Power transmission line conductor galloping early warning method and system based on unmanned aerial vehicle edge identification

Abstract

The invention discloses a power transmission line conductor galloping early warning method and system based on unmanned aerial vehicle edge identification, and relates to the technical field of conductor galloping early warning, and the method comprises the steps: planning a flight route of an unmanned aerial vehicle, deploying an edge calculation device and a neural network model group, enabling the unmanned aerial vehicle to fly along the flight route to collect image data and environment data, extracting a pixel mask of the wire and identifying a wire icing phenomenon through the neural network model group; generating a pixel-level displacement time sequence based on the pixel mask of the lead, converting the pixel-level displacement time sequence into a vibration displacement sequence, and extracting multi-scale vibration features from the vibration displacement sequence by using short-time Fourier transform and continuous wavelet transform to obtain a two-dimensional dual-channel vibration feature map; and outputting a galloping confidence score through a deep network based on the two-dimensional dual-channel vibration feature map. According to the method, high-precision and low-delay identification of conductor galloping is realized through multi-scale feature fusion and edge intelligent identification, and the false alarm rate of early warning is effectively reduced by combining propagation situation analysis and multi-source criterion graded early warning.

Description

Technical Field

[0001] This invention relates to the field of conductor galloping early warning technology, and in particular to a method and system for early warning of conductor galloping in power transmission lines based on edge recognition by unmanned aerial vehicles. Background Technology

[0002] With the continuous expansion of power systems and the increase in the length of transmission lines, ensuring the safe and stable operation of transmission lines has become a crucial task. In recent years, due to frequent extreme weather conditions, the phenomenon of transmission line galloping has become increasingly serious, posing a significant threat to the safety of the power grid. Traditional monitoring methods mainly rely on manual inspections or real-time monitoring based on fixed sensor networks; however, these methods have many limitations. On the one hand, manual inspections are inefficient and costly; on the other hand, while fixed sensors can provide real-time data, their deployment and maintenance costs are high, and they are difficult to cover long-distance transmission lines in complex terrain. Therefore, using unmanned aerial vehicles (UAVs) for remote monitoring and identification has gradually become a research hotspot.

[0003] While existing UAV-based power transmission line monitoring systems have improved upon traditional methods to some extent, they still face challenges in processing non-stationary vibration signals. First, in the feature extraction stage, there is a lack of an effective way to integrate different types of time-frequency features to more comprehensively describe the characteristics of the vibration signal. Second, existing machine learning models often do not adequately consider the time-series characteristics of vibration signals, which may lead to insufficient capture of long-term dependencies. Most early warning systems fail to fully utilize the influence of meteorological factors such as wind direction and other external conditions, thus affecting the accuracy of early warning results. Therefore, current technologies still suffer from problems such as insufficient feature integration, inadequate consideration of time-series characteristics, and underutilization of the influence of external conditions. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method and system for early warning of conductor galloping in power transmission lines based on UAV edge recognition, which solves the problems of insufficient feature integration, insufficient consideration of time series characteristics, and underutilization of the influence of external conditions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for early warning of conductor galloping in power transmission lines based on edge recognition by unmanned aerial vehicles (UAVs), comprising, Plan the flight path of the drone and deploy edge computing devices and neural network model groups. The drone flies along the flight path to collect image data and environmental data, and extracts the pixel mask of the conductor and identifies the icing phenomenon of the conductor through the neural network model group. The neural network model group consists of a semantic segmentation model and a convolutional neural network. Pixel-level displacement time series are generated based on the pixel mask of the wire and converted into vibration displacement sequence. The vibration displacement sequence is then subjected to short-time Fourier transform and continuous wavelet transform to extract multi-scale vibration features and obtain a two-dimensional dual-channel vibration feature map. Based on the two-dimensional dual-channel vibration feature map, the galloping confidence score is output through a deep network, and the galloping propagation early warning index is calculated to divide the galloping stage. The graded early warning is then implemented in conjunction with the conductor icing phenomenon.

[0007] As a preferred embodiment of the transmission line conductor galloping early warning method based on UAV edge recognition described in this invention, the method involves: extracting multi-scale vibration features from the vibration displacement sequence using short-time Fourier transform and continuous wavelet transform to obtain a two-dimensional dual-channel vibration feature map; performing boundary extension processing and interpolation on the vibration displacement sequence to obtain a continuous-time signal; performing short-time Fourier transform and modulus operation to obtain the amplitude spectrum; using continuous wavelet transform for supplementary analysis; and selecting the optimal bandwidth parameter based on the Shannon entropy criterion. The complex Morlet wavelet basis function is reconstructed, and continuous wavelet transform is performed to obtain wavelet coefficients. The middle part corresponding to the time period of the original signal is extracted from the wavelet coefficients, and the coefficients corresponding to the extended regions at both ends are discarded to obtain an effective wavelet coefficient segment with the same time period as the original vibration signal. The wavelet energy spectrum is obtained by taking the modulus square of the wavelet coefficients of the effective segment. The amplitude spectrum is used as the first channel and the wavelet energy spectrum is used as the second channel. Normalization and linear interpolation are performed uniformly. After normalization and alignment, the data of the two channels are stitched together along the channel dimension to form a two-dimensional dual-channel vibration feature map.

