A power transmission channel intelligent retrieval and classification method

Abstract

The application provides a power transmission channel intelligent retrieval and classification method, which comprises the following steps: acquiring multi-modal data of a power transmission channel, inputting a hidden danger identification model, and outputting a hidden danger label sequence; taking a tower as a node and a channel segment between adjacent towers as an edge to construct a topological graph, and taking the embedding mean of the hidden danger label sequence in the corresponding channel segment as the initial feature of the edge to obtain a topological feature graph; performing hyperbolic space embedding on the topological feature graph to obtain a hyperbolic embedding vector of each edge, and performing hash coding to generate a hyperbolic topological hash code; receiving a retrieval intention of a user, and analyzing the retrieval intention into a differentiable operator chain through a large model; performing end-to-end differentiable back propagation mask processing on the hyperbolic topological hash code based on the differentiable operator chain to obtain a subgraph satisfying a retrieval condition, and simultaneously outputting a hidden danger category and a risk level corresponding to the power transmission channel. The application can quickly identify and locate weak links of a power transmission channel, support preventive maintenance, and reduce the probability of failure.

Description

Technical Field

[0001] This invention relates to the field of power transmission channel technology, and in particular to a method for intelligent retrieval and classification of power transmission channels. Background Technology

[0002] With the continuous expansion of my country's power grid and the acceleration of its intelligent transformation, the operation and maintenance management of power transmission channels is facing severe challenges brought by massive, multi-source, and heterogeneous inspection data. Currently, the inspection of transmission lines widely adopts various methods such as drones, online monitoring devices, and satellite remote sensing, generating tens of thousands of images, videos, laser point clouds, and sensor time-series data every day. This data contains key information such as equipment defects, external hazards, and environmental risks, and is the foundation for achieving condition-based maintenance and risk early warning. However, existing technologies have significant bottlenecks in terms of data utilization efficiency, depth of intelligent analysis, and agility of business response.

[0003] First, multimodal data fusion is difficult. Visible light images, infrared data, and inspection reports are processed independently, creating information silos that lack spatiotemporal alignment and semantic association, making comprehensive analysis difficult. Second, traditional retrieval capabilities are limited, relying on keywords and metadata, and cannot understand complex semantic queries, such as natural language commands combining spatial relationships and multiple attributes. Third, the automation level of hazard identification is low; existing models mostly target single defects, lacking a unified, fine-grained classification framework, and the results lack in-depth characterization of risk intensity and evolution trends. More importantly, existing methods lack modeling of channel network topology and risk propagation characteristics, failing to understand the clustering and transmission of risks from a network perspective, thus limiting preventative maintenance decision-making capabilities. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent retrieval and classification method for power transmission channels. By integrating artificial intelligence technology, this invention enables data-driven intelligent retrieval, accurate classification, and risk warning for the operation and maintenance management of power transmission channels, providing reliable technical support for the safe operation of the power grid.

[0005] To achieve the above objectives, the present invention provides an intelligent retrieval and classification method for power transmission channels, the method comprising: S11. Acquire multimodal data of the power transmission channel, input the multimodal data into the hazard identification model, and output the hazard label sequence; S12. Construct a topology graph with towers as nodes and passage segments between adjacent towers as edges, and use the embedding mean of the hazard label sequence within the corresponding passage segment as the initial feature of the edge to obtain the topology feature graph. S13. Perform hyperbolic space embedding on the topological feature map to obtain the hyperbolic embedding vector of each edge, and perform hash encoding on the hyperbolic embedding vector to generate hyperbolic topological hash code. S14. Receive the user's search intent and parse the search intent into a chain of differentiable operators using the large model; S15. Perform end-to-end differentiable backpropagation masking on the hyperbolic topological hash code based on the differentiable operator chain to obtain the subgraph that meets the retrieval conditions. Return the subgraph as the retrieval result and simultaneously output the hidden danger category and risk level corresponding to the power transmission channel.

[0006] Furthermore, step S11 specifically includes: S21. Acquire multimodal data of the target power transmission channel, wherein the multimodal data includes at least visible light images, infrared data and text description data from the same spatiotemporal correlation framework; S22. Perform spatiotemporal alignment and spatial slicing on the multimodal data, and divide the channel into several segments with a fixed spatial spacing; S23. Based on the text description data, extract the motion region mask for the visible light image in each segment, and filter the infrared data using the temperature threshold mask. S24. The multimodal data with completed masking is fused, input into the hazard identification model for processing, and output a hazard label with the corresponding risk intensity value. S25. Based on the risk intensity value, filter out the hazard labels that are higher than the preset threshold. Use a classifier to map the filtered hazard labels to hazard category labels, and map the corresponding section location information and time sequence information to generate the final hazard label sequence.

[0007] Furthermore, a topological feature map is constructed, specifically including: S31. Based on the hazard label sequence, obtain the location information of multiple towers and use them as nodes; S32. Group the towers according to the lines they belong to, and within each group, sort the towers according to their order on the lines. S33. Calculate the spatial distance between adjacent towers in the sorted order. If the spatial distance exceeds the preset distance threshold, mark it as abnormal or interrupt the connection. Connect the adjacent towers in each group that do not exceed the preset distance threshold to form an edge and construct a topology graph. S34. Map each label in the hidden danger label sequence to an embedding vector, calculate the mean of all embedding vectors, so that each channel segment obtains a corresponding feature vector, and take the corresponding maximum risk intensity value as the risk intensity attribute of the edge to obtain the topological feature map.

[0008] Furthermore, hyperbolic space embedding is performed on the topological feature map, specifically including: S41. Take the feature vector of each edge in the topological feature graph and transfer it from Euclidean space to hyperbolic space with a preset curvature through exponential mapping to obtain the initial hyperbolic embedding vector. S42. Based on the connection structure of the topological feature map, the initial hyperbolic embedding vector of all adjacent edges on the corresponding associated node is aggregated in the hyperbolic space for each edge to obtain the aggregated vector; S43. After weighting the aggregated vector through the hyperbolic attention mechanism, update the hyperbolic embedding vector of the current edge through hyperbolic transformation.

