Causal graph neural network based multi-scale failure root cause tracing and retirement reason explanation method for power battery

Abstract

The application provides a power battery multi-scale failure tracing and retirement reason explanation method based on a causal diagram neural network, and the method comprises the following steps: collecting data from a battery management system, combining a physical level connection relationship to construct a directed edge, forming a multi-scale causal diagram; meanwhile, a graph neural network is used for causal propagation and aggregation to generate embedding vectors of each node; a retirement event is obtained, and a corresponding retirement node is located in the multi-scale causal diagram; based on the local causal sub-diagram and a node embedding set thereof, a candidate causal path from a potential cause node to the retirement node is identified, and a key node set is extracted; the key causal node is subjected to counterfactual intervention simulation, the influence degree of the retirement node if the state of the node is not abnormal is evaluated, whether the key causal node is a necessary reason for the retirement is judged, and a structured explanation result is generated. The application realizes systematization and refinement of power battery retirement cause analysis.

Description

Technical Field

[0001] This invention belongs to the technical field of causal graph neural networks, and particularly relates to a method for tracing the source of multi-scale failures and explaining the reasons for retirement of power batteries based on causal graph neural networks. Background Technology

[0002] With the large-scale application of new energy vehicles, the state evolution, decommissioning process, and key factors leading to decommissioning of power batteries, as the core energy carrier, have become key concerns for battery management systems and operation and maintenance platforms. Power battery decommissioning exhibits typical cross-scale, multi-variable coupling characteristics, typically starting with weak anomalies at the cell level and gradually spreading to the module and pack levels, triggering macroscopically observable capacity decay, voltage differential anomalies, or thermal runaway trends. However, existing technologies mainly focus on monitoring macroscopic characteristics such as voltage, current, temperature, and SOC, or rely on statistical models and empirical rules for fault identification. Their modeling structures are often limited to single-scale or weakly correlated expressions, making it difficult to establish the correlation paths between cells, modules, and packs. Regarding the decommissioning formation mechanism, existing methods can often only detect anomalies, but struggle to identify the complete evolutionary chain from "local weak signals" to "system decommissioning." Under complex operating conditions, the internal thermal, electrical, and chemical processes of power batteries exhibit significant nonlinearity and time lag. Local cell anomalies may be masked by the overall performance at the module level, making it difficult for existing methods to trace the true starting point in time and space. While deep learning methods can extract complex features, their closed structure and lack of causal path representation capabilities prevent them from identifying the causal direction between different variables, and they cannot answer critical engineering questions such as "which variable becomes the necessary factor leading to retirement at what time?" Furthermore, the inconsistent sampling frequencies of variables in power batteries, their long operating cycles, and the scarcity of retirement samples make traditional correlation analysis-based methods susceptible to noise disturbances and structurally redundant variables, hindering the identification of critical nodes with practical engineering significance within the complex battery system.

[0003] In summary, existing solutions lack a technical system that combines structured multi-scale representation, path-level causal identification, and verifiable counterfactual reasoning capabilities, making it difficult to meet the requirements of interpretability, traceability, and engineering feasibility for power battery retirement analysis. Summary of the Invention

[0004] The purpose of this invention is to propose a method for tracing the causes of multi-scale failures and explaining the reasons for retirement of power batteries based on causal graph neural networks, so as to solve the above-mentioned problems.

[0005] To achieve the above objectives, the present invention provides a method for tracing the causes of multi-scale failures and explaining the reasons for decommissioning of power batteries based on causal graph neural networks, comprising the following steps: S1. Collect cell temperature sequence, cell voltage sequence, module temperature difference, module discharge rate, and module internal resistance change rate from the battery management system, combine them into node attributes, and construct directed edges by combining physical hierarchy connection relationships to form a multi-scale causal graph; at the same time, use graph neural network to perform causal propagation and aggregation on the node states of the multi-scale causal graph to generate the embedding vector of each node. S2. Obtain the decommissioning event that has occurred, and locate the corresponding decommissioning node in the multi-scale causal graph; based on the structural hierarchy distance between decommissioning nodes, the state fluctuation characteristics within the time window, and the similarity of the node embedding vectors, prune the multi-scale causal graph, and extract the local causal subgraphs related to the decommissioning event and their node embedding sets. S3. Based on the local causal subgraph and its node embedding set, combined with the node embedding vector offset direction, structural hierarchy weight and path accumulation features, identify candidate causal paths from potential cause nodes to decommissioned nodes, and extract key node sets. S4. Perform counterfactual intervention simulation on the key causal nodes, assess the impact on the decommissioning nodes if the node's state is not abnormal, determine whether it is a necessary reason for decommissioning, and generate structured explanation results.