[0008] As a preferred embodiment of the transmission line conductor galloping early warning method based on UAV edge recognition described in this invention, the method involves: using a two-dimensional dual-channel vibration feature map, outputting a galloping confidence score through a deep network, dividing the galloping stages, and calculating a propagation early warning index. The two-dimensional dual-channel vibration feature map is then input into a lightweight CNN-BiLSTM-Attention network model to obtain the final galloping confidence score, and a confidence threshold is set. When the dance confidence score S is greater than or equal to the threshold If a signal is detected as having a clear galloping trend, it is determined that the conductor exhibits a significant galloping tendency; otherwise, it is considered normal. When a clear galloping trend is detected, the number of main peaks, energy concentration, and frequency drift rate are extracted from the amplitude spectrum to represent the spectral characteristics. These are then input into a support vector machine classifier for galloping stage identification. The classifier outputs a current galloping state label and performs cross-correlation analysis on the vibration displacement sequences of any two adjacent sampling points to obtain the cross-correlation function. By finding the relative time delay that maximizes the cross-correlation function, the actual propagation time delay is obtained by combining it with the sampling interval. Based on the spatial distance between two sampling points of the transmission line and the corresponding actual propagation time delay, the propagation speed of the vibration signal in the section is calculated. Based on the vibration propagation speed between all sampling points, the propagation speed of the entire line is vector-fitted to obtain the overall propagation direction and average propagation speed vector. The average propagation speed vector is corrected with the real-time wind direction, and a galloping propagation early warning index is defined, with a threshold set for the propagation early warning index. It assesses the risk of gossamer spread and outputs the propagation speed, direction, and early warning index of each sampling point and section, automatically generates a gossamer spread trend map, and marks high-risk sections.

[0009] As a preferred embodiment of the transmission line conductor galloping early warning method based on UAV edge recognition described in this invention, the step of combining conductor icing phenomenon with graded early warning refers to performing graded early warning based on the galloping propagation trend map and icing phenomenon. After triggering any level of early warning, it is encapsulated into a standardized early warning data packet, and the standardized early warning data packet is uploaded to the power grid dispatch center and operation and maintenance management platform in real time, triggering audible and visual alarms, generating disposal work orders, or activating anti-galloping device linkage control.

[0010] As a preferred embodiment of the power transmission line conductor galloping early warning method based on UAV edge recognition described in this invention, the step of generating pixel-level displacement time sequence based on the conductor pixel mask and converting it into a vibration displacement sequence refers to further processing the image using morphological closing operation after obtaining the conductor pixel mask, extracting the conductor skeleton line, selecting N equally spaced sampling points on the conductor skeleton line, constructing a space-time displacement matrix based on the coordinates of all sampling points, and calculating the lateral spatial resolution by measuring the pixel width of the conductor in the image and using monocular visual geometric relationships. Based on the lateral spatial resolution, the pixel offset is converted into vibration displacement to construct a vibration displacement sequence.

[0011] As a preferred embodiment of the power transmission line conductor galloping early warning method based on UAV edge recognition described in this invention, the UAV flies along the flight path to collect image data and environmental data, and extracts the pixel mask of the conductor and identifies icing phenomena through a neural network model group. After takeoff, the UAV begins to cruise along the planned route. After entering the target area, it takes pictures. While collecting images, the UAV uses an external micro-meteorological sensor to collect wind speed and direction in real time, and records the GPS position and attitude angle during flight. At the same time, on the Jetson Xavier NX edge computing platform, the neural network model group is run synchronously to extract the pixel mask of the conductor and icing phenomena. All collected image data and environmental data are stored synchronously with the UTC timestamps during flight.

[0012] As a preferred embodiment of the power transmission line conductor galloping early warning method based on UAV edge recognition described in this invention, the following steps are taken: planning the UAV flight path and deploying edge computing devices and neural network model groups refer to acquiring the geographical coordinates and historical galloping section data of the power transmission line to be monitored, using aerial survey software to automatically calculate a flight path covering the target area based on the distribution of the power transmission line and key areas, and after completing the planning, selecting Jetson Xavier NX as the edge computing platform and deploying the neural network model group on Jetson Xavier NX.

[0013] Secondly, the present invention provides a power transmission line conductor galloping early warning system based on UAV edge recognition, comprising, The flight path planning and model deployment module is used to plan the flight path of the UAV and deploy edge computing devices and neural network model groups to support real-time monitoring of task execution. The displacement extraction module is used to generate a sequence of equally spaced sampling points based on the wire pixel mask and convert it into physical vibration displacement time series data; The feature fusion module is used to combine the results of short-time Fourier transform and optimal wavelet basis function analysis to construct a dual-channel normalized time-frequency feature map; The dance recognition module is used to input dual-channel feature maps to a lightweight network to output confidence scores, and combines SVM to achieve dance stage classification. The vibration early warning module is used to calculate the vibration propagation speed and direction, integrate wind direction consistency to generate a propagation early warning index, and trigger a graded response.

[0014] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the method for early warning of conductor galloping of power transmission lines based on UAV edge recognition as described in the first aspect of the present invention.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the method for early warning of conductor galloping of transmission lines based on UAV edge recognition as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By employing multi-scale time-frequency analysis technology, including short-time Fourier transform (STFT) and continuous wavelet transform (CWT) based on the optimal bandwidth adaptive selection mechanism, this invention generates first and second channel feature maps containing rich hierarchical information. This not only enhances the ability to capture non-stationary vibration signals but also improves the energy focusing effect, making it particularly suitable for identifying weak oscillation characteristics under icing conditions. Furthermore, by combining a lightweight CNN-BiLSTM-Attention network model, it effectively extracts multi-level features from vibration signals and utilizes a bidirectional long short-term memory network (BiLSTM) to capture the time-series dependencies, thereby improving the accuracy of abnormal fluctuation detection. Finally, by comprehensively considering the dance confidence score, propagation speed, and direction propagation warning index, it achieves more accurate dance risk warning. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a transmission line conductor galloping early warning method based on UAV edge recognition in Example 1.

