Early warning intelligent positioning method and system for internal short circuit of lithium battery under multi-data fusion

Abstract

The application discloses a kind of early warning intelligent positioning method and system under lithium battery internal short circuit early warning of multiple data fusion.The method comprises: creating by bridge node and terminal node the directed topology library of lithium battery abnormal image feature;Obtain the multivariate vector expression of target lithium battery current image;Based on the vector expression, search the multiple section link to each terminal node in the directed topology library, calculate the total link cost of each link, and select the minimum cost link as the cost optimal chain;According to the cost optimal chain, determine the target terminal node and calculate the confidence, when the confidence is higher than threshold, issue intelligent early warning;Further judge whether there is significant mutation in the cost distribution in link, if there is, the current multivariate vector expression is constructed as new supplement unit, as independent node is inserted into directed topology library, complete the adaptive evolution of atlas structure.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault early warning, specifically to an intelligent location method and system for early warning of internal short circuits in lithium batteries based on multi-data fusion. Background Technology

[0002] With the rapid development of the new energy industry, lithium-ion batteries are widely used in electric vehicles, energy storage systems, and consumer electronics due to their high energy density and long cycle life. However, lithium batteries have a certain probability of internal short circuits during manufacturing, transportation, and use. Early identification of soft short circuits, in particular, is difficult, and failure to provide timely warnings can easily lead to safety accidents such as thermal runaway, fires, and explosions. Therefore, how to achieve early warning and accurate location of internal short circuits in lithium batteries has become a key technical problem that urgently needs to be solved in the field of lithium battery safety management.

[0003] Existing lithium battery fault detection methods mainly rely on the following means: (1) abnormal monitoring of electrical parameters such as voltage, current, and temperature. However, such methods can only detect the fault after a short circuit has caused significant impact, and do not have the ability to detect it early. (2) thermal image analysis or ultrasonic tomography technology. These methods can sense the surface and internal temperature or structural abnormalities through infrared or acoustic means. Although they can provide some visualization information, they often have problems such as large image morphology diversity, noise interference, and uncertain defect expression. This leads to the model training relying on idealized defect reference images. However, the on-site images have significant differences from the reference samples in terms of color, texture, and distribution, which can easily cause misjudgment or missed judgment.

[0004] Therefore, how to better perform early fault warning and simultaneously optimize the reference sample when the reference sample is not comprehensive is an urgent problem to be solved. Summary of the Invention

[0005] This disclosure provides an intelligent location method for early warning of internal short circuits in lithium batteries based on multi-data fusion.

[0006] In a first aspect, this disclosure provides an intelligent location method for early warning of internal short circuits in lithium batteries based on multi-data fusion, characterized by comprising:

[0007] A lithium battery anomaly image feature orientation topology library is created, consisting of bridging nodes and terminal nodes. The bridging nodes are used to describe the intermediate forms of image morphology during the anomaly evolution process, while the terminal nodes correspond to the specific anomaly types that have actually occurred.

[0008] Connection relationships are used to describe the directional dependencies and transition paths between different nodes during the anomalous evolution process, and the differences in image features between nodes are quantified by allocating connection costs;

[0009] Obtain the multivariate vector representation of the current image of the target lithium battery; find the cost-optimal chain from the multivariate vector representation to the terminal node in the directional topology library;

[0010] The target terminal node is determined based on the cost-optimal chain, and an intelligent early warning is issued;

[0011] Optimize the directional topology library based on the multivariate vector representation and the cost-optimal chain;

[0012] The optimized directional topology library contains a local subgraph constructed based on the optimal chain node set and one-hop neighborhood, as well as a height label based on the local subgraph that has passed consistency verification.

[0013] Furthermore, the step of finding the cost-optimal chain from the multivariate vector representation to the terminal node in the directional topology library further includes:

[0014] Obtain the multi-segment links from the multi-vector representation to all terminal nodes, and sum the link costs of the multi-segment links through each node;

[0015] The multi-section link with the minimum cost after summation is taken as the cost-optimal link.

[0016] Furthermore, the step of determining the target terminal node based on the cost-optimal chain and issuing an intelligent early warning also includes:

[0017] The terminal node at the end of the cost-optimal chain is taken as the target terminal node;

[0018] Calculate the confidence level between the target terminal node and the multivariate vector representation;

[0019] If the confidence level is higher than the threshold, then intelligent early warning is performed based on the anomaly type corresponding to the target terminal node and the cost-optimal chain.

[0020] Furthermore, optimizing the directional topology library based on the multi-vector representation and the cost-optimal chain further includes:

[0021] Determine whether the link cost between adjacent units in the cost-optimal chain of the multivariate vector representation exceeds a preset difference threshold;

[0022] If the threshold is exceeded, a new supplementary unit containing the multivariate vector expression is constructed and inserted into the directional topology library as an independent node.

[0023] Furthermore, the method also includes:

[0024] Based on the directional topology library, multivariate vector representation, and cost-optimal chain, a capacity-based local subgraph and capacity matrix are constructed to determine the initial height label and starting point set of the current sample as the virtual source point.

[0025] Initial residual network parameters are generated based on the initial height labels and starting point set, and then the residual network and traffic matrix after termination are iteratively updated through Push and Relabel to determine the consistency verification index.

[0026] Obtain the candidate path cost set and the minimum total cost for each terminal in the path search, and perform linear normalization on the candidate path cost set and the minimum total cost to calculate the confidence score and suboptimal competitiveness of the candidate path cost and the minimum total cost.

[0027] Furthermore, the construction of a capacity-optimized local subgraph and capacity matrix based on the directional topology library, multi-vector representation, and cost-optimal chain to determine the initial height label and starting point set of the current sample as a virtual source point includes:

[0028] The optimal chain node set and one-hop neighborhood are defined based on the directional topology library, and the union of the optimal chain node set and one-hop neighborhood is used as the local node set of the local subgraph. At the same time, the local edge set of the local subgraph is determined according to the one-hop neighborhood.

[0029] The feature difference degree between different nodes is defined in the Euclidean distance normalization space. A capacity matrix and a difference degree matrix are constructed through a continuously monotonically decaying capacity mapping. The virtual source point is then connected to a bridging node that meets a set threshold to obtain an initialized height label vector.

[0030] Furthermore, the method also includes:

[0031] Based on the residual network and flow matrix after termination, the source-side reachable set and the sink-side unreachable set are determined to generate the cut edge set and the main chain saturated edge set. The cut edge set consists of all edges between the source-side reachable set and the sink-side unreachable set.

[0032] The spatial alignment information of thermal images and ultrasound in a unified coordinate system and a pre-established node region template library are obtained. A weight map is generated based on the cut edge set and the main chain saturated edge set. The coordinates of thermal images and ultrasound are unified onto the image grid according to the weight map and the spatial alignment information to obtain the risk corridor region.

[0033] Secondly, this disclosure provides an intelligent positioning system for early warning of internal short circuits in lithium batteries based on multi-data fusion, the system being used to execute the method described in the first aspect, the system comprising:

[0034] The topology library server is used to create a lithium battery anomaly image feature orientation topology library composed of bridging nodes and terminal nodes. Bridging nodes are used to describe the intermediate form of image morphology in the anomaly evolution process, while terminal nodes correspond to the specific anomaly types that have actually occurred. Connection relationships are used to describe the directional dependence and transition path between different nodes in the anomaly evolution process, and to quantify the differences in image features between nodes by allocating connection costs.

[0035] The image parsing unit is used to obtain the multi-vector representation of the current image of the target lithium battery;

[0036] An optimal chain calculation unit is used to find the cost-optimal chain from the multivariate vector expression to the terminal node in the directional topology library;

[0037] The intelligent early warning unit is used to determine the target terminal node based on the cost-optimal chain and issue an intelligent early warning.

