Power transmission line fault identification method and system fusing weather semantics and harmonic image

Abstract

The invention discloses a power transmission line fault identification method and system fusing weather semantics and a harmonic image. The method comprises the following steps: acquiring current fault recording data and weather information; converting the weather information into a current weather semantic vector containing fault associated semantics; processing the recording data to generate a harmonic radar image representing fault harmonic characteristics; and the fault identification model carries out dynamic modulation and fusion based on cross attention on deep electrical characteristics extracted from the harmonic radar image by taking the weather semantic vector as a query, and finally outputs a fault cause identification result. According to the method, deep dynamic fusion of weather semantic information and electrical image features is creatively realized, high-precision, strong-robustness and full-automatic identification can be carried out on power transmission line faults caused by typical external factors such as lightning stroke, forest fire and icing in an actual scene with a huge power grid scale and a complex operation environment, and the method is suitable for popularization and application. And the accuracy of fault reason identification and the engineering practicability are obviously improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power transmission line fault identification, and in particular to a power transmission line fault identification method and system fusing weather semantics and harmonic images. BACKGROUND

[0002] With the rapid development of smart grids, massive fault recording data provides a solid data foundation for the rapid and accurate identification of power transmission line faults. Overhead transmission lines cover a wide range and are exposed to complex and variable natural environments for a long time, which are easily affected by adverse weather factors such as lightning, forest fires, heavy snow, and the like, thereby inducing various types of line faults, which seriously threaten the safe and stable operation of the power grid. Under this background, accurately identifying fault causes has important engineering application value for guiding dispatching power transmission decisions, optimizing operation and maintenance repair strategies, and formulating power grid emergency response measures. Therefore, how to make full use of multi-source information to build a high-precision and strong generalization capability of power transmission line fault cause identification model has become a key technical problem to be broken through in the current smart grid fault diagnosis field.

[0003] In recent years, deep learning technology has made significant progress in fault identification due to its strong automatic feature learning ability. Researchers have converted one-dimensional fault recording data into various two-dimensional image forms, such as Gram angle field, Markov transition field, and transient waveform image, and learned the spatial structure features in them using various deep learning network models, which greatly improves the identification ability of fault causes.

[0004] However, most existing methods mainly focus on electrical quantities themselves, ignoring external environmental information, especially the influence of weather conditions on fault causes. However, in fact, there is a strong correlation between many power transmission line faults and specific weather conditions. For example, lightning faults often occur in rainy weather; windage faults are related to strong wind weather; and icing faults are directly related to low temperature, snow and freezing weather. Simply relying on electrical waveform features sometimes makes it difficult to distinguish the root cause of the fault. For example, lightning and forest fires can both cause line tripping, and their transient waveforms may be similar in some cases; the electrical characteristics of a pollution flashover fault may also be confused with other insulation faults. If effective use of weather information is lacking, relying solely on electrical quantities for judgment will greatly reduce the accuracy and reliability of the identification results, which may lead to operation and maintenance decision errors.

[0005] Therefore, fusing multi-source information such as weather and season, which is non-electric quantity, becomes an effective way to improve the recognition accuracy. However, the existing methods still have limitations in the representation of non-electric quantity, such as One-Hot Encoding which cannot reflect the relevance and physical similarity between different weather types, and is difficult to fully reveal the deep semantic relationship between environmental factors and fault mechanism. Specifically, One-Hot Encoding can only convert weather categories such as "sunny", "rainy", and "snowy" into binary vectors that are isolated from each other and sparse in dimension. This representation method has obvious defects:

[0006] 1) Unable to express semantic relevance: it cannot reflect the similarity of "thunder shower" and "heavy rain" in causing lightning fault risk, nor can it distinguish the huge difference between "breeze" and "gale" in causing wind deviation fault probability. The physical similarity between weather types and the strength of fault association are completely ignored.

[0007] 2) Lack of intensity and probability information: One-Hot Encoding cannot incorporate important knowledge such as "the historical probability of a certain type of fault occurring under a certain weather condition". For example, "moderate snow" and "heavy snow" have different effects on icing faults, but simple category encoding cannot reflect this difference.

[0008] 3) Difficulty in fusion with electrical features: due to the shallow and semantic lack of encoding method, this weather representation is difficult to effectively and meaningfully fuse with deep features extracted from complex electrical waveforms. The usual fusion method (such as early concatenation) is difficult to model the deep interaction between weather conditions and electrical features.

[0009] In summary, there is an urgent need for a power line fault cause recognition method that can effectively fuse weather semantic information and electrical features and model their deep interaction mechanism, to overcome the limitations of existing technology in image feature and weather information fusion. SUMMARY

[0010] Therefore, the present application provides a method that can effectively fuse weather environment information with rich semantics and electrical transient features, and establish a deep interaction mechanism between them, thereby achieving high-precision, high-robustness, and automated cause recognition of power line faults induced by typical external environmental factors such as lightning, forest fire, wind deviation, icing, pollution flashover, and foreign matter.

[0011] The first aspect of the present application discloses a power line fault recognition method that fuses weather semantics and harmonic images, comprising the following steps:

[0012] S1, acquire the current fault waveform data of the transmission line to be identified and the weather information at the time of the current fault occurrence; S2, perform semantic encoding processing on the weather information of the current fault to generate a low-dimensional dense vector representing the semantic association between it and the fault cause, as the current weather semantic vector; S3, process the current fault waveform data to generate a current harmonic radar image representing the harmonic characteristics of the fault electrical quantity; S4, input the current weather semantic vector and the current harmonic radar image into a pre-trained fault identification model, wherein the fault identification model is configured to extract the deep electrical features of the harmonic radar image, and use the current weather semantic vector as a query vector to dynamically modulate the deep electrical features to obtain a fused feature vector that incorporates weather information, and output the identification result of the current transmission line fault cause based on the fused feature vector.

[0013] Regarding step S1, it can retrieve waveform files in formats such as COMTRADE corresponding to the target fault event from the power grid information protection system or fault information management system, and obtain specific weather descriptions (such as "thunderstorm", "strong wind", "icing") of the fault occurrence time and location from fault reports or meteorological databases. Based on step S1, the alignment and acquisition of multi-source heterogeneous data are completed, providing a foundation for subsequent fusion processing.

