A Smart Early Warning Method and System for Protecting Transmission Lines from External Damage

Abstract

The present application relates to the technical field of power system intelligent operation and maintenance, and particularly relates to a power transmission line external damage prevention intelligent early warning method and system. The method comprises the following steps: collecting power transmission line acoustic data through an acoustic sensor; performing acoustic fingerprint feature sparse reconstruction according to the power transmission line acoustic data to obtain acoustic fingerprint feature data; performing sparse causal graph matching on the acoustic fingerprint feature data to obtain acoustic source behavior causal graph atlas data; performing double-domain attention cross identification according to the acoustic source behavior causal graph atlas data to obtain acoustic source separation identification data; and generating a scene atlas from the acoustic source separation identification data to obtain acoustic source scene atlas data. By constructing a nested causal chain graph and fusing a propagation weight mechanism, the present application can accurately capture the development and evolution of power transmission line external damage events and multi-level triggering relationships. Compared with traditional single-point detection methods, the present application has stronger chain reasoning ability and early risk identification ability, effectively improving the response foresight and causal explainability of the early warning system.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power systems, and in particular to an intelligent early warning method and system for preventing external damage to transmission lines. Background Technology

[0002] During the long-term operation of power transmission lines, external destructive factors such as intrusion of construction machinery, illegal connections, and encroachment of vegetation along power corridors, whether human-induced or environmental disturbances, can easily lead to line tripping, power outages, or even large-scale safety accidents. Currently, mainstream anti-external damage technologies primarily rely on video surveillance, laser ranging, or image recognition. However, these are limited by visibility conditions, deployment range, and interference from obstructions, often resulting in numerous blind spots, high response delays, and high false alarm rates. Acoustic sensing technology refers to a system that uses acoustic sensors (such as microphone arrays) to collect, analyze, and model sound wave signals generated in the environment to obtain the spatial location, behavioral characteristics, or state changes of target events. This technology converts sound wave signals into calculable digital information and, combined with signal processing, spectrum reconstruction, and pattern recognition methods, can achieve accurate identification and dynamic tracking of specific sound sources (such as mechanical vibration, human activity, or environmental disturbances). Therefore, combining acoustic sensing with anti-external damage technology for power transmission lines has become a crucial issue. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes an intelligent early warning method and system for preventing external damage to transmission lines, thereby resolving at least one of the aforementioned technical issues.

[0004] This application provides an intelligent early warning method for preventing external damage to transmission lines, the method comprising:

[0005] S1. Acoustic data of power transmission lines are collected using acoustic sensors;

[0006] S2. Based on the acoustic data of the transmission line, perform sparse reconstruction of the voiceprint features to obtain voiceprint feature data; perform sparse causal graph matching on the voiceprint feature data to obtain sound source behavior causal graph data.

[0007] S3. Perform dual-domain attention cross-identification based on the sound source behavior causal graph data to obtain sound source separation and identification data;

[0008] S4. Generate a scene map from the sound source separation and identification data to obtain sound source scene map data.

[0009] In this invention, step S1 acquires environmental disturbance signals non-contactly using acoustic sensors, exhibiting high robustness in all weather conditions; step S2 employs sparse reconstruction and causal modeling to extract key voiceprint driving features, significantly improving the ability to identify the causes of behavior; step S3 combines a dual-domain attention mechanism to achieve joint time-frequency analysis, effectively improving the accuracy of multi-source separation; and step S4 constructs a visual representation based on structural graphs, containing causal chains, propagation paths, and behavior types, providing a clear decision-making basis for intelligent response and risk level assessment.

[0010] Optionally, S1 includes:

[0011] Acquire power transmission load data;

[0012] Based on the power transmission load data, sampling frequency processing and directional sensitivity processing are performed to obtain sampling sensitivity data and directional sensitivity data, respectively.

[0013] Based on the sampling sensitivity data and directional sensitivity data, the distributed sound field monitoring data is obtained by sampling through a dual-layer microphone array unit preset along the power transmission line. The dual-layer microphone array unit consists of a first layer for environmental sound pressure perception and a second layer for target frequency domain capture.

[0014] Acoustic disturbance latent space is constructed based on distributed sound field monitoring data to obtain acoustic disturbance latent space data.

[0015] Based on the acoustic disturbance hidden space data, the sound source propagation structure is modeled to obtain the acoustic data of the transmission line.

[0016] This invention introduces a dynamic sensing strategy driven by power transmission load data, enabling adaptive adjustment of the acoustic sampling process and effectively improving the system's response to abnormal sound sources under different current load conditions. Through dual control of sampling frequency and directional sensitivity, combined with a pre-installed dual-layer microphone array along the power transmission line, layered sensing of ambient background sound and target sound source frequency bands is achieved, thereby enhancing sound source capture capabilities in low signal-to-noise ratio environments. Simultaneously, the construction of an acoustic perturbation latent space based on distributed sound field monitoring data not only improves the sparsity and discriminative power of acoustic signature features but also provides a high-dimensional structural representation foundation for subsequent sound source propagation modeling.

[0017] Optionally, the sparse reconstruction of the voiceprint features includes:

[0018] Sparse Bayesian regression was performed on the electrical data of the transmission lines to obtain electrical sensing data.

[0019] Frequency-modulated drive spectrum construction is performed on electrical sensing data to obtain audio spectrum data;

[0020] Obtain transmission line current data;

[0021] Based on the transmission line current data and the audio spectrum data, an electrically driven phase disturbance is performed to obtain the disturbance spectrum data.

[0022] Transferable voiceprint embedding is performed on the perturbation map data to obtain voiceprint feature data.

[0023] This invention utilizes sparse Bayesian regression to uncover the potential driving factors of electrical state changes on acoustic response, effectively improving the modeling sensitivity to weak external disturbance signals. Frequency-modulated drive spectrum construction enhances the frequency resolution of acoustic signature representation, aiding in the identification of specific mechanical disturbance patterns. By introducing current data for electro-driven phase perturbation processing, the phase perturbation effect under electro-acoustic coupling can be simulated, eliminating non-physically consistent pseudo-signals and thus improving the physical plausibility of sound source identification. Through transferable acoustic signature embedding, perturbation features are mapped to low-dimensional sparse vector representations, exhibiting good generalization ability and structural discriminative power.

[0024] Optionally, the sparse causal graph matching includes:

[0025] Perturbation point detection is performed on the voiceprint feature data to obtain voiceprint perturbation point data;

[0026] Sparse principal components are extracted from the voiceprint perturbation point data to obtain sparse principal component data.

[0027] Acoustic sequence graphs are constructed based on sparse principal component data to obtain acoustic sequence graph data;

[0028] Asymmetric transfer weights are calculated based on acoustic sequence graph data to obtain transfer weight graph data;

[0029] Environmental electrical weighting is performed on the transfer weight map data to obtain the causal spectrum data of sound source behavior.

[0030] This invention effectively improves the accuracy of identifying the causes and propagation paths of external disturbances through multi-stage disturbance analysis and causal modeling. Disturbance point detection can accurately locate abnormal response fragments from acoustic signature features, and combined with sparse principal cause extraction methods, it can filter out the truly significant core events from a large number of redundant disturbances. Through acoustic sequence graph construction and asymmetric transfer weight calculation, the system can quantify the causal dependencies between events, forming a dynamic event spectrum with clear directions. Further weight correction by incorporating electrical information from the power transmission environment ensures that the causal path not only depends on the inherent changes in the acoustic signature signal but also integrates the actual state of the electrical system, enhancing the physical consistency and engineering credibility of the causal chain.

[0031] Optionally, the sparse principal factor extraction includes:

[0032] A perturbation point map is constructed from the voiceprint perturbation point data to obtain perturbation point map data;

[0033] Local variation clustering is performed on the perturbation point map data to obtain perturbation point submap data;

[0034] Graph variational coding is performed on the perturbation point graph data to obtain sparse principal cause selection data;

[0035] The sparse main factor data is obtained by filtering the counterfactual confidence level of the selected data.

[0036] This invention effectively improves the accuracy and robustness of identifying core driving factors in voiceprint perturbations through graph structure modeling and variational inference. By constructing a perturbation point graph, the temporal, frequency, or morphological similarities between voiceprint perturbation segments are encoded into a graph structure, preserving the contextual relationships of perturbation events. Local change clustering can identify sub-graph regions with significant changes in perturbation intensity, highlighting potentially high-influence anomalous clusters. Graph variational encoding is used to perform low-dimensional embedding and sparse selection on candidate perturbation sub-graphs, enabling the structural extraction of the most behaviorally significant perturbation factors. This invention avoids the problem of traditional methods being overly sensitive to single-point amplitudes, and can extract high-value information with causal driving force from different noise backgrounds. Based on counterfactual confidence screening, perturbation response verification is performed on principal cause candidates to ensure that the retained perturbation points have actual causal driving effects on system behavior, thereby improving the accuracy and interpretability of causal modeling.

[0037] Optionally, S3 includes:

[0038] Intra-domain feature focusing processing is performed on the sound source behavior causal spectrum data to obtain intra-domain feature focusing data;

[0039] Perform dual-domain cross-attention mapping on the feature-focused data within the domain to obtain dual-domain cross-attention data;

[0040] Multi-channel dynamic anomaly enhancement is performed on dual-domain cross-attention data to obtain feature-enhanced data;

[0041] Semantic source separation is performed on the feature-enhanced data to obtain source separation and recognition data.