[0009] Furthermore, the hyperbolic embedding vector is hash-encoded, specifically including: S51. Project the hyperbolic embedding vector onto the Euclidean tangent space through a logarithmic mapping. S52. The projected hyperbolic embedding vector is mapped to a continuous representation of the target dimension using a learnable hash function; S53. Perform differentiable binarization on the continuous representation to generate a binary hash code.

[0010] Furthermore, the continuous representation undergoes differentiable binarization, specifically including: S61. During the training phase, each element in the continuous representation is treated as a binary classification problem, and the corresponding logical value is constructed. S62. In the forward propagation, preset noise is added to the logic value, and a continuous probability distribution is obtained through the softmax function with temperature parameter. S63. Calculate the expected value based on the probability distribution to obtain a continuously relaxed approximate binary code, and calculate the loss function to optimize the model parameters through backpropagation; S64. Based on the model parameters, each scalar element in the continuous representation is processed by a symbolic function during the inference phase to obtain the corresponding discrete binary element. S65. After mapping the discrete binary elements from the first value domain to the target binary value domain, combine all the mapped elements to generate a binary hash code.

[0011] Furthermore, the retrieval intent is parsed into a chain of differentiable operators through a large model, specifically including: S71. Use a large model to encode the search intent and identify constraints and logical relationships; S72. Based on a predefined library of differentiable operators, constraints and logical relationships are mapped to operator sequences. S73. Generate learnable parameter configurations for each operator in the operator sequence to form a chain of differentiable operators.

[0012] Furthermore, the constraints and logical relationships are mapped to a sequence of operators, specifically including: S81. Based on a predefined library of differentiable operators, semantically match each constraint with the operators in the library to obtain a set of candidate operators. S82. Based on the type of logical relationship, select operators from the candidate operator set and determine the execution order to generate an operator sequence framework; S83. Bind each operator in the operator sequence framework to a preset specific parameter to form an operator sequence specification; S84. Optimize and verify the operator sequence specification to obtain the operator sequence.

[0013] Furthermore, subgraphs are returned as search results, specifically including: S91. Input the hyperbolic topological hash code into the chain of differentiable operators, calculate the matching probability of each channel segment through forward propagation, and obtain the mask vector; S92. Based on the preset optimization objective, the probability values ​​in the mask vector are adjusted through backpropagation to obtain the channel segments that meet the preset retrieval conditions; S93. Threshold the optimized mask vector and extract channel segments with probabilities exceeding a preset threshold as candidate edges. S94. Based on graph connectivity analysis, extract the maximum connected component from the candidate edges to form a subgraph; S95. Calculate the hazard label sequence of all channel segments in the sub-map, calculate the distribution ratio of hazard categories, and determine the comprehensive risk level of the sub-map based on the preset maximum risk intensity and hazard combination rules. S96. Returns search results including node and edge information of the subgraph, distribution of hazard categories, and comprehensive risk level.

[0014] Furthermore, the categories and risk levels of potential hazards corresponding to power transmission channels specifically include: S101. Statistically analyze the hazard label sequence of all channel segments in the returned subgraph, and calculate the distribution ratio, severity distribution and spatiotemporal distribution characteristics of hazard categories; S102. Based on the hazard label sequence and subgraph topology, calculate the comprehensive risk score and determine the risk level through a preset multi-dimensional risk assessment model; S103. An operation and maintenance suggestion report is generated based on the distribution ratio, severity distribution, spatiotemporal distribution characteristics and risk level of the hidden danger categories.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention provides an intelligent retrieval and classification method for power transmission channels. It comprehensively utilizes multimodal data such as images, text, and sensor data, generating a hazard label sequence through a hazard identification model to fully capture the complex state of power transmission channels and improve the accuracy of hazard detection. A topology graph is constructed using towers as nodes and channel segments as edges, preserving the spatial connectivity and topological structure of the transmission channels, facilitating the analysis of hazard propagation and associated impacts. Hyperbolic space is used to represent the hierarchy and tree structure of the power transmission network. Embedding the topology feature graph further enhances the capture of power grid hierarchical relationships and improves the accuracy of similarity retrieval. Converting the hyperbolic embedding vector into binary hash codes significantly reduces storage overhead. A large model is used to parse the user's natural language search intent into a chain of differentiable operators, enabling flexible and complex semantic queries. End-to-end backpropagation masking is performed on the hyperbolic topology hash codes using the differentiable operator chain to accurately extract subgraphs that meet the conditions. While returning the search results, the method automatically outputs the hazard category and risk level, assisting maintenance personnel in quickly locating high-risk sections and developing targeted inspection or maintenance plans. This invention integrates artificial intelligence technology to achieve data-driven intelligent retrieval, accurate classification, and risk warning for the operation and maintenance management of power transmission channels, providing reliable technical support for the safe operation of the power grid. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort. Figure 1 This is a schematic diagram of a method for intelligent retrieval and classification of power transmission channels provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the process for generating a hazard label sequence provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the process for constructing a topological feature map provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the hyperbolic space embedding process for a topological feature map provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the hash encoding process for hyperbolic embedding vectors provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the process for performing differentiable binarization on a continuous representation, as provided in an embodiment of the present invention. Figure 7 This is a schematic diagram illustrating the process of parsing retrieval intent into a chain of differentiable operators using a large model, as provided in an embodiment of the present invention. Figure 8This is a schematic diagram illustrating the process of mapping constraints and logical relationships into operator sequences, provided as an embodiment of the present invention. Figure 9 This is a schematic diagram illustrating the process of returning sub-graphs as search results, provided by an embodiment of the present invention. Figure 10 This is a flowchart illustrating the categories of hidden dangers and risk levels corresponding to the output power transmission channels provided in this embodiment of the invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0018] Reference Figure 1 This embodiment provides a method for intelligent retrieval and classification of power transmission channels, the method comprising: S11. Obtain multimodal data of the power transmission channel, input the multimodal data into the hazard identification model, and output the hazard label sequence.