[0014] Regarding step S2, it involves classifying the current weather description into a standard weather type and directly mapping it to a vector using a pre-trained knowledge graph embedding model. This vector is not a simple label; its value and position in the vector space encode the historical correlation strength between the weather type and various fault causes. Those skilled in the art will understand that "weather semantics" does not refer to the general meaning of the weather description text, but specifically to the empirical and probabilistic relationships between weather types and fault causes statistically mined from historical fault data. This "semantic encoding" is achieved by constructing a "weather-fault" knowledge graph and training it using embedding models such as TransE, encoding this statistical correlation into the geometric relationships of low-dimensional dense vectors. Therefore, the "semantics" of the "weather semantic vector" is specifically manifested in the relative position and distance relationships between the weather vector and various fault cause vectors in the vector space, reflecting the conditional probability strength in historical statistics. That is, the weather "semantics" and "semantic correlation" described in this invention specifically refer to the conditional probabilistic correlation between weather conditions and specific fault causes obtained through statistical analysis of historical fault data, and this relationship is encoded into a vector representation using knowledge graph embedding technology.

[0015] Regarding step S3, it can be achieved by extracting the recorded waveform data, performing FFT harmonic analysis and normalization, and then plotting the harmonic radar sub-maps of each phase voltage and current through polar coordinate mapping, which are then combined into a multi-channel image (such as an RGB image). Based on step S3, a one-dimensional time-series signal is converted into a two-dimensional image that can comprehensively and intuitively display the relative distribution of each phase and each harmonic, providing an optimized input form for the convolutional neural network that is tailored to the characteristics of fault harmonics.

[0016] Regarding step S4, the fault identification model can process two inputs in parallel or sequentially: on the one hand, it extracts deep electrical features from harmonic radar images through its internal image feature extraction unit (such as CNN); on the other hand, through a multimodal fusion unit, using weather semantic vectors as the "query," it dynamically modulates the extracted electrical features based on cross-attention to generate fused features; finally, through a classification output unit, it calculates and outputs the probability distribution of fault causes based on the fused features. This achieves high-precision end-to-end identification. Through the dynamic fusion mechanism, weather information can proactively guide the model to focus on the electrical features most relevant to the current weather, significantly improving the model's ability to distinguish complex faults and the interpretability of its decisions.

[0017] According to the method disclosed in the first aspect of the present invention, the fault identification model dynamically modulates the deep electrical features using the current weather semantic vector as the query vector, including: structurally reconstructing the deep electrical features to convert them into a key matrix and a value matrix that can be adapted to the calculation of the cross-attention mechanism; calculating the similarity between the weather semantic vector and the key matrix using the weather semantic vector as the query vector to generate cross-attention weights; and performing a weighted summation of the value matrix using the cross-attention weights to complete the dynamic modulation of the deep electrical features by the weather semantics and outputting a fused feature vector. Here, the specific implementation of the dynamic modulation is defined as a cross-attention mechanism, which achieves true dynamic and conditional fusion. The attention weights are calculated in real time and reflect the importance of different parts of the electrical features under the "question" of the current specific weather semantics, thereby achieving focused information fusion rather than simple mixing.

[0018] According to the method disclosed in the first aspect of the present invention, the fault identification model outputs the identification result of the current transmission line fault cause based on the fused feature vector, comprising: inputting the fused feature vector into at least one fully connected layer of the fault identification model, performing feature dimension mapping and nonlinear transformation; converting the transformed features into a probability distribution corresponding to multiple fault causes through a normalized exponential function; and taking the fault cause with the largest value in the probability distribution as the final fault cause identification result. Specifically, the fused feature vector is input into one or more fully connected layers for nonlinear transformation and dimensionality reduction, and finally the output value is converted into a probability distribution over all preset fault cause categories through a Softmax function. The output probability provides richer decision-making information than a single label, not only giving the most likely cause, but also showing the confidence level of other possibilities, supporting maintenance personnel to make more robust decisions, and prompting manual review when the probabilities are close.

[0019] According to the method disclosed in the first aspect of the present invention, step S3 includes: extracting three-phase voltage and three-phase current waveform data within a preset number of cycles before and after the fault; extracting the amplitude of each harmonic in each phase from the waveform data based on fast Fourier transform to form a harmonic amplitude sequence; normalizing each harmonic amplitude sequence; mapping each amplitude in each normalized harmonic amplitude sequence sequentially to a preset corresponding angle in a polar coordinate system to obtain a polar coordinate point set; converting the polar coordinate point set to a rectangular coordinate system and drawing connecting lines to form a single-phase harmonic radar sub-map; combining multiple single-phase harmonic radar sub-maps to generate a multi-channel current harmonic radar image. Here, by normalizing each harmonic amplitude sequence, the influence of the absolute dimension of the amplitude caused by different line voltage levels and fault severity is eliminated, allowing the model to focus on learning the relative proportion pattern of each harmonic content, which is one of the essential characteristics of a fault, greatly enhancing the model's generalization ability. The resulting RGB harmonic radar image contains both spatial (harmonic distribution shape) and channel (three-phase information) features, making it extremely suitable as input for CNNs and capable of efficiently transmitting composite harmonic mode information of faults.

[0020] According to the method disclosed in the first aspect of the present invention, the fault identification model is used to extract deep electrical features from a harmonic radar image, comprising the following steps performed sequentially: initial feature extraction of the harmonic radar image based on the max pooling layer of the fault identification model; feature sampling at different time lengths completed stepwise by multiple bottleneck layer modules of the fault identification model; spatial dimension compression and channel dimensionality reduction based on the average pooling layer of the fault identification model, outputting the final deep electrical features. Specifically, the image feature extraction unit of the fault identification model preferably employs a lightweight convolutional neural network such as MobileNetV3. The process includes: 1) Head: rapid downsampling through convolutional layers and max pooling layers to extract preliminary features; 2) Body: deep feature extraction through multiple inverse residual bottleneck layer modules, each module including operations such as dimensionality increase-depth convolution-SE attention-dimensionality reduction, wherein the SE attention module can adaptively calibrate the importance of channel features; 3) Tail: spatial information is compressed through a global average pooling layer, and a one-dimensional deep electrical feature vector is finally output. The use of a lightweight network here significantly reduces model complexity and computational overhead while ensuring feature extraction capabilities, making the proposed solution easy to deploy on edge devices with limited computing resources and meeting the requirements of engineering practicality.

[0021] According to the method disclosed in the first aspect of the present invention, before step S1 or step S2, the method further includes a step of pre-constructing a knowledge graph embedding model for semantic encoding, which includes: standardizing and classifying different weather information based on a historical fault sample set to form a finite set of standard weather categories; calculating the conditional probability of each fault cause occurring under each standard weather category; classifying the association strength level between weather categories and fault causes based on a preset association strength threshold; constructing a weighted knowledge graph in which each group consists of a triplet composed of a weather category, an association strength level, and a fault cause; and training the weighted knowledge graph using a knowledge graph embedding algorithm to obtain a knowledge graph embedding model, which enables it to map the weather information at the time of the fault into a low-latitude dense weather semantic vector.