[0019] Figure 2 This is a structural diagram of a power transmission line conductor galloping early warning system based on UAV edge recognition in Example 1.

[0020] Figure 3 This is a flowchart of the dancing warning process in Example 1. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] Example 1, referring to Figures 1-3 This is the first embodiment of the present invention, which provides a method for early warning of conductor galloping on power transmission lines based on UAV edge recognition, including the following steps: S1. Plan the flight path of the UAV and deploy edge computing devices and neural network model groups. The UAV flies along the flight path to collect image data and environmental data, and extracts the pixel mask of the conductor and identifies the icing phenomenon of the conductor through the neural network model group. Specifically, planning drone flight paths and deploying edge computing devices and neural network models are crucial for accurate monitoring of power transmission lines. This involves acquiring the geographical coordinates (such as latitude and longitude) of the transmission lines to be monitored, as well as historical data on galloping sections (based on records of past vibration or icing events). This data is used to set the drone's flight path, ensuring that the flight path covers all areas where faults may occur. Using aerial surveying software, a flight path covering the target area is automatically calculated based on the distribution of the power transmission lines and key areas (sections where conductor galloping or icing accidents have occurred). During flight, the drone's frontal view angle (i.e., vertical angle) is set to ensure that the drone can photograph the power transmission lines from a vertical perspective, maximizing the quality and accuracy of image acquisition. After setup, the planned flight path is simulated to ensure complete coverage, and verified through an aerial survey system. If any areas are missed, the flight path is adjusted until the entire monitoring area is covered. After the flight path is planned, Jetson Xavier NX is selected as the edge computing platform, and a neural network model group is deployed on Jetson Xavier NX to perform both conductor semantic segmentation and conductor icing detection simultaneously. This includes a semantic segmentation model (conductor semantic segmentation) and a convolutional neural network (CNN) (icing detection). The semantic segmentation model adopts the U-Net architecture as the conductor semantic segmentation model, with MobileNetV2 as the encoder part. During training, it uses an annotated transmission line image dataset and performs end-to-end learning by minimizing the Dice+BCE loss function. After training, the model is converted to TensorRT format for deployment and operation. The convolutional neural network uses ResNet-18 as the icing detection convolutional neural network. The input is the image of the conductor region extracted by the semantic segmentation model. The model is trained on labeled icing / normal image pairs through supervised learning. The cross-entropy loss function is used to optimize the classification performance. After training, the model is deployed and run.

[0025] By automatically planning flight paths and controlling the vertical perspective, the integrity and geometric stability of image acquisition are ensured. Pixel-level accurate segmentation of conductors is achieved with the help of edge-end U-Net / MobileNetV2, and icing status classification is completed by combining ResNet-18. A lightweight, efficient, and hierarchical dual-model collaborative mechanism is constructed, which not only reduces the dependence on communication bandwidth, but also realizes real-time identification of abnormal conductor status in complex environments, significantly improving monitoring efficiency and timely early warning.

[0026] Furthermore, the UAV flies along the flight path to collect image and environmental data, and extracts pixel masks of the conductors and identifies icing phenomena through a neural network model group. After takeoff, the UAV begins to cruise along the planned route. After entering the target area, the UAV will take high-definition visible light images at a fixed frequency to ensure the time alignment of the image sequence, which facilitates subsequent data processing and analysis. While acquiring images, the UAV uses external micro-meteorological sensors (such as ultrasonic anemometers or wind speed meters) to collect wind speed and direction in real time. The UAV's flight control system records the GPS position (longitude, latitude, and altitude information) and attitude angles (heading angle, pitch angle, and roll angle) in real time during the flight. At the same time, the neural network model group runs synchronously on the Jetson XavierNX edge computing platform. Each frame of the image is processed by a semantic segmentation model to extract the pixel mask of the wire (identifying the pixel region in the image that belongs to the wire). The model performs real-time classification analysis on the conductor images using a convolutional neural network (CNN). The model outputs a binary criterion: 0 indicates no obvious icing and 1 indicates the presence of icing. All collected image data and environmental data (GPS, attitude angle, wind speed, wind direction) are synchronized and stored with the UTC timestamps (based on a globally unified time standard) during the flight.

[0027] By using multi-source sensor collaborative acquisition and real-time edge processing, high-precision time synchronization and spatial matching of image, meteorological, and pose data are achieved. Semantic segmentation is used to accurately extract conductor masks, and CNN is combined to complete the icing status discrimination. While ensuring the timeliness of processing, the robustness of recognition is improved. The established multimodal synchronous storage mechanism supports the joint analysis of subsequent vibration characteristics and environmental factors, significantly enhancing the integrity and traceability of the monitoring system.