[0019] S12. Construct a topology graph with towers as nodes and passage segments between adjacent towers as edges, and use the embedding mean of the hazard label sequence within the corresponding passage segment as the initial feature of the edge to obtain the topology feature graph.

[0020] S13. Perform hyperbolic space embedding on the topological feature map to obtain the hyperbolic embedding vector of each edge, and perform hash encoding on the hyperbolic embedding vector to generate hyperbolic topological hash code.

[0021] S14. Receive the user's search intent and parse the search intent into a chain of differentiable operators using a large model.

[0022] S15. Perform end-to-end differentiable backpropagation masking on the hyperbolic topological hash code based on the differentiable operator chain to obtain the subgraph that meets the retrieval conditions. Return the subgraph as the retrieval result and simultaneously output the hidden danger category and risk level corresponding to the power transmission channel.

[0023] In this embodiment, multimodal data of the power transmission channel (such as UAV visible / infrared images, laser point clouds, online monitoring text reports, meteorological data, etc.) are collected, input into a trained hazard identification model, and output a hazard label sequence. A topology graph of the power transmission network is constructed using transmission towers as nodes and line segments between adjacent towers as edges. All hazard label sequences corresponding to each line segment are converted into numerical vectors using embedding technology, and their mean is calculated as the initial feature vector for the corresponding edge, ultimately obtaining the topology feature graph. Since the power transmission network itself has hierarchical and tree-like expansion characteristics, the topology feature graph is embedded into a hyperbolic space representing the hierarchical structure to obtain a hyperbolic vector representation of each edge. Subsequently, the hyperbolic vector is compressed into a binary hyperbolic topology hash code using a hash function. The system receives user retrieval requests expressed in natural language and automatically parses and translates them into a chain of differentiable operators using a large language model. Differentiable backpropagation masking is performed on the hyperbolic topology hash code using the differentiable operator chain to quickly and accurately filter out line segments that match the user's intent, i.e., the target subgraph. Finally, the system returns to the target submap, and automatically summarizes and outputs statistics on the types of hazards involved in the target route segment and an overall risk level assessment.

[0024] As a preferred embodiment, step S11 specifically includes: S21. Acquire multimodal data of the target power transmission channel, wherein the multimodal data includes at least visible light images, infrared data and text description data from the same spatiotemporal correlation framework.

[0025] S22. Perform spatiotemporal alignment and spatial slicing on the multimodal data, and divide the channel into several segments with a fixed spatial spacing.

[0026] S23. Based on the text description data, extract the motion region mask for the visible light image in each segment, and perform temperature threshold mask filtering on the infrared data.

[0027] S24. The multimodal data with completed masking is fused, input into the hazard identification model for processing, and output a hazard label with the corresponding risk intensity value.

[0028] S25. Based on the risk intensity value, filter out the hazard labels that are higher than the preset threshold. Use a classifier to map the filtered hazard labels to hazard category labels, and map the corresponding section location information and time sequence information to generate the final hazard label sequence.

[0029] In this embodiment, inspection tasks are planned for the same transmission line to ensure that visible light cameras and infrared thermal imagers at airborne or fixed locations acquire data synchronously or near synchronously. Text data (such as inspection reports and work orders) are associated with image data through timestamps and tower numbers. Using a geographic information system and tower spatial coordinates, all data are unified under the same geographic reference system. Temporal interpolation and spatial registration algorithms ensure that data from different modalities are aligned in time and space. Using the centerline of the transmission line corridor as a reference, buffer zones are divided along the line direction at preset fixed intervals to generate a series of continuous, non-overlapping regular segments.

[0030] By extracting motion region masks from visible light images within each segment and filtering infrared data using temperature threshold masks, attention is focused. This involves using high-level semantics of one modality to guide and constrain the extraction of low-level features from two other modalities, significantly reducing data redundancy and irrelevant background interference input to the subsequent fusion model. Specifically, computer vision algorithms extract motion region masks from visible light images using textual descriptions (e.g., cranes, excavators, construction). For example, using motion target detection models based on optical flow or background subtraction, or directly using pre-trained target detection models to detect dynamic objects such as construction vehicles and machinery, the generated masks mark dynamic risk areas in the image that are related to the textual semantics. Temperature threshold masks are then filtered from infrared data based on textual descriptions (e.g., heat, overheating) or prior knowledge (e.g., the normal temperature rise range of different equipment). One or more temperature thresholds are set, and the infrared thermal image is traversed pixel-by-pixel, marking areas with temperatures exceeding the thresholds as masks. Adaptive thresholds can be used to adapt to environmental changes.

[0031] The visible light image patch with masking, the infrared temperature map patch, and the text description vector with word embedding are fused and input into the hazard identification model for processing. A multimodal Transformer encoder can be used, with the masking information used as spatial attention weights or directly applied to the feature map. The hazard identification model ultimately outputs a multi-label classification probability vector, with each dimension corresponding to a type of hazard (e.g., tree obstruction, foreign object, wildfire, insulator damage), and also outputs a continuous risk intensity value (e.g., between 0 and 1). The risk intensity value integrates the visual salience of the hazard, the degree of temperature anomaly, and the severity of the text description. By setting an empirical or statistically based confidence threshold, hazard predictions with risk intensity values ​​higher than the confidence threshold are retained, thereby controlling the false alarm rate. The classifier maps the filtered hazard predictions (e.g., insulator damage, intensity 0.92) to standardized hazard category labels (e.g., equipment defect_insulator). The output of each segment (hazard category label, risk intensity value) is bound to the unique spatial identifier of the corresponding segment (e.g., start-end tower number) and the collection timestamp to form a structured data unit. Arrange them in spatial order according to the segments to generate the final sequence of hazard labels.