[0022] Step S2 is based on the final knowledge graph obtained for semantic encoding.

[0023] According to the method disclosed in the first aspect of the present invention, the knowledge graph embedding algorithm is a translational embedding model, and the steps of training the weighted knowledge graph include: setting a loss function for training the translational embedding model; and embedding the initial knowledge graph so that the resulting vector is used to represent the semantic association between weather information and various fault causes.

[0024] According to the method disclosed in the first aspect of the present invention, the method further includes the step of pre-constructing a fault identification model, which includes: acquiring a historical fault sample set containing historical fault waveform data, historical fault weather information, and confirmed fault cause labels; processing the historical fault weather information using the pre-trained knowledge graph embedding model to obtain historical weather semantic vectors; processing the historical fault waveform data to generate historical harmonic radar images; and constructing an initial fault identification model, which includes at least an image feature extraction module for extracting deep electrical features of the image, a multi-modal fusion module for dynamically modulating and fusing deep electrical features using weather semantic vectors as query vectors, and a classification output module for outputting fault identification results.

[0025] Using the historical harmonic radar images and corresponding historical weather semantic vectors as training inputs, and the confirmed fault cause labels as training targets, the initial fault identification model is trained until convergence, resulting in the pre-trained fault identification model. This is an end-to-end supervised training process. Specifically, a historical sample set is used, where each sample has been generated with "historical harmonic radar images" and "historical weather semantic vectors" as input pairs using the aforementioned method, and the fault causes confirmed by line inspection are used as labels. The constructed initial model includes a CNN feature extractor, a cross-attention fusion layer, and a classifier. During training, the aforementioned input pairs are used as model inputs, and the aforementioned labels are used as supervision signals. Iterative optimization is performed using the cross-entropy loss function and optimizers such as Adam until the model converges on the validation set. Through this training process, the model not only learns to identify electrical fault patterns from harmonic radar images, but also learns how to use weather semantic vectors as context to dynamically adjust its interpretation focus of electrical patterns, ultimately obtaining an intelligent model that deeply understands the mapping relationship between "weather-electrical" joint patterns and fault causes.

[0026] The second aspect of this invention discloses a system for implementing the transmission line fault identification method that integrates weather semantics and harmonic images disclosed in the first aspect of this invention, comprising: a data acquisition unit configured to acquire current fault waveform data of the transmission line to be identified and weather information at the time of the current fault occurrence; a semantic vector conversion unit configured to perform semantic encoding processing on the weather information of the current fault to generate a low-dimensional dense vector representing the semantic association between the current fault and the fault cause, as the current weather semantic vector; and an image generation unit configured to process the waveform data of the current fault to generate an image representing the fault. The system comprises: a current harmonic radar image showing the harmonic characteristics of electrical quantities; and a fault identification model, including: an input unit configured to receive inputs of a current weather semantic vector and the harmonic radar image; an image feature extraction unit configured to extract deep electrical features from the harmonic radar image; a multimodal cross-attention fusion unit configured to dynamically modulate the deep electrical features using the current weather semantic vector as a query vector to obtain a fused feature vector incorporating weather information; and a classification and recognition unit configured to output the identification result of the current transmission line fault cause based on the fused feature vector. This system modularizes and productizes the method of the first aspect of the invention, clarifies the boundaries and collaborative relationships of each functional unit, and provides clear architectural guidance for software implementation or integrated hardware and software implementation.

[0027] The third aspect of the present invention discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the transmission line fault identification method that integrates weather semantics and harmonic images disclosed in the first aspect of the present invention.

[0028] Beneficial Effects: In the transmission line fault identification method and system integrating weather semantics and harmonic images of this invention, by deeply fusing weather information containing rich semantic associations with specialized harmonic image features, and utilizing cross-attention to achieve dynamic modulation, the accuracy and recall of identifying faults induced by complex weather conditions are significantly improved. Specifically, the harmonic radar image specifically highlights the harmonic distribution characteristics of the fault, while the weather semantic vector generated by knowledge graph embedding goes beyond simple labels, including historical statistical associations; both provide higher-quality feature representations for downstream tasks. Furthermore, the cross-attention fusion mechanism using weather as a query is the core innovation of this invention, realizing conditional deep interaction of cross-modal information, overcoming the limitations of static stitching, and making model decision-making closer to human expert logic.

[0029] The following describes in detail the transmission line fault identification method and system of the present invention, which integrates weather semantics and harmonic images, with reference to the embodiments shown in the accompanying drawings and the reference numerals. Attached Figure Description

[0030] Figure 1 This is a flowchart of the steps in the transmission line fault identification method that integrates weather semantics and harmonic images in this invention.

[0031] Figure 2 This is a schematic diagram of the transient waveform and harmonic radar diagram corresponding to the fault cause type in this invention.

[0032] Figure 3 This is a schematic diagram of the fault identification model in this invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0034] It should be noted that all directional indicators (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a specific posture (as shown in the figure). If the specific posture changes, the directional indicator will also change accordingly.

[0035] Furthermore, the use of terms such as "first" and "second" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. If the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0036] Existing technologies for identifying the causes of transmission line faults have several shortcomings in practical applications: traditional methods rely on manual feature extraction or analysis of single electrical quantities, making it difficult to capture the deep features of faults; the processing of weather information during multi-source information fusion is relatively crude, failing to fully explore its semantic relationships; the image feature representation is not optimized enough, limiting its ability to distinguish easily confused faults; and the model structure is too complex, hindering practical deployment. To address these issues, this application provides a method and system for identifying the causes of transmission line faults through multimodal cross-fusion of weather semantics and harmonic images. It achieves efficient expression of electrical features by constructing harmonic radar images, mines weather semantic information using knowledge graph embedding technology, combines a cross-attention mechanism to achieve deep fusion of multimodal information, and finally achieves accurate identification of fault causes through a lightweight neural network model.