[0038] The topology library server is also used to optimize the directional topology library based on the multivariate vector representation and the cost-optimal chain.

[0039] Furthermore, the system also includes a capacity matrix generation module, a residual flow update module, and an evaluation module;

[0040] The capacity matrix generation module is used to construct a capacity-based local subgraph and capacity matrix based on the directional topology library, multi-vector representation, and cost-optimal chain, so as to determine the initial height label and starting point set of the current sample as a virtual source point;

[0041] The residual traffic update module is used to generate initial residual network parameters based on the initial height label and starting point set, and iteratively update the residual network and traffic matrix after termination through Push and Relabel to determine the consistency verification index.

[0042] The evaluation module is used to obtain the candidate path cost set and the minimum total cost for each terminal in the path search, and to linearly normalize the candidate path cost set and the minimum total cost to calculate the confidence score and suboptimal competitiveness of the candidate path cost and the minimum total cost.

[0043] Furthermore, the system also includes a detection module and an identification module;

[0044] The detection module is used to determine the source-side reachable set and the sink-side unreachable set based on the residual network and traffic matrix after termination, so as to generate the cut edge set and the main chain saturation edge set. The cut edge set is composed of all edges between the source-side reachable set and the sink-side unreachable set.

[0045] The identification module is used to acquire the spatial alignment information of thermal images and ultrasound in a unified coordinate system and a pre-established node region template library, and generate a weight map based on the cut edge set and the main chain saturated edge set, so as to unify the coordinates of thermal images and ultrasound onto the image grid according to the weight map and the spatial alignment information to obtain the risk corridor region.

[0046] The beneficial effects of this disclosure are that, compared with the prior art, this disclosure has the following advantages:

[0047] 1) By constructing a directional topology library and integrating the multi-vector representation of the current image of the target lithium battery, the system can identify and judge abnormal features before they reach a severe level through multi-modal information (such as thermal imaging and ultrasound), which has a stronger early perception capability than traditional voltage and current monitoring methods.

[0048] 2) This invention proposes using a "cost-optimal chain" as a multi-vector representation of the reasoning channels leading to each terminal node. This not only determines the most matching abnormal terminal node but can also be understood as the most likely direction of fault evolution for the target lithium battery in its current state, thus achieving an integrated design for early warning and path prediction. This link construction mechanism integrates historical topology and current image representation, featuring strong directionality and clear contextual relationships, providing structural guarantees for early fault location and trend analysis.

[0049] 3) By constructing a lightweight flow network near the cost-optimal chain and employing label relabeling for consistency verification, the system can output three key indicators while maintaining the original decision: evidence score, suboptimal competitiveness, and location stability. The evidence score helps assess the reliability of the main path's conclusion, the suboptimal competitiveness reveals the degree of interference from other potential paths, and location stability enhances spatial interpretability through risk corridor output. In this way, the early warning system is upgraded from a "single result" to a "result + credibility + risk interpretation" system, which not only enhances the reliability of early warnings but also improves engineers' decision-making confidence and review efficiency under complex operating conditions. Ultimately, it reduces false alarms and false negatives, ensuring that early warning results better reflect the actual operating state of the battery. Attached Figure Description

[0050] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0051] Figure 1 This is a schematic diagram of an intelligent positioning method for early warning of internal short circuit in lithium batteries under multi-data fusion, provided in an embodiment of this disclosure.

[0052] Figure 2 This is a schematic diagram of an intelligent positioning system for early warning of internal short circuit in lithium batteries under multi-data fusion, provided in an embodiment of this disclosure.

[0053] The accompanying drawings have illustrated specific embodiments of this disclosure, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this disclosure to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0054] The present disclosure will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present disclosure and should not be used to limit the scope of protection of the present disclosure.

[0055] Figure 1 This disclosure provides a schematic diagram of an intelligent location method for early warning of internal short circuits in lithium batteries based on multi-data fusion. See also... Figure 1 The following is a detailed discussion of each step in conjunction with this embodiment.

[0056] S100. Create a lithium battery anomaly image feature orientation topology library composed of bridging nodes and terminal nodes.

[0057] In this embodiment, the lithium battery anomaly image feature directional topology library is used to characterize the image structural relationships of different types of short-circuit anomalies. Its basic structure consists of multiple bridging nodes and terminal nodes, organized through directional connections (i.e., directional links). Bridging nodes describe the intermediate forms of image morphology during the anomaly evolution process, while terminal nodes correspond to specific anomaly types that have actually occurred (such as tab short circuits, electrode contact, inner layer collapse, etc.).

[0058] In some embodiments, in the infrared thermographic images of pouch cells, tab short circuits typically manifest as localized high-temperature concentrations. The system collected samples showing a transitional state where the annular hotspot gradually shrinks before the short circuit forms. Clustering results revealed a class of samples whose thermal image vectors exhibited a high degree of consistency in the characteristic of "hotspots distributed in annular bands." These images differ from uniformly heated normal samples and do not reach the terminal state of concentrated high temperatures in the tab region. Therefore, the broken annular hotspot can be defined as a bridging node, its function being to characterize the intermediate morphology from linear hotspots gradually evolving into a tab short circuit.

[0059] The system integrates 1200 sets of anomalous image samples, including infrared thermograms and ultrasonic tomography. Each set of samples corresponds to a standardized image vector representation with a uniform 128-dimensional dimension. The image representation includes various statistical features such as heat distribution, texture direction, and structural attenuation to comprehensively describe potential anomalous features in the image. First, all image samples are grouped using an image clustering algorithm, obtaining 40 image representation centers with an initial cluster size of 40. The clustering results are then reviewed using manual annotation and empirical knowledge, ultimately selecting 18 bridging nodes and 6 terminal nodes as the initial topology. Each bridging node is connected to at least one forward node and one backward node, forming a clear path for the evolution of anomalous morphology. For example, bridging node B7 represents a "fractured ring hotspot," with its forward node being B3 (linear hotspot) and its backward-connected terminal node being T2 (tab short circuit).

[0060] In the topology, any two nodes may have directional connections. The system assigns a fixed connection cost to each pair of connected nodes. The connection relationship describes the directional dependence and transition path between different nodes during anomalous evolution, and quantifies the differences in image features between nodes by assigning connection costs. The value of this connection cost depends on the overall difference in the image representation between the two nodes, typically ranging from 0 to 1, with smaller values ​​indicating more similar image representations. For example, the connection cost from node B3 to B7 is 0.42; the connection cost from node B7 to terminal node T2 is 0.38; and the connection cost from B3 directly to T2 is 0.91. Therefore, the system prefers the path B3→B7→T2. The final directional topology library contains 24 nodes (18 bridging nodes and 6 terminal nodes) and 46 directional connections. The entire structure is stored in the graph database as an adjacency matrix, supporting efficient subsequent retrieval and path reasoning.

[0061] S200: Obtain the multivariate vector representation of the current image of the target lithium battery.

[0062] The system deploys an infrared thermal imaging camera and an ultrasonic transducer array, mounted above and to the side of the battery pack, to simultaneously acquire images of the battery's surface temperature and internal structure. The thermal images undergo non-uniformity correction and Gaussian filtering to remove edge noise and background thermal field interference; the ultrasonic images undergo time-domain averaging and dynamic gain adjustment to enhance the echo structure boundaries; the two types of images are aligned in spatial coordinates and uniformly cropped into a 256×256 region for use by the subsequent vector extraction module.

[0063] A unified image encoding module is used to encode multimodal images into a 128-dimensional vector representation with a unified structure. The system uses normalization processing to control all dimensions between 0 and 1, which facilitates direct comparison with node representations in the directional topology library.