[0006] Furthermore, the graph neural network is specifically as follows: The first layer involves acquiring nodes as input nodes to the graph neural network, aggregating information from adjacent nodes, and obtaining the first layer embedding of the nodes. The second layer, based on the first layer embedding of the nodes, introduces time window features and uses a gated time series encoder to embed the historical state of each node, thereby obtaining the embedding vector of each node.

[0007] Furthermore, the gated time series encoder is implemented using a single-layer gated recurrent network, with the input being the nodes' past performance. The state sequence at each time point is output as historical trend features.

[0008] Furthermore, the pruning process of the multi-scale causal graph specifically includes: Structure filtering is performed based on the graph structure distance between nodes and decommissioned nodes and the node type weight, and time filtering is performed based on the intensity of the state fluctuation of nodes within the time window before decommissioning and the similarity of their embedding vectors with those of decommissioned nodes.

[0009] Furthermore, the structural radius of the graph structure distance is defined by the shortest path distance of the graph structure. At the same time, the characteristic of local thermal chain diffusion in the actual operation of the power battery is taken into account. That is, when the node is a module temperature difference node or a cell temperature node, its correlation weight with retirement is higher.

[0010] Furthermore, in S3, when identifying candidate causal paths, a node score is performed on the causal significance of each node. The node score integrates the offset of the node embedding vector relative to the decommissioned node, the projection along the decommissioning evolution trend direction, and the structural weight of the node's level. Furthermore, in S3, the candidate causal path is calculated based on the node score, specifically as follows: Starting from the decommissioning node, a back-diffusion search is performed on all possible upstream structural paths. For any path, its causal significance is defined as the weighted cumulative average of the causal significance of the nodes. Different weights are assigned to nodes at different levels in the path to increase the additional amplification factor for deeper nodes, so that the score of the cell-level nodes in the path has a more significant contribution to the overall score, thus obtaining the causal significance of the path. Path diffusion search is performed on the local causal subgraph to generate a set of candidate paths sorted from high to low significance, and a set of key nodes with scores higher than a threshold is selected from all nodes.

[0011] Furthermore, the cell-level nodes have a higher weight than the module-level and pack-level nodes.

[0012] Furthermore, in S4, the counterfactual intervention simulation is achieved by replacing the embedding vector of the key node with a preset health status reference vector and re-propagating the graph neural network to calculate the degree of change in the embedding vector of the retired node before and after the intervention.

[0013] Furthermore, in S4, when determining whether a node is a necessary cause, the co-occurrence frequency of the node in the candidate path or the stability of the edges connected to it is also considered to correct the score of the influence of a single node.

[0014] The beneficial technical effects of the present invention are at least as follows: This invention proposes a multi-scale causal path tracing and cause explanation method for power battery retirement scenarios. By constructing a nested causal graph neural network covering cell, module, and pack levels, it achieves a structured causal expression of operating state variables and utilizes the temporal evolution characteristics of the embedding space to improve the accuracy of causal modeling. Based on this, by combining the location information of retirement events in the graph, structural distance constraints, and pre-retirement temporal fluctuation characteristics, a pruning mechanism conforming to the thermal diffusion and electrochemical degradation characteristics of power batteries is designed to extract highly correlated local subgraphs from the large causal graph. Furthermore, this invention proposes a joint causal scoring method based on node embedding offset direction, hierarchical structure weight, and path cumulative significance to identify candidate causal paths and key nodes that may lead to retirement, ensuring that path identification takes into account embedding trends, structural levels, and physical propagation patterns. Finally, this invention constructs a counterfactual intervention mechanism oriented towards graph structures to simulate the retirement embedding impact "if the node state does not change abnormally" without altering the original causal topology, to determine whether a node constitutes a necessary cause of the retirement event and outputs a structured explanation result. The four-stage system constructed by this invention covers the entire process of modeling, localization, identification, and verification. It can not only reconstruct the real causal chain formed by retirement from multi-scale structures, but also output interpretable conclusions with engineering credibility. It overcomes the problems of weak causal expression ability, ambiguous path tracing, and unverifiable explanation in the prior art, and realizes the systematization and refinement of the analysis of the causes of power battery retirement. Attached Figure Description