[0037] Specifically, in combination Figure 1 As shown, this invention discloses a method for identifying transmission line faults by fusing weather semantics and harmonic images, comprising the following steps: S1, acquiring the current fault waveform data of the transmission line to be identified and the weather information at the time of the current fault occurrence; S2, performing semantic encoding processing on the weather information of the current fault to generate a low-dimensional dense vector representing the semantic association between the current fault and the fault cause, as the current weather semantic vector; S3, processing the current fault waveform data to generate a current harmonic radar image representing the harmonic characteristics of the fault electrical quantities; S4, inputting the current weather semantic vector and the current harmonic radar image into a pre-trained fault identification model, wherein the fault identification model is configured to extract deep electrical features from the harmonic radar image, and dynamically modulate the deep electrical features using the current weather semantic vector as a query vector to obtain a fused feature vector that incorporates weather information, and output the identification result of the current transmission line fault cause based on the fused feature vector.

[0038] Specifically, in step S1, fault-related data of the transmission line is acquired, including fault waveform data and weather information. Specifically, this refers to acquiring the three-phase voltage and three-phase current waveform data corresponding to the transmission line fault event, along with the corresponding weather information. In step S2, the weather information is semantically encoded, that is, converted into a low-dimensional dense vector that retains the semantic correlation between weather and transmission line faults and is easily understood by computers. In step S3, a harmonic radar image is generated based on the fault waveform data. In step S4, the fault identification model includes an image feature extraction unit and a multimodal cross-attention fusion unit. The harmonic radar image is input into the image feature extraction unit to extract deep electrical features. Then, through the multimodal cross-attention fusion unit, using the weather semantic vector as a query, the deep electrical features are dynamically modulated to achieve multimodal information fusion. That is, through the cross-attention fusion unit, the electrical features are dynamically modulated using the weather semantic vector as a query. Finally, based on the fused features, the fault identification model outputs the identified fault cause and its corresponding probability, completing the fault cause identification. Specifically, the fused features are input into the classification and recognition unit, and the probability distribution of preset fault causes such as lightning strikes, wildfires, and icing is output through a fully connected layer and a normalized exponential function (Softmax). The category corresponding to the highest probability is taken as the final identification result.

[0039] Through the above steps, this invention realizes a complete process for automatic identification of fault causes by integrating weather semantic information, overcomes the limitations of traditional methods, and improves the accuracy and engineering practicality of identifying power transmission line faults caused by severe weather factors such as lightning strikes, wildfires, and icing.

[0040] In one specific embodiment, the fault identification model dynamically modulates the deep electrical features using the current weather semantic vector as the query vector. This includes: structurally reconstructing the deep electrical features, converting them into a key matrix and value matrix suitable for cross-attention mechanism calculation; calculating the similarity between the weather semantic vector and the key matrix using the weather semantic vector as the query vector to generate cross-attention weights; and performing a weighted summation of the value matrix using the cross-attention weights to complete the dynamic modulation of the deep electrical features by weather semantics, outputting a fused feature vector. Specifically, the deep electrical features are reconstructed into key-value pairs; the similarity between the query vector and the key vector is calculated using the weather semantic vector as the query vector to obtain cross-attention weights; and the value vector is weighted and summed using the cross-attention weights to output a fused feature vector, thus realizing the dynamic guidance of electrical features by weather semantic information.

[0041] In a specific embodiment, the fault identification model outputs the identification result of the current transmission line fault cause based on the fused feature vector, including: inputting the fused feature vector into at least one fully connected layer of the fault identification model, performing feature dimension mapping and nonlinear transformation; converting the transformed features into a probability distribution corresponding to multiple fault causes through a normalization exponent (e.g., Softmax) function; taking the fault cause with the largest value in the probability distribution as the final fault cause identification result, that is, selecting the category with the largest value in the probability distribution as the final fault cause identification result, and outputting the complete probability distribution for decision reference.

[0042] In a specific embodiment, step S3 includes: extracting three-phase voltage and three-phase current waveform data within a preset number of cycles before and after the fault; extracting the amplitude of each harmonic in each phase from the waveform data based on fast Fourier transform to form a harmonic amplitude sequence; normalizing each harmonic amplitude sequence; mapping each amplitude in each normalized harmonic amplitude sequence to a preset corresponding angle in the polar coordinate system to obtain a polar coordinate point set; converting the polar coordinate point set to a rectangular coordinate system and drawing connecting lines to form a single-phase harmonic radar sub-map; combining multiple single-phase harmonic radar sub-maps to generate the multi-channel current harmonic radar image.

[0043] That is, the three-phase voltage and three-phase current sequences before and after a preset number of cycles are extracted from the recorded waveform data, and a harmonic radar image is generated through polar coordinate harmonic mapping. The specific implementation process can be as follows: the three-phase voltage and three-phase current recorded waveform data within one cycle before the fault time to two cycles after the fault time are extracted, and then the amplitude sequence of each harmonic of each phase in the recorded waveform data is extracted through fast Fourier transform. Then, the harmonic amplitude is normalized to eliminate the dimensional differences between different signals. Then, the uniform angle in the polar coordinates corresponding to each harmonic is converted to rectangular coordinates. Finally, the harmonic radar image is drawn based on the converted coordinate data. The harmonic radar image is used as the core component of the fault electrical quantity feature set. The harmonic radar image is an RGB three-channel image, and the image size of each channel is 224×224 pixels.

[0044] In one specific embodiment, the fault identification model is used to extract deep electrical features from harmonic radar images, including the following steps performed sequentially: initial feature extraction of the harmonic radar image based on the max pooling layer of the fault identification model; feature sampling under different time lengths is completed step by step based on multiple bottleneck layer modules of the fault identification model; spatial dimension is compressed based on the average pooling layer of the fault identification model, and channel dimensionality reduction is performed to output the final deep electrical features.

[0045] The image feature extraction unit of the fault identification model can extract electrical features from a classic convolutional neural network such as MobileNetV3, i.e., by inputting the harmonic radar image into a lightweight convolutional neural network. Specifically, the extraction of deep electrical features by the image feature extraction unit includes: the head consists of a 3×3 convolution and a 3×3 max pooling layer to perform initial feature extraction on the harmonic radar image feature map; the main body consists of multiple bottleneck layer (bneck) modules, which perform feature downsampling step by step by setting different phase lengths; specifically, each bneck module first performs channel dimensionality upscaling on the input features through a 1×1 convolution, then performs spatial filtering on the upscaled features using a depthwise convolution while preserving the independence between channels, then dynamically calibrates the channel dimension through an SE attention module to enhance the weight of effective features, and finally performs channel dimensionality reduction through a 1×1 convolution. The nonlinear activation within the module uses HardSwish or ReLU activation functions according to the feature dimension; the tail compresses the spatial dimension through average pooling, and then performs channel dimensionality reduction through a 1×1 convolution to output deep electrical features.