[0064] Taking a real image acquisition task as an example, the system acquires data from a 3.2V, 35Ah pouch battery. The extracted current image vector fragment is as follows: the main hot spot area of ​​the thermal image appears in the second quadrant of the image, with a maximum temperature difference of 0.6°C; the ultrasound image shows discontinuous echo bands and blurred delay boundaries; the corresponding first 6-dimensional vector is: [0.28, 0.34, 0.12, 0.08, 0.43, 0.51, ...]. This 6-dimensional vector, i.e., multi-dimensional vector expression, is used to perform cost calculation and path construction with all terminal nodes in the directional topology library.

[0065] S300. Find the cost-optimal chain from the multivariate vector representation to the terminal node in the directional topology library.

[0066] In this embodiment, the goal of step S300 is to search for a minimum-cost link to the terminal node based on the multivariate vector representation of the current lithium battery image obtained in step S200, combined with the directional topology library constructed in step S100, as the most likely evolution direction of the current abnormal state.

[0067] For example, the directional topology library can contain 24 nodes, of which 18 are bridging nodes and 6 are terminal nodes. The connections between each pair of nodes are directional, and a preset connection cost is assigned. Specifically, the system first temporarily inserts the current multi-vector representation as a starting node into the graph structure. This temporary node does not establish connections with all nodes in the graph, but only with some unidirectional connections to bridging nodes. The establishment of these connections is based on the degree of vector difference between the current vector and each bridging node. When the difference between the current vector and a bridging node is less than a preset similarity threshold (e.g., 0.75), the system establishes a temporary connection and assigns a corresponding connection cost.

[0068] Subsequently, the system uses this temporary starting point as the source point and traverses all reachable terminal node paths. Each path consists of several bridging nodes and the final terminal node. The system accumulates the connection costs between each segment of the path to obtain the total cost of the entire path. Among all feasible paths, the path with the minimum total cost is determined as the cost-optimal chain. For example, the current vector establishes temporary connections with bridging nodes B3, B5, and B9, with costs of 0.28, 0.36, and 0.42, respectively. The system continues to calculate the paths from these nodes to each terminal node, forming the following candidate chains: Path 1: Current node → B3 → B7 → T2, total cost 0.28 + 0.29 + 0.33 = 0.90; Path 2: Current node → B5 → T3, total cost 0.36 + 0.41 = 0.77; Path 3: Current node → B9 → B14 → T5, total cost 0.42 + 0.36 + 0.22 = 1.00. Finally, the system selects the path with the lowest total cost (path 2) as the cost-optimal chain for the current image, and passes the terminal node T3 at the end of the path as the inference result corresponding to the current abnormal state to the next step for confidence judgment and intelligent early warning processing.

[0069] In early warning schemes for lithium battery short circuits, existing methods mainly rely on the "cost-optimal chain" to determine the target terminal node. While this can find the path with the minimum cost, it lacks quantitative verification of the reliability of this conclusion and cannot reflect the potential interference of suboptimal paths on the judgment. This single-result output method can easily lead users to question the credibility of the conclusion in the early stages of abnormal evolution, when image features are blurred, or when signal quality is insufficient. Users cannot judge how robust the result is or perceive whether there are competing suboptimal interpretations. If a local consistency verification and uncertainty quantification mechanism are not introduced on the basis of the cost-optimal chain, three defects will occur: First, the warning result remains at the conclusion of a single link, lacking credibility quantification, making it difficult for users to trust when faced with ambiguous signals; second, the system cannot reveal the competitive relationship of suboptimal paths. When multiple abnormal trends coexist, it may only output a superficial "optimal" one, thus ignoring potential high-risk evolution directions and posing a risk of missed detection; third, the spatial positioning interpretation is insufficient. Relying solely on single-link hotspot positioning makes it difficult to accurately present the abnormal diffusion area, leaving engineers without intuitive evidence during on-site investigations. These combined issues not only weaken the interpretability and reliability of the early warning system, but also limit its application value in actual battery management and safety monitoring.

[0070] In this example, the process includes local subgraph capacity construction and label initialization. First, based on the directional topology library, the current multivariate vector, the cost-optimal chain, and its terminal nodes, an optimal chain node set and a one-hop neighborhood are defined. The union of the optimal chain node set and the one-hop neighborhood is taken as the local node set. Simultaneously, a local edge set is constructed based on the directed edges between different nodes in the one-hop neighborhood. Finally, a local subgraph is constructed based on the local node set and the local edge set. Next, based on the constructed local subgraph, the node vector library, and the current multivariate vector, spatial normalization is performed using Euclidean distance. The feature dissimilarity between nodes and the dissimilarity between the current sample and the node are defined. Then, a capacity matrix and a dissimilarity matrix are constructed through a continuously monotonically decaying capacity mapping. Finally, the current sample is used as a virtual source point, and a starting point connection is established between it and a bridging node that satisfies that the elements of the capacity matrix are greater than 0. Height label initialization is performed to obtain a height label vector and a starting point set (containing the index corresponding to the virtual source point).

[0071] After local subgraph capacity building and label initialization, based on the capacity matrix, height label vector, starting point set, and local subgraph, the initial flow matrix is ​​aligned with the local edge set, residual capacity is calculated, and the inherited height label of the virtual source node is determined with a unit preflow of 1. Then, nodes meeting the set requirements are iteratively updated using push conditions and push volume, and simultaneously, nodes meeting the set conditions are iteratively updated using the Relabel rule, with "stopping after unit preflow reaches a certain terminal node" as the termination condition, to obtain the residual network and flow matrix after termination. Finally, based on the residual network and flow matrix after termination, combined with the candidate path cost set and the minimum total cost for each terminal in the path search, three indicators for consistency verification are determined: cut gap (measuring the primary-secondary gap using the minimum path cost of candidate terminals), saturation ratio (the proportion of edges "fully loaded" on the optimal chain), and residual throughput.

[0072] In this embodiment, the cut gap and minimum residual flux of the minimum total cost for each terminal in the candidate path cost set and path search are normalized to the interval [0,1] to obtain normalized cost and normalized gap. Then, based on the normalized cost and normalized gap, combined with the saturation ratio and residual flux, a confidence score is calculated to determine the suboptimal competitiveness. Finally, based on the residual network and flow matrix after termination, the source-side reachable set and sink-side unreachable set are determined to generate cut edge set and main chain saturated edge set. The cut edge set consists of all edges between the source-side reachable set and the sink-side unreachable set. Spatial alignment information of thermal imaging and ultrasound in a unified coordinate system and a pre-established node region template library are obtained. A weight graph is generated based on the cut edge set and main chain saturated edge set. The coordinates of thermal imaging and ultrasound are unified onto the image grid according to the weight graph and spatial alignment information to obtain the risk corridor region.

[0073] S400. Determine the target terminal node based on the cost-optimal chain and issue an intelligent warning.

[0074] In this embodiment, the objective of step S400 is to further identify the target terminal node corresponding to the end of the cost-optimal chain determined in step S300, and generate an intelligent early warning based on this. The cost-optimal chain not only represents the similarity path between the current image state and a certain terminal anomaly in the map, but also reflects the trend process at the structural level that the current state may evolve into a certain known anomaly.

[0075] The system first extracts the last node in the cost-optimal chain. This node is the terminal node that the current vector representation is most likely to point to in the graph structure, representing the known anomaly type that is closest to the current anomalous state on the evolution path. This terminal node is usually bound to anomaly labels (such as "bar short circuit", "structural collapse", "internal contact", etc.) and the number of historical sample support.