[0032] As a preferred embodiment, constructing a topological feature map specifically includes: S31. Based on the hazard label sequence, obtain the location information of multiple towers and use them as nodes.

[0033] S32. Group the towers according to the lines they belong to, and within each group, sort the towers according to their order on the lines.

[0034] S33. Calculate the spatial distance between adjacent towers in the sorted order. If the spatial distance exceeds the preset distance threshold, mark it as abnormal or interrupt the connection. Connect the adjacent towers in each group that do not exceed the preset distance threshold to form an edge and construct a topology graph.

[0035] S34. Map each label in the hidden danger label sequence to an embedding vector, calculate the mean of all embedding vectors, so that each channel segment obtains a corresponding feature vector, and take the corresponding maximum risk intensity value as the risk intensity attribute of the edge to obtain the topological feature map.

[0036] In this embodiment, the tower locations in the hazard label sequence are merely a series of discrete points in space. The essence of a transmission line is a network with towers as support points and lines as connecting edges. To analyze the distribution, association, and impact of hazards in the corridor, the inherent topological connections must be restored, i.e., a topological feature map is constructed based on spatial location and line affiliation. All unique tower identifiers (such as IDs or coordinates) are parsed from the hazard label sequence, with each tower as a graph node. The basic attributes of each node include the corresponding geographical coordinates. Towers are grouped according to their line affiliation. For tower sets within the same line group, they are sorted according to their inherent tower numbers. In power grid management, tower numbers themselves reflect their physical order on the line. If no number is provided, the order is determined using a spatial curve sorting algorithm along the approximate route of the line, based on geographical coordinates, thus outputting multiple ordered tower lists, each representing the tower order on a single line.

[0037] For each line's ordered tower list, the Euclidean spatial distance between adjacent ordered tower nodes is calculated, and then compared with a preset distance threshold. This preset distance threshold is typically set based on line design specifications and historical data statistics. If the Euclidean spatial distance is less than or equal to the preset distance threshold, an undirected edge is created between adjacent ordered tower nodes, and its attributes are initialized to empty. If the Euclidean spatial distance is greater than the preset distance threshold, it indicates a potential data breakpoint, large span, or data error; therefore, an abnormal connection is marked, and no edge is created, thus creating a topology interruption. This ensures that the constructed graph reflects reliable, regular, and continuous channel segments, ultimately generating a basic, featureless topology graph.

[0038] For each edge of the topology graph, all hazard labels whose spatial locations are within the corresponding segment of the topology graph are searched in the hazard label sequence. Each label is mapped to a high-dimensional semantic vector through a pre-trained word embedding model. The arithmetic mean of the vectors corresponding to all labels on each edge is calculated to obtain the comprehensive feature vector of each edge. The mean operation can smooth noise and fuse the semantic information of multiple labels. From the same set of labels, the risk intensity values ​​corresponding to all labels are extracted, and the maximum value is taken as the risk intensity attribute of that edge. Using the maximum value ensures that the overall risk level of a channel segment is determined by its most serious hazard. The feature vector and risk attribute are assigned to the corresponding edge, and finally the topology feature map is obtained.

[0039] As a preferred embodiment, hyperbolic space embedding is performed on the topological feature map, specifically including: S41. Take the feature vector of each edge in the topological feature graph and use an exponential mapping to transfer it from Euclidean space to a hyperbolic space with a preset curvature to obtain the initial hyperbolic embedding vector.

[0040] S42. Based on the connection structure of the topological feature map, the initial hyperbolic embedding vector of all adjacent edges on the corresponding associated node is aggregated in the hyperbolic space for each edge to obtain the aggregated vector.

[0041] S43. After weighting the aggregated vector through the hyperbolic attention mechanism, update the hyperbolic embedding vector of the current edge through hyperbolic transformation.

[0042] In this embodiment, traditional GNNs perform information transmission and aggregation in Euclidean space. However, power transmission networks are inherently radial, multi-level structures, approaching a tree-like topology. Euclidean space is inefficient in characterizing this type of exponentially growing hierarchical data, requiring extremely high dimensionality and struggling to maintain inherent metric relationships. Hyperbolic space, with its negative curvature, is naturally suitable for embedding tree-like or hierarchical data, maintaining hierarchical similarity between nodes / edges in a lower dimension and more compactly. In hyperbolic space, the geodesic (shortest path) from one point to another can be considered a potential path for risk or fault propagation in the network. Therefore, analyzing the distribution and distance of hyperbolic embedding vectors can be used to identify: high-risk clusters, dense clusters of points formed in hyperbolic space, which may correspond to geographically adjacent or topologically connected high-risk concentration areas; vulnerable links in the network, edges located at bridge positions between multiple high-risk clusters, whose hyperbolic representations may have specific vector characteristics; and critical system hubs, points near the hyperbolic geometric center, which may correspond to critical channel segments connecting multiple sub-networks, with a wide impact range due to their failure.

[0043] The graph embedding method of this invention focuses on channel segments (i.e., edges), directly assigning hazard characteristics and risk attributes to edge representations. The state of an edge depends not only on its own hazard characteristics but also on the states of its adjacent edges (towers); for example, multiple channel segments connecting the same high-risk tower may have risk associations. Therefore, edge representation needs to aggregate local neighborhood information (adjacent edges within one or two hops) to capture the topological context. Furthermore, the influence weights of different adjacent edges are not equal: adjacent edges of high-risk, similar hazard types have a much greater influence than adjacent edges of low-risk, unrelated hazard types.