[0046] In this application, the so-called deep electrical features refer to the abstract feature representations obtained by multiple convolutions and nonlinear activations after being output by the main and tail layers of the lightweight convolutional neural network (such as MobileNetV3); more specifically, they refer to the high-dimensional abstract feature vectors extracted by the last one or the last few bottleneck layers in the image feature extraction unit (such as the MobileNetV3 network) for final classification.

[0047] In a specific embodiment, before step S1 or step S2, a step of pre-constructing a knowledge graph embedding model for semantic encoding is further included. That is, based on the statistics of weather descriptions and fault causes in historical fault samples, a "weather condition-fault cause" knowledge graph containing correlation strength is constructed. This includes: standardizing and classifying different weather information based on the historical fault sample set to form a finite set of standard weather categories; calculating the conditional probability of each fault cause occurring under each standard weather category; classifying the correlation strength level between weather categories and fault causes based on a preset correlation strength threshold; constructing a weighted knowledge graph where each group consists of a triplet composed of a weather category, a correlation strength level, and a fault cause; training the weighted knowledge graph using a knowledge graph embedding algorithm to obtain a knowledge graph embedding model, which can map the weather information at the time of the fault occurrence into a low-latitude dense weather semantic vector; step S2 is performed based on the obtained final knowledge graph for semantic encoding.

[0048] Specifically, based on historical fault records, the probability of occurrence of various fault causes under different weather types is statistically analyzed. According to a preset probability threshold, the relationship between weather type and fault cause is divided into four levels: no impact, weak correlation, moderate correlation, and strong correlation, constructing a "weather condition-fault cause" knowledge graph. A knowledge graph embedding model, such as Translating Embeddings (TransE), is used to embed the knowledge graph, generating weather semantic vectors that satisfy geometric constraints. These weather semantic vectors are dense vectors, implying the inherent semantic relationships between weather conditions. That is, firstly, the weather descriptions in the fault samples are standardized and classified to form a finite set of standard weather categories; then, the conditional probability of each fault cause occurring under each weather category is statistically analyzed; and based on a set association strength threshold, frequently co-occurring "weather-fault cause" combinations are defined as strong associations, while low-frequency or non-co-occurring combinations are defined as weak associations or no associations; each pair of relationships is represented as a triple of weather, relationship strength, and fault cause, forming a weighted knowledge graph.

[0049] The knowledge graph embedding algorithm used is a translational embedding model. The steps for training the weighted knowledge graph include: setting a loss function for training the translational embedding model; and training the initial knowledge graph to embed it so that the resulting vectors are used to represent the semantic association between weather information and various fault causes. Specifically, the initial knowledge graph is trained using a translational embedding model (TransE), and the trained knowledge graph embedding model can map the weather description of the current fault into a low-dimensional, dense weather semantic vector.

[0050] In one specific embodiment, this application further includes a step of pre-constructing a fault identification model, which includes: acquiring a historical fault sample set containing historical fault waveform data, historical fault weather information, and confirmed fault cause labels; processing the historical fault weather information using the pre-trained knowledge graph embedding model to obtain historical weather semantic vectors; processing the historical fault waveform data to generate historical harmonic radar images; constructing an initial fault identification model, which at least includes an image feature extraction module for extracting deep electrical features of the image, a multi-modal fusion module for dynamically modulating and fusing deep electrical features using weather semantic vectors as query vectors, and a classification output module for outputting fault identification results; training the initial fault identification model using the historical harmonic radar images and corresponding historical weather semantic vectors as training inputs and the confirmed fault cause labels as training targets until the model converges, thereby obtaining the pre-trained fault identification model.

[0051] The acquisition of historical fault sample sets includes: collecting fault waveform files from real power grids and corresponding weather information at the time and location; standardizing and labeling the fault causes described in the briefings according to preset fault cause classification standards; and pairing the labeled fault events with their waveform data to form a structured fault sample set. Specifically, the process involves collecting fault samples containing complete fault waveforms, line inspection results, and corresponding weather conditions from the power grid information protection system; classifying the fault causes of the fault samples and determining fault cause tags, which include lightning strikes, wildfires, wind deflection, pollution flashover, icing, and foreign object faults.

[0052] In this embodiment, fault samples are divided into training and testing sets according to a certain ratio; a suitable optimizer is used, certain initial training parameters and dynamic learning rate scheduling strategies are set, and the verification loss is monitored during the training process.

[0053] Based on the above, the identification method provided in this application sequentially executes the following steps: acquiring transmission line fault recording data and natural language weather descriptions; constructing a "weather condition - fault cause" knowledge graph with fused correlation strength; mapping weather descriptions into dense vectors that retain the semantic association between weather and transmission line faults; generating harmonic radar images of three-phase voltage and current; extracting electrical features using a lightweight convolutional network and combining it with a cross-attention mechanism to achieve dynamic modulation of electrical features by weather semantics; and outputting the probability distribution of faults caused by various external environmental factors. This achieves a complete technical process from multi-source raw data input to automatic fault cause identification. It can perform high-precision, robust, and fully automated identification of transmission line faults caused by typical external factors such as lightning strikes, wildfires, and icing in real-world scenarios with large-scale power grids and complex operating environments. It effectively overcomes the limitations of traditional methods, such as reliance on manual feature engineering, neglect of environmental semantic association, and shallow fusion methods, and significantly improves the accuracy and engineering practicality of fault cause identification.

[0054] The construction process of the fault identification model and knowledge graph embedding model in this application can be specifically implemented through the following methods:

[0055] First, obtain sampling data of various electrical quantities before and after historical faults in transmission lines, along with corresponding natural language weather descriptions. Collect and label real power grid data to form a structured historical fault sample set. This includes the following steps A1-A2.

[0056] A1: Collect fault sample data.

[0057] Fault samples, including complete fault waveforms, weather conditions, and fault reports, are collected from the insurance information system. The complete fault waveform data includes time-domain sampling data of three-phase voltage and three-phase current; the fault reports record information such as the time and location of the fault, weather description, and inspection results.

[0058] A2: Standardization of fault cause labeling.

[0059] The collected fault samples were categorized and organized. The original fault cause labels included lightning strikes, wildfires, tree flashovers, external damage, bird droppings, bird bodies, wind deflection, and foreign objects. These original fault cause labels were then standardized and classified. Since some faults had similar mechanisms and transient waveform characteristics, they were merged from the perspective of fault mechanism. Preferably, six typical fault cause labels could be identified: lightning strikes, wildfires, wind deflection, pollution flashovers, icing, and foreign object faults, ensuring the labels' distinguishability and representativeness.