[0076] Subsequently, the system calculates the similarity score between the current multivariate vector and the target terminal node vector to determine whether the current state has sufficient credibility to be classified as an anomaly of this type. If the similarity score is higher than a set threshold (e.g., 0.82), an intelligent early warning process is triggered.

[0077] Upon triggering the alert, the system encapsulates the anomaly in a structured manner into an alert information package, which includes: anomaly type identifier, details of the optimal cost chain path, image timestamp, visual thermal image screenshot, and the battery module number. This information will be pushed to the upper-level management system or edge security gateway in real time, enabling on-duty personnel or automated control systems to respond quickly.

[0078] In this embodiment, the core of step S400 is that it not only uses the end node of the cost-optimal chain to identify the current anomaly type, but also uses the structural features of the entire chain to further infer the evolution stage of the current anomaly, thereby realizing intelligent early warning with trend judgment capability.

[0079] Specifically, the system first extracts the terminal node at the end of the cost-optimal chain and uses the anomaly type corresponding to that node as the result of this identification, such as "tab short circuit" or "inner layer rupture". However, at the same time, the system is not limited to the judgment of the terminal node itself, but further analyzes information such as the number of bridging nodes, path depth, and cost change gradient in the cost-optimal chain to assess the relative stage of the current anomaly state in the entire evolution path.

[0080] For example, if the cost-optimal chain reaches the terminal node with only one bridging node, and the cost of the initial connection is much lower than the cost of the final connection, it indicates that the current state is highly close to the terminal anomaly, and the warning level can be set as "severe anomaly, entering the final stage". Conversely, if the current vector needs to gradually transition through 3 to 4 bridging nodes to reach the terminal node, and the cost generally shows an increasing trend, it indicates that the current state is still in the early or middle stage of anomaly development. At this time, the system will automatically issue a "trend warning" to remind the user that the lithium battery has the potential risk of developing into a specific anomaly.

[0081] S500. Optimize the directional topology library based on the multivariate vector representation and the cost-optimal chain.

[0082] In this embodiment, the goal of step S500 is to dynamically optimize the structure of the directional topology library by using the current multivariate vector representation obtained in step S200 and the cost-optimal chain determined in steps S300 and S400, so as to enhance the map's ability to express and adapt to novel anomalies in the field.

[0083] Specifically, the system first determines whether there is a sudden change in the connection cost between adjacent units of the current multivariate vector in its cost-optimal chain. If the connection cost of a certain segment in the chain is significantly higher than that of other segments, and the difference in the characteristics of the connection nodes in that segment is greater than a preset structural difference threshold (e.g., similarity below 0.7), the system considers that the existing graph lacks transitional expression between that segment, resulting in an expression blind spot. In this case, the system constructs the current multivariate vector expression as a new supplementary unit and inserts it into the graph structure as an independent node. During the insertion process, the original connection relationships between nodes are not modified; instead, two new directional links are added, connecting the current vector node to the two original nodes before and after the segment, respectively. The assigned connection cost is calculated from the actual distance between the current vector and the two nodes.

[0084] For example, in a certain early warning process, the current cost-optimal chain is: B4→B11→T5, where the connection cost between B4 and B11 is 0.79, which is much higher than the average cost of 0.36 for other segments of the path. The system determines that there is a semantic gap in this segment, so it inserts the new node N1 represented by the current vector into this segment to form an updated chain: B4→N1→B11→T5, while retaining the original path.

[0085] In addition, the system will also associate and store metadata such as the current vector's label information, trigger time, image summary and confidence level into the newly added node to ensure that it can be identified, referenced and updated in subsequent path searches.

[0086] Furthermore, the intelligent positioning method for early warning of internal short circuit in lithium battery under multi-data fusion disclosed in this embodiment, for step S300, the step of finding the cost-optimal chain from the multi-vector expression to the terminal node in the directional topology library, further includes: obtaining the multi-segment links from the multi-vector expression to all terminal nodes, and summing the link costs of the multi-segment links through each node; and taking the multi-segment link with the minimum summation cost as the cost-optimal chain.

[0087] First, the system uses the current multi-vector representation as a temporary starting node and quickly locates the set of all target nodes identified as terminal nodes by consulting the structural index of the directional topology library. This set can be filtered by the "type label" field in the node meta-attributes. The system pre-sets all abnormal final-state nodes as terminal node identifiers.

[0088] Subsequently, starting from the current starting point, the system uses a breadth-first traversal algorithm or an A* heuristic search algorithm to construct all valid paths leading to each terminal node. To limit search complexity and avoid ineffective expansion, the system sets the maximum number of hops to no more than 5, meaning each path passes through a maximum of 4 bridging nodes and 1 terminal node.

[0089] For each path, the system extracts the connection cost between each adjacent node segment, performs a segment-by-segment summation operation, and obtains the cumulative cost of the entire path. The summation method is linear superposition. For example, a path consists of nodes A→B→C→D, where A, B, and C are bridging nodes and D is a terminal node. If the cost of A→B is 0.25, the cost of B→C is 0.38, and the cost of C→D is 0.32, then the cumulative cost of the entire path is 0.95.

[0090] To further improve path search efficiency, especially in scenarios with a large number of terminal nodes in the directional topology library, this embodiment introduces an efficient filtering and pruning strategy during the search for the cost-optimal chain. This strategy is used to quickly remove terminal nodes that are weakly related to the current multi-vector representation, avoid invalid path construction, and reduce the system's computational load.

[0091] First, during the topology library construction phase, the system pre-classifies all terminal nodes semantically according to anomaly type or image structural features, such as dividing them into multiple subclasses like "pole ears," "interlayer contact," and "local rupture," and calculates a representative center vector for each subclass. During the actual path search, the current vector representation is first compared with the center vectors of each subclass using coarse-grained similarity (e.g., cosine similarity or Euclidean distance). Only the terminal nodes in the top two or three subclasses with the highest similarity are selected as candidate targets, while the remaining terminal nodes are eliminated in this round of search, thus effectively compressing the search space.

[0092] Secondly, before constructing the path, the system performs a quick preliminary similarity assessment between the current vector and each terminal node. If the basic similarity between a terminal node and the current vector is lower than a preset threshold (e.g., less than 0.5), the node is considered too far from the current state and its path participation is directly excluded to avoid constructing redundant paths later.

[0093] Furthermore, the system introduces a "total cost threshold control" mechanism during path traversal. When the current cumulative cost of a path exceeds the system's set total cost limit (e.g., 1.2) during the construction process, further expansion of the path is immediately terminated, and Early-Stopping pruning is performed to ensure that only path branches that may constitute the optimal chain are retained for the next round of comparison.

[0094] By applying the above strategies in a coordinated manner, the system can quickly filter large-scale terminal node sets and perform real-time pruning during path construction, significantly improving overall operating efficiency and ensuring that the generation of the cost-optimal chain is computable, stable, and real-time, making it suitable for high-frequency early warning needs in large-scale lithium battery monitoring application scenarios.

[0095] Furthermore, the intelligent positioning method for early warning of internal short circuit in lithium battery under multi-data fusion disclosed in this embodiment, wherein determining the target terminal node according to the cost-optimal chain and issuing an intelligent warning, further includes: taking the terminal node at the end of the cost-optimal chain as the target terminal node; calculating the confidence degree between the target terminal node and the multivariate vector expression; and if the confidence degree is higher than a threshold, then issuing an intelligent warning according to the anomaly type corresponding to the target terminal node.

[0096] First, the system extracts the terminal node of the cost-optimal chain as the target terminal node corresponding to the current image state. This terminal node is bound to a preset anomaly type label, such as "pole short circuit", "structural collapse" or "interlayer contact", and represents the final state of a certain anomaly development path in the topology.