[0042] This invention utilizes a sound source behavior causal graph for intra-domain feature focusing, effectively enhancing the system's perception of key sound source regions and suppressing interference from invalid information. Subsequently, through dual-domain cross-attention mapping, deep collaborative modeling of temporal and frequency domain features is achieved, enhancing the model's adaptability to asynchronous and non-uniform signals. Multi-channel dynamic anomaly enhancement dynamically adjusts feature response weights based on anomaly confidence, improving the sensitivity to high-risk behaviors. Finally, a semantic sound source separation method maps the enhanced feature results into a structured output with semantic labels such as behavior category and spatial location.

[0043] Optionally, the dual-domain cross-attention mapping includes:

[0044] The dual-domain heterogeneous graph is constructed based on the feature-focused data within the domain, resulting in dual-domain heterogeneous graph data. The construction of the dual-domain heterogeneous graph involves time-domain graph transformation and frequency-domain graph transformation based on the feature-focused data within the domain.

[0045] Graph cooperative adversarial learning is performed on dual-domain heterogeneous graph data to obtain graph cooperative adversarial data.

[0046] A dual-domain attention weight field is calculated on graph cooperative adversarial data to obtain graph-weighted data;

[0047] Nonparametric streaming activation mapping is performed on graph-weighted data to obtain dual-domain cross-attention data.

[0048] This invention achieves deep fusion and precise alignment of temporal and frequency domain acoustic features through the construction of heterogeneous graph structures, graph collaborative learning, and nonparametric activation, significantly improving recognition accuracy and model robustness in industrial sound source environments. The dual-domain heterogeneous graph construction models temporal features and frequency structures as graph structures, effectively preserving local dependencies and topological relationships within different domains. Through a graph collaborative adversarial learning mechanism, cross-domain embedding alignment is achieved while maintaining the integrity of individual semantic features, enhancing the model's understanding of semantic relationships between heterogeneous information. Dual-domain attention weight field calculation generates a structure-aware response distribution in the graph embedding space, used to accurately capture key time-frequency collaborative activation regions. The use of a nonparametric streaming activation mapping method instead of the traditional softmax mechanism can adaptively generate high-confidence attention regions in the presence of strong interference or abnormally dense regions, achieving more interpretable attention distribution output.

[0049] Optionally, S4 includes:

[0050] Based on the sound source separation and identification data, a sound source behavior map and a sound source trend map are constructed to obtain sound source behavior map data and sound source trend map data, respectively.

[0051] Graph fusion is performed on the sound source behavior graph data and the sound source trend graph data to obtain the context graph data;

[0052] By performing causal nesting edge processing on the scenario graph data, we obtain the propagation structure graph data;

[0053] Based on the propagation structure diagram data, a contextual reasoning path is generated to obtain the sound source contextual map data.

[0054] This invention achieves a deep understanding of the context and reasonable expression of abnormal sound source events by constructing a multi-dimensional graph structure and fusing behavioral and trend information, significantly improving the system's intelligent analysis and early warning decision-making capabilities. Based on sound source separation and identification data, sound source behavior graphs and trend graphs are constructed separately, enabling the system to simultaneously capture the structural relationships and temporal evolution trends between event types, enhancing its three-dimensional modeling capabilities for multi-source anomalies. A unified context graph is generated through graph fusion, integrating static behavioral structures and dynamic trend features into a unified semantic space, providing an integrated graph foundation for reasoning. By introducing multi-hop driving relationships and propagation probabilities through causal nested edge processing, a propagation structure graph with causal chains and propagation levels is constructed, improving the accuracy of logical modeling for complex chain events. A traceable contextual reasoning path is generated based on the propagation structure graph, enabling structured causal tracking and impact range prediction of abnormal events.

[0055] Optionally, the causal nested edge processing includes:

[0056] Event time mapping is performed on the scenario diagram data to obtain scenario diagram mapping data;

[0057] Candidate causal edges are generated based on the scenario graph mapping data to obtain candidate causal graph data;

[0058] Nested causal chain processing is performed on the candidate causal graph data to obtain the causal graph data;

[0059] Propagation weights are calculated on the causal graph data to obtain the propagation structure graph data.

[0060] This invention achieves high-precision construction of the temporal relationship and causal logic of sound source behavior, significantly enhancing the system's reasoning ability and propagation path simulation capability for complex events. By normalizing the temporal order of nodes in the context graph through event time mapping, it provides a clear temporal sequence for causal judgment. Through candidate causal edge generation, potential causal links are constructed based on multi-dimensional features such as structural proximity and frequency similarity, effectively expanding the causal search space. Nested causal chain processing can identify multi-hop and multi-level triggering relationships, revealing potential higher-order causal paths and enhancing the system's depth of identification and modeling of chain-like abnormal behaviors. By calculating propagation weights, the driving strength and influence of each causal path are quantified, generating propagation structure graph data, providing a quantifiable and traceable structural foundation for contextual reasoning.

[0061] Optionally, this application also provides an intelligent early warning system for preventing external damage to transmission lines, used to execute the intelligent early warning method for preventing external damage to transmission lines as described above, wherein the intelligent early warning system for preventing external damage to transmission lines includes:

[0062] The acoustic sensing and acquisition module is used to acquire acoustic data of power transmission lines through acoustic sensors;

[0063] The voiceprint modeling and causal recognition module is used to perform sparse reconstruction of voiceprint features based on the acoustic data of transmission lines to obtain voiceprint feature data; and to perform sparse causal graph matching on the voiceprint feature data to obtain sound source behavior causal graph data.

[0064] The multi-domain attention sound source recognition module is used to perform dual-domain attention cross-recognition based on sound source behavior causal spectrum data to obtain sound source separation and recognition data.

[0065] The sound source scene map construction module is used to generate scene maps from sound source separation and identification data to obtain sound source scene map data.

[0066] The purpose of this invention is to acquire sound field data along a route using acoustic sensors combined with a dual-layer microphone array, enabling high-sensitivity sound source perception in environments such as low light and dense fog. It also effectively reconstructs sound source propagation characteristics through acoustic perturbation latent space modeling. The system performs sparse Bayesian reconstruction and transferable voiceprint embedding, combined with electrical signal-driven perturbations, to obtain more discernible multimodal voiceprint feature data. A sparse causal graph matching mechanism is used to extract the main driving perturbation points with strong driving force, constructing a sound source causal graph with temporal and electromagnetic response backgrounds. Utilizing dual-domain heterogeneous graph construction and graph collaborative adversarial learning techniques, cross-correlation attention weights between the frequency and time domains are extracted, achieving accurate semantic separation and feature enhancement of sound sources in high-interference scenarios. By fusing sound source behavior graphs and trend graphs, a causal propagation graph is established and a structured contextual reasoning path is generated, enabling not only the tracing of abnormal events but also the judgment of early warning levels and the deduction of response strategies. Attached Figure Description

[0067] Other features, objects, and advantages of this application will become more apparent from the following detailed description of the non-limiting embodiments, taken with reference to the accompanying drawings:

[0068] Figure 1 A flowchart illustrating the steps of an intelligent early warning method for preventing external damage to transmission lines according to an embodiment is shown.

[0069] Figure 2 A flowchart illustrating the steps of an acoustic sensing acquisition method according to an embodiment is shown.

[0070] Figure 3 A flowchart illustrating the steps of a voiceprint modeling and causal recognition method according to an embodiment is shown.

[0071] Figure 4 A flowchart illustrating the steps of a multi-domain attention sound source recognition method according to an embodiment is shown.

[0072] Figure 5 A flowchart illustrating the steps of a sound source scene map construction method according to an embodiment is shown.

[0073] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0074] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0075] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0076] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0077] Please see Figures 1 to 5 This application provides an intelligent early warning method for preventing external damage to transmission lines, the method comprising:

[0078] S1. Acoustic data of power transmission lines are collected using acoustic sensors;

[0079] In one embodiment, the system deploys a dual-layer acoustic sensor array at equal intervals along the transmission line. Each array consists of two layers of microphones. The first layer of microphones is used to collect ambient sound pressure signals; the second layer of microphones is configured as a frequency domain acquisition channel with directional response characteristics to enhance the capture capability of sound sources propagating in specific directions (such as low-to-mid frequency events like construction pile driving or heavy object impacts). The number of channels in each microphone layer can be set from 8 to 16 depending on the site deployment requirements. The sampling frequency is set to no less than 48 kHz. The system introduces a dynamic sampling mechanism based on a noise threshold: when the current amplitude of the acquired signal exceeds the background mean plus 2.5 times the standard deviation, an acoustic data sampling window is triggered. The system dynamically adjusts the gain parameters of the microphone array based on real-time transmission load data to ensure that the system can still clearly capture key acoustic events, such as low-to-mid frequency mechanical vibration signals, under operating conditions.

[0080] S2. Based on the acoustic data of the transmission line, perform sparse reconstruction of the voiceprint features to obtain voiceprint feature data; perform sparse causal graph matching on the voiceprint feature data to obtain sound source behavior causal graph data.