[0060] The next step is to construct a knowledge graph, which is based on the conditional probabilities of weather descriptions and causes of malfunctions from historical samples, classifying standard weather categories, setting association strength thresholds, and constructing "weather condition - cause of malfunction" combinations to form a weighted knowledge graph. Specifically, this includes the following steps B1-B2.

[0061] B1: Statistical analysis of the probability of association between weather and malfunctions.

[0062] Based on historical fault inspection records, the probability of occurrence of various fault causes under different weather conditions is statistically analyzed. Preferably, the corresponding conditional probability is calculated using the following formula. (1);

[0063] In the formula: Indicates the i-th weather type; Indicates the cause of the j-th type of failure; This indicates that in the historical fault records, there was a weather-related issue. The following malfunction occurred The sample size. Statistical results are expressed as a percentage for easy comparison with a preset relationship strength threshold.

[0064] B2: Construct a knowledge graph of "weather conditions - causes of failure".

[0065] Based on preset probability thresholds, the relationship between weather type and fault cause is divided; each relationship is represented as a triple of "weather type - relationship strength - fault cause", and all triples together constitute a knowledge graph with semantic strength. Preferably, in this example, a probability of less than 5% is defined as no impact, 5%-15% as weak correlation, 15%-30% as moderate correlation, and 30% and above as strong correlation.

[0066] The training of the knowledge graph involves setting a loss function for the translation embedding model to map the current fault weather description into a low-dimensional, dense weather semantic vector. Specifically, this includes the following steps C1-C2.

[0067] C1: Triple set representation.

[0068] The knowledge graph obtained in step B2 is represented as a set of triples. A knowledge graph embedding model is used to train the constructed knowledge graph, mapping weather type, fault cause and the relationship between them to a low-dimensional vector space composed of embedding vectors.

[0069] The generated embedding vectors satisfy the following geometric constraints: (2);

[0070] In the formula: The embedding vector for the weather of the head entity. For the embedding vector of the relation, Let be the embedding vector of the cause of the tail entity's failure. This equation shows that, ideally, the vector representation of the tail entity should be approximated after translating along the relational direction from the head entity. This translation structure allows the model to capture the semantic relationships between entities and maintain logical consistency in a low-dimensional continuous space.

[0071] C2: Optimize the model using a loss function.

[0072] Preferably, the model is optimized using the marginal ranking loss function;

[0073] (3);

[0074] In the formula: The set of positive triples in the training set; It is a set of negative triples, which is usually generated by randomly replacing the tail entity; The preset boundary is set to a positive value to control the degree of separation between positive and negative samples. The max function guarantees that the negative sample score will only be higher than the positive sample score by more than [a certain percentage]. Losses only occur at that time.

[0075] After the knowledge graph embedding model is built and trained, the process returns to building and training the fault identification model. Specifically, based on the historical fault sample set, electrical sequences with a preset number of cycles before and after each historical fault are extracted. After normalization via Fast Fourier Transform, historical fault harmonic radar images are generated through polar coordinate harmonic mapping. This includes the following steps D1-D5.

[0076] D1: Steps for capturing historical fault transient data.

[0077] Based on the historical fault time markers in the historical fault recording file, a certain number of cycles of sampled data are extracted; preferably, the three-phase voltage and three-phase voltage sequence are extracted from the cycle before the historical fault time to the cycle after the historical fault time.

[0078] D2: Steps for extracting harmonic amplitude.

[0079] Harmonic analysis is performed on the extracted electrical quantity sequences of each phase to extract the amplitude sequences of the 0th to 7th harmonics. Preferably, this example uses Fast Fourier Transform (FFT) for analysis. FFT can convert time-domain signals into frequency-domain signals, allowing for sampling and preprocessing of the three-phase signals.

[0080] Preferably, the amplitude sequence of the 0th to 7th harmonics of each phase is obtained by FFT decomposition, and the harmonic amplitude extraction formula is as follows: (4);

[0081] In the formula: Given a discrete sampled signal of a single phase; N is the number of sampling points; k corresponds to the harmonic order; the actual amplitude of each harmonic is obtained by taking the modulus of the FFT transform result and multiplying it by the amplitude correction factor 2 / N. .

[0082] D3: The steps for harmonic amplitude normalization.

[0083] The amplitudes of the 0th to 7th harmonics of each phase are normalized, and the amplitude range is uniformly mapped to the [0,1] interval. This operation eliminates the dimensional differences caused by different voltage levels and current ranges, enabling the harmonic characteristics of different fault samples to have lateral comparison capabilities.

[0084] Preferably, the maximum value of each harmonic amplitude is used as the benchmark, and the normalization formula is as follows:

[0085] (5);

[0086] In the formula: This is a function to find the maximum value. This represents the normalized amplitude of the kth harmonic.

[0087] D4: Polar coordinates to rectangular coordinates conversion.

[0088] By uniformly distributing the polar angles using equation (6), each harmonic is sequentially mapped to one of the eight uniform angles in polar coordinates, ensuring the symmetry of the harmonic distribution and the clarity of its visualization. (6);

[0089] Then, a Cartesian coordinate transformation is performed. Using the normalized harmonic amplitude as the polar radius and the corresponding polar angle, the coordinates in the Cartesian coordinate system are calculated. The corresponding formula is as follows: (7).

[0090] D5: Steps for drawing harmonic radar images.

[0091] Based on the discrete point data after Cartesian coordinate transformation, the points are connected to form a closed harmonic radar map. The three-phase voltage and three-phase current are mapped into six sub-maps respectively, and combined to form a fault sample image as shown in Figure 2. Preferably, this example uses a two-row, three-column sub-map arrangement. The first row of sub-maps contains the three-phase voltage images of A, B, and C in sequence, and the second row contains the three-phase current images of A, B, and C in sequence.

[0092] Next, the step of inputting the harmonic radar image into a lightweight convolutional neural network to extract electrical features is performed. Specifically, this includes the following steps E1-E5.

[0093] E1: Image preprocessing steps. That is, initial feature extraction is performed on the harmonic radar image feature map through a 3×3 convolutional layer and a 3×3 max pooling layer in the header.

[0094] E2: Bottleneck layer sampling step. That is, the main body consists of multiple bottleneck layer (Bottleneck, bneck) modules, which perform feature downsampling step by step by setting different step lengths.