[0015] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the steps of the method for tracing the source of multi-scale failures and explaining the reasons for retirement of power batteries based on causal graph neural networks, as proposed in this invention. Detailed Implementation

[0017] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0018] In one or more embodiments, such as Figure 1 As shown, a method for tracing the causes of multi-scale failures and explaining the reasons for retirement of power batteries based on causal graph neural networks is disclosed. The method includes the following steps: S1. Collect cell temperature sequence, cell voltage sequence, module temperature difference, module discharge rate, and module internal resistance change rate from the battery management system, combine them into node attributes, and construct directed edges based on physical hierarchy connections to form a multi-scale causal graph; at the same time, use a graph neural network to perform causal propagation and aggregation on the node states of the multi-scale causal graph to generate embedding vectors for each node.

[0019] Specifically, this step involves constructing a multi-scale nested causal graph neural network model for a power battery system. This model maps the system's operating state variables to graph nodes and constructs directed edges in the graph structure based on the structural connections between cells, modules, and packs. A graph neural network module is embedded in this graph structure to achieve causal propagation and aggregation of node states, forming a unified causal feature representation. The modeling output of this step will support subsequent decommissioned node location, causal path identification, and causal verification.

[0020] All input data comes from the battery management system acquisition module and the system structure configuration table. The state variables used include five categories: (1) cell temperature sequence (2) Cell voltage sequence (3) Module temperature difference (4) Module discharge rate The ratio of the sampling current to the rated capacity of the module is used to determine the following: (5) the rate of change of the module's internal resistance. The value is obtained by periodically injecting small currents and collecting the voltage response. Among the variables, Indicates the cell number. Indicates the module number. Indicates time.

[0021] Furthermore, the graph structure connection information is obtained through the BMS configuration file, specifying the module to which each cell belongs, the connection methods between modules, and the thermal channel structure. Graph structure construction. , where the set of nodes Includes all state variables, edge set It includes physical layer edges and variable causal edges. Each variable node will serve as the input node of the graph neural network, and the edge directions are determined by physical connections and expert experience.

[0022] A graph neural network module is deployed on a graph structure, with the first layer implementing information aggregation of adjacent nodes. The first layer of embedding is represented as: ; in, For nodes The set of adjacent nodes, For the edge Causal weights This is the first-level linear transformation matrix. For nodes Current time input features, The activation function is LeakyReLU. Each node's input features include the raw variables collected at the current time, such as temperature and voltage.

[0023] The second layer introduces time window features. A gated time series encoder is used to embed the historical state of each node. The final embedding is: ; in, This is the second-level linear transformation matrix. This represents a vector concatenation operation. This represents the historical state embedding function, with nodes as input. in the past The module outputs historical trend characteristics based on the state sequence at each time point. It is implemented using a single-layer gated recurrent network to extract the dynamic evolution characteristics of variables such as temperature and voltage.

[0024] Before being fed into the graph neural network, all variables are normalized within their respective modules. Taking temperature as an example, the mean and standard deviation of the temperature for each cell are calculated within a 5-minute sliding time window, and the sequence is normalized to zero mean and unit variance to avoid inconsistencies in the units or numerical ranges of different variables affecting network training and inference. After model training is complete, each node receives an embedding vector. This vector encodes the node's current observation state, historical trends, and causal influences from neighboring nodes. (Graph structure) and embedded collections This forms the input basis for the next stage of decommissioning node identification and causal path inversion.