[0044] Specifically, the edge representation learning process of the hyperbolic graph attention network uses the Lorenz model as a concrete implementation of the hyperbolic space. First, the feature vector of each edge in the topological feature graph is projected onto the tangent space of the hyperbolic space (Lorenz model). The feature vector of each edge in the topological feature graph is considered as a tangent space vector. An exponential mapping is used to map the tangent space vector onto the hyperbolic manifold, obtaining the initial hyperbolic embedding vector of each edge in the hyperbolic space. For each edge in the hyperbolic space, the associated first and second nodes are found. All other edges connected to the first node are collected, denoted as the first neighborhood edge set, and all other edges connected to the second node are denoted as the second neighborhood edge set. That is, the local neighborhood of each edge is the union of the first and second neighborhood edge sets. For each edge in the neighborhood, its current hyperbolic embedding vector is taken and aggregated in the hyperbolic space. The hyperbolic embedding vector is projected back to the base point of the tangent space through a logarithmic mapping, then weighted and averaged in the tangent space, and finally mapped back to the hyperbolic space through an exponential mapping, resulting in an aggregated vector that incorporates the local topological context.

[0045] An attention scoring function defined in hyperbolic space or its tangent space assigns different importance weights to each edge in the neighborhood. The neighbor vectors in the tangent space are then weighted and summed using these importance weights to obtain a weighted tangent space aggregation vector. This is then used to obtain the hyperbolic space attention-weighted aggregation vector through an exponential mapping. The hyperbolic embedding and attention aggregation vector of the current edge are mapped to the tangent space, where a linear combination and nonlinear activation are applied. Finally, the vectors are mapped back to the hyperbolic space, and the output is used as the updated hyperbolic embedding vector for the edge.

[0046] As a preferred embodiment, hash encoding of the hyperbolic embedding vector specifically includes: S51. Project the hyperbolic embedding vector onto the Euclidean tangent space through a logarithmic mapping.

[0047] S52. The projected hyperbolic embedding vector is mapped to a continuous representation of the target dimension using a learnable hash function.

[0048] S53. Perform differentiable binarization on the continuous representation to generate a binary hash code.

[0049] In this embodiment, the input to the hash encoding is a hyperbolic embedding vector. This vector encodes the inherent risk characteristics, local topological context, and hierarchical position of a channel segment within the network. Driven by training data, a learnable hash function learns to map the most relevant and discriminative feature dimensions of the high-dimensional hyperbolic vector to a finite number of binary bits. The optimization objective of the training process is to minimize the Hamming distance between hash codes generated from similar channel segments in the hyperbolic space, and conversely, to maximize the Hamming distance between dissimilar channel segments. Therefore, each hash code is equivalent to a semantically meaningful fingerprint of the corresponding channel segment, and the similarity of these fingerprints directly reflects the similarity of the channel segments in terms of risk-topological features.

[0050] Hamming distance is calculated extremely quickly, enabling the system to instantly find all other segments in a massive channel segment database that have similar risk patterns or topological roles to the target segment. This directly corresponds to critical tasks in operations and maintenance: finding all historical segments with similar characteristics to the currently faulty segment to aid in diagnosis, or discovering all segments in the entire network with a specific high-risk pattern. Binary hash codes can be easily used for bitwise operations, providing efficient underlying execution logic during the inference phase for parsing natural language queries into a chain of differentiable operators. For example, a query for an area with tree barriers and adjacent construction can be compiled into a bitwise AND operation on the hash code subspaces representing the two types of features, tree barriers and construction, thereby quickly locating a set of channel segments that simultaneously meet multiple conditions.

[0051] As a preferred embodiment, the continuous representation is subjected to differentiable binarization processing, specifically including: S61. During the training phase, each element in the continuous representation is treated as a binary classification problem, and the corresponding logical value is constructed.

[0052] S62. In the forward propagation, preset noise is added to the logic value, and a continuous probability distribution is obtained through the softmax function with temperature parameter.

[0053] S63. Calculate the expected value based on the probability distribution to obtain a continuously relaxed approximate binary code, and calculate the loss function to optimize the model parameters through backpropagation.

[0054] S64. Based on the model parameters, each scalar element in the continuous representation is processed by a symbolic function during the inference phase to obtain the corresponding discrete binary element.

[0055] S65. After mapping the discrete binary elements from the first value domain to the target binary value domain, combine all the mapped elements to generate a binary hash code.

[0056] In this embodiment, the random noise added to the logical values ​​during the training phase can be seen as modeling various uncertainties in the real world, such as sensor acquisition errors, feature ambiguity caused by environmental interference, or fluctuations in the confidence level of the hazard label itself. Through softening decisions with temperature parameters, the system initially allows the generation of hash codes to maintain a certain degree of ambiguity and exploration space, corresponding to the initial assessment stage of complex and novel hazards by maintenance personnel. As the temperature parameter decreases during annealing, the model output gradually tends towards deterministic binarization. This process simulates the solidification of decision-making regarding the risk status of a channel from suspected to confirmed, as evidence (data) becomes more abundant and understanding deepens. Each binary hash code output during the inference phase is a discrete mapping of the comprehensive risk-topological characteristics of a channel segment at a specific moment. Each binary hash code is determined by features learned by the upstream model that have discriminative power to distinguish different risk patterns. For example, some bits may jointly encode the "association pattern between wildfire risk and vegetation cover," while others may characterize the "spatiotemporal characteristics of nearby construction activities."

[0057] The differentiability of the entire binarization process allows the gradient to propagate back from the final retrieval performance evaluation (e.g., whether the returned subgraph fully matches the intent of the natural language query) all the way to the feature extraction network at this stage and before. In other words, the hash code generation rules are "trained" to serve the purpose of "quickly and accurately finding high-risk sections." The system automatically learns how to generate hash codes to most effectively support complex combined queries such as "finding all sections with wildfire risk and construction within the transmission corridor." Therefore, the hash code generation mechanism inherently encapsulates the business logic and retrieval intent of transmission corridor operation and maintenance.

[0058] As a preferred embodiment, the retrieval intent is parsed into a chain of differentiable operators using a large model, specifically including: S71. Use a large model to encode the search intent and identify constraints and logical relationships.

[0059] S72. Based on a predefined library of differentiable operators, constraints and logical relationships are mapped to operator sequences.