[0028] S2. Pixel-level displacement time series are generated based on the pixel mask of the wire and converted into vibration displacement sequence. The vibration displacement sequence is then subjected to short-time Fourier transform and continuous wavelet transform to extract multi-scale vibration features and obtain a two-dimensional dual-channel vibration feature map. Specifically, pixel-level displacement time series are generated based on the pixel mask of the conductor and converted into a vibration displacement sequence. After obtaining the pixel mask of the conductor, morphological closing operations are used to further process the image. Erosion and dilation operations are used to effectively fill the holes in the conductor mask and remove small noise points to extract the skeleton line of the conductor. N equally spaced sampling points are selected on the skeleton line of the conductor. The position of each sampling point is determined by the image coordinate system (x, y coordinates) to ensure that all important parts of the conductor are covered. A space-time displacement matrix is ​​constructed based on the coordinates of all sampling points. Where Y is a pixel-level displacement sequence, It is the pixel offset of the i-th sampling point; To achieve the conversion between image data and the physical world, the lateral spatial resolution is calculated by measuring the pixel width of the lines in the image and utilizing monocular vision geometric relationships: in, is the horizontal spatial resolution, w is the pixel width of the wire in the image, and D is the actual diameter of the wire; Based on horizontal spatial resolution The pixel offset is converted into vibration displacement, and a vibration displacement sequence is constructed: in, It is a vibration displacement sequence. It is the vibration displacement of the i-th sampling point.

[0029] By optimizing the morphology and skeletalization process, the geometric integrity of the conductor mask was improved. By constructing a space-time displacement matrix using equidistant sampling, a structured characterization of the vibration process was achieved. Combined with a monocular vision scale mapping model, the accurate conversion from pixel to physical displacement was completed, solving the measurement inaccuracy problem caused by changes in shooting distance in UAV inspection. The generated physical vibration displacement sequence has the advantages of being quantifiable, traceable, and capable of participating in dynamic analysis, thus improving the accuracy of non-contact vibration monitoring of transmission lines.

[0030] Furthermore, short-time Fourier transform and continuous wavelet transform are used to extract multi-scale vibration features from the vibration displacement sequence to obtain a two-dimensional dual-channel vibration feature map. Boundary extension processing is performed to resolve signal edge effects and ensure the accuracy of subsequent transformations. A symmetric extension method is used for the vibration displacement sequence. The beginning and end of the vibration signal are extended by mirroring the starting segment forward with the first sampling point as the center of symmetry, and mirroring the ending segment backward with the last sampling point as the center of symmetry, and then linear interpolation is performed to obtain the continuous time signal. ,in, It is a continuous-time variable, and performs a Short-Time Fourier Transform (STFT) to analyze the variation of frequency components over time. The formula for calculating the Short-Time Fourier Transform is: Where STFT(t,f) is the complex amplitude of frequency f at time t. It is a continuous-time signal. It is the Hanning window function. , Here, e is the limit of integration, e is the natural constant, and j is the imaginary unit. Pi is the mathematical constant, and f is the frequency. It is the time infinitesimal element of the integral variable; The STFT is calculated segment by segment by sliding the window along the time axis, and the amplitude spectrum |F(t,f)| is obtained by modulo operation. To improve the resolution of non-stationary vibrations, continuous wavelet transform (CWT) is used for supplementary analysis to enhance the analytical capability of non-stationary vibration signals, especially the capture of local features. The complex Morlet wavelet is used as the mother function, combined with an adjustable bandwidth parameter. Construct complex Morlet wavelet basis functions to balance time-frequency resolution, the expression of which is: in, The bandwidth parameter is Complex Morlet wavelet basis functions, It is the center frequency. It is a time variable, and j is the imaginary unit; To achieve optimal energy focusing, the best bandwidth parameter was selected based on the Shannon entropy criterion, and the complex Morlet wavelet basis function was reconstructed to adjust the time-frequency resolution balance. Based on the characteristics of the transmission line vibration signal, a set of discrete candidate values ​​was defined. That is, the set of bandwidth parameters, using the complex Morlet wavelet basis functions of each bandwidth parameter to calculate the continuous wavelet transform coefficients: in, It is the complex Morlet wavelet basis function constructed from the k-th candidate bandwidth parameter. The wavelet coefficients are obtained by performing wavelet transform on the vibration signal at scale a and translation position b, reflecting the local vibration characteristics. 'a' is the scale factor, which is approximately inversely proportional to the frequency, and 'b' is the translation factor, representing the time position. It is a complex conjugate. It is an integral variable. It is the k-th candidate bandwidth parameter; The energy distribution is obtained by taking the squared modulus of the continuous wavelet transform result for each candidate bandwidth parameter. The probability distribution is then obtained by normalizing the energy distribution using standard energy normalization techniques. And calculate the Shannon entropy: in, It is Shannon entropy; Choose the bandwidth parameter that minimizes the Shannon entropy as the optimal bandwidth parameter: in, It is the optimal bandwidth parameter; Use optimal bandwidth parameters Constructing the final wavelet basis functions Perform a continuous wavelet transform to obtain the wavelet coefficients: Where CWT(a,b) are the wavelet coefficients at scale a and translation b. It is the i-th sampling point in the extended vibration displacement sequence in continuous time. Vibration displacement at the location; After the transformation, the middle portion corresponding to the original signal time period is extracted from the wavelet coefficients CWT(a,b), and the coefficients corresponding to the extended regions at both ends are discarded to obtain an effective wavelet coefficient segment with the same time period as the original vibration signal. The wavelet energy spectrum Q(a,b) is obtained by taking the modulus and square of the wavelet coefficients of the effective segment. The amplitude spectrum |F(t,f)| is used as the first channel, and the wavelet energy spectrum Q(a,b) is used as the second channel. Min-Max normalization is performed uniformly. Since the time resolution of the Short Time Fourier Transform (STFT) and the Continuous Wavelet Transform (CWT) is different (depending on their respective parameter settings), linear interpolation is used to unify the time axis of the two channels to the same number of time frames, and the frequency axis to the same number of frequency points. After normalization and alignment, the data of the two channels are stitched together along the "channel" dimension to form a two-dimensional dual-channel vibration feature map. .