[0025] S2. Obtain the decommissioning event that has occurred, and locate the corresponding decommissioning node in the multi-scale causal graph; based on the structural hierarchy distance between decommissioning nodes, the state fluctuation characteristics within the time window, and the similarity of the node embedding vectors, prune the multi-scale causal graph, and extract the local causal subgraphs related to the decommissioning event and their node embedding sets.

[0026] Specifically, this step, based on the multi-scale causal graph neural network model constructed in the first step, maps the actual decommissioning events that occur during the operation of power batteries from "real decommissioning phenomena" to "graph structure inference entry points." Through structural pruning and time window filtering, a local causal subgraph is constructed that covers key causal paths without introducing interference noise. This step essentially compresses the large causal graph model into a "causally significant region" surrounding the decommissioning event, and its input depends entirely on the output of the first step: the nested causal graph. and node embedding set Both are indispensable and will be used in this step. Since power battery retirement often occurs at the pack level, and the causes of retirement often stem from hidden cumulative changes at the module and cell levels, the pruning mechanism in this step must simultaneously consider three dimensions: structural hierarchy, temporal evolution trajectory, and operational fluctuation intensity. Based on these dimensions, a local subgraph with the characteristics of a power battery retirement scenario will be constructed. .

[0027] The input comes from the causal graph structure in the first step. Its node set Including cell temperature Cell voltage Module temperature difference Module internal resistance change rate With discharge rate Each node has an embedding vector obtained through a graph neural network. edge set This characterizes the causal relationship between nodes; for example, cell temperature points to module temperature difference, and module temperature difference points to Pack-level temperature rise risk. Actual decommissioning events are triggered by the BMS system, including various forms such as capacity degradation exceeding limits, sustained high temperature, and abnormal voltage drop. In these cases, the Pack node that experienced decommissioning can be directly located based on the system logs, denoted as... In the nested graph structure, it corresponds to a Pack-level state node, which has already been embedded in the first step. .

[0028] Furthermore, in order to construct a system with This step involves considering the set of nodes with potential causal relationships. It requires integrating factors such as the structural nesting relationships of the power battery, the evolution of temperature / voltage over time, and short-term surges in fluctuations before decommissioning into the pruning judgment. First, following the structural hierarchy of Pack→Module→Cell, a structural radius constraint is applied to all nodes from top to bottom. The structural radius is defined using the shortest path distance in a graph structure. Furthermore, considering the localized heat chain diffusion characteristic of power batteries during actual operation, a simple distance threshold is not used. Instead, a structural weighting factor with characteristics of power battery retirement scenarios is designed. That is, when a node is a module temperature difference node or a cell temperature node, its correlation with decommissioning is higher. A candidate structure set is formed by combining structural distance and weighting factors, and its selection formula is as follows: ; in, This represents the shortest path distance between nodes in the graph; The upper limit of the structural radius is generally taken as... ; As the structural weighting factor, when When it belongs to the cell temperature node or the module temperature difference node, take When it belongs to the module electrical parameters, take If it belongs to a Pack node, then take .because The smaller the value, the more likely it is to be an upstream cause in a real system. Therefore, this design is equivalent to automatically expanding the structural sampling range of thermal chain-related variables while narrowing the sampling range of Pack-level variables, making the structural pruning more in line with the actual characteristics of the power battery before retirement, which is "spreading from local cells".

[0029] Furthermore, after structural pruning, it is necessary to further determine the candidate nodes based on the time dimension. The decommissioning of power batteries is a phenomenon that accumulates over time, but the key triggers often occur before decommissioning. Hours to Sudden fluctuations occurred within hours, such as localized temperature spikes in some cells or short-term increases in module internal resistance. Therefore, this step incorporates a time-dependent scoring system to prioritize nodes with abnormal fluctuations during time pruning. Node Definition During the retirement window Time activity level is: ; in, For nodes In time The observed values, This represents the average value of the node during its normal operation over the past three days. This represents the number of sampling points within the window. and They are nodes The embedding vector obtained from step one for the retired node; Weights are embedded for similarity terms, used to measure the similarity between the dynamic features of a node and the features of a retired node. When... When the size is large, more emphasis is placed on the embedding similarity between nodes and retired nodes, so that the pruning region can accurately capture those nodes that may "indirectly lead to retirement".