[0081] In one embodiment, the system is based on the collected acoustic data of the transmission line and synchronous electrical monitoring data (such as three-phase current). The system constructs a coupling mapping relationship between acoustic signature features and current perturbation, and generates sparse acoustic signature feature vectors and causal spectrum data of sound source behavior based on this relationship. The system performs acoustic-electrical coupling modeling. Sparse Bayesian learning is used to construct the input feature matrix based on the rate of change of current, with the objective of predicting the corresponding acoustic spectrum envelope signal. The objective function of the coupling modeling includes an error term and a sparse regularization term, i.e. ,in For parameters This is the operator for minimizing variables. Let be the target acoustic spectrum envelope signal vector. The input feature matrix is ​​constructed from the rate of change of current. This is a sparse regression coefficient vector. These are the weighting coefficients (regularization parameters) for the sparse regularization term. The system performs a Fast Fourier Transform (FFT) on the current signal in the frequency domain to extract phase perturbation features at current abrupt changes (such as those caused by driver noise or unsteady-state discharge). The system constructs a perturbation-driven FM acoustic map. In the high-dimensional feature extraction stage, a speaker embedding network trained under a transfer learning structure (such as the SpeakerNet model trained based on the Triplet Loss function) is used to extract the embedding representation of each acoustic signal segment. The final output is a sparse speaker feature vector sequence with 128 or 256 dimensions. The system performs sparse causal graph matching processing on the above speaker feature sequence. Abnormal perturbation points in the speaker sequence are identified based on local outlier factors or isolated forest algorithms; these points are considered potential sound source behavior nodes. A directed causal graph G=(V,E) is constructed using the identified perturbation points, where each node represents a sound source perturbation event, and the direction of the edges indicates the temporal causal relationship between perturbations. The system performs asymmetric weighting calculation on the edge weights in the graph. Specifically, for any two nodes, the initial weight of the edge is calculated from the Kullback-Leibler divergence between the corresponding speaker distributions, and a weighted time difference is introduced. The transition probability weights are defined using the Sigmoid function for the combined variable with the rate of energy change. : ,in This is an adjustment coefficient used to control the impact of disturbance levels on edge weight calculation. The system performs causal correction based on a current threshold criterion. If the current fluctuation within a time period corresponding to a causal edge... If the value exceeds 15%, it is determined that it has a strong electromagnetic interference correlation, and the system will increase the causal weight of that side accordingly.

[0082] S3. Perform dual-domain attention cross-identification based on the sound source behavior causal graph data to obtain sound source separation and identification data;

[0083] In one embodiment, the system performs feature focusing operations in two feature subspaces: the time domain and the frequency domain. The time-domain features mainly include the intensity and duration of energy transitions in the sound source behavior; the frequency-domain features include frequency stability (e.g., frequency shift variance) and frequency quality factor (Q value). For each feature subset within the aforementioned domains, the system uses max pooling to extract the most representative feature values, forming a time-domain feature vector. With frequency domain eigenvectors The system constructs a dual-domain heterogeneous graph structure. Specifically, it will respectively... and Mapped to independent graph structures (Time Domain Graph) and (Frequency domain graph), where each node represents an independent sound source feature point. Edges in the graph are constructed based on the cosine similarity between feature vectors: if the similarity between two nodes is greater than 0.7, a connection edge is established. This method ensures that only high-similarity associations are retained in the graph structure, effectively improving the accuracy of subsequent clustering. The system performs cross-domain semantic alignment on the constructed dual-graph structure. The system uses a dual-discriminator generative adversarial network (GAN) structure, setting discriminators for the time-domain graph and the frequency-domain graph respectively. The discriminator aims to distinguish the source domain of the graph, while the generator learns how to map the graph embedding representations of the two domains to a unified semantic space, thereby achieving cross-domain adversarial alignment.

[0084] After semantic alignment, the system performs graph attention processing within the graph structure. Attention weights are assigned to the connections between any two nodes in the graph, generating a cross-domain attention distribution matrix. Represents a node To the node The system emphasizes the credibility and importance of propagated information. It employs a non-parametric density-driven mechanism instead of the traditional softmax normalization method. Based on node density distribution and adjacency feature overlap, a density response function is constructed to determine the confidence propagation path and activation level of each node in the overall graph. The system generates feature masks based on cross-attention maps to guide the separation and classification of sound sources. The masks are applied to the original spectral data and temporal structure, corresponding to sound source categories such as "mechanical vibration," "footsteps," and "lightning strikes." Through this attention-guided mechanism, the system achieves accurate extraction and labeling of multiple target sound sources.

[0085] S4. Generate a scene map from the sound source separation and identification data to obtain sound source scene map data.

[0086] In one embodiment, a sound source behavior graph is constructed. The system establishes an adjacency matrix for the behavior graph based on the sound source identification label sequence (e.g., continuously identified labels such as "mechanical noise," "knocking sound," and "thunder sound"). The elements of the adjacency matrix represent the co-occurrence relationship or transition probability of different types of sound sources in adjacent time periods, thus forming a directed behavior graph with sound source categories as nodes and behavior transitions as edges. The system then constructs a sound source trend graph. By modeling the sound source identification time series, a recurrent neural network structure (such as a gated recurrent unit, GRU) is introduced to capture the dynamic evolution trend of the sound source signal. The system labels the trend attributes of each node based on the trend encoding results output by the neural network, mainly including trend types such as "continuous enhancement," "periodic fluctuation," and "relative stability." After completing the construction of the behavior graph and trend graph, the system performs a graph structure fusion operation. To maintain the integrity and semantic consistency of the original graph structure information, a structure-preserving graph attention fusion is adopted. Based on node feature similarity and graph topology consistency, fusion weights are calculated, and the edge weights and node attributes of the behavior graph and trend graph are weighted and merged to generate a unified sound source context graph structure. The system performs nested causal edge processing on the fused graph structure. Timestamps are added to edge relationships in the graph based on the event occurrence time sequence. Based on temporal sequence and frequent co-occurrence strength, the system filters a set of candidate causal edges. Propagation potential is calculated for each pair of candidate causal nodes. and The causal strength score between them is calculated using the following formula: ,in For nodes With nodes The causal strength score between them Represents a node and The time interval between This represents the co-occurrence frequency of the event pair in historical data. Represents a node The event entropy (used to measure its specificity as a causal outcome) is calculated. A higher score indicates a higher confidence level that the edge is a nested causal edge. After the nested causal edges are constructed, the system generates causal inference paths in the fusion graph. A depth-first search (DFS) is used to traverse the propagation paths, and a weight gating mechanism is used to filter the propagation intensity in each path, retaining only the top-K inference paths with the highest scores as the output scenario causal graph. The output sound source scenario graph data includes the following core elements: timestamps of sound source events; identified sound source labels, such as "hitting" or "high-frequency howling"; the order and causality between events; and automatic assignment based on the inference path weights and propagation patterns, such as "low risk," "medium risk," or "high risk."

[0087] Optionally, S1 includes:

[0088] S11. Obtain power transmission load data;

[0089] In one embodiment, by interfacing with a power transmission monitoring and data acquisition system, the operating parameters of the transmission line, including three-phase current, are collected in real time. ,Voltage Power factor and load change rate Data is recorded every second to create a dynamic window. ,in For time The load dynamic window, For time Active load value at any given time. For time Active load value at any given time.

[0090] S12. Based on the power transmission load data, perform sampling frequency processing and directional sensitivity processing to obtain sampling sensitivity data and directional sensitivity data respectively.

[0091] In one embodiment, the sampling frequency processing (time-domain adaptive adjustment) sets the system's base sampling frequency to 48kHz as the default sampling configuration under normal operating conditions. This is based on the load fluctuation dynamic factor. The real-time values ​​are used to determine if there are drastic fluctuations in power load. When Exceeding the preset sensitivity threshold Time (e.g.) (=0.08), the system triggers the sampling frequency increase mechanism. Specifically, the adjusted sampling frequency is calculated as: Sampling Frequency = Basic Sampling Frequency (1+adjustment coefficient) Volatility factor), i.e. ,in Sampling frequency, Based on the sampling frequency (48kHz). This is an adjustment coefficient, and its value ranges from 1.5 to 2.0. This represents the load fluctuation factor at the current moment. Directional sensitivity processing (spatial domain weight adjustment) involves constructing a multi-directional sensing weight model in the spatial dimension, combining the actual route of the transmission line, surrounding construction locations, road distribution, and terrain undulation information. This model is based on directional vectors and defines several directional units. to Each directional unit corresponds to one main lobe of the acoustic array. The system assigns a weight value to each directional unit, forming a normalized directional weight vector: These weight values ​​are used to adjust the sensitivity control of the microphone array in different directions. The higher the weight value, the higher the sensing priority of the direction. The adjustment of directional sensitivity is achieved through hardware methods (such as adjustable MEMS arrays) or software algorithms (such as beamforming-based directivity enhancement algorithms).

[0092] S13. Based on the sampling sensitivity data and directional sensitivity data, sampling is performed through a dual-layer microphone array unit preset along the transmission line to obtain distributed sound field monitoring data. The dual-layer microphone array unit is a microphone array unit with the first layer for environmental sound pressure perception and the second layer for target frequency domain capture.

[0093] In one embodiment, the system deploys an array unit every 100 meters. Each group includes two layers of microphone arrays. The first layer (upper array) is mainly used to sense ambient background sound pressure and is suitable for capturing low-frequency signals between 20Hz and 1kHz, a frequency band commonly found in wind noise, electromagnetic noise, or traffic background noise. The second layer (lower array) focuses on the high-frequency region between 1kHz and 8kHz and is used to detect localized instantaneous high-frequency disturbance signals such as construction knocks or metal impacts. Each array layer is configured with 8 microphone elements, forming a uniformly distributed or directionally controllable array structure, which can be programmably beam-adjusted according to directional sensitivity weights. The system organizes the raw sound pressure signals from all microphone arrays into a tensor structure according to their spatial coordinates, time frame sequence, and frequency distribution.

