters loaded for both hardware and software; For the downstream decoder module, used to convert the embedding vector Mapped to actual predicted values This includes the output of health status (SOH), remaining lifespan (RUL), and probability of anomalies.

[0137] The online predictions for SOH and RUL can be expressed as follows:

[0138]

[0139]

[0140] in, and The task decoding function is implemented as a lightweight fully connected layer or a multilayer perceptron (MLP).

[0141] Meanwhile, to achieve precise location of abnormal cells, a node-level fault probability output is constructed:

[0142]

[0143] in, For the first time after migrating to the target battery pack Each node Embedded representation of time, This indicates the predicted probability of an anomaly for the battery cell.

[0144] This embodiment significantly improves the intelligent monitoring and response capabilities of the battery pack at the edge while ensuring prediction accuracy, realizing a closed loop of health status perception and fault early warning throughout the entire life cycle.

[0145] This invention collects multi-source operational data of battery packs under different operating conditions, constructs a graph neural network with each individual cell in the battery pack as a graph node, and builds an initial graph topology based on their physical series and parallel electrical connections and numbering order, forming an electrical connection adjacency matrix and a structural coupling adjacency matrix respectively. A static graph adjacency matrix reflecting prior structural information is generated through a weighted fusion strategy. Multi-source state data during battery operation is collected, including key operating parameters such as cell temperature, voltage, current, and AC internal resistance. The collected structured numerical data undergoes missing value filling, outlier removal, normalization, and standardization, and is divided into multiple time segments based on a sliding window mechanism. Sine and cosine position encoding is introduced to construct time-aware node feature representations. Instead of directly introducing a node importance weighting mechanism into the graph structure, edge weights are dynamically updated based on the Euclidean distance between node states. An exponential decay function is used to adaptively adjust the adjacency matrix at the time step level, generating a dynamic graph structure sequence. Considering the strong temporal dependence of the battery degradation process, a dynamic graph neural network combining graph convolution and gated recurrent units (GRUs) is adopted. The system employs a graph neural network (D-GNN) structure to model the evolution of cell node states over time. By fusing graph snapshot sequences with time encoding, it achieves a fine-grained characterization of state evolution paths and degradation trends. Building upon this, a transfer learning mechanism is introduced. The system is first pre-trained on dynamic graph data constructed from multiple source domain battery packs to capture generalized spatiotemporal degradation patterns and structural evolution characteristics. Then, by freezing structural modeling parameters and fine-tuning only task decoding parameters to adapt to the target battery pack, the prediction accuracy and generalization ability under small sample conditions are improved. Finally, the trained dynamic graph neural network model is deployed in the battery management system (BMS) to receive real-time operating data and perform joint online prediction of SOH (State of Health) and RUL (Relative Lifetime). Simultaneously, by combining the node-level residual analysis results of the model output and observations, potentially degraded cells are further identified, enabling intelligent health diagnosis and risk warning for the battery pack.

[0146] This invention also provides a smart health management device for battery packs; the specific device may include:

[0147] The source domain data acquisition module is used to acquire multi-source operation history data of the source domain battery pack and the corresponding battery pack structure information data;

[0148] The time-series feature extraction module is used to preprocess and extract features from the multi-source operation history data of the source domain battery pack to generate a time-series cell state feature representation.

[0149] The dynamic graph structure construction module is used to construct an initial static graph structure based on the battery pack structure information data, calculate the state difference degree between cell nodes in the initial static graph structure within each time step based on the time-series cell state feature representation, and introduce an edge weight update function to adaptively adjust the static edge weights to generate a dynamic graph structure sequence.

[0150] The model training module is used to construct a source domain dataset based on the time-series cell state feature representation and dynamic graph structure sequence, pre-train the pre-constructed dynamic graph neural network model based on the source domain dataset, and transfer the trained model to the target domain for fine-tuning and optimization.

[0151] The online prediction module is used to load the optimized model into the battery management system and output the battery health status and remaining service life prediction results in real time based on the target battery pack's operating data.

[0152] The battery pack intelligent health management device of this embodiment is used to implement the aforementioned battery pack intelligent health management method. Therefore, the specific implementation of the battery pack intelligent health management device can be found in the embodiment section of the battery pack intelligent health management method above. For example, the source domain data acquisition module, the time series feature extraction module, the dynamic graph structure construction module, the model training module, and the online prediction module are respectively used to implement steps S101, S102, S103, S104, and S105 in the above-mentioned battery pack intelligent health management method. Therefore, its specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0153] A specific embodiment of the present invention also provides a battery pack intelligent health management device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described battery pack intelligent health management method.

[0154] A specific embodiment of the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described intelligent health management method for a battery pack.