[0060] S73. Generate learnable parameter configurations for each operator in the operator sequence to form a chain of differentiable operators.

[0061] In this embodiment, during power transmission channel operation and maintenance, experts often use composite conditions to describe risks, such as "finding line segments susceptible to wildfires and with a history of tree obstruction." The role of the differentiable operator chain is to dynamically parse such descriptions into a series of basic operators. The construction process of the operator chain is to compile unstructured business experience into structured operation programs on topological feature graphs and hash codes in real time. Through parameter learning and structural combination, the differentiable operator chain forms a continuously evolving and active rule system that can adapt to new risk patterns more quickly. Each successful retrieval and subsequent verification will feed back into the parameters and structure selection of the operator chain through gradient updates, realizing the automatic deposition of domain knowledge into the machine model. For example, after successfully retrieving "lines prone to wind deflection flashover in strong wind areas" multiple times, the system will more accurately allocate weights to features such as wind zone level and line sag.

[0062] Specifically, firstly, domain-specific corpora (such as inspection reports, hazard dictionaries, and operating procedures) are used to fine-tune and enhance the large language model, enabling it to deeply understand terms like tree obstruction, wind deflection, and insulator damage, as well as entity concepts such as towers, channel sections, and risk levels. Specific output templates are designed to guide the large model in generating structured intermediate representations; for example, queries to find all high-risk tree obstruction sections are parsed into JSON structures. A library of operators for graph structure hash code retrieval is pre-built, including: Filter(attribute, operator, value), filtering based on edge attributes (such as risk level and hazard category); Nearby(node_type, radius), finding edges within a certain range of a certain type of node (such as a construction point); TemporalFilter(time_range), filtering based on timestamps; and And / Or / Not(logical_op), logical combination operators.

[0063] Based on the parsed structured representation, each semantic unit is transformed into a corresponding operator through a rule engine or neural network mapping, preserving their logical order to form an initial operator sequence. Each operator in the operator sequence is instantiated and associated with learnable parameters. For example, the threshold in the Filter operator can be set as a learnable scalar; the radius in the Nearby operator can be set as a learnable parameter; even the weights of logical combinations can be learned. The parameterized operators are then combined in logical order into a complete differentiable computation graph, i.e., a differentiable operator chain.

[0064] As a preferred embodiment, the constraints and logical relationships are mapped to a sequence of operators, specifically including: S81. Based on a predefined library of differentiable operators, semantically match each constraint with the operators in the library to obtain a set of candidate operators.

[0065] S82. Based on the type of logical relationship, select operators from the candidate operator set and determine the execution order to generate an operator sequence framework.

[0066] S83. Bind each operator in the operator sequence framework to a preset specific parameter to form an operator sequence specification.

[0067] S84. Optimize and verify the operator sequence specification to obtain the operator sequence.

[0068] In this embodiment, a typical query in power transmission channel operation and maintenance (such as retrieving all line segments where wildfires have occurred and towers are located in mining subsidence areas) includes three key business elements: constraints, corresponding to specific risk attributes or states (such as wildfires, mining subsidence areas); logical relationships, corresponding to the combination of risk elements (and representing risk superposition); and analytical intent, corresponding to the overall risk assessment objective. The process of mapping and generating an operator sequence is precisely the process of translating these three business elements into a standardized analysis program: constraints correspond to basic analysis operators, and each constraint is mapped to basic operators such as Filter, Nearby, and TemporalRange, which is equivalent to transforming unstructured risk descriptions (wildfires) into structured data fields (hazard label = wildfire); logical relationships correspond to program control flow, and logical relationships determine the combination and execution order of operators. For example, A and B are mapped to the intersection operation of two filtering operators, corresponding to the composite judgment logic of "multiple factors at the same time" in risk management; the analysis intent corresponds to a complete analysis pipeline, and the final operator sequence is a complete data processing and screening pipeline for a specific risk pattern, defining a deterministic calculation path to locate the target subset from the full channel data.

[0069] Specifically, each operator in the predefined operator library is accompanied by a meta-description, including its functional text description, the types of input attributes it accepts, and the data type of its output. Using a sentence encoder fine-tuned by the domain text, each constraint in the user query (e.g., high risk of wildfires) is encoded into a semantic vector. Simultaneously, the meta-description of each operator in the operator library is also encoded into a vector. By calculating cosine similarity or a more refined cross-attention score, the top K operators with the highest similarity for each constraint are retrieved, forming a candidate operator set. The identified logical relationships are transformed into logical expression trees. The system translates these logical expression trees into a data flow graph framework for operator execution according to predefined rules. For example, AND(A, B) typically means that two operators A and B can be executed in parallel, and their results are then merged through an Intersection operator; OR(A, B) might correspond to results merged through a Union operator; and the logical NOT(A) might be mapped to a Complement operator.

[0070] Based on the operator output / input data type contract, the system selects the most suitable operator for each logical leaf node (constraint) from the candidate operator set and initially determines the data flow connection order between them, forming an unparalleled computational graph framework. Each operator type has its parameter signature, binding each operator in the operator sequence framework with preset specific parameters. Each operator is a parameter-complete instance, forming an operator sequence specification. The operator sequence specification is optimized and verified, outputting the final differentiable operator sequence.

[0071] As a preferred embodiment, returning subgraphs as search results specifically includes: S91. Input the hyperbolic topological hash code into the chain of differentiable operators, calculate the matching probability of each channel segment through forward propagation, and obtain the mask vector.

[0072] S92. Based on the preset optimization objective, the probability values ​​in the mask vector are adjusted through backpropagation to obtain the channel segments that meet the preset retrieval conditions.

[0073] S93. Threshold the optimized mask vector and extract the channel segments with probabilities exceeding the preset threshold as candidate edges.

[0074] S94. Based on graph connectivity analysis, extract the maximum connected component from the candidate edges to form a subgraph.