[0030] The first part of this formula reflects the fluctuation intensity of the variable before retirement. The second part uses the graph neural network embedding obtained in the first step for similarity detection, enabling pruning to incorporate deep causal features in the power battery scenario, demonstrating the inventiveness of this patent. The final time pruning result is to retain... The set of nodes that exceed the threshold is denoted as .

[0031] Through filtering based on both structural and temporal dimensions, the final set of nodes in the local subgraph is: edge set Original image All connected nodes in A restricted subset. Node embedding set. The embedded representation of the local graph will serve as input for the next step, causal path identification. The output of this step is: the local causal subgraph. Its corresponding node embedding set This result identifies regions with strong structural and temporal correlation to decommissioning events extracted from large-scale, multi-scale causal graphs. This ensures that causal path reasoning is based on real system behavior rather than blindly searching the entire graph, thus significantly improving the accuracy and computational efficiency of causal identification in subsequent steps.

[0032] S3. Based on the local causal subgraph and its node embedding set, and combining the node embedding vector offset direction, structural hierarchy weight and path accumulation features, identify candidate causal paths from potential cause nodes to decommissioned nodes, and extract key node sets.

[0033] Specifically, this step builds upon the local causal subgraph constructed in the previous stage. and its node embedding set Building upon this foundation, potential causal paths to decommissioning are identified through the propagation results of graph neural networks, and key nodes that may directly lead to decommissioning events are further extracted. This step acts as a "causal significant pathway amplifier" within the entire patent chain. Through joint analysis of structure, embedding, and temporal evolution, a large number of candidate nodes are filtered into a small number of core nodes with genuine causal influence, and a set of candidate paths with structural continuity is output, laying the foundation for the next step of counterfactual verification. This step relies entirely on the output of the second step; therefore, all variables, including... , , , and retirement node Both are used explicitly.

[0034] In the specific operation process, the first step of path recognition is to start from the decommissioned nodes based on the graph structure. The algorithm reverse-engineers all possible upstream nodes. However, the retirement of power batteries is often not caused by a single variable abrupt change, but rather by the gradual accumulation of multiple variables such as temperature, voltage, and internal resistance of the underlying cells over several hours to days. Therefore, the algorithm designed in this step is not only based on structural topology, but also needs to combine the directional characteristics of node embedding to identify nodes that exhibit a "near-retirement trend" in the embedding space, i.e., nodes whose embedding vectors continuously drift from the normal region to the abnormal region. This embedding drift has clear physical significance in the power battery retirement scenario: phenomena such as long-term high node temperature, long-term low voltage, and continuous increase in internal resistance will all cause significant shifts in the embedding space. Therefore, this step proposes a "structure-embedded joint directional causal scoring" designed for the power battery scenario to quantify the causal significance of nodes on retirement events and further guide path selection.

[0035] Therefore, this step first constructs a causal score for the decommissioning direction for each node. This score consists of three parts: the offset magnitude of the embedding space, the embedding projection along the decommissioning direction, and a structural regularization term incorporating the characteristics of the power battery. The comprehensive score design is as follows: ; in, For nodes Embedding; Embedding for decommissioning nodes; The direction of embedding changes of retired nodes within past time windows is obtained by performing a linear fit on their historical embedding sequences, which can reflect the trend of their evolution from "healthy" to "abnormal". For node type weights, a larger coefficient is selected for the cell temperature node to emphasize the causal role of the thermal chain in decommissioning. To embed the directional weights, nodes that are embedded along the retirement trend direction will receive higher scores, reflecting the strong trend anomalies that occur before the retirement of power batteries. This is a structural regularization term. When a node is located at the module or cell level and has multiple physical connections with decommissioned nodes, this term takes a larger value to reflect the credibility of the structural path. The weights are the structural regularization terms.

[0036] The formula consists of three terms: the first term measures the node offset, the second term measures the projection along the decommissioning direction, and the third term reflects the physical relationships of the multi-scale structures within the battery. This formula is specifically designed for power battery decommissioning scenarios, combining the physical evolution laws and structural propagation characteristics of the decommissioning process, thus making it highly targeted.