[0095] E2.1: Steps for Channel Upscaling of Feature Maps. Specifically, the preprocessed feature map is upscaled through 1×1 convolutions, with an expansion factor t set to 2, meaning the number of channels after upscaling is twice the number of input channels. Channel upscaling increases the expressive power of the features. Specifically, channel upscaling is performed using the following formula:

[0096] (8);

[0097] In the formula: for Convolution, used for channel upscaling; input feature map H and W represent the height and width of the space, respectively. This represents the number of input channels. The expansion factor is usually set to t∈{2,4,6}, i.e. .

[0098] E2.2: Steps for depthwise convolution and spatial feature extraction.

[0099] A 3×3 depthwise convolution is used to perform spatial filtering on the upscaled features. The depthwise convolution is performed independently on each channel, which can effectively extract the spatial information of the feature map while preserving the independence between channels and reducing computational overhead.

[0100] Specifically, the formula for depthwise convolution of the extended features calculated by E2 is as follows:

[0101] (9);

[0102] In the formula: k represents the kernel size. This operation performs spatial filtering independently on each channel, preserving the independence between channels to reduce computational overhead.

[0103] E2.3: Steps for SE attention module modulation. Specifically, the SE attention module performs channel weight calibration on the features after depthwise convolution. The specific operations are as follows: Global average pooling is used to compress the spatial dimension into a scalar, followed by ReLU activation, a fully connected layer, and a sigmoid function to obtain the weight vector for each channel. The corresponding weight vector calculation formula is as follows:

[0104] (10);

[0105] In the formula: This indicates global average pooling, which compresses the spatial dimension into a scalar. It is the ReLU activation function. For the Sigmoid function; and Here are the weight matrices for the two fully connected layers.

[0106] Furthermore, the weight vector is multiplied channel-by-channel with the depthwise convolutional features to enhance effective feature channels and suppress ineffective feature channels. At the tail end, average pooling is used to compress the spatial dimension, followed by 1×1 convolution for channel dimensionality reduction, outputting deep electrical features.

[0107] E3: Steps for channel dimensionality reduction and nonlinear enhancement.

[0108] Channel compression is performed using 3x3 convolution, and the HS (HardSwish) activation function is introduced to enhance nonlinear expressive power. The formula is as follows, and the final output is the extracted image features.

[0109] (11).

[0110] Next, the electrical features are dynamically modulated using the weather semantic vector as the query through the cross-attention fusion unit of the fault identification model. Specifically, this includes the following steps F1-F3.

[0111] F1: Steps for feature structure reconstruction.

[0112] The output image features are reconstructed into key-value pairs, where the key... ,value .

[0113] F2: Steps for calculating attention weights.

[0114] The formula for calculating the cross-attention weights using weather semantic vectors and key matrices is as follows:

[0115] ;

[0116] In the formula: The cross-attention weights represent the degree of attention that weather semantics pays to each spatial location of the waveform; The weather semantic vector is used to query the key through multiplication. Similarity; This is a scaling factor used to alleviate gradient saturation issues caused by high-dimensional dot products.

[0117] F3: Steps for weighted fusion and feature output.

[0118] Furthermore, by weighted summation of the value matrices, the fused feature vector is output. :

[0119] .

[0120] After completing the feature fusion steps described above, the fused features are input into the classification and recognition unit. The unit outputs the probability distribution of preset fault causes such as lightning strikes, wildfires, and icing through a fully connected layer and a normalized exponential function. The category corresponding to the highest probability is used as the final identification result. Specifically, this includes the following steps G1-G3.

[0121] G1: The steps of feature mapping.

[0122] The fused feature vector z is input into a fully connected layer and a non-linear activation function to map the high-dimensional fused features into a 6-dimensional feature vector.

[0123] G2: Steps for calculating probability distribution.

[0124] The mapped 6-dimensional feature vector is input into the Softmax output layer to calculate the probability of occurrence of each fault cause. This function transforms the scores of different fault causes into probabilistic form through exponential operations, and its calculation expression is as follows: ;

[0125] In the formula: Let i be the probability value of class i. is the j-th result output by the fully connected layer; k is the total number of categories. This function converts the feature vector into a probability distribution through exponential operation and normalization. Each probability value is in the interval [0,1], and the sum of all probability values ​​is 1, which intuitively reflects the probability of various faults occurring.

[0126] G3: Fault Cause Determination and Output

[0127] The fault category with the largest value in the probability distribution is selected as the final fault cause identification result.

[0128] Based on the above steps, the initial fault identification model is trained using historical harmonic radar images and corresponding historical weather semantic vectors as training inputs, and confirmed fault cause labels as training targets, until the model converges, thus obtaining the pre-trained fault identification model. That is, through the steps A1-A2, B1-B2, C1-C2, D1-D5, E1-E5, F1-F3, and G1-G3 described above, the construction and training of the fault identification model can be achieved.

[0129] Furthermore, it is important to emphasize that the process of identifying faults in the current transmission line based on a pre-built and trained fault identification model also involves steps C1-C2, D1-D5, E1-E5, F1-F3, and G1-G3. However, unlike the fault identification model building and training process, these steps target the current fault waveform data and weather information at the time of the fault, rather than historical fault sample data. That is, in executing S1—obtaining the current fault waveform data and weather information of the current transmission line—the steps focus on the current fault waveform data and weather information at the time of the fault. After obtaining the weather information at the time of the incident, steps S3-S4 are implemented in the manner described in steps C1-C2, D1-D5, E1-E5, F1-F3, and G1-G3. Furthermore, it should be emphasized that during the process of identifying the fault of the transmission line to be identified based on the pre-built and trained fault identification model, the steps related to the construction and training of the knowledge graph embedding model are no longer executed. Instead, the pre-built and trained knowledge graph embedding model is directly called. That is, step S2 is implemented based on calling the pre-built and trained knowledge graph embedding model.

[0130] This application also discloses a system for implementing the transmission line fault identification method that integrates weather semantics and harmonic images as described in this application, comprising: a data acquisition unit configured to acquire current fault waveform data of the transmission line to be identified and weather information at the time of the current fault occurrence; a semantic vector conversion unit configured to perform semantic encoding processing on the weather information of the current fault to generate a low-dimensional dense vector representing the semantic association between the current fault and the fault cause, as the current weather semantic vector; and an image generation unit configured to process the waveform data of the current fault to generate a current harmonic radar image representing the harmonic characteristics of the fault electrical quantities.