[0031] By processing the vibration displacement sequence through boundary extension, the boundary effect problem in non-stationary signal analysis is solved, and the accuracy of frequency domain feature extraction is improved. The time-frequency characteristics of the signal are captured by short-time Fourier transform (STFT), laying the foundation for subsequent feature fusion. Combined with complex Morlet wavelet and its optimal bandwidth selection mechanism, the analytical capability for local singularities and transient vibration features is significantly enhanced, overcoming the limitations of traditional fixed-bandwidth wavelet transform. Finally, by combining STFT and CWT to construct a two-dimensional dual-channel vibration feature map, the effective integration and enhanced expression of multi-level features are realized, providing high-quality input for deep learning models and significantly improving the recognition accuracy and reliability of transmission line galloping state.

[0032] S3. Based on the two-dimensional dual-channel vibration feature map, the galloping confidence score is output through a deep network, and the galloping propagation early warning index is calculated to divide the galloping stage. The graded early warning is then implemented in conjunction with the conductor icing phenomenon.

[0033] Specifically, based on the two-dimensional dual-channel vibration feature map, a goblins confidence score is output through a deep network to divide the goblins into stages, and a propagation warning index is calculated. Input a lightweight CNN-BiLSTM-Attention network model; The lightweight CNN-BiLSTM-Attention network model is composed of a convolutional neural network (CNN) module, a bidirectional long short-term memory network (BiLSTM) module, and an attention mechanism (Attention) module connected in series. The model is trained offline based on a large number of labeled historical transmission line galloping samples. The input is the vibration displacement sequence or time-frequency feature map, and the output is the galloping confidence score. During the training process, the classification cross-entropy loss function is used to measure the prediction error, and the Adam optimization algorithm is used to iteratively update the network parameters. Training is stopped when the validation set loss no longer decreases for several consecutive cycles. After training, the model is structurally pruned to remove redundant neurons. High-level features of vibration signals in the frequency domain and local time segments are extracted using a convolutional neural network (CNN), resulting in a two-dimensional dual-channel vibration feature map. Frame segmentation is performed according to the time dimension to obtain a time frame sequence. ,in, Let be the frequency band energy distribution of the vibration signal in the two feature channels at time frame t, where T is the total number of time frames and corresponds to the temporal length of the vibration signal. A one-dimensional convolution operation (using a convolution kernel of length 3 and a stride of 1) is performed along the frequency dimension on each frame in the time frame sequence to extract local correlation features between adjacent frequency bands. The calculation formula is as follows: in, It is the feature mapping after the first convolutional layer. This is the convolution kernel weight matrix, set through a random initialization method. * represents a one-dimensional convolution operation. It is the bias vector, and ReLU(·) is the activation function of the linear rectified unit; To reduce feature dimensionality and preserve significant responses, the feature maps after the first convolutional layer are... Perform a 2×2 max pooling operation. The pooling operation takes the maximum value within a local region. The calculation formula is as follows: in, It is the feature map after pooling. It is an operation that takes the maximum value within a window of two adjacent data points; Based on the feature map after pooling Perform a second layer of one-dimensional convolution to capture higher-level frequency band correlation features: in, It is the output feature of the second convolutional layer. It is the weight matrix of the second-layer convolutional kernel, which is set through a random initialization method. It is the second-level bias term; Based on the output features of the second convolutional layer Perform a 2×2 max pooling operation to obtain the deep local feature representation. The pooling output of each frame is flattened to obtain a set of low-dimensional local feature vector sequences. ,in, This is the local feature vector of frame t at time t. All frame-level local feature vectors are concatenated in chronological order to form a complete time series matrix. Where Z is the complete time series matrix, containing complete time-dimensional structural information; To avoid the influence of different feature dimensions on subsequent modeling, the eigenvalues ​​of each dimension in the complete time series matrix Z are subjected to min-max normalization to unify the numerical range to the [0,1] interval, thus obtaining a standardized feature sequence. ,in, It is the local feature vector of the standardized time frame t, and the standardized feature sequence The input is fed into a bidirectional long short-term memory network (BiLSTM, which contains two LSTM layers, each with 64 hidden units to extract long-range feature associations in the time series) to simultaneously learn the forward and backward dependencies of the vibration signal in the time dimension. In each direction, the LSTM updates its internal state by performing forward operations according to a standard gating mechanism to obtain the forward hidden state. in, It is the forward hidden state at time t. It is the forward operation of LSTM. It is the forward hidden state of the previous moment; In a bidirectional structure, the backward hidden state is obtained by reversing the time sequence: in, It is in a backward hidden state. It is a backward operation of LSTM. It is the forward hidden state of the next moment; The forward and backward hidden states are concatenated to form a complete bidirectional output representation. and bidirectional output representation A nonlinear transformation is performed to calculate the attention score. The calculation formula is as follows: in, It's an attention score. It is the attention weight matrix, set using a random initialization method, used to map the hidden state to the attention space. It is an attention vector used to convert the result of a nonlinear transformation into a scalar score. It is the transpose operation, and tanh(·) is the hyperbolic tangent function. It is a bias term; Softmax normalization is applied to the attention scores at all times to obtain the attention weights at each time step. The hidden states at all time steps are weighted and summed according to the attention weights to obtain the global context vector: Where c is the global context vector; The global context vector c is input into the fully connected layer to map the vibration state vector to the confidence score. The fully connected layer obtains the predicted value through a linear transformation and normalizes it using the Sigmoid function, outputting the final dance confidence score. Where S is the dance confidence score. It is the Sigmoid function. These are the weight vectors of the fully connected layer, set using a random initialization method. It is the bias term of the fully connected layer, and exp(·) is the exponentiation operator; Confidence thresholds were set using statistical analysis. When the dance confidence score S is greater than or equal to the threshold If the conductor exhibits a clear galloping tendency, it is considered normal. When the conductor exhibits a clear galloping tendency, in order to further refine the division of the galloping development stage, the number of main peaks Z, energy concentration E, and frequency drift rate P are extracted from the amplitude spectrum F(t,f) to represent the spectral characteristics. The number of main peaks Z is in the power spectrum. The number of local maxima is identified to characterize the number of main vibration frequencies present in the signal; The energy concentration measure is whether the vibrational energy is concentrated near the dominant frequency, and the calculation formula is: Where E is the energy concentration. It is the sum of the spectral energy. It is the maximum value in the power spectrum, corresponding to the dominant frequency energy; The frequency drift rate reflects the trend of the main frequency over time and is used to determine whether the system is stable. The calculation formula is as follows: Where P is the frequency drift rate. It is the average frequency of the previous period. It is the average frequency of the main frequency in the subsequent time period. It is the difference in the start time between two consecutive time periods; The extracted number of main peaks Z, energy concentration E, and frequency drift rate P are input into a support vector machine (SVM) classifier for dance phase identification. The classifier outputs the current dance state label: Among them, Natural vibration refers to natural vibration, critical flutter refers to the critical flutter, and strong flutter refers to strong flutter. The Support Vector Machine (SVM) classifier is a multi-class support vector machine based on radial basis function kernels. It adopts a "one-to-many" classification strategy. During training, it uses labeled historical vibration datasets to extract the number of main peaks Z, energy concentration E, and frequency drift rate P as feature vectors. The features are then normalized, and the regularization parameters and kernel parameters are optimized through grid search to obtain the optimal classification model.