[0037] To elevate node scores to path scores, this step starts with the decommissioned node and performs a back-diffusion search across all possible upstream structural paths. For any path... The overall significance is defined as the weighted cumulative average of the node significances. Since the retirement mechanism of power batteries typically exhibits the characteristic of "cumulative amplification of underlying anomalies," an additional amplification factor for deeper nodes is added to the path scoring design, making the scores of cell-level nodes in the path contribute more significantly to the overall score. The path scoring is as follows: ; in, The node-level weights are set as follows: larger values ​​(e.g., 2) for cell-level nodes, medium values ​​(e.g., 1) for module-level nodes, and smaller values ​​(e.g., 0.5) for pack-level nodes, reflecting the engineering pattern that the retirement path of power batteries usually spreads from the bottom layer. Score the causal relationship of the path; Indicates the causal significance of the path.

[0038] This scoring method emphasizes both node saliency and structural hierarchy, causing the algorithm to favor paths exhibiting typical power battery retirement patterns, such as "individual cells continuously heating up → module temperature difference widening → pack triggering retirement." Through the two scoring formulas mentioned above, this step, in the local graph... Perform a path diffusion search to generate a set of candidate paths sorted by significance from high to low. And select the set of key nodes with scores higher than a threshold from all nodes. In real-world scenarios involving power batteries, it's common to encounter situations where certain paths are structurally sound but lack a clear trend in their embedding, or where embedding changes significantly but the structural paths are weak. The joint scoring mechanism in this step effectively eliminates these two types of false paths, ensuring that the final causal candidate paths better reflect the operational characteristics of real systems.

[0039] S4. Perform counterfactual intervention simulation on the key causal nodes, assess the impact on the decommissioning nodes if the node's state is not abnormal, determine whether it is a necessary reason for decommissioning, and generate structured explanation results.

[0040] Specifically, this step aims to build upon the set of candidate causal paths identified in the previous stage. With key node set While maintaining the graph structure With node embedding collection Under the premise of no change, counterfactual verification is performed on each potential causal node. By simulating the logic of "whether the decommissioning event would still be triggered if the state of this node did not change abnormally," the necessity of decommissioning is assessed, and a structured and feasible causal explanation is generated accordingly. This step is a key transitional stage from modeling to explanation in the entire scheme, requiring the simultaneous maintenance of causal rigor, operational feasibility, and engineering understandability of the conclusions.

[0041] In terms of data usage, the input comes entirely from the output of the previous step: the path set. Each path in the graph consists of a sequence of nodes, and each node corresponds to a graph structure. A state variable node, the node type includes cell temperature node. Cell voltage node Module internal resistance node Module temperature difference node Each node has an embedded representation. The embedding, obtained in the first step through training a graph neural network, includes the state features of its current time slice, the aggregation results of neighboring node information, and historical time series trend features. Specifically, the embedding, which incorporates dynamic temporal features, is constructed by using a single-layer gated recurrent network (such as a GRU unit) to analyze the nearest neighbors of each node. The sequence is generated by encoding the input at each time point and then concatenating the sequences.

[0042] Furthermore, in order to perform counterfactual verification, it is necessary to construct an alternative node embedding representation in the "healthy state". to replace the original embedding Perform a simulation. In a real system, There are two ways to generate the embedding vector: (1) Extract it from the historical stable operation window of the node, for example, the time period when the node state does not fluctuate within 48 to 72 hours before the retirement, and obtain the time average representation of the embedding vector by averaging through sliding window; (2) Find nodes with similar structural positions from similar nodes (such as healthy modules or cells of the same type), and take the average of their embedding vectors as the replacement representation of the current node. In the actual operation of the power battery pack, multiple modules usually have structural symmetry and thermal management consistency, so they can be horizontally compared and embedded within a vehicle to complete the replacement generation.