[0155] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0156] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0157] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0158] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0159] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for intelligent health management of battery packs based on dynamic spatiotemporal graph neural networks, characterized in that, include: Acquire historical operating data of the source domain battery pack and corresponding battery pack structure information; The multi-source operation history data of the source domain battery pack is preprocessed and feature extracted to generate a time-series cell state feature representation; Based on the battery pack structure information data, an initial static graph structure is constructed. Based on the time-series cell state feature representation, the state difference degree between cell nodes in the initial static graph structure within each time step is calculated. An edge weight update function is introduced to adaptively adjust the static edge weights to generate a dynamic graph structure sequence. A source domain dataset is constructed based on time-series cell state feature representation and dynamic graph structure sequence. A pre-built dynamic graph neural network model is pre-trained based on this source domain dataset, and the trained model is then transferred to the target domain for fine-tuning and optimization, including: Based on the source domain dataset, the spatiotemporal degradation patterns and structural evolution characteristics of battery packs from different sources are learned according to the multi-task loss joint optimization strategy, and the pre-constructed dynamic graph neural network model is pre-trained. Collect historical operating data of the target battery pack and corresponding battery pack structural information data, generate corresponding time-series cell state feature representations and dynamic graph structure sequences, and construct the target domain dataset; Based on the target domain dataset, keep the shared parameters for structural modeling frozen and fine-tune and optimize the task parameters; The construction of the dynamic graph neural network includes: A dynamic graph neural network is constructed, and a gated recurrent unit is used to model the temporal dependency relationship of the cell nodes. For each time step, the neighbor state information of each cell node is aggregated based on the dynamic graph structure, the hidden state of the cell node is updated, and graph convolution operation is performed on the updated hidden state of the cell node to extract the embedded features of the cell node. The embedded features of the cell node are then mapped to the predicted value. The optimized model is loaded into the battery management system, which outputs the battery health status and remaining lifespan prediction results in real time based on the target battery pack's operating data.

2. The intelligent health management method for battery packs according to claim 1, characterized in that, The multi-source operation history data of the source domain battery pack includes key state parameters of cell operation under different operating conditions from different source battery packs. The battery pack structure information data includes a diagram of the series and parallel connection relationship of cells within the battery pack and the cell structure arrangement numbering sequence.

3. The intelligent health management method for battery packs according to claim 2, characterized in that, The step of preprocessing and feature extraction of the multi-source operation history data of the source domain battery pack to generate a time-series cell feature representation includes: Preprocessing operations, including missing value imputation, outlier removal, normalization, and standardization, are performed on the multi-source operation history data of the source domain battery pack to obtain continuous time series data; The continuous time series data is divided into multiple consecutive time segments based on a sliding window mechanism; A position encoding mechanism is introduced for the cell state feature vector matrix in each time segment, which maps the time position to a vector representation and concatenates it with the corresponding cell state feature vector matrix to generate a time-series cell state feature representation.

4. The intelligent health management method for battery packs according to claim 2, characterized in that, The step of constructing the initial static diagram structure based on the battery pack structure information data includes: An electrical connection adjacency matrix is ​​constructed based on the series-parallel connection relationship diagram of the cells in the battery pack; Calculate the structural coupling edge weights based on the cell structure layout numbering order, and construct the structural coupling adjacency matrix; Based on the electrical connection adjacency matrix and the structural coupling adjacency matrix, an initial static graph adjacency matrix is ​​generated using a weighted fusion strategy, and a static graph structure is constructed.

5. The intelligent health management method for battery packs according to claim 1, characterized in that, Based on the time-series cell state feature representation, the state difference degree between cell nodes in the initial static graph structure within each time step is calculated, and an edge weight update function is introduced to adaptively adjust the static edge weights to generate a dynamic graph structure sequence, including: Based on the time-series cell state characteristic representation, Euclidean distance is used to calculate the state difference between cell nodes in the initial static graph structure within each time step, and an exponential decay function is introduced to adaptively adjust the static edge weights to generate a dynamic graph adjacency matrix under time evolution. The adjacency matrix of the dynamic graph under the time evolution is symmetric and normalized to construct a dynamic graph structure sequence.

6. A smart health management device for a battery pack, characterized in that, include: The source domain data acquisition module is used to acquire multi-source operation history data of the source domain battery pack and the corresponding battery pack structure information data; The time-series feature extraction module is used to preprocess and extract features from the multi-source operation history data of the source domain battery pack to generate a time-series cell state feature representation. The dynamic graph structure construction module is used to construct an initial static graph structure based on the battery pack structure information data, calculate the state difference degree between cell nodes in the initial static graph structure within each time step based on the time-series cell state feature representation, and introduce an edge weight update function to adaptively adjust the static edge weights to generate a dynamic graph structure sequence. The model training module is used to construct a source domain dataset based on time-series cell state feature representation and dynamic graph structure sequence, pre-train a pre-constructed dynamic graph neural network model based on the source domain dataset, and transfer the trained model to the target domain for fine-tuning and optimization, including: Based on the source domain dataset, the spatiotemporal degradation patterns and structural evolution characteristics of battery packs from different sources are learned according to the multi-task loss joint optimization strategy, and the pre-constructed dynamic graph neural network model is pre-trained. Collect historical operating data of the target battery pack and corresponding battery pack structural information data, generate corresponding time-series cell state feature representations and dynamic graph structure sequences, and construct the target domain dataset; Based on the target domain dataset, keep the shared parameters for structural modeling frozen and fine-tune and optimize the task parameters; The construction of the dynamic graph neural network includes: A dynamic graph neural network is constructed, and a gated recurrent unit is used to model the temporal dependency relationship of the cell nodes. For each time step, the neighbor state information of each cell node is aggregated based on the dynamic graph structure, the hidden state of the cell node is updated, and graph convolution operation is performed on the updated hidden state of the cell node to extract the embedded features of the cell node. The embedded features of the cell node are then mapped to the predicted value. The online prediction module is used to load the optimized model into the battery management system and output the battery health status and remaining lifespan prediction results in real time based on the target battery pack's operating data.

7. A smart health management device for battery packs, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the intelligent health management method for a battery pack based on a dynamic spatiotemporal graph neural network as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the intelligent health management method for a battery pack based on a dynamic spatiotemporal graph neural network as described in any one of claims 1 to 5.

Images