[0034] Cross-correlation analysis is performed on the vibration displacement sequences of any two adjacent sampling points to determine the propagation delay of the vibration signal between the two points. The cross-correlation function is defined as follows: in, It is the cross-correlation function between the i-th sampling point and the l-th sampling point. It is the relative time delay (in units of sampling points), where t is the time sampling frame number. It is the vibration displacement of the i-th sampling point in time frame t. Is the j-th sampling point in time frame The vibration displacement value, where T is the total number of time frames; The actual propagation time delay is obtained by finding the relative time delay that maximizes the cross-correlation function and combining it with the sampling interval: in, It is the actual propagation time delay of the vibration signal from the i-th sampling point to the l-th sampling point. It is the delayed index that maximizes the cross-correlation function. It is the sampling interval; Based on the spatial distance between two sampling points on the transmission line and the corresponding actual propagation time delay, the propagation speed of the vibration signal in the section is calculated: in, It is the speed of vibration propagation. It is the spatial distance between two sampling points on a transmission line, calculated using line design drawings or GPS coordinates. It is the absolute value of the actual propagation time delay; Based on the vibration propagation velocity among all sampling points, a vector fitting is performed on the propagation velocity of the entire line to obtain the overall propagation direction and average propagation velocity vector: in, It is the average propagation speed vector of the entire line. It is the unit vector of direction from sampling point i to sampling point l, and N is the total number of equally spaced sampling points; To comprehensively consider the influence of wind direction on vibration propagation, the average propagation velocity vector is... With real-time wind direction To correct this, we define the dance propagation early warning index: Where I is the transmission warning index, It is the magnitude of the average propagation speed. It represents the consistency between the propagation direction and the wind direction (the value is close to 1 when the propagation direction is consistent with the wind direction), and S is the dance confidence score. It is a preset critical dancing threshold coefficient, set based on historical data; Setting a propagation early warning index threshold through grid search. When the propagation warning index I is greater than the threshold If a significant vibration energy propagation trend is detected in a given section, indicating a potential risk of galloping spread, an early warning mechanism should be activated. The system will then output the vibration propagation speed, direction, and propagation warning index for each sampling point and section, automatically generating a "galloping propagation trend map" and marking high-risk sections as the basis for subsequent anti-galloping control or emergency response.

[0035] By employing a Convolutional Neural Network (CNN) module, one-dimensional convolution and pooling operations are used to extract local correlations between frequency bands layer by layer, effectively compressing redundant information and enhancing the response of key frequency bands. A Bidirectional Long Short-Term Memory (BiLSTM) network is used to process the standardized frame-level feature sequences, enabling joint learning of the contextual relationships of vibration signals. This overcomes the limitation of unidirectional RNNs in being unable to perceive future states, improving the ability to capture the evolutionary trend of vibration. Combined with an attention mechanism, the hidden states at key time steps are dynamically weighted, focusing on abnormal fluctuation intervals and enhancing the interpretability and robustness of the model's criteria. A fully connected layer and a sigmoid function output a vibration confidence score, setting a threshold to determine the existence of a vibration trend, realizing the transformation from continuous probability to discrete decision-making. Further, the amplitude... The number of main peaks, energy concentration, and frequency drift rate extracted from the value spectrum are used as physically interpretable features. These features are input into an RBF kernel SVM classifier optimized by grid search to achieve a fine division of the three stages of natural vibration, critical galloping, and strong galloping. This enhances the stage-specific nature and guiding value of the early warning. By performing cross-correlation analysis on the vibration displacement sequences of adjacent sampling points, the signal propagation delay is accurately estimated. Combined with spatial distance, the segment propagation velocity is calculated. Then, the overall propagation direction and average velocity are obtained through vector fitting. This achieves a leap from "point monitoring" to "regional propagation situation awareness," providing a dynamic basis for judging whether galloping has a risk of diffusion. An early warning index that integrates propagation dynamics and environmental wind field is constructed to generate a visualized "galloping propagation trend map," realizing a closed loop from state identification to risk diffusion prediction.