[0094] S14. Construct the acoustic disturbance hidden space based on the distributed sound field monitoring data to obtain the acoustic disturbance hidden space data.

[0095] In one embodiment, the original sound field data tensor ( Each local segment in the image is extracted as a high-dimensional sound field feature vector. These vectors are then mapped using a nonlinear dimensionality reduction algorithm (such as t-distributed random neighborhood embedding t-SNE or unified manifold approximation mapping UMAP) to a low-dimensional space representation of dimension 3, denoted as the local embedding vector: , indicating the first The system maps the positions of each sound field segment in the latent space. It constructs a perturbed connected graph based on the dimensionality-reduced embedding vectors. For any two sound field segments... and The connectivity strength weight is defined as ,in and Each is a fragment With fragments Embedded vectors in the latent space; Its Euclidean distance represents the similarity of perturbations or the degree of similarity in their propagation paths; The bandwidth parameter controls the rate at which adjacency weights decay with distance. The system integrates the embedding results of all local perturbations into a three-dimensional tensor structure.

[0096] S15. Based on the acoustic disturbance hidden space data, model the sound source propagation structure to obtain the acoustic data of the transmission line.

[0097] In one embodiment, the system constructs a structured model describing the propagation path of sound source disturbances in space and time based on the aforementioned acoustic disturbance latent space data. The system constructs a sound source propagation graph based on the acoustic disturbance latent space tensor, defined as follows: ,in This represents the set of perturbation nodes, where each node corresponds to a local perturbation block in the acoustic latent space (i.e., an anomalous sound source segment observed at a specific time and spatial location). Edges represent perturbation propagation and are used to connect perturbation blocks with spatiotemporal adjacency, such as establishing an edge between two nodes that are temporally continuous and spatially close. The system assigns each edge in the graph... Introduce a propagation weight function, such as propagation weight. ,in and Representing nodes respectively and nodes The acoustic energy value corresponding to the perturbation segment; Represents a node and nodes The difference on the time axis reflects the propagation delay; This is the time decay coefficient (positive value); Indicates the disturbance originating from the node Passed to node The relative intensity of the sound source disturbance is determined. Based on the aforementioned propagation map, the system identifies the main path of the sound source disturbance in the spatiotemporal map. Through path analysis and structure factor calculation, regions with high energy coupling and strong inter-node connectivity in the propagation path are identified. These highly coupled paths are then connected in series to form the main propagation chain, such as the acoustic trajectory formed by continuous hammer blows propagating along the ground surface. The system outputs a structured acoustic data volume.

[0098] Optionally, the sparse reconstruction of the voiceprint features includes:

[0099] S21. Perform sparse Bayes regression based on the electrical data of the transmission line to obtain electrical sensing data;

[0100] In one embodiment, the sparse Bayesian regression takes the form of: ,in Let time be the target variable. The predicted output (electric sensing value). This is an index variable used to iterate through the feature basis functions. This represents the total number of basis functions (number of features). For the first The weight coefficients corresponding to each basis function. For the first Each basis function in the input The response value on In time The electrical input feature vector, The Gaussian noise term represents the model error. With a mean of 0 and a variance of Gaussian distribution, The noise standard deviation is used to control the fluctuation of the regression model. Representative change patterns after sparsification are extracted from high-dimensional, multi-source electrical perturbation variables, and then used as potential factors driving acoustic changes to obtain electrical sensing data.

[0101] S22. Construct a frequency-modulated drive spectrum from the electrical sensing data to obtain audio spectrum data;

[0102] In one embodiment, electrical sensing data is mapped to an acoustic frequency feature domain (such as the Bark band) to construct an electric drive spectrogram. ,in In time and frequency The electrical drive spectral response value at that location, For time parameters, For frequency parameters, Index of potential electrical driving factors (from the dimensions of the electrical perception tensor). To drive factors The corresponding modulation weights reflect the strength of their influence on the target frequency components. This is a frequency-modulated kernel function used to simulate the factor. In time For frequency The system utilizes modulation response models, such as multi-channel Gabor filter kernels. It performs multi-scale decomposition of the spectral signal, such as wavelet packet decomposition, to obtain complete low-frequency and high-frequency sub-bands. Correlation analysis is performed on all decomposed frequency bands to calculate their response correlation index with electrical driving factors. For frequency bands with significant correlation (e.g., correlation coefficients exceeding a threshold), they are retained as target frequency feature segments and incorporated into the acoustic spectrum construction. The system outputs the acoustic spectrum data generated by modulation modeling.

[0103] S23. Obtain transmission line current data;

[0104] In one embodiment, the three-phase current value is acquired in real time using a current transformer (CT) or a digital energy meter. , synthesized into , To synthesize the equivalent current, representing the root mean square value of the three-phase current, the instantaneous phase is obtained through Hilbert transform. The data is standardized to current tensor input. , This refers to the phase component.

[0105] In one embodiment, data is collected through a power transmission line cloud platform to obtain real-time current data.

[0106] S24. Perform electric-driven phase perturbation based on transmission line current data and audio spectrum data to obtain perturbation spectrum data;

[0107] In one embodiment, a joint modulation model is constructed based on transmission line current data and audio spectrum data: ,in For perturbation map data, representing time... and frequency The acoustic response after being subjected to an electrically driven disturbance. The raw audio spectrum data is used as the acoustic drive input. The time variable represents the horizontal axis coordinate in the graph. For frequency channels, the vertical axis coordinates in the spectrum are represented. Let be the modulation function, representing the modulation effect of the current phase on the audio spectrum. The instantaneous phase function is calculated based on the current signal. `<perturbine phase offset constant>` is used to control the initial offset of the phase perturbation. A fine-grained perturbation kernel function is constructed to perform perturbation modulation on the high-frequency components. This perturbation kernel function is defined as follows: ,in This is a perturbation kernel function, a function structure that modulates high-frequency components. For time variables, For frequency channels, It is a cosine function. Pi is a constant, with a value of approximately 3.1416. The small time drift caused by current disturbance originates from disturbance behaviors such as current abrupt changes and switching. The system outputs disturbance spectrum data under electric drive.

[0108] S25. Perform transferable voiceprint embedding on the perturbation map data to obtain voiceprint feature data.

[0109] In one embodiment, a deep embedding network model trained based on Triplet Loss, such as ECAPA-TDNN, is used to process the perturbed audio spectrum. Embedding operations are performed to generate semantically consistent voiceprint features. The system then performs transfer regularization, such as... ,in The transfer consistency loss function is used to measure the spatial distance between the acoustic signature vectors of the strongly constrained perturbation state and the baseline state, ensuring that the acoustic representation still has transfer consistency under electrical disturbance scenarios. This represents the embedding model used for voiceprint extraction; Input the perturbation map; A reference speaker embedding is constructed under normal operating conditions; the system undergoes domain adversarial training, and a discriminator module is introduced to constrain the speaker embedding to maintain consistent distribution across different electrical disturbance domains. The system operates in the embedding space. Extract a fixed-length low-dimensional voiceprint feature vector, set to 192 dimensions, to obtain voiceprint feature data.

[0110] Optionally, the sparse causal graph matching includes:

[0111] Perturbation point detection is performed on the voiceprint feature data to obtain voiceprint perturbation point data;

[0112] In one embodiment, the system constructs a set of disturbance points. Each perturbation point Include The voiceprint feature vector of the perturbation point represents its semantic features in the embedding space; The timestamp corresponding to each disturbance point indicates the time and location of its occurrence. For each disturbance point... The system maps it to a node in the graph. and the corresponding voiceprint vector As an embedded attribute of node features, the system constructs a graph structure based on edge weight functions of feature similarity and temporal proximity. For any two perturbed points... , The edge weights are defined as follows: ,in For nodes and Edge weights between them For the disturbance point The voiceprint feature vector, For the disturbance point The voiceprint feature vector, This represents the squared Euclidean distance between the voiceprint feature vectors of the nodes. For feature similarity control parameters, Represents distance on the timeline. For the disturbance point timestamp, For the disturbance point timestamp, This is a time decay factor; the edge weight function integrates the content similarity and temporal proximity of the perturbations, reflecting potential propagation or coupling trends. The system only retains edge weights greater than a threshold. The connecting edges, i.e., only retain those that satisfy... edge The system outputs graph data with node attributes and a sparse edge structure.

[0113] Sparse principal components are extracted from the voiceprint perturbation point data to obtain sparse principal component data.

[0114] In one embodiment, the system constructs a perturbation graph structure. Each identified perturbation point is used as a graph node, and nodes are defined as follows: ,in Indicates the time index of the disturbance occurrence. This represents the voiceprint feature vector at that moment. The edge weight between any two nodes in the graph is determined by the decrease in information entropy between their feature sequences or the Kullback-Leibler (KL) divergence. The system employs a graph variational autoencoder to perform embedded reconstruction and sparse principal component extraction on the perturbation graph. The perturbation graph is encoded to learn the latent space representation of each perturbation node; Gaussian distribution sampling is performed on the embedding of each node to generate a latent representation with perturbation uncertainty; during the reconstruction process, the system calculates the attribution weight of each node to the overall graph perturbation result and ranks the nodes by importance; the top nodes are selected from the ranking of attribution weights. The system identifies several high-weight nodes as potential causal primary causes. A counterfactual confidence screening process is then performed. For each candidate primary cause node... Construct a "deactivation" version graph structure, that is, nodes The influence of graph structure is masked; the difference in predicted loss between the original graph structure and the deactivated graph structure is compared, and the counterfactual influence is defined as... ,in Represents a node The counterfactual influence This represents the model prediction error or reconstruction loss. To estimate the expected value of the loss under the variational autoencoder model, For nodes Deactivate the version graph structure. This represents the original, complete perturbation diagram structure. If... If a node has a significant negative impact on the perturbation result, then that node is retained. The system outputs a sparse set of principal causes.