[0097] Subsequently, the system calculates the corresponding confidence score based on the similarity relationship between the current multivariate vector representation and the vector representation of the target terminal node. The confidence score reflects the degree of matching between the current state and the target anomaly, and can be calculated using cosine similarity, the inverse of weighted Euclidean distance, or a standardized score-based model. The calculation result is a real value between 0 and 1, with higher values ​​indicating a closer match.

[0098] The system sets a confidence threshold (e.g., 0.75) to determine whether the warning conditions are met. If the current confidence score is higher than this threshold, the system triggers the intelligent warning process. The warning content includes key information such as the anomaly type indicated by the target terminal node, the current path structure (cost-optimal chain), image acquisition time, and anomaly severity assessment level (e.g., early, mid, critical). This information is pushed to maintenance personnel or the upper-level scheduling system through the local control system or remote management platform.

[0099] It is worth mentioning that the intelligent positioning method for early warning of internal short circuits in lithium batteries under multi-data fusion disclosed in the embodiment, which determines the target terminal node based on the cost-optimal chain and issues an intelligent warning, further includes: not only using the terminal node at the end of the cost-optimal chain as the basis for judging the target anomaly type, but also further combining the structural composition and cost distribution characteristics of the entire link to evaluate the development stage of the current anomaly state, thereby forming a more trend-based warning conclusion. Specifically, after obtaining the cost-optimal chain, the system not only analyzes the anomaly type indicated by the end of the link, but also performs structural analysis on factors such as the number of bridging nodes, path length, and connection cost change trends along the entire path. For example, when the link is short, there are few bridging nodes, and the connection cost is low, it usually indicates that the current state is highly similar to the terminal anomaly morphology, and the system judges this state as a "critical anomaly" and issues a high-level warning. When the link is long, there are many bridging nodes, and the path cost gradually increases, it indicates that the current anomaly state is still in a slow evolution stage, and the system judges it as a "trend anomaly" or "early anomaly," prompting the user to pay close attention but not to take immediate action. Furthermore, the system supports confidence scoring at the link level, which calculates the overall credibility of the link as an abnormal evolution trend based on the stability and continuity of the connection cost of each segment in the path. For example, if there are significant abrupt changes in path cost or skipped connection segments, it may indicate a lack of structural consistency in the path. The system can reduce the final warning level of the path by lowering its weights to avoid false alarms due to structural incompleteness.

[0100] Furthermore, the intelligent positioning method for early warning of internal short circuit in lithium battery under multi-data fusion disclosed in this embodiment, wherein optimizing the directional topology library based on the multivariate vector expression and the cost-optimal chain, further includes: determining whether the link cost between adjacent units of the multivariate vector expression in the cost-optimal chain exceeds a preset difference threshold; if it exceeds the threshold, constructing a new supplementary unit containing the multivariate vector expression and inserting it as an independent node into the directional topology library.

[0101] The system first determines whether there is a significant abrupt change in the link cost between adjacent units in the cost-optimal chain of the current multivariate vector expression. Specifically, it checks whether the cost of a certain connection segment exceeds a difference threshold set by the system (e.g., 0.65). If a segment is found in the path with a significantly higher connection cost than other segments and a significant structural jump, it is considered that there is an expression blind spot in the current topology at that point.

[0102] At this point, the system constructs the current multivariate vector representation as a new supplementary unit and inserts it into the topology library as an independent node. This node simultaneously establishes directional connections with the two existing nodes before and after the mutation segment to fill structural gaps and enhance link continuity and graph representation integrity. Typically, the newly added supplementary node is a bridging node by default.

[0103] Furthermore, the system will perform a structural evaluation of the image representation features corresponding to the newly added supplementary unit to determine whether it possesses typical abnormal final state features. The judgment rules include, but are not limited to: the path cost between the current node and any terminal node is greater than the warning threshold; the current node's own confidence level is higher than the historical reference average; or its image morphology possesses highly abnormal features (such as concentrated high-temperature areas, structural faults, abnormal texture distribution, etc.). If the terminal feature judgment conditions are met, the system marks the newly added node as a new terminal node and registers its anomaly type, occurrence conditions, and risk level in the topology for subsequent path reasoning.

[0104] Figure 2 This is a schematic diagram of an intelligent positioning system for early warning of internal short circuit in lithium batteries under multi-data fusion, provided in an embodiment of this disclosure.

[0105] The topology library server is used to create a lithium battery anomaly image feature-oriented topology library composed of bridging nodes and terminal nodes.

[0106] In this embodiment, the lithium battery anomaly image feature directional topology library is used to characterize the image structural relationships of different types of short-circuit anomalies. Its basic structure consists of multiple bridging nodes and terminal nodes, organized through directional connections (i.e., directional links). Bridging nodes describe the intermediate forms of image morphology during the anomaly evolution process, while terminal nodes correspond to specific anomaly types that have actually occurred (such as tab short circuits, electrode contact, inner layer collapse, etc.).

[0107] The system integrates 1200 sets of anomalous image samples, including infrared thermograms and ultrasonic tomography. Each set of samples corresponds to a standardized image vector representation with a uniform 128-dimensional dimension. The image representation includes various statistical features such as heat distribution, texture direction, and structural attenuation to comprehensively describe potential anomalous features in the image. First, all image samples are grouped using an image clustering algorithm, obtaining 40 image representation centers with an initial cluster size of 40. The clustering results are then reviewed using manual annotation and empirical knowledge, ultimately selecting 18 bridging nodes and 6 terminal nodes as the initial topology. Each bridging node is connected to at least one forward node and one backward node, forming a clear path for the evolution of anomalous morphology. For example, bridging node B7 represents a "fractured ring hotspot," with its forward node being B3 (linear hotspot) and its backward-connected terminal node being T2 (tab short circuit).

[0108] In the topology, any two nodes may be connected directionally. The system assigns a fixed connection cost to each pair of connected nodes. This cost depends on the overall difference in the image representations of the two nodes, typically ranging from 0 to 1, with smaller values ​​indicating more similar image representations. For example, the connection cost from node B3 to B7 is 0.42; the connection cost from node B7 to terminal node T2 is 0.38; and the connection cost from B3 directly to T2 is 0.91. Therefore, the system prefers the path B3→B7→T2. The final directional topology library contains 24 nodes (18 bridging nodes and 6 terminal nodes) and 46 directional connections. The entire structure is stored in the graph database as an adjacency matrix, supporting efficient subsequent retrieval and path reasoning.

[0109] The image parsing unit is used to obtain the multivariate vector representation of the current image of the target lithium battery.

[0110] The system deploys an infrared thermal imaging camera and an ultrasonic transducer array, mounted above and to the side of the battery pack, to simultaneously acquire images of the battery's surface temperature and internal structure. The thermal images undergo non-uniformity correction and Gaussian filtering to remove edge noise and background thermal field interference; the ultrasonic images undergo time-domain averaging and dynamic gain adjustment to enhance the echo structure boundaries; the two types of images are aligned in spatial coordinates and uniformly cropped into a 256×256 region for use by the subsequent vector extraction module.

[0111] A unified image encoding module is used to encode multimodal images into a 128-dimensional vector representation with a unified structure. The system uses normalization processing to control all dimensions between 0 and 1, which facilitates direct comparison with node representations in the directional topology library.

[0112] Taking a real image acquisition task as an example, the system acquires data from a 3.2V, 35Ah pouch battery. The extracted current image vector fragment is as follows: the main hot spot area of ​​the thermal image appears in the second quadrant of the image, with a maximum temperature difference of 0.6°C; the ultrasound image shows discontinuous echo bands and blurred delay boundaries; the corresponding first 6-dimensional vector is: [0.28, 0.34, 0.12, 0.08, 0.43, 0.51, ...]. This vector will be directly used as the input in the subsequent step S300 for cost calculation and path construction with all terminal nodes in the directional topology library.