[0036] Furthermore, the implementation of tiered early warning systems based on conductor icing phenomena refers to the use of "galloping propagation trend charts" and icing phenomena to trigger tiered early warning systems. A primary warning is triggered when icing exists on the line and wind speeds are within ranges that easily induce low-frequency galloping. in, The line surface is covered with ice. It is the wind speed perpendicular to the direction of the line. It is a wind speed range, a typical wind speed range that is prone to inducing initial vibration, determined based on engineering experience; A medium-level warning is triggered when the dancing state reaches the critical state or when the propagation warning index shows a clear energy diffusion trend. The judgment conditions are as follows: Here, phase refers to a stage, and critical flutter refers to the critical flutter state. When the power line is in a state of strong galloping and the energy propagation intensity is high, a high-level warning is triggered, and the judgment conditions are as follows: Among them, strong flutter is strong dancing; After any level of warning is triggered, a standardized warning data packet is encapsulated using the JSON standard structure (including timestamp, monitoring point number and geographical coordinates, current warning level (primary / intermediate / advanced), and corresponding key indicators). The standardized warning data packet is then uploaded to the power grid dispatch center and operation and maintenance management platform in real time, triggering audible and visual alarms, generating a work order, or initiating the linkage control of the anti-galloping device.

[0037] By integrating icing conditions, wind speed conditions, and propagation dynamics, a three-tiered progressive early warning mechanism is constructed to achieve refined responses to the galloping evolution stage. The primary early warning provides forward-looking prevention, the intermediate early warning captures the transition state, and the advanced early warning identifies high-risk situations, improving the accuracy and guidance of the early warning. Through standardized JSON encapsulation and multi-channel linkage response, the data link between the monitoring system and the power grid control terminal is opened up to form a closed-loop handling process, significantly enhancing the proactive defense capability and intelligent operation and maintenance level of transmission line galloping disasters.

[0038] This embodiment also provides a power transmission line conductor galloping early warning system based on UAV edge recognition, including: The flight path planning and model deployment module is used to plan the flight path of the UAV and deploy edge computing devices and neural network model groups to support real-time monitoring of task execution. The displacement extraction module is used to generate a sequence of equally spaced sampling points based on the wire pixel mask and convert it into physical vibration displacement time series data; The feature fusion module is used to combine the results of short-time Fourier transform and optimal wavelet basis function analysis to construct a dual-channel normalized time-frequency feature map; The dance recognition module is used to input dual-channel feature maps to a lightweight network to output confidence scores, and combines SVM to achieve dance stage classification. The vibration early warning module is used to calculate the vibration propagation speed and direction, integrate wind direction consistency to generate a propagation early warning index, and trigger a graded response.

[0039] This embodiment also provides a computer device applicable to a method for early warning of conductor galloping on power transmission lines based on UAV edge recognition, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method for early warning of conductor galloping on power transmission lines based on UAV edge recognition as proposed in the above embodiment.

[0040] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0041] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements a method for early warning of conductor galloping on power transmission lines based on UAV edge recognition, as proposed in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