[0115] Acoustic sequence graphs are constructed based on sparse principal component data to obtain acoustic sequence graph data;

[0116] In one embodiment, a sparse principal set is used. Each principal node in the sequence is represented as a set of nodes in the acoustic sequence graph, denoted as . Each node Corresponding to a specific time index This represents a key voiceprint perturbation event. The node set is sorted according to the timestamp order; for each pair of key nodes... If their time difference satisfies ,in To set the maximum disturbance propagation interval threshold (based on scenario experience, such as 10 seconds), add a line from... point to Directed edge This is used to represent the perturbation propagation relationship. Each edge... initial weights Based on the statistical dependencies between perturbation events, candidate indices include covariance mobility, representing the covariance trend of perturbation intensity between two events within a short time window; and cross-spectral coherence, representing the degree of coordination between two perturbation signals in the frequency domain (the cross-spectral density ratio calculated by the Welch method). The constructed graph is a directed acyclic graph to avoid cyclic logic in perturbation relationships and ensure causal order; the output acoustic sequence graph is represented as a triplet structure.

[0117] Asymmetric transfer weights are calculated based on acoustic sequence graph data to obtain transfer weight graph data;

[0118] In one embodiment, the system uses transfer information gain or Granger causality strength as a modeling tool to measure the difference in the directionality of information flow between any two nodes. For each pair of directed edges existing in the graph... Calculate from node respectively arrive Information transfer intensity From node arrive Information transfer intensity The difference between the two is used as the asymmetric transition weight of the edge, such as... ,in Representing an edge The net transfer direction weight. If If , it indicates the existence of a clear positive causal propagation relationship, and that edge is retained; if ,in For setting a causal significance threshold (e.g.) If the edge value is less than or equal to 0.01, it is considered to have no significant transition relationship, and the system removes this edge to reduce causal noise interference in the graph. To prevent the graph's propagation ability from being weakened due to extremely sparse structure, the system employs weak diffusion correction, applying this correction to each retained edge. Redefine its weighted edge weights The calculation method is as follows ,in The edge weights are adjusted for diffusion, representing the weights actually used for propagation in the graph. The diffusion factor controls the fusion ratio between the original asymmetry and the average distribution (e.g., a value of...). ), For asymmetric transfer weights, For nodes The set of outgoing adjacent nodes, The size of the set, The term represents a uniform diffusion term, ensuring that all nodes in the graph have at least a weak propagation capability. The output is a weighted directed graph structure, i.e., transition weight graph data.

[0119] Environmental electrical weighting is performed on the transfer weight map data to obtain the causal spectrum data of sound source behavior.

[0120] In one embodiment, the system constructs an electrical weighting graph. , used to describe at each moment The intensity of environmental electrical disturbances is used as a background disturbance control factor; electrical disturbance indicators may include, but are not limited to, voltage fluctuation rate, current anomaly ratio, harmonic distortion, etc., which are uniformly normalized by the system to generate an environmental electrical disturbance influence factor function. Its value range is [0,1], and it is used to characterize the real-time impact of electrical disturbances on the system. For each edge in the acoustic transfer weight graph... The system adjusts its weighting to incorporate the edge weights of electrical interference, as shown in the specific calculation formula. ,in Indicates the node in the previous stage (acoustic transfer diagram) arrive The transfer weight; Represents a node Current Time The influence factor of electrical disturbance; The edge weights are the values ​​after incorporating electrical factors. This ensures that the propagation intensity of each hop in the causal propagation chain reflects the superimposed effects of acoustic and electrical disturbances. The system performs multimodal embedding processing on the fused weighted directed graph. Through a tensor product expansion mechanism, the acoustic causal node representation and the electrical background state are modally fused to generate a joint embedding vector. This embedding vector, combined with structural adjacency relationships, preserves the dual causal characteristics of nodes under both acoustic topology and electrical disturbance distribution. Based on the joint embedding representation between nodes in the fused graph, the system employs structure-preserving clustering algorithms (such as spectral clustering and graph convolutional clustering) to partition the causal structure, forming several propagation clusters with causal consistency characteristics. Each cluster represents a set of similar disturbance sources or behavioral paths. The system outputs the constructed causal graph structure of sound source behavior.

[0121] Optionally, the sparse principal factor extraction includes:

[0122] A perturbation point map is constructed from the voiceprint perturbation point data to obtain perturbation point map data;

[0123] In one embodiment, the set of input disturbance points Each perturbation point , including its voiceprint feature vector and timestamp Each perturbation point As a graph node Node characteristics are Edges are constructed using temporal distance and feature similarity: ,in For nodes and The edge weights between them It is an exponential function. For the disturbance point The voiceprint feature vector, For the disturbance point The voiceprint feature vector, The scaling parameter for the feature distance. For the disturbance point timestamp, For the disturbance point timestamp, The scaling parameter for the time difference retains only the edge weights. The edges form a sparse perturbation graph. Output graph structure data containing node features and sparse edges. .

[0124] Local variation clustering is performed on the perturbation point map data to obtain perturbation point submap data;

[0125] In one embodiment, the system in the disturbance graph The above performs a graph-based clustering algorithm, using graph embedding algorithms (such as DeepWalk or GraphSAGE) to cluster each node in the graph. The mapping is represented as a low-dimensional vector, denoted as This preserves the structural adjacency relationships and feature context for all embedded vectors. Clustering algorithms, such as spectral clustering or K-means clustering, are executed to divide the perturbation subgroups. The system performs Laplacian spectral entropy analysis to evaluate the structural information content of the perturbation graph. High spectral entropy indicates complex structure and loose distribution, while low spectral entropy indicates strong separability. If the spectral entropy is below a set threshold, K is increased to refine the cluster boundaries; if the spectral entropy is high, K is appropriately decreased to preserve dense cluster regions. The obtained cluster number K represents the number of potential perturbation sub-patterns in the perturbation graph. For each cluster... The system extracts the corresponding node set. and its internal connected edge set Constructing subgraph structure ,in Indicates the first A local perturbation submap For the first A set of nodes in a perturbation subgraph. For the first The set of edges of a perturbed subgraph; nodes retain their original voiceprint feature embeddings, and edge weights inherit the connection strength in the perturbed point graph.

[0126] Graph variational coding is performed on the perturbation point graph data to obtain sparse principal cause selection data;

[0127] In the embodiment, for each perturbation subgraph The system uses it as input to a graph variational autoencoder (GraphVAE) for each node. Encoded as a latent variable vector: ,in For nodes The latent variable vector representation, For the encoder module of a graph variational autoencoder, latent variables are extracted from the nodes. For the first A perturbation subgraph, For the perturbation subgraph, the first There are 10 nodes; the encoder output corresponds to the potential distribution: ,in For nodes Potential distribution, and Representing nodes respectively The mean vector and covariance matrix of the corresponding latent variables; latent variables This expresses the high-level semantic features of nodes within the perturbed subgraph structure. The system aims to maximize the lower bound of evidence (ELBO) and optimizes the following loss function: ,in The loss function (negative ELBO) of the graph variational autoencoder. For the potential distribution With prior distribution of divergence, For nodes Potential distribution, Let the prior distribution of the latent variable be the standard normal distribution. , For the potential distribution Expectations For the first A perturbation subgraph structure. For the latent variable vector, Let represent the reconstruction probability of the perturbation subgraph given the latent variables. The first term represents the distribution of the latent variables and the prior. of The first term, divergence, encourages sparsity and interpretability; the second term represents the likelihood of structural reconstruction of the perturbed subgraph under latent variable conditions, reflecting the fidelity of node embedding. The system performs a multiplication table for each node. Calculate its structural restructuring impact data, such as ,in This represents the adjacency structure of the original perturbed subgraph; Indicates the removal of a node. The subgraph obtained after reconstruction; The Frobenius norm is used to quantify the degree of change in the overall graph structure. For nodes Structural restructuring impact data, scoring The larger the value, the more likely it is to be a node. The higher the contribution to the reconstruction of the original image structure, the more likely it is to be the main cause of the perturbation. The system outputs a sparse candidate set of main causes: Each element in the set includes Candidate main cause node; : Its corresponding importance score.

[0128] The sparse main factor data is obtained by filtering the counterfactual confidence level of the selected data.

[0129] In one embodiment, for each candidate main cause node The system constructs its corresponding counterfactual graph structure. Specifically, from the original perturbation subgraph Remove node The modified graph structure is obtained by combining the graph structure with all its associated edges. ,in To remove a node The graph structure following its associated edges (counterfact graph). To remove a node The subsequent set of nodes, To remove nodes The set of edges following all adjacent edges. Using a pre-defined recognition model (e.g., a voiceprint classifier to identify sound source type, action category, etc.; a causal inference module to estimate the causal influence between nodes; the aforementioned recognition model is constructed by performing graph convolution, graph attention weighting, pooling layers, and fully connected layers and softmax processing based on historical data) to perform predictions on the original graph and the counterfactual graph respectively, obtaining the corresponding outputs. and Calculate the predicted difference value ,in This represents the model's output function on the input graph structure, such as classification probability or causal score. If the node... Differences in the impact on model output If so, then retain that node as the high-confidence primary cause, where This is a preset confidence threshold. If the number of nodes meeting the conditions exceeds the maximum capacity... Then according to Sort the values ​​from highest to lowest and select the top ones. The system outputs a sparse set of primary cause nodes after filtering, identifying the most influential primary cause node.