[0113] The optimal chain calculation unit is used to find the cost-optimal chain from the multivariate vector expression to the terminal node in the directional topology library.

[0114] The directional topology library contains 24 nodes, including 18 bridging nodes and 6 terminal nodes. Each pair of nodes has a directional connection with a preset connection cost. Specifically, the system first temporarily inserts the current multi-vector representation as a starting node into the graph structure. This temporary node does not establish connections with all nodes in the graph, but only with some unidirectional connections to bridging nodes. The establishment of these connections is based on the degree of vector difference between the current vector and each bridging node. When the difference between the current vector and a bridging node is less than a preset similarity threshold (e.g., 0.75), the system establishes a temporary connection and assigns a corresponding connection cost.

[0115] Subsequently, the system uses this temporary starting point as the source point and traverses all reachable terminal node paths. Each path consists of several bridging nodes and the final terminal node. The system accumulates the connection costs between each segment of the path to obtain the total cost of the entire path. Among all feasible paths, the path with the minimum total cost is determined as the cost-optimal chain. For example, the current vector establishes temporary connections with bridging nodes B3, B5, and B9, with costs of 0.28, 0.36, and 0.42, respectively. The system continues to calculate the paths from these nodes to each terminal node, forming the following candidate chains: Path 1: Current node → B3 → B7 → T2, total cost 0.28 + 0.29 + 0.33 = 0.90; Path 2: Current node → B5 → T3, total cost 0.36 + 0.41 = 0.77; Path 3: Current node → B9 → B14 → T5, total cost 0.42 + 0.36 + 0.22 = 1.00. Finally, the system selects the path with the lowest total cost (path 2) as the cost-optimal chain for the current image, and passes the terminal node T3 at the end of the path as the inference result corresponding to the current abnormal state to the next step for confidence judgment and intelligent early warning processing.

[0116] The intelligent early warning unit is used to determine the target terminal node based on the cost-optimal chain and issue an intelligent early warning.

[0117] The system first extracts the last node in the cost-optimal chain. This node is the terminal node that the current vector representation is most likely to point to in the graph structure, representing the known anomaly type that is closest to the current anomalous state on the evolution path. This terminal node is usually bound to anomaly labels (such as "bar short circuit", "structural collapse", "internal contact", etc.) and the number of historical sample support.

[0118] Subsequently, the system calculates the similarity score between the current multivariate vector and the target terminal node vector to determine whether the current state has sufficient credibility to be classified as an anomaly of this type. If the similarity score is higher than a set threshold (e.g., 0.82), an intelligent early warning process is triggered.

[0119] Upon triggering the alert, the system encapsulates the anomaly in a structured manner into an alert information package, which includes: anomaly type identifier, details of the optimal cost chain path, image timestamp, visual thermal image screenshot, and the battery module number. This information will be pushed to the upper-level management system or edge security gateway in real time, enabling on-duty personnel or automated control systems to respond quickly.

[0120] In this embodiment, the core of the intelligent early warning unit lies in: not only using the end node of the cost-optimal chain to identify the current anomaly type, but also using the structural characteristics of the entire chain to further infer the evolution stage of the current anomaly, thereby achieving intelligent early warning with trend judgment capabilities.

[0121] Specifically, the system first extracts the terminal node at the end of the cost-optimal chain and uses the anomaly type corresponding to that node as the result of this identification, such as "tab short circuit" or "inner layer rupture". However, at the same time, the system is not limited to the judgment of the terminal node itself, but further analyzes information such as the number of bridging nodes, path depth, and cost change gradient in the cost-optimal chain to assess the relative stage of the current anomaly state in the entire evolution path.

[0122] For example, if the cost-optimal chain reaches the terminal node with only one bridging node, and the cost of the initial connection is much lower than the cost of the final connection, it indicates that the current state is highly close to the terminal anomaly, and the warning level can be set as "severe anomaly, entering the final stage". Conversely, if the current vector needs to gradually transition through 3 to 4 bridging nodes to reach the terminal node, and the cost generally shows an increasing trend, it indicates that the current state is still in the early or middle stage of anomaly development. At this time, the system will automatically issue a "trend warning" to remind the user that the lithium battery has the potential risk of developing into a specific anomaly.

[0123] The topology library server is also used to optimize the directional topology library based on the multivariate vector representation and the cost-optimal chain.

[0124] Specifically, the system first determines whether there is a sudden change in the connection cost between adjacent units of the current multivariate vector in its cost-optimal chain. If the connection cost of a certain segment in the chain is significantly higher than that of other segments, and the difference in the characteristics of the connection nodes in that segment is greater than a preset structural difference threshold (e.g., similarity below 0.7), the system considers that the existing graph lacks transitional expression between that segment, resulting in an expression blind spot. In this case, the system constructs the current multivariate vector expression as a new supplementary unit and inserts it into the graph structure as an independent node. During the insertion process, the original connection relationships between nodes are not modified; instead, two new directional links are added, connecting the current vector node to the two original nodes before and after the segment, respectively. The assigned connection cost is calculated from the actual distance between the current vector and the two nodes.

[0125] For example, in a certain early warning process, the current cost-optimal chain is B4→B11→T5, where the connection cost between B4 and B11 is 0.79, which is much higher than the average cost of 0.36 for other segments of the path. The system determines that there is a semantic gap in this segment, so it inserts the new node N1 represented by the current vector into this segment, forming the updated chain B4→N1→B11→T5, and retains the original path.

[0126] In addition, the system will also associate and store metadata such as the current vector's label information, trigger time, image summary and confidence level into the newly added node to ensure that it can be identified, referenced and updated in subsequent path searches.

[0127] Furthermore, the intelligent positioning system for early warning of internal short circuit in lithium battery under multi-data fusion disclosed in this embodiment further includes, in the step of finding the cost-optimal chain from the multivariate vector expression to the terminal node in the directional topology library, obtaining the multi-segment links from the multivariate vector expression to all terminal nodes, and summing the link costs of the multi-segment links through each node; and taking the multi-segment link with the minimum summation cost as the cost-optimal chain.

[0128] First, the system uses the current multi-vector representation as a temporary starting node and quickly locates the set of all target nodes identified as terminal nodes by consulting the structural index of the directional topology library. This set can be filtered by the "type label" field in the node meta-attributes. The system pre-sets all abnormal final-state nodes as terminal node identifiers.

[0129] Subsequently, starting from the current starting point, the system constructs all valid paths leading to each terminal node using either a breadth-first traversal algorithm with a directed graph structure or an A* heuristic search algorithm. To limit search complexity and avoid ineffective expansion, the system sets the maximum number of hops to no more than 5, meaning each path passes through a maximum of 4 bridging nodes and 1 terminal node.

[0130] For each path, the system extracts the connection cost between each adjacent node segment, performs a segment-by-segment summation operation, and obtains the cumulative cost of the entire path. The summation method is linear superposition. For example, a path consists of nodes A→B→C→D, where A, B, and C are bridging nodes and D is a terminal node. If the cost of A→B is 0.25, the cost of B→C is 0.38, and the cost of C→D is 0.32, then the cumulative cost of the entire path is 0.95.

[0131] To further improve path search efficiency, especially in scenarios with a large number of terminal nodes in the directional topology library, this embodiment introduces an efficient filtering and pruning strategy during the search for the cost-optimal chain. This strategy is used to quickly remove terminal nodes that are weakly related to the current multi-vector representation, avoid invalid path construction, and reduce the system's computational load.