[0131] The fault identification model includes: an input unit configured to receive inputs of the current weather semantic vector and the harmonic radar image; an image feature extraction unit configured to extract deep electrical features from the harmonic radar image; a multimodal cross-attention fusion unit configured to dynamically modulate the deep electrical features using the current weather semantic vector as a query vector to obtain a fused feature vector incorporating weather information; and a classification and recognition unit configured to output the identification result of the current transmission line fault cause based on the fused feature vector.

[0132] This application also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the transmission line fault identification method that integrates weather semantics and harmonic images disclosed in the first aspect of the present invention.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying power transmission line faults by integrating weather semantics and harmonic images, characterized in that, Includes the following steps: S1, acquire the current fault waveform data of the transmission line to be identified and the weather information at the time of the current fault; S2, perform semantic encoding on the weather information of the current fault to generate a low-dimensional dense vector that represents the semantic association between the fault and its cause, as the current weather semantic vector; S3, process the current fault recording data to generate a current harmonic radar image characterizing the harmonic characteristics of the fault electrical quantities; S4, input the current weather semantic vector and the current harmonic radar image into the pre-trained fault identification model, wherein the fault identification model is configured to extract the deep electrical features of the harmonic radar image, and use the current weather semantic vector as a query vector to dynamically modulate the deep electrical features to obtain a fused feature vector that incorporates weather information, and output the identification result of the current transmission line fault cause based on the fused feature vector.

2. The method according to claim 1, characterized in that, The fault identification model uses the current weather semantic vector as the query vector to dynamically modulate the deep electrical features, including: The deep electrical features are restructured and converted into key and value matrices that can be computed using the cross-attention mechanism. Using weather semantic vectors as query vectors, calculate their similarity to the key matrix to generate cross-attention weights; By using cross-attention weights to perform weighted summation of the value matrix, the weather semantics are dynamically modulated to the deep electrical features, and a fused feature vector is output.

3. The method according to claim 1, characterized in that, The fault identification model, based on the fused feature vector, outputs the identification result of the cause of the current transmission line fault, including: The fused feature vector is input into at least one fully connected layer of the fault identification model for feature dimension mapping and nonlinear transformation. The transformed features are converted into probability distributions corresponding to multiple fault causes by using a normalized exponential function; The fault cause with the largest value in the probability distribution is taken as the final fault cause identification result.

4. The method according to claim 1, characterized in that, Step S3 includes: Extract the three-phase voltage and three-phase current waveform data within a preset number of cycles before and after the fault; The amplitude of each phase and harmonic in the recorded waveform data is extracted based on the Fast Fourier Transform, forming a harmonic amplitude sequence. Normalize each harmonic amplitude sequence; Each amplitude in the normalized harmonic amplitude sequence is mapped sequentially to a preset corresponding angle in the polar coordinate system to obtain a set of polar coordinate points. Transform the polar coordinate point set to a rectangular coordinate system and draw connecting lines to form a single-phase harmonic radar sub-map; Multiple single-phase harmonic radar sub-images are combined to generate a multi-channel current harmonic radar image.

5. The method according to claim 1, characterized in that, The fault identification model is equipped with deep electrical features extracted from harmonic radar images, including the following steps performed sequentially: Initial feature extraction of harmonic radar images is performed using a maximum pooling layer based on a fault identification model. Based on the fault identification model, multiple bottleneck layer modules gradually complete feature sampling under different time lengths. The average pooling layer based on the fault identification model compresses the spatial dimension and performs channel dimensionality reduction to output the final deep electrical features.

6. The method according to any one of claims 1-5, characterized in that, Prior to step S1 or step S2, a step of pre-constructing a knowledge graph embedding model for semantic encoding is also included, which includes: Based on historical fault sample sets, different weather information is standardized and classified to form a finite set of standard weather categories; Calculate the conditional probability of each fault cause occurring under each standard weather category; Based on a preset correlation strength threshold, the correlation strength level between weather categories and fault causes is classified. Each group is constructed as a weighted knowledge graph consisting of a triplet of weather category, correlation strength level, and fault cause; A knowledge graph embedding algorithm is used to train a weighted knowledge graph to obtain a knowledge graph embedding model, which enables the model to map weather information at the time of a fault into a low-latitude dense weather semantic vector. Step S2 is based on the final knowledge graph obtained for semantic encoding.

7. The method according to claim 6, characterized in that, The knowledge graph embedding algorithm is a translation embedding model, and its training steps for the weighted knowledge graph include: Set the loss function for training the translation embedding model; The initial knowledge graph is embedded and trained so that the resulting vectors can be used to represent the semantic association between weather information and various causes of failure.

8. The method according to claim 6, characterized in that, It also includes the step of pre-building a fault identification model, which includes: Obtain a historical fault sample set that includes historical fault waveform data, historical fault weather information, and confirmed fault cause tags; The historical fault weather information is processed using the pre-trained knowledge graph embedding model to obtain historical weather semantic vectors; The historical fault recording data is processed to generate historical harmonic radar images; An initial fault identification model is constructed, which includes at least an image feature extraction module for extracting deep electrical features from images, a multimodal fusion module for dynamically modulating and fusing deep electrical features using weather semantic vectors as query vectors, and a classification output module for outputting fault identification results. Using the historical harmonic radar image and the corresponding historical weather semantic vector as training input, and the confirmed fault cause label as training target, the initial fault identification model is trained until the model converges, thus obtaining the pre-trained fault identification model.

9. A system for implementing the transmission line fault identification method based on the fusion of weather semantics and harmonic images according to any one of claims 1-8, characterized in that, include: The data acquisition unit is configured to acquire the current fault waveform data of the transmission line to be identified and the weather information at the time of the current fault. The semantic vector transformation unit is constructed by performing semantic encoding processing on the weather information of the current fault to generate a low-dimensional dense vector representing the semantic association between the fault and the cause of the fault, which serves as the current weather semantic vector. The image generation unit is configured to process the recorded waveform data of the current fault and generate a current harmonic radar image characterizing the harmonic characteristics of the fault electrical quantities. The fault identification model includes: The input unit is configured to receive the current weather semantic vector and the harmonic radar image as input; The image feature extraction unit is constructed to extract deep electrical features from harmonic radar images; The multimodal cross-attention fusion unit is constructed to dynamically modulate the deep electrical features using the current weather semantic vector as the query vector, thereby obtaining a fused feature vector that incorporates weather information. The classification and identification unit is configured to output the identification result of the cause of the current transmission line fault based on the fused feature vector.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the transmission line fault identification method that integrates weather semantics and harmonic images as described in any one of claims 1 to 8.

Images