[0130] Optionally, S3 includes:

[0131] S31. Perform intra-domain feature focusing processing on the sound source behavior causal spectrum data to obtain intra-domain feature focusing data;

[0132] In one embodiment, for any node If the number of connected edges within the preceding and following 30-second time windows is less than a preset threshold If it is, then it is considered a low-activity node and will be removed. That is, if ,but It is not included in the focus graph structure. Represents a node exist Instant Number of connected edges per second for The corresponding time parameters. For each node. The system calculates the salience of its behavioral labels. , defined as ,in This represents the average connection strength of a node (e.g., the average weight of the connected edges). This represents the causal direction of a node (i.e., the absolute value of the difference between outgoing and incoming edges, reflecting the degree of causal dominance of that node). The average connection strength weight for nodes is 0.55. This represents the causal orientation weight data for the nodes, with values ​​such as 0.45. For label saliency data, the system is categorized as follows: Sort by score in descending order and select the top score. (Preset node threshold) nodes are used to construct a feature-focused subgraph within the domain. The system will then score the nodes. Node set and its corresponding connecting edges Feature-focused plot within the constituent domain: This subgraph preserves the key interaction chains in the causal relationship of sound source behavior.

[0133] S32. Perform dual-domain cross-attention mapping on the feature-focused data within the domain to obtain dual-domain cross-attention data;

[0134] In one embodiment, the system constructs two paired graph structures from the domain-focused data, namely a time series graph. Node set The temporal feature representation corresponding to the sound source behavior event; edge set The edge weights are the time interval differences between nodes, representing the chronological order of events. Frequency graph Node set Corresponding to the feature representation of the same node in the frequency domain; edge set The edge weights are spectral similarity, such as cosine similarity calculated based on spectral feature vectors. The system maps nodes of the two graphs one-to-one, meaning each node... It also has a time-domain representation A frequency domain representation The system supports all cross-domain nodes. The attention weights are calculated above, such as ,in Represents time-domain nodes The query vector is a linear mapping matrix. Acting on its eigenvectors get, Represents frequency domain nodes The key vector is formed by the linear mapping matrix. Frequency domain characteristics Multiplying them together, we get For frequency domain nodes The key vector, For frequency domain nodes, The normalized attention coefficient measures the number of nodes in the time domain. For frequency domain nodes The degree of dependence. The system uses a non-parametric streaming activation method to weight and aggregate frequency domain features, generating cross-domain activation representations, calculated as follows: ,in This refers to the attention weight coefficient (cross-domain). It is the frequency domain feature mapping matrix. For activation function, This is the fused dual-domain representation, used to replace or enhance the original temporal domain node representation. The system output, combined with the dual-domain attention mechanism, yields a graph structure, defined as... ,in This represents the updated set of node features, which integrates the moderating effect of frequency domain attention on time domain features; It is the original set of temporal and frequency domain edges, preserving the structural context of both domains.

[0135] S33. Perform multi-channel dynamic anomaly enhancement on the dual-domain cross-attention data to obtain feature-enhanced data;

[0136] In one embodiment, the system extracts time-series features from the behavior and structural attributes of graph nodes across multiple channels, specifically including channel 1, the voiceprint intensity change sequence. Channel 1: Reflects the temporal change in the amplitude of the node's voiceprint feature vector; Channel 2: Causal propagation path length sequence Records the dynamic length of the propagation chain of a node along a causal path; Channel 3: Time-varying attention entropy fluctuations. Entropy at each time step is calculated based on cross-domain attention distribution; Channel 4: Spectral energy shift index This measures the temporal offset of the node's spectral centroid, used to detect high-frequency anomalies. The system performs this for each channel. Calculate the maximum mutation rate within the time window , ,in For channel In time eigenvalues, To obtain Maximum value handler, For channel In time eigenvalues, This indicates the magnitude of the largest behavioral mutation detected in this channel. Set the anomaly detection threshold for each channel. If satisfied Then the channel Nodes are marked as abnormal channels. For nodes falling into abnormal channels, the system performs weight enhancement processing on the nodes. Original label or embedding weight Update ,in For nodes Weights after abnormal enhancement For nodes The original label or embedding weight, For channel The abnormal enhancement factor controls its amplification factor in the overall weight update; if a node falls into multiple abnormal channels, multiple channels can be accumulated or weighted and merged. The system outputs a feature dataset enhanced with multi-channel dynamic anomalies.

[0137] S34. Perform semantic source separation on the feature-enhanced data to obtain source separation and recognition data.

[0138] In one embodiment, the system uses a graph attention network as the sound source separation model, and the input is enhanced graph structure data. The node features have been integrated with anomaly detection results from multiple channels, including time domain, frequency domain, and attention dimension. Each node... The feature vectors include multimodal behavior embeddings and weight enhancement results. For each node The system outputs the probability distribution of each sound source type. in , For nodes Belongs to the The probability (confidence level) of a sound source. For nodes Sound source category tags, This is a sound source type category index, with a value range of [value range missing]. , representing a node Belongs to the The system assigns confidence scores to sound sources, which can include mechanical impacts, human voices, animal activity, and environmental noise. Based on the sound source classification results, the system further performs behavior chain separation to identify the propagation structure of each independent sound source in the graph. Nodes with the same classification label and continuous path relationships in the graph structure are grouped into the same sound source behavior chain, i.e., if the set of nodes... satisfy And a path exists. If the sound sources are similar, they are grouped into the same subgraph. The system uses the following metrics to jointly evaluate the sound source separation quality: information entropy, which reflects the confidence distribution of sound source identification; connectivity, which measures the completeness and clustering of the sound source behavior chain in the graph; and propagation consistency, which determines whether the identification results conform to the cause-effect propagation path. The system outputs a structured sound source separation and identification dataset.

[0139] Optionally, the dual-domain cross-attention mapping includes:

[0140] The dual-domain heterogeneous graph is constructed based on the feature-focused data within the domain, resulting in dual-domain heterogeneous graph data. The construction of the dual-domain heterogeneous graph involves time-domain graph transformation and frequency-domain graph transformation based on the feature-focused data within the domain.

[0141] In one embodiment, each sample is considered as a time-domain node, denoted as... , representing the acoustic events within the i-th time window; for any two nodes , The edge weights are defined based on their time difference. ,in Nodes in the time-domain graph and The boundary rights between them For the natural index term, For nodes timestamp, For nodes timestamp, The hyperparameters are adjusted for scale in the time domain; the larger the edge weight, the greater the temporal similarity between two nodes, thus constructing a time-domain graph. The dominant frequency band of each sample is represented as a frequency-domain node. , used to represent the focal center of the sample in the frequency dimension; frequency domain edge weights are constructed based on the Euclidean distance of the spectral features between samples. ,in For nodes in the frequency domain graph and The boundary rights between them For the natural index term, For nodes Spectral feature vector, For nodes Spectral feature vector, This is a scaling factor for spectral distance; the closer the distance, the higher the frequency domain similarity, and the larger the edge weight. By merging the two types of graphs mentioned above, a unified bi-domain graph structure is defined. ,in For all time-domain graph nodes; For all frequency domain graph nodes; These are the edge sets of the time-domain graph and the frequency-domain graph, respectively. To establish corresponding edges between nodes of the same data sample in two graphs, we represent the nodes. A one-to-one mapping is used to achieve cross-domain semantic alignment. The output is a two-domain heterogeneous graph structure. .

[0142] Graph cooperative adversarial learning is performed on dual-domain heterogeneous graph data to obtain graph cooperative adversarial data.

[0143] In one embodiment, the system respectively analyzes the time-domain graphs. With frequency domain diagram Construct a graph convolutional neural network encoder, denoted as the temporal encoder. Frequency domain encoder ,in , Let represent the high-dimensional feature embedding results after encoding the time-domain and frequency-domain graphs, respectively. An adversarial discriminator module D is defined to determine which graph domain (time domain or frequency domain) the input embedded features originate from. The discriminator's training objective is to maximize the graph domain recognition accuracy, while the encoder's training objective is to make the encoding results of the two graph domains as similar as possible, thereby confusing the discriminator and achieving feature domain alignment. The system uses the following adversarial loss function: ,in To counteract the loss function, To The expected operator, For the discriminator pair The discriminant output, This represents the high-dimensional embedding features of the time-domain graph. To The expected operator, For the discriminator pair The discriminant output, For the high-dimensional embedding feature representation of the frequency domain map, the first term encourages the discriminator to correctly identify embeddings from the time domain map, and the second term encourages the discriminator to correctly identify embeddings from the frequency domain map. During training, the encoder and discriminator alternately update their weights: the encoder's objective is to minimize the adversarial loss, making the embeddings in the time and frequency domains indistinguishable, thus promoting semantic alignment; the discriminator's objective is to maximize the adversarial loss to distinguish the embedding sources. This "min-maximum" game process drives the feature distribution to converge to a unified embedding space. The system performs alignment operations on the encoding results. ,in Represents collaborative feature representations in a unified embedding space; The operation is a learnable linear or nonlinear mapping function used to fuse the encoding results of two graph domains. This represents the high-dimensional embedding features of the time-domain graph. This represents the high-dimensional embedding feature representation of the frequency domain graph.