[0132] First, during the topology library construction phase, the system pre-classifies all terminal nodes semantically according to anomaly type or image structural features, such as dividing them into multiple subclasses like "pole ears," "interlayer contact," and "local rupture," and calculates a representative center vector for each subclass. During the actual path search, the current vector representation is first compared with the center vectors of each subclass using coarse-grained similarity (e.g., cosine similarity or Euclidean distance). Only the terminal nodes in the top two or three subclasses with the highest similarity are selected as candidate targets, while the remaining terminal nodes are eliminated in this round of search, thus effectively compressing the search space.

[0133] Secondly, before constructing the path, the system performs a quick preliminary similarity assessment between the current vector and each terminal node. If the basic similarity between a terminal node and the current vector is lower than a preset threshold (e.g., less than 0.5), the node is considered too far from the current state and its path participation is directly excluded to avoid constructing redundant paths later.

[0134] Furthermore, the system introduces a "total cost threshold control" mechanism during path traversal. When the current cumulative cost of a path exceeds the system's set total cost limit (e.g., 1.2) during the construction process, further expansion of the path is immediately terminated, and Early-Stopping pruning is performed to ensure that only path branches that may constitute the optimal chain are retained for the next round of comparison.

[0135] By applying the above strategies in a coordinated manner, the system can quickly filter large-scale terminal node sets and perform real-time pruning during path construction, significantly improving overall operating efficiency and ensuring that the generation of the cost-optimal chain is computable, stable, and real-time, making it suitable for high-frequency early warning needs in large-scale lithium battery monitoring application scenarios.

[0136] Furthermore, the intelligent positioning system for early warning of internal short circuit in lithium battery under multi-data fusion disclosed in this embodiment, wherein determining the target terminal node according to the cost-optimal chain and issuing an intelligent warning further includes: taking the terminal node at the end of the cost-optimal chain as the target terminal node; calculating the confidence degree between the target terminal node and the multivariate vector expression; and if the confidence degree is higher than a threshold, then issuing an intelligent warning according to the anomaly type corresponding to the target terminal node.

[0137] First, the system extracts the terminal node of the cost-optimal chain as the target terminal node corresponding to the current image state. This terminal node is bound to a preset anomaly type label, such as "pole short circuit", "structural collapse" or "interlayer contact", and represents the final state of a certain anomaly development path in the topology.

[0138] Subsequently, the system calculates the corresponding confidence score based on the similarity relationship between the current multivariate vector representation and the vector representation of the target terminal node. The confidence score reflects the degree of matching between the current state and the target anomaly, and can be calculated using cosine similarity, the inverse of weighted Euclidean distance, or a standardized score-based model. The calculation result is a real value between 0 and 1, with higher values ​​indicating a closer match.

[0139] The system sets a confidence threshold (e.g., 0.75) to determine whether the warning conditions are met. If the current confidence score is higher than this threshold, the system triggers the intelligent warning process. The warning content includes key information such as the anomaly type indicated by the target terminal node, the current path structure (cost-optimal chain), image acquisition time, and anomaly severity assessment level (e.g., early, mid, critical). This information is pushed to maintenance personnel or the upper-level scheduling system through the local control system or remote management platform.

[0140] It is worth mentioning that the intelligent positioning method for early warning of internal short circuits in lithium batteries under multi-data fusion disclosed in the embodiment, which determines the target terminal node based on the cost-optimal chain and issues an intelligent warning, further includes: not only using the terminal node at the end of the cost-optimal chain as the basis for judging the target anomaly type, but also further combining the structural composition and cost distribution characteristics of the entire link to evaluate the development stage of the current anomaly state, thereby forming a more trend-based warning conclusion. Specifically, after obtaining the cost-optimal chain, the system not only analyzes the anomaly type indicated by the end of the link, but also performs structural analysis on factors such as the number of bridging nodes, path length, and connection cost change trends along the entire path. For example, when the link is short, there are few bridging nodes, and the connection cost is low, it usually indicates that the current state is highly similar to the terminal anomaly morphology, and the system judges this state as a "critical anomaly" and issues a high-level warning. When the link is long, there are many bridging nodes, and the path cost gradually increases, it indicates that the current anomaly state is still in a slow evolution stage, and the system judges it as a "trend anomaly" or "early anomaly," prompting the user to pay close attention but not to take immediate action. Furthermore, the system supports confidence scoring at the link level, which calculates the overall credibility of the link as an abnormal evolution trend based on the stability and continuity of the connection cost of each segment in the path. For example, if there are significant abrupt changes in path cost or skipped connection segments, it may indicate a lack of structural consistency in the path. The system can reduce the final warning level of the path by lowering its weights to avoid false alarms due to structural incompleteness.

[0141] Furthermore, the intelligent positioning system for early warning of internal short circuit in lithium battery under multi-data fusion disclosed in this embodiment further includes optimizing the directional topology library based on the multivariate vector expression and the cost-optimal chain, and further includes: determining whether the link cost between adjacent units of the multivariate vector expression in the cost-optimal chain exceeds a preset difference threshold; if it exceeds the threshold, constructing a new supplementary unit containing the multivariate vector expression and inserting it into the directional topology library as an independent node.

[0142] The system first determines whether there is a significant abrupt change in the link cost between adjacent units in the cost-optimal chain of the current multivariate vector expression. Specifically, it checks whether the cost of a certain connection segment exceeds a difference threshold set by the system (e.g., 0.65). If a segment is found in the path with a significantly higher connection cost than other segments and a significant structural jump, it is considered that there is an expression blind spot in the current topology at that point.

[0143] At this point, the system constructs the current multivariate vector representation as a new supplementary unit and inserts it into the topology library as an independent node. This node simultaneously establishes directional connections with the two existing nodes before and after the mutation segment to fill structural gaps and enhance link continuity and graph representation integrity. Typically, the newly added supplementary node is a bridging node by default.

[0144] Furthermore, the system will perform a structural evaluation of the image representation features corresponding to the newly added supplementary unit to determine whether it possesses typical abnormal final state features. The judgment rules include, but are not limited to: the path cost between the current node and any terminal node is greater than the warning threshold; the current node's own confidence level is higher than the historical reference average; or its image morphology possesses highly abnormal features (such as concentrated high-temperature areas, structural faults, abnormal texture distribution, etc.). If the terminal feature judgment conditions are met, the system marks the newly added node as a new terminal node and registers its anomaly type, occurrence conditions, and risk level in the topology for subsequent path reasoning.

[0145] Furthermore, the system also includes a capacity matrix generation module, a residual flow update module, and an evaluation module;

[0146] The capacity matrix generation module is used to construct a capacity-based local subgraph and capacity matrix based on the directional topology library, multi-vector representation, and cost-optimal chain, so as to determine the initial height label and starting point set of the current sample as a virtual source point;

[0147] The residual traffic update module is used to generate initial residual network parameters based on the initial height label and starting point set, and iteratively update the residual network and traffic matrix after termination through Push and Relabel to determine the consistency verification index.

[0148] The evaluation module is used to obtain the candidate path cost set and the minimum total cost for each terminal in the path search, and to linearly normalize the candidate path cost set and the minimum total cost to calculate the confidence score and suboptimal competitiveness of the candidate path cost and the minimum total cost.

[0149] Furthermore, the system also includes a detection module and a recognition module;

[0150] The detection module is used to determine the source-side reachable set and the sink-side unreachable set based on the residual network and traffic matrix after termination, so as to generate the cut edge set and the main chain saturation edge set. The cut edge set is composed of all edges between the source-side reachable set and the sink-side unreachable set.