[0075] S95. Calculate the hazard label sequence of all channel segments in the statistical sub-map, calculate the distribution ratio of hazard categories, and determine the comprehensive risk level of the sub-map based on the preset maximum risk intensity and hazard combination rules.

[0076] S96. Returns search results including node and edge information of the subgraph, distribution of hazard categories, and comprehensive risk level.

[0077] In this embodiment, the continuous probability values ​​(mask vectors) output by the differentiable operator chain can only reflect the probability that each channel segment independently meets the query intent. However, the actual risks of transmission channels often exhibit significant spatial continuity (e.g., multiple consecutive tree barriers) or topological correlation (e.g., multiple lines originating from the same power source node). If only high-probability discrete edge sets are returned, the clustering effect of risks in the network structure may be ignored. Therefore, graph structure analysis needs to be introduced on top of the discrete probability results to extract connected subgraph regions based on network connectivity, thereby identifying meaningful and complete risk clusters. Furthermore, the system automatically performs risk feature statistics and quantitative rating on the subgraphs to directly generate structured conclusions indicating the risk location, dominant hazard type, and overall severity level, supporting subsequent work order generation and resource scheduling.

[0078] Specifically, the hyperbolic topological hash code is input into the constructed differentiable operator chain computation graph. Each operator in the operator chain performs a differentiable operation on the hash code, ultimately outputting a vector with the same number of edges. Each element in the vector represents the matching probability of the i-th edge satisfying the entire composite query condition. The mask vector itself is also treated as an optimizable parameter, and a loss function related to the retrieval target is set (e.g., encouraging the mask probability to polarize towards 0 or 1 while satisfying certain overall constraints). Through several rounds of backpropagation, the mask vector is fine-tuned to more clearly separate matching and non-matching edges while maintaining differentiability. A fixed or adaptive threshold is applied to the optimized mask vector. All edges whose optimized mask vectors are greater than the threshold are marked as candidate edges, forming the initial result set. Based on the nodes and edges of the original topological feature graph, only the edge set is retained, resulting in a sparse and potentially disconnected initial subgraph.

[0079] A connected component analysis algorithm is run on the initial subgraph to find the largest connected component with the most edges, representing the most significant risk areas connected spatially or topologically, and this is denoted as the final output subgraph. The hazard label sequence associated with all edges in the subgraph is traversed, the frequency of each hazard category is counted, and their proportional distribution is calculated. The highest risk intensity is the maximum value among all preset maximum risk intensity values ​​of all edges in the subgraph. Hazard combination rules are based on pre-defined rules from the domain, such as "if both wildfire risk and tree obstruction exist simultaneously, the risk level automatically increases by one level." Combining the highest risk intensity and hazard combination rules, the overall risk level of the subgraph is determined through a predefined mapping table or rule engine. Finally, a structured object is returned, including: subgraph structure information, a list of nodes, a list of edges and their spatial coordinates; a risk analysis summary, hazard category distribution, and overall risk level; and optional metadata, such as search criteria, timestamps, and covered route names.

[0080] As a preferred embodiment, the hazard categories and risk levels corresponding to the output transmission channel specifically include: S101. Statistically return the hazard label sequence of all channel segments in the subgraph, and calculate the distribution ratio, severity distribution and spatiotemporal distribution characteristics of hazard categories.

[0081] S102. Based on the hazard label sequence and subgraph topology, calculate the comprehensive risk score through a preset multi-dimensional risk assessment model to determine the risk level.

[0082] S103. An operation and maintenance suggestion report is generated based on the distribution ratio, severity distribution, spatiotemporal distribution characteristics and risk level of the hidden danger categories.

[0083] In this embodiment, the hazard label sequence of all edges of the subgraph is traversed, the frequency of each category is counted, and the percentage of each category in the total number of hazards is calculated. The proportion of different risk intensity levels in the label sequence is statistically analyzed, and statistical quantities such as average intensity and intensity variance are calculated to describe the risk concentration. Based on the geographical coordinates of the channel segment, the spatial density of hazard points is calculated, or the uniformity of their distribution along the route is analyzed. The timestamps of the hazard labels are extracted, and the first appearance time and most recent update time of the hazard are analyzed to identify whether it is in an active growth period. The above statistical features and subgraph topological attributes are input into a preset multi-dimensional risk assessment model to obtain a comprehensive risk score. The calculated comprehensive risk score is mapped to a preset discrete risk level range to determine the risk level. A predefined report template is used, including fixed fields and dynamic content areas filled with data. Based on the above analysis results, suggestions are generated through a predefined business rule base, ultimately generating a structured document including regional location, statistical data, risk level, and specific suggestions.

[0084] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent retrieval and classification of power transmission channels, characterized in that, The method includes: S11. Acquire multimodal data of the power transmission channel, input the multimodal data into the hazard identification model, and output the hazard label sequence; S12. Construct a topology graph with towers as nodes and passage segments between adjacent towers as edges, and use the embedding mean of the hazard label sequence within the corresponding passage segment as the initial feature of the edge to obtain the topology feature graph. S13. Perform hyperbolic space embedding on the topological feature map to obtain the hyperbolic embedding vector of each edge, and perform hash encoding on the hyperbolic embedding vector to generate hyperbolic topological hash code. S14. Receive the user's search intent and parse the search intent into a chain of differentiable operators using the large model; S15. Perform end-to-end differentiable backpropagation masking on the hyperbolic topological hash code based on the differentiable operator chain to obtain the subgraph that meets the retrieval conditions. Return the subgraph as the retrieval result and simultaneously output the hidden danger category and risk level corresponding to the power transmission channel.