[0043] When performing counterfactual simulations, the embedded representation of a single node is changed from... Replace with The remaining nodes remain unchanged. Then, without altering the original graph structure... Under the premise of re-exercising the graph neural network, the decommissioned nodes are obtained. Embedded representation under this intervention And then its actual embedding By creating a counterfactual effect: ; in Represents a node The counterfactual impact of a decommissioning node is indicated by a higher value, signifying a more significant influence of the node's current state on the decommissioning outcome. Intervention in the node's state leads to a noticeable change in decommissioning characteristics, indicating stronger causality. This indicator balances physical operability (intervention only in the node's state) with the fidelity of the model's causal propagation mechanism, making it suitable for complex industrial systems with high variable coupling, such as power batteries.

[0044] To avoid oversimplifying the determination of causal necessity based solely on a single influence value, a confidence level adjustment term was further designed. Used to measure at the node After intervention, is it possible for the impact to be compensated by other paths? This value consists of two parts: first, the node's position in the path set; and second, whether the impact can be compensated by other paths. The co-occurrence frequency in the data is considered; if a node appears only in one path, its substitutability is low. Secondly, the stability of the edges between a node and its direct upstream nodes is assessed, including whether the nodes repeatedly appear across multiple historical windows and whether the edge weights remain consistent. This item is mapped to a standardized score. The interval. The final causal necessity score, combining impact and substitutability, is as follows: ; in To adjust the weights, they are generally set to arrive This formula balances the impact of a single node with the causal weakening effect of redundant paths in the system structure. It eliminates a large number of "covariate paths" in the complex structure of a power battery, making the final explanatory node more logically rigorous, emphasizing its essentiality. For example, in a real battery pack, if multiple cells are at high temperatures, but only one cell has an irreplaceable impact (e.g., coinciding with a blockage in the cooling pipe), then that node... It will score significantly higher than other cell nodes.

[0045] In actual implementation, the system will analyze the set of key nodes. Each node in the process performs a counterfactual simulation one by one and generates... The values ​​are retained, while information such as the state before and after the embedding substitution, path number, and original embedding offset are preserved for subsequent structured output descriptions. Finally, all... Value higher than the set threshold The nodes are defined as counterfactual necessary nodes, forming a set. .in The value can be set according to the system's requirements for the clarity of the cause; a typical value can be taken as follows. to .

[0046] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0047] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0048] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0049] The various illustrative logic blocks, modules, and circuits described in conjunction with the embodiments disclosed herein can be implemented or performed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. The general-purpose processor may be a microprocessor, but in alternatives, it may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration.

[0050] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of both. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor such that the processor can read and write information to / from the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and storage medium may reside as discrete components in the user terminal.

[0051] In one or more exemplary embodiments, the described functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functionality may be stored or transmitted as one or more instructions or code on or through a computer-readable medium. A computer-readable medium includes both computer storage media and communication media, encompassing any medium that facilitates the transfer of a computer program from one location to another. A storage medium may be any available medium accessible to a computer. By way of example and not limitation, such a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and is accessible to a computer. Any connection is also legitimately referred to as a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of a medium. As used in this article, disk and disc include compact discs (CDs), laser discs, optical discs, digital multi-purpose discs (DVDs), floppy disks, and Blu-ray discs. Disks typically reproduce data magnetically, while discs reproduce data optically using lasers. Combinations of these should also be included within the scope of computer-readable media.

[0052] The prior description of this disclosure is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not intended to be limited to the examples and designs described herein, but should be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for tracing the causes of multi-scale failures and explaining the reasons for retirement of power batteries based on causal graph neural networks, characterized in that, Includes the following steps: S1. Collect cell temperature sequence, cell voltage sequence, module temperature difference, module discharge rate, and module internal resistance change rate from the battery management system, combine them into node attributes, and construct directed edges by combining physical hierarchy connection relationships to form a multi-scale causal graph; at the same time, use graph neural network to perform causal propagation and aggregation on the node states of the multi-scale causal graph to generate the embedding vector of each node. S2. Obtain the decommissioning event that has occurred, and locate the corresponding decommissioning node in the multi-scale causal graph; based on the structural hierarchy distance between decommissioning nodes, the state fluctuation characteristics within the time window, and the similarity of the node embedding vectors, prune the multi-scale causal graph, and extract the local causal subgraphs related to the decommissioning event and their node embedding sets. S3. Based on the local causal subgraph and its node embedding set, combined with the node embedding vector offset direction, structural hierarchy weight and path accumulation features, identify candidate causal paths from potential cause nodes to decommissioned nodes, and extract key node sets. S4. Perform counterfactual intervention simulation on the key causal nodes, assess the impact on the decommissioning nodes if the node's state is not abnormal, determine whether it is a necessary reason for decommissioning, and generate structured explanation results.