[0151] The identification module is used to acquire the spatial alignment information of thermal images and ultrasound in a unified coordinate system and a pre-established node region template library, and generate a weight map based on the cut edge set and the main chain saturated edge set, so as to unify the coordinates of thermal images and ultrasound onto the image grid according to the weight map and the spatial alignment information to obtain the risk corridor region.

[0152] According to embodiments of this disclosure, an electronic device is also provided, which may include a processor, a communications interface, a memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions in the memory to execute a configuration software-based software licensing implementation method.

[0153] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0154] On the other hand, this disclosure also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the configuration software-based software licensing implementation methods provided by the above methods.

[0155] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0156] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0157] It should be understood that the above embodiments are only used to illustrate the technical solutions of this disclosure, and not to limit them; although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure.

Claims

1. A method for intelligent localization of early warning of internal short circuit in lithium batteries based on multi-data fusion, characterized in that, include: A lithium battery anomaly image feature orientation topology library is created, consisting of bridging nodes and terminal nodes. The bridging nodes are used to describe the intermediate forms of image morphology during the anomaly evolution process, while the terminal nodes correspond to the specific anomaly types that have actually occurred. Connection relationships are used to describe the directional dependencies and transition paths between different nodes during the anomalous evolution process, and the differences in image features between nodes are quantified by allocating connection costs; Obtain the multivariate vector representation of the current image of the target lithium battery; find the cost-optimal chain from the multivariate vector representation to the terminal node in the directional topology library; The target terminal node is determined based on the cost-optimal chain, and an intelligent early warning is issued; Optimize the directional topology library based on the multivariate vector representation and the cost-optimal chain; In the optimized directional topology library, a local subgraph is constructed based on the optimal chain node set and one-hop neighborhood, and a height label is generated based on the local subgraph through consistency verification. The method further includes: Based on the directional topology library, multivariate vector representation, and cost-optimal chain, a capacity-based local subgraph and capacity matrix are constructed to determine the initial height label and starting point set of the current sample as the virtual source point. Initial residual network parameters are generated based on the initial height labels and starting point set, and then the residual network and traffic matrix after termination are iteratively updated through Push and Relabel to determine the consistency verification index. Obtain the candidate path cost set and the minimum total cost for each terminal in the path search, and perform linear normalization on the candidate path cost set and the minimum total cost to calculate the confidence score and suboptimal competitiveness of the candidate path cost and the minimum total cost. The process of constructing a capacity-optimized local subgraph and capacity matrix based on the directional topology library, multivariate vector representation, and cost-optimal chain to determine the initial height label and starting point set of the current sample as the virtual source point includes: The optimal chain node set and one-hop neighborhood are defined based on the directional topology library, and the union of the optimal chain node set and one-hop neighborhood is used as the local node set of the local subgraph. At the same time, the local edge set of the local subgraph is determined according to the one-hop neighborhood. The feature difference degree between different nodes is defined in the Euclidean distance normalization space. A capacity matrix and a difference degree matrix are constructed through a continuously monotonically decaying capacity mapping. The virtual source point is then connected to a bridging node that meets a set threshold to obtain an initialized height label vector.

2. The intelligent positioning method for early warning of internal short circuit in lithium batteries under multi-data fusion according to claim 1, characterized in that, The step of finding the cost-optimal chain from the multivariate vector representation to the terminal node in the directional topology library further includes: Obtain the multi-segment links from the multi-vector representation to all terminal nodes, and sum the link costs of the multi-segment links through each node; The multi-section link with the minimum cost after summation is taken as the cost-optimal link.

3. The intelligent positioning method for early warning of internal short circuit in lithium batteries under multi-data fusion according to claim 2, characterized in that, The step of determining the target terminal node based on the cost-optimal chain and issuing an intelligent early warning also includes: The terminal node at the end of the cost-optimal chain is taken as the target terminal node; Calculate the confidence level between the target terminal node and the multivariate vector representation; If the confidence level is higher than the threshold, then intelligent early warning is performed based on the anomaly type corresponding to the target terminal node and the cost-optimal chain.

4. The intelligent positioning method for early warning of internal short circuit in lithium battery under multi-data fusion according to claim 1, characterized in that, The step of optimizing the directional topology library based on the multivariate vector representation and the cost-optimal chain further includes: Determine whether the link cost between adjacent units in the cost-optimal chain of the multivariate vector representation exceeds a preset difference threshold; If the threshold is exceeded, a new supplementary unit containing the multivariate vector expression is constructed and inserted into the directional topology library as an independent node.

5. The intelligent positioning method for early warning of internal short circuit in lithium battery under multi-data fusion according to claim 1, characterized in that, The method further includes: Based on the residual network and flow matrix after termination, the source-side reachable set and the sink-side unreachable set are determined to generate the cut edge set and the main chain saturated edge set. The cut edge set consists of all edges between the source-side reachable set and the sink-side unreachable set. The spatial alignment information of thermal images and ultrasound in a unified coordinate system and a pre-established node region template library are obtained. A weight map is generated based on the cut edge set and the main chain saturated edge set. The coordinates of thermal images and ultrasound are unified onto the image grid according to the weight map and the spatial alignment information to obtain the risk corridor region.

6. A multi-data fusion-based intelligent positioning system for early warning of internal short circuits in lithium batteries, used to execute the method described in any one of claims 1-5, characterized in that, The system includes: The topology library server is used to create a lithium battery anomaly image feature orientation topology library composed of bridging nodes and terminal nodes. Bridging nodes are used to describe the intermediate form of image morphology in the anomaly evolution process, while terminal nodes correspond to the specific anomaly types that have actually occurred. Connection relationships are used to describe the directional dependence and transition path between different nodes in the anomaly evolution process, and to quantify the differences in image features between nodes by allocating connection costs. The image parsing unit is used to obtain the multi-vector representation of the current image of the target lithium battery; An optimal chain calculation unit is used to find the cost-optimal chain from the multivariate vector expression to the terminal node in the directional topology library; The intelligent early warning unit is used to determine the target terminal node based on the cost-optimal chain and issue an intelligent early warning. The topology library server is also used to optimize the directional topology library based on the multivariate vector representation and the cost-optimal chain.

7. The intelligent positioning system for early warning of internal short circuit in lithium batteries under multi-data fusion as described in claim 6, characterized in that, The system also includes a capacity matrix generation module, a residual flow update module, and an evaluation module; The capacity matrix generation module is used to construct a capacity-based local subgraph and capacity matrix based on the directional topology library, multi-vector representation, and cost-optimal chain, so as to determine the initial height label and starting point set of the current sample as a virtual source point; The residual traffic update module is used to generate initial residual network parameters based on the initial height label and starting point set, and iteratively update the residual network and traffic matrix after termination through Push and Relabel to determine the consistency verification index. The evaluation module is used to obtain the candidate path cost set and the minimum total cost for each terminal in the path search, and to linearly normalize the candidate path cost set and the minimum total cost to calculate the confidence score and suboptimal competitiveness of the candidate path cost and the minimum total cost.

8. The intelligent positioning system for early warning of internal short circuit in lithium battery under multi-data fusion as described in claim 7, characterized in that, The system also includes a detection module and an identification module; The detection module is used to determine the source-side reachable set and the sink-side unreachable set based on the residual network and traffic matrix after termination, so as to generate the cut edge set and the main chain saturation edge set. The cut edge set is composed of all edges between the source-side reachable set and the sink-side unreachable set. The identification module is used to acquire the spatial alignment information of thermal images and ultrasound in a unified coordinate system and a pre-established node region template library, and generate a weight map based on the cut edge set and the main chain saturated edge set, so as to unify the coordinates of thermal images and ultrasound onto the image grid according to the weight map and the spatial alignment information to obtain the risk corridor region.

Images