2. The intelligent retrieval and classification method for power transmission channels according to claim 1, characterized in that, Step S11 specifically includes: S21. Acquire multimodal data of the target power transmission channel, wherein the multimodal data includes at least visible light images, infrared data and text description data from the same spatiotemporal correlation framework; S22. Perform spatiotemporal alignment and spatial slicing on the multimodal data, and divide the channel into several segments with a fixed spatial spacing; S23. Based on the text description data, extract the motion region mask for the visible light image in each segment, and filter the infrared data using the temperature threshold mask. S24. The multimodal data with completed masking is fused, input into the hazard identification model for processing, and output a hazard label with the corresponding risk intensity value. S25. Based on the risk intensity value, filter out the hazard labels that are higher than the preset threshold. Use a classifier to map the filtered hazard labels to hazard category labels, and map the corresponding section location information and time sequence information to generate the final hazard label sequence.

3. The intelligent retrieval and classification method for power transmission channels according to claim 2, characterized in that, Constructing a topological feature map specifically includes: S31. Based on the hazard label sequence, obtain the location information of multiple towers and use them as nodes; S32. Group the towers according to the lines they belong to, and within each group, sort the towers according to their order on the lines. S33. Calculate the spatial distance between adjacent towers in the sorted order. If the spatial distance exceeds the preset distance threshold, mark it as abnormal or interrupt the connection. Connect the adjacent towers in each group that do not exceed the preset distance threshold to form an edge and construct a topology graph. S34. Map each label in the hidden danger label sequence to an embedding vector, calculate the mean of all embedding vectors, so that each channel segment obtains a corresponding feature vector, and take the corresponding maximum risk intensity value as the risk intensity attribute of the edge to obtain the topological feature map.

4. The intelligent retrieval and classification method for power transmission channels according to claim 1, characterized in that, Hyperbolic space embedding of the topological feature map specifically includes: S41. Take the feature vector of each edge in the topological feature graph and transfer it from Euclidean space to hyperbolic space with a preset curvature through exponential mapping to obtain the initial hyperbolic embedding vector. S42. Based on the connection structure of the topological feature map, the initial hyperbolic embedding vector of all adjacent edges on the corresponding associated node is aggregated in the hyperbolic space for each edge to obtain the aggregated vector; S43. After weighting the aggregated vector through the hyperbolic attention mechanism, update the hyperbolic embedding vector of the current edge through hyperbolic transformation.

5. The intelligent retrieval and classification method for power transmission channels according to claim 1, characterized in that, Hash encoding of hyperbolic embedding vectors specifically includes: S51. Project the hyperbolic embedding vector onto the Euclidean tangent space through a logarithmic mapping. S52. The projected hyperbolic embedding vector is mapped to a continuous representation of the target dimension using a learnable hash function; S53. Perform differentiable binarization on the continuous representation to generate a binary hash code.

6. The intelligent retrieval and classification method for power transmission channels according to claim 5, characterized in that, Differentiable binarization of continuous representations includes: S61. During the training phase, each element in the continuous representation is treated as a binary classification problem, and the corresponding logical value is constructed. S62. In the forward propagation, preset noise is added to the logic value, and a continuous probability distribution is obtained through the softmax function with temperature parameter. S63. Calculate the expected value based on the probability distribution to obtain a continuously relaxed approximate binary code, and calculate the loss function to optimize the model parameters through backpropagation; S64. Based on the model parameters, each scalar element in the continuous representation is processed by a symbolic function during the inference phase to obtain the corresponding discrete binary element. S65. After mapping the discrete binary elements from the first value domain to the target binary value domain, combine all the mapped elements to generate a binary hash code.

7. The intelligent retrieval and classification method for power transmission channels according to claim 1, characterized in that, The large model parses the retrieval intent into a chain of differentiable operators, specifically including: S71. Use a large model to encode the search intent and identify constraints and logical relationships; S72. Based on a predefined library of differentiable operators, constraints and logical relationships are mapped to operator sequences. S73. Generate learnable parameter configurations for each operator in the operator sequence to form a chain of differentiable operators.

8. The intelligent retrieval and classification method for power transmission channels according to claim 7, characterized in that, Mapping constraints and logical relationships into a sequence of operators specifically includes: S81. Based on a predefined library of differentiable operators, semantically match each constraint with the operators in the library to obtain a set of candidate operators. S82. Based on the type of logical relationship, select operators from the candidate operator set and determine the execution order to generate an operator sequence framework; S83. Bind each operator in the operator sequence framework to a preset specific parameter to form an operator sequence specification; S84. Optimize and verify the operator sequence specification to obtain the operator sequence.

9. The intelligent retrieval and classification method for power transmission channels according to claim 1, characterized in that, Returning subgraphs as search results specifically includes: S91. Input the hyperbolic topological hash code into the chain of differentiable operators, calculate the matching probability of each channel segment through forward propagation, and obtain the mask vector; S92. Based on the preset optimization objective, the probability values ​​in the mask vector are adjusted through backpropagation to obtain the channel segments that meet the preset retrieval conditions; S93. Threshold the optimized mask vector and extract channel segments with probabilities exceeding a preset threshold as candidate edges. S94. Based on graph connectivity analysis, extract the maximum connected component from the candidate edges to form a subgraph; S95. Calculate the hazard label sequence of all channel segments in the sub-map, calculate the distribution ratio of hazard categories, and determine the comprehensive risk level of the sub-map based on the preset maximum risk intensity and hazard combination rules. S96. Returns search results including node and edge information of the subgraph, distribution of hazard categories, and comprehensive risk level.

10. The intelligent retrieval and classification method for power transmission channels according to claim 1, characterized in that, The categories and risk levels of potential hazards corresponding to power transmission channels specifically include: S101. Statistically analyze the hazard label sequence of all channel segments in the returned subgraph, and calculate the distribution ratio, severity distribution and spatiotemporal distribution characteristics of hazard categories; S102. Based on the hazard label sequence and subgraph topology, calculate the comprehensive risk score and determine the risk level through a preset multi-dimensional risk assessment model; S103. An operation and maintenance suggestion report is generated based on the distribution ratio, severity distribution, spatiotemporal distribution characteristics and risk level of the hidden danger categories.

Images