Claims

1. A power transmission line conductor galloping early warning method based on unmanned aerial vehicle edge recognition, characterized in that: comprising, planning a flight route for the unmanned aerial vehicle, deploying an edge computing device and a neural network model group, the unmanned aerial vehicle flying along the flight route to collect image data and environmental data, and extracting a pixel mask of the conductor and identifying icing phenomena of the conductor through the neural network model group; the neural network model group is a semantic segmentation model and a convolutional neural network; based on the pixel mask of the conductor, a pixel-level displacement time sequence is generated and converted into a vibration displacement sequence, and a two-dimensional double-channel vibration feature map is obtained by extracting multi-scale vibration features of the vibration displacement sequence using short-time Fourier transform and continuous wavelet transform; based on the two-dimensional double-channel vibration feature map, a dancing confidence score is output through a deep network, a dancing propagation early warning index is calculated to divide the dancing stage, and a hierarchical early warning is performed in combination with the icing phenomena of the conductor. 2.The power line conductor galloping early warning method based on UAV edge recognition of claim 1, wherein: The two-dimensional double-channel vibration feature map is obtained by using the short-time Fourier transform and the continuous wavelet transform to extract multi-scale vibration features from the vibration displacement sequence, which comprises the following steps: performing boundary extension processing and interpolation on the vibration displacement sequence to obtain a continuous time signal, performing short-time Fourier transform on the continuous time signal, performing modulus operation to obtain an amplitude spectrum, performing supplementary analysis by using continuous wavelet transform, and selecting an optimal bandwidth parameter according to a Shannon entropy criterion , reconstructing a complex Morlet wavelet basis function, performing continuous wavelet transform to obtain wavelet coefficients, cutting off the middle part of the wavelet coefficients corresponding to the original signal time period, discarding the coefficients corresponding to the two end extension regions, obtaining an effective wavelet coefficient segment with the same length as the original vibration signal, taking modulus square of the wavelet coefficients of the effective segment to obtain a wavelet energy spectrum, taking the amplitude spectrum as a first channel and the wavelet energy spectrum as a second channel, uniformly performing normalization processing and linear interpolation, and after normalization and alignment, splicing the data of the two channels along the channel dimension to form a two-dimensional double-channel vibration feature map. 3.The power line conductor galloping early warning method based on UAV edge recognition of claim 2, wherein: The two-dimensional double-channel vibration feature map is input into a lightweight CNN-BiLSTM-Attention network model to obtain a final dancing confidence score, and a confidence threshold is set When the dancing confidence score S is greater than or equal to the threshold , it is determined that the conductor has a significant dancing trend, otherwise, it is normal. When it is determined that the conductor has a significant dancing trend, the number of main peaks, energy concentration degree and frequency drift rate are extracted from the amplitude spectrum to represent the spectral feature, and are input into a support vector machine classifier for dancing stage recognition. The classifier outputs a current dancing state label, and the cross-correlation function of the vibration displacement sequences of any two adjacent sampling points is obtained through cross-correlation analysis. The relative time delay that makes the cross-correlation function reach a maximum value is found, and the actual propagation time delay is obtained in combination with the sampling interval. According to the spatial distance between the two sampling points of the transmission line and the corresponding actual propagation time delay, the propagation speed of the vibration signal in the section is calculated. Based on the vibration propagation speeds between all sampling points, the propagation speed of the entire line is vector fitted to obtain the overall propagation direction and the average propagation speed vector. The average propagation speed vector is corrected with the real-time wind direction to define a dancing propagation early warning index, set a propagation early warning index threshold , judge the dancing diffusion risk, and finally output the propagation speed, direction and propagation early warning index of each sampling point and section, automatically generate a dancing propagation trend chart, and mark the high-risk section. 4.The power line conductor galloping early warning method based on UAV edge recognition of claim 3, wherein: The hierarchical early warning in combination with the icing phenomena of the conductor is based on a dancing propagation trend map and icing phenomena to perform hierarchical early warning. After triggering any level of early warning, a standardized early warning data packet is packaged, the standardized early warning data packet is uploaded to the power grid dispatching center and operation and maintenance management platform in real time, and sound and light alarms are triggered, treatment work orders are generated, or anti-dancing device linkage control is started.

5. The power line conductor galloping early warning method based on UAV edge recognition of claim 4, wherein: The pixel mask of the conductor is obtained, a morphological closing operation method is used to further process the image, the skeleton line of the conductor is extracted, N equally spaced sampling points are selected on the skeleton line of the conductor, a space-time displacement matrix is formed based on the coordinates of all sampling points, the horizontal spatial resolution is calculated using monocular vision geometry by measuring the pixel width of the conductor in the image, the pixel offset is converted into vibration displacement based on the horizontal spatial resolution, and the vibration displacement sequence is constructed.

6. The power line conductor galloping early warning method based on UAV edge recognition of claim 5, wherein: After the unmanned aerial vehicle takes off, it starts cruising according to the planned route, takes pictures after entering the target interval, and simultaneously collects wind speed and direction through the external micro-meteorological sensor equipped on the unmanned aerial vehicle, records the GPS position and attitude angle during flight, and synchronously runs the neural network model group on the Jetson Xavier NX edge computing platform to extract the pixel mask of the conductor and the icing phenomena, and synchronously stores all collected image data and environmental data with the UTC time stamp during flight.

7. The power line conductor galloping early warning method based on UAV edge recognition of claim 6, wherein: The planning of the flight route for the unmanned aerial vehicle and the deployment of the edge computing device and the neural network model group are based on the geographical coordinates and historical dancing section data of the monitored power transmission line, using aerial survey software, a flight path covering the target area is automatically calculated according to the distribution of the power transmission line and the key areas, after the planning is completed, Jetson Xavier NX is selected as the edge computing platform, and the neural network model group is deployed on the Jetson Xavier NX.

8. A power transmission line conductor galloping early warning system based on UAV edge recognition, based on any one of claims 1-7, a power transmission line conductor galloping early warning method based on UAV edge recognition, characterized in that: comprising, a route planning and model deployment module for planning a flight path for the unmanned aerial vehicle and deploying an edge computing device and a neural network model group to support real-time monitoring task execution; A displacement extraction module is configured to generate an equidistant sampling point sequence based on the conductor pixel mask and convert the equidistant sampling point sequence into physical vibration displacement time series data. A feature fusion module is configured to combine the short-time Fourier transform and the optimal wavelet base function analysis result to construct a double-channel normalized time-frequency feature map. A galloping recognition module is configured to input the double-channel feature map into a lightweight network to output a confidence score and implement galloping stage classification in combination with an SVM. A galloping early warning module is configured to calculate a vibration propagation speed and direction, fuse wind direction consistency to generate a propagation early warning index, and trigger a hierarchical response. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to implement the steps of the power line conductor galloping early warning method based on the UAV edge recognition according to any one of claims 1-7.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to implement the steps of the power line conductor galloping early warning method based on the UAV edge recognition according to any one of claims 1-7.

Images