[0144] A dual-domain attention weight field is calculated on graph cooperative adversarial data to obtain graph-weighted data;

[0145] In one embodiment, the system uses a learnable linear transformation to generate query vectors and key vectors for time-domain nodes and frequency-domain nodes, respectively. ,in For trainable weight matrix, For the first Embedding features of each temporal node For the first Embedding features of frequency domain nodes For the original embedding dimension, For attention space dimension, Indicates the first The query vector for each time-domain node. Indicates the first The key vector of each frequency domain node. For any time domain node... With frequency domain nodes Its attention weight is defined as: ,in For frequency domain nodes For time domain nodes Attention weights It is a natural exponential function. For the index variable of the frequency domain node, For the first The key vector of each frequency domain node. For the first The key vector of each frequency domain node (used to normalize the denominator), the formula represents the frequency domain node among all frequency domain nodes. relatively The level of attention, meeting the normalization conditions. .exist Significant weights are propagated along the edges, if a certain frequency domain center node If a node is followed by multiple time-domain nodes (with high weight), forward diffusion is performed among its strongly correlated frequency-domain neighbors. During diffusion, the weight accumulation of distant nodes is controlled by a dynamic propagation coefficient attenuation factor to avoid attention saturation. The system generates a dual-domain weighted connection matrix. , representing the attention connection strength between all time-domain nodes and frequency-domain nodes; in the cross-domain edge set The edge weights are iteratively updated to reflect the attention focusing structure and diffusion dynamics. The output is a weighted graph connectivity matrix.

[0146] Nonparametric streaming activation mapping is performed on graph-weighted data to obtain dual-domain cross-attention data.

[0147] In one embodiment, for the first For each target node, its cross-attention activation is represented as: ,in This represents the attention weights from graph-weighted data. For frequency domain nodes Embedding features, This is a nonparametric activation function used to highlight characteristic channels with significant responses. Activation function Soft threshold activation or adaptive ReLU can be used, such as ,in For the output of the activation function, To find the maximum value function, For input feature values, Indicates an adaptive threshold. To activate the boundary fine-tuning factor, Frequency domain embedding features The median; facing a real-time feature stream input, the system performs the following dynamic processing: when the latest data arrives, it identifies its nearest neighbor node and obtains the corresponding attention weight; a sliding window (of length K) is used to cache the K most recent activation outputs. Average pooling is performed on the activation results within the window. The system output is a set of dual-domain cross-attention features.

[0148] Optionally, S4 includes:

[0149] S41. Construct a sound source behavior map and a sound source trend map based on the sound source separation and identification data, and obtain sound source behavior map data and sound source trend map data respectively.

[0150] In one embodiment, the input sound source separation and recognition dataset ,in For timestamps, For position coordinates, Action labels (e.g., "construction", "impact"). Each time-location-action triplet is a graph node. ; Construction ,in For nodes timestamp, For nodes timestamp, For nodes and The distance between them For nodes Location coordinates, For nodes Location coordinates, Time connection threshold ， Spatial connectivity threshold; output sound source behavior graph This method records sound source activity maps at different times and locations. The sound source trend map is constructed by dividing the continuous time series into equal-length windows (e.g., 5 minutes per time slice); each time window corresponds to a graph node, and its node characteristics are represented by a histogram of the distribution of sound source behavior types within that time period. Establish edges between adjacent time window nodes The similarity of their behavioral distributions is calculated; the Jensen-Shannon distance (JS distance) is used as a measure of distribution difference. If the JS distance between the behavioral histograms of two time periods is less than a preset threshold, the similarity is calculated. If the edge is similar to or continuous with the behavioral patterns of the two time periods, then the edge is constructed. This results in the completed sound source trend map.

[0151] S42. Perform graph fusion on the sound source behavior graph data and the sound source trend graph data to obtain the context graph data;

[0152] In one embodiment, for the sound source behavior graph Each node in Find its corresponding time window Map this node to the trend chart. The corresponding window node in And establish cross-graph reference relationships; if multiple behavioral graph nodes belong to the same trend window, virtual edges can be created pointing from behavioral graph nodes to trend nodes. Strengthen key connection paths through the structural similarity between behavioral and trend graphs; if in the behavioral graph... There is an edge in This indicates that two acoustic behaviors are closely related in time and space; determine their corresponding time windows. In the trend graph, determine whether the nodes are connected by a strongly correlated edge, i.e., calculate the similarity of behavioral distributions between trend nodes (such as Jensen-Shannon distance or co-occurrence frequency similarity); if the similarity exceeds a set threshold... If the connections between behavioral nodes are supported by temporal evolution, then the weights of the corresponding edges in the behavioral graph are enhanced, or a new fusion edge is added. Construct the fused context diagram ,in It is the combined set of all nodes in the behavior graph and the trend graph; Let be the set of edges of a self-behavioral graph; This is the set of edges from the trend graph; This is an additional set of reinforcing edges generated for cross-attention mechanisms or graph similarity measures.

[0153] S43. Perform causal nested edge processing on the scenario graph data to obtain the propagation structure graph data;

[0154] In one embodiment, the scenario diagram is... Each node timestamp Mapping to a unified relative timeline (e.g., with the first event as the origin) standardizes the event order; only considers events that satisfy... node pairs As a candidate causal edge; determine the action label pair Whether a logical reasoning path (e.g., "disturbance" → "alarm triggered") exists in the predefined knowledge graph serves as one of the prior criterion for establishing a causal relationship. For all node pairs that satisfy the time sequence condition... Determine whether it satisfies the following joint conditions, and the action conditional probability. That is, in historical statistics or knowledge graphs, subsequent behaviors From the preceding behavior The probability of triggering exceeds a threshold; time delay between events. If the propagation delay between the two events is within an acceptable range, then a new causal edge is added to the graph. , indicating an event Regarding the event Constitutes direct causal triggering; identifies ternary causal sequences (such as...) ), determine whether it forms a closed chain or has multiple intersecting paths; if the link stably exists in multiple windows or has high support, mark it as a nested causal chain. For each causal edge Calculate its propagation strength weight ,like ,in For propagation strength weight, Semantic similarity between behavioral labels (which can be modeled based on Word2Vec or label co-occurrence). For the proximity of spatial locations of items (such as Euclidean distance normalization); In a knowledge graph, action pairs Logical path support (such as path weight or logical confidence). This is the semantic similarity weighting coefficient, with a value of 0.46. This is the spatial proximity weighting coefficient, with a value of 0.39. The weighting coefficient for the knowledge graph support is set to 0.15. The weights of all causal edges are normalized. The propagation structure graph is then constructed.

[0155] S44. Generate a contextual reasoning path based on the propagation structure diagram data to obtain sound source contextual map data.

[0156] In one embodiment, a propagation structure diagram is used. The nodes with high perturbation influence or leading event nodes are used as the starting points of the path (e.g., selected from the sparse principal cause node set obtained in the previous step); a depth-first traversal (DFS) is used, and the weights are propagated along the outgoing edges of the current node in each step. Continue moving forward along the longest edge; generate a sequence of nodes. , represents a potential causal transmission path, with the preceding and following nodes representing the path from the initial node, intermediate nodes, and termination node, respectively, where all adjacent nodes satisfy . For each causal path, calculate its path credibility score. , defined as the product of the weights of all edges on the path: ,in To score the path credibility, For node indexing, Indexing another node, For potential causal transmission path node sequences, For node pairs The propagation weight reflects the overall strength of the causal relationship in the entire path; the higher the value, the more credible the causal chain. This can be considered a confidence probability estimate that the event chain is a true causal sequence. All candidate paths are filtered as follows: only paths whose length (i.e., number of nodes) meets the following criteria are retained. ,in This represents the path length (number of nodes). To preset a minimum propagation chain length, in order to avoid random or short-cycle event chains; the average propagation strength of all edges in the path must satisfy... ,in This represents the number of edges in the path. A minimum average confidence threshold is used to ensure that each causal chain in the path has strong explanatory power. All path results that pass the screening are uniformly organized and output as structured sound source context map data.

[0157] Optionally, the causal nested edge processing includes:

[0158] Event time mapping is performed on the scenario diagram data to obtain scenario diagram mapping data;

[0159] In one embodiment, for each node Extract its event timestamp Construct a normalized time axis ,in For nodes Normalized timestamp, For nodes Event timestamp, The earliest event time, Set the time for the last event; set the time window parameters. Slide groups according to window granularity to generate event time groups. ,in For the first Event time groups For event nodes, For nodes Normalized timestamp, Index the time window and output context graph mapping data. The normalized time attribute of each node is preserved.

[0160] Candidate causal edges are generated based on the scenario graph mapping data to obtain candidate causal graph data;

[0161] In one embodiment, a potential causal edge set is constructed, provided that the time sequence constraint is satisfied. For all node pairs... satisfy Only then can directional candidate edges be established; for each pair of candidate nodes, the following causal support indicators are calculated, including behavioral precedence. If the event type Trigger The value is set to 1 if it is not 1, otherwise it is 0; spatial conductivity The value is set to 1 if the Euclidean distance is within the propagation radius R, and 0 otherwise; Voiceprint similarity. The parameters are set using cosine similarity or KL divergence of sound source features; if the following conditions are met: , The threshold for determining causal edges is a preset value, set by experience or expert knowledge, to establish the edge. Add to the candidate causal edge set; output the candidate causal graph. .