2. The method for tracing the source of multi-scale failures and explaining the causes of retirement of power batteries based on causal graph neural networks according to claim 1, characterized in that, The graph neural network is specifically described as follows: The first layer involves acquiring nodes as input nodes to the graph neural network, aggregating information from adjacent nodes, and obtaining the first layer embedding of the nodes. The second layer, based on the first layer embedding of the nodes, introduces time window features and uses a gated time series encoder to embed the historical state of each node, thereby obtaining the embedding vector of each node.

3. The method for tracing the source of multi-scale failures and explaining the causes of retirement of power batteries based on causal graph neural networks according to claim 2, characterized in that, The gated time series encoder is implemented using a single-layer gated recurrent network, with the input being the node's past performance. The state sequence at each time point is output as historical trend features.

4. The method for tracing the source of multi-scale failures and explaining the causes of retirement of power batteries based on causal graph neural networks according to claim 1, characterized in that, The pruning process of the multi-scale cause-effect graph specifically includes: Structure filtering is performed based on the graph structure distance between nodes and decommissioned nodes and the node type weight, and time filtering is performed based on the intensity of the state fluctuation of nodes within the time window before decommissioning and the similarity of their embedding vectors with those of decommissioned nodes.

5. The method for tracing the source of multi-scale failures and explaining the causes of retirement of power batteries based on causal graph neural networks according to claim 4, characterized in that, The structural radius of the graph structure distance is defined by the shortest path distance of the graph structure. At the same time, the characteristic of local thermal chain diffusion in the actual operation of the power battery is taken into account. That is, when the node is the module temperature difference node or the cell temperature node, its correlation weight with retirement is higher.

6. The method for tracing the source of multi-scale failures and explaining the causes of retirement of power batteries based on causal graph neural networks according to claim 1, characterized in that, In S3, when identifying candidate causal paths, a node score is performed on the causal significance of each node. The node score integrates the offset of the node embedding vector relative to the decommissioned node, the projection along the decommissioning evolution trend direction, and the structural weight of the node's level.

7. The method for tracing the source of multi-scale failures and explaining the causes of retirement of power batteries based on causal graph neural networks according to claim 6, characterized in that, In S3, the candidate causal path is calculated based on the node score, specifically as follows: Starting from the decommissioning node, a back-diffusion search is performed on all possible upstream structural paths. For any path, its causal significance is defined as the weighted cumulative average of the causal significance of the nodes. Different weights are assigned to nodes at different levels in the path to increase the additional amplification factor for deeper nodes, so that the score of the cell-level nodes in the path has a more significant contribution to the overall score, thus obtaining the causal significance of the path. Path diffusion search is performed on the local causal subgraph to generate a set of candidate paths sorted from high to low significance, and a set of key nodes with scores higher than a threshold is selected from all nodes.

8. The method for tracing the source of multi-scale failures and explaining the causes of retirement of power batteries based on causal graph neural networks according to claim 7, characterized in that, The cell-level nodes have a higher weight than the module-level and pack-level nodes.

9. The method for tracing the source of multi-scale failures and explaining the causes of retirement of power batteries based on causal graph neural networks according to claim 1, characterized in that, In S4, the counterfactual intervention simulation is achieved by replacing the embedding vector of the key node with a preset health status reference vector and re-propagating the graph neural network to calculate the degree of change in the embedding vector of the retired node before and after the intervention.

10. The method for tracing the source of multi-scale failures and explaining the causes of retirement of power batteries based on causal graph neural networks according to claim 9, characterized in that, In S4, when determining whether a node is a necessary cause, the co-occurrence frequency of the node in the candidate path or the stability of the edges connected to it are also considered to correct the score of the influence of a single node.

Images