[0162] Nested causal chain processing is performed on the candidate causal graph data to obtain the causal graph data;

[0163] In one embodiment, causal chains with multi-level triggering relationships are identified, and nested causal chain groups are constructed. All edges in the candidate causal graph are traversed. To check if a long chain exists: , This is the starting node of a nested causal chain. It is the intermediate node (first hop) of a nested causal chain. It is an intermediate node in a nested causal chain. The terminator of the nested causal chain is identified by a time-incrementing condition: ,in For nodes timestamp, For nodes timestamp, For nodes Timestamps, cumulative path confidence , This is the cumulative product of the propagation confidence scores of all edges in the chain. The path confidence threshold for a nested causal chain, where the chain is considered a nested causal chain unit. , To obtain a set of nested causal chain units (node ​​sequences) that meet the conditions, and to construct hyperedges or metapaths for them; output the nested causal graph. ,in It includes basic causal edges and nested chain super edges.

[0164] Propagation weights are calculated on the causal graph data to obtain the propagation structure graph data.

[0165] In one embodiment, the propagation capability is calculated for each edge in the computation graph. For each causal edge... Define propagation weights For weighted combination ,in To propagate weight values, representing nodes To the node The score of the ability to spread, This is the semantic similarity weight coefficient, with a value ranging from 0.3 to 0.5. This is a semantic similarity metric that measures the degree of semantic content similarity between event pairs. This is the continuity weighting coefficient for behavior, with a value ranging from 0.2 to 0.3. Behavioral consistency scoring measures the logical coherence between two events in terms of behavioral type. This is the spatial proximity weighting coefficient, with a value ranging from 0.1 to 0.3. Spatial distance similarity score measures how close two events are geographically. This is the frequency co-occurrence weighting coefficient, with a value ranging from 0.05 to 0.2. Represents the co-occurrence frequency of event pairs in historical scenarios; weights are standardized (e.g., softmax, min-max scaling) to adapt to the path propagation model; outputs a propagation structure graph. ,in Let be the propagation score matrix for the edges.

[0166] Optionally, this application also provides an intelligent early warning system for preventing external damage to transmission lines, used to execute the intelligent early warning method for preventing external damage to transmission lines as described above, wherein the intelligent early warning system for preventing external damage to transmission lines includes:

[0167] The acoustic sensing and acquisition module is used to acquire acoustic data of power transmission lines through acoustic sensors;

[0168] The voiceprint modeling and causal recognition module is used to perform sparse reconstruction of voiceprint features based on the acoustic data of transmission lines to obtain voiceprint feature data; and to perform sparse causal graph matching on the voiceprint feature data to obtain sound source behavior causal graph data.

[0169] The multi-domain attention sound source recognition module is used to perform dual-domain attention cross-recognition based on sound source behavior causal spectrum data to obtain sound source separation and recognition data.

[0170] The sound source scene map construction module is used to generate scene maps from sound source separation and identification data to obtain sound source scene map data.

[0171] Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended application documents rather than the foregoing description. Thus, it is intended that all changes falling within the meaning and scope of the equivalents of the application documents be incorporated into the invention.

[0172] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A power transmission line external damage intelligent early warning method, characterized in that, The method comprises: S1, collecting power transmission line acoustic data by an acoustic sensor; S2, performing acoustic fingerprint feature sparse reconstruction according to the power transmission line acoustic data to obtain acoustic fingerprint feature data; performing disturbance point detection on the acoustic fingerprint feature data to obtain acoustic fingerprint disturbance point data; performing sparse main cause extraction on the acoustic fingerprint disturbance point data to obtain sparse main cause data; performing acoustic sequence graph construction according to the sparse main cause data to obtain acoustic sequence graph data; performing asymmetric transition weight calculation according to the acoustic sequence graph data to obtain transition weight graph data, wherein the asymmetric transition weight calculation comprises the process of calculating the information transition intensity of a node pair in two directions based on any two nodes in the acoustic sequence graph data, taking the difference value as the asymmetric transition weight of the directed edge, and generating a weighted directed graph structure after threshold screening and weak diffusion correction; and performing environmental electrical weight processing according to the transition weight graph data to obtain a sound source behavior causal graph atlas data; S3, performing intra-domain feature focusing processing according to the sound source behavior causal graph atlas data to obtain intra-domain feature focusing data; performing double-domain cross-attention mapping on the intra-domain feature focusing data to obtain double-domain cross-attention data; performing multi-channel dynamic abnormality strengthening on the double-domain cross-attention data to obtain feature strengthening data; and performing semantic sound source separation on the feature strengthening data to obtain sound source separation identification data; S4, generating a scenario graph atlas from the sound source separation identification data to obtain sound source scenario graph data.

2. The method of claim 1, wherein, S1 comprises: acquiring power transmission load data; performing sampling frequency processing and direction sensitivity processing according to the power transmission load data to obtain sampling sensitivity data and direction sensitivity data, respectively; sampling through a double-layer microphone array unit preset along the power transmission line according to the sampling sensitivity data and the direction sensitivity data to obtain distributed sound field monitoring data, wherein the double-layer microphone array unit is a microphone array unit that performs environmental sound pressure sensing in the first layer and target frequency domain capture in the second layer; constructing an acoustic disturbance hidden space according to the distributed sound field monitoring data to obtain acoustic disturbance hidden space data, wherein the acoustic disturbance hidden space construction comprises the process of obtaining a high-dimensional sound field feature vector from a local segment of the distributed sound field monitoring data, mapping the high-dimensional sound field feature vector into a low-dimensional embedding vector using a nonlinear dimension reduction algorithm, calculating the connection strength weight based on the distance between the low-dimensional embedding vectors, and generating the acoustic disturbance hidden space data; modeling the sound source propagation structure according to the acoustic disturbance hidden space data to obtain the power transmission line acoustic data.

3. The method of claim 1, wherein, The acoustic fingerprint feature sparse reconstruction comprises: performing sparse Bayesian regression according to power transmission line electrical data to obtain electrical sensing data; constructing a frequency modulation driving graph atlas according to the electrical sensing data to obtain a sound frequency graph atlas data; acquiring power transmission line current data; performing electrical driving phase disturbance according to the power transmission line current data and the sound frequency graph atlas data to obtain disturbance graph atlas data; performing transmissible acoustic fingerprint embedding on the disturbance graph atlas data to obtain acoustic fingerprint feature data.

4. The method of claim 1, wherein, The sparse main cause extraction comprises: constructing a disturbance point graph according to the acoustic fingerprint disturbance point data to obtain disturbance point graph data; The perturbation point subgraph data is obtained by performing local change clustering on the perturbation point graph data. The sparse cause selection data is obtained by performing graph variational encoding on the perturbation point subgraph data. The sparse cause data is obtained by performing counterfactual confidence screening according to the sparse cause selection data, wherein the counterfactual confidence screening comprises the following processes: constructing a counterfactual graph after the sparse cause selection data is removed from the perturbation point subgraph data, respectively predicting the perturbation point subgraph data and the counterfactual graph through a preset identification model, calculating a difference between prediction outputs of the two as a counterfactual confidence, and obtaining the sparse cause data according to a preset threshold and a sorting strategy.

5. The method of claim 1, wherein, The dual-domain cross-attention mapping comprises: The dual-domain heterogeneous graph data is obtained by performing dual-domain heterogeneous graph construction according to the intra-domain feature focusing data, wherein the dual-domain heterogeneous graph construction is time-domain graph conversion and frequency-domain graph conversion according to the intra-domain feature focusing data. The graph collaborative adversarial data is obtained by performing graph collaborative adversarial learning on the dual-domain heterogeneous graph data. The graph weighted data is obtained by performing dual-domain attention weight field calculation on the graph collaborative adversarial data. The dual-domain cross-attention data is obtained by performing non-parametric streaming activation mapping according to the graph weighted data, wherein the non-parametric streaming activation mapping comprises the following processes: weighting the graph weighted data according to cross-attention activation, performing soft threshold activation on the weighted features by using an activation function, and performing average pooling on soft threshold activation results in a sliding window to generate the dual-domain cross-attention data.

6. The method of claim 1, wherein, S4 comprises: The source behavior graph data and the source trend graph data are respectively obtained by performing source behavior graph construction and source trend graph construction according to the source separation identification data. The situational graph data is obtained by performing graph fusion on the source behavior graph data and the source trend graph data. The propagation structure graph data is obtained by performing causal nested edge processing on the situational graph data. The source situational graph atlas data is obtained by performing situational reasoning path generation according to the propagation structure graph data.

7. The method of claim 6, wherein, The causal nested edge processing comprises: The situational graph mapping data is obtained by performing event time mapping on the situational graph data. The candidate causal graph data is obtained by performing candidate causal edge generation according to the situational graph mapping data. The causal graph data is obtained by performing nested causal chain processing according to the candidate causal graph data. The propagation structure graph data is obtained by performing propagation weight calculation on the causal graph data.

8. A power transmission line external damage intelligent early warning system, characterized in that, The power line anti-external damage intelligent early warning system comprises: An acoustic perception acquisition module is configured to acquire power line acoustic data through an acoustic sensor. A voiceprint modeling and causal identification module is configured to perform sparse reconstruction on voiceprint features based on the power line acoustic data to obtain voiceprint feature data, and perform sparse causal graph matching on the voiceprint feature data to obtain a source behavior causal graph atlas data. A multi-domain attention source identification module is configured to perform dual-domain attention cross-identification based on the source behavior causal graph atlas data to obtain source separation identification data. A source situational graph atlas construction module is configured to generate a source situational graph atlas data by performing situational graph atlas generation on the source separation identification data.

Images