A method for monitoring power line galloping based on virtual point tracking

Abstract

This invention discloses a method for monitoring power line galloping based on virtual point tracking, comprising: acquiring a full-coverage monitoring video stream of the power transmission line; decomposing the video stream into an ordered frame sequence while retaining the original resolution, and generating a semantic mask through a pre-trained semantic segmentation network; uniformly deploying virtual points along the line; initializing the tracking feature template based on bilinear interpolation sampling, and constructing an input token by combining multi-scale feature matching and sine / cosine position encoding; introducing a Transformer architecture with a decomposed attention mechanism, alternately performing temporal and spatial attention operations, and achieving accurate virtual point tracking and visibility prediction through multiple rounds of position and feature updates; obtaining the galloping amplitude and dominant frequency through instantaneous displacement calculation and FFT spectrum analysis, and determining the galloping type and stability by combining trajectory morphology analysis and virtual point motion characteristics. This method achieves comprehensive, non-contact, and accurate monitoring of power line galloping characteristics without the need for physical targets.

Description

Technical Field

[0001] This invention relates to a method for monitoring power line galloping based on virtual point tracking, belonging to the field of image / video pattern recognition and processing technology. Background Technology

[0002] Since the phenomenon of power transmission line galloping due to icing was first discovered in the 1930s, scholars both domestically and internationally have conducted extensive research on the mechanism modeling, vibration control, and online monitoring of power transmission line galloping. Traditional monitoring methods are mostly contact-based, relying on the installation of physical sensors to directly collect vibration data from the transmission lines for parameter calculation. This method suffers from difficulties in installation and maintenance, high costs, and susceptibility to interference from high-voltage electromagnetic environments, making it difficult to meet the widespread need for real-time monitoring in complex terrains in China. With the rapid development of optoelectronic technology and computer vision, non-contact measurement technologies have gradually matured. Non-contact monitoring methods are represented by optical measurements, but traditional optical flow methods and feature point matching techniques are prone to trajectory tracking drift in complex background interference (such as vegetation obstruction and changes in lighting) and scenarios where local details of the transmission lines are blurred; they also lack the ability to systematically analyze the multi-dimensional characteristics (amplitude, frequency, type, and stability) of power transmission line galloping, making it difficult to meet the comprehensive data requirements for power transmission line safety assessment.

[0003] With the development of artificial intelligence and computer vision technologies, researchers have paid more attention to the measured response analysis and real-time monitoring technology during power transmission line galloping in recent years. Deep learning-based target tracking and feature extraction methods have provided a new technical path for power transmission line galloping monitoring: semantic segmentation networks can achieve accurate separation of the transmission line from the background environment, laying the foundation for subsequent feature extraction; the Transformer architecture, with its powerful global feature modeling capabilities, shows significant advantages in sequence data processing and long-distance dependency capture, making it possible to perform spatiotemporal correlation analysis of virtual points in multi-frame images. Summary of the Invention

[0004] This invention provides a method for monitoring power line galloping based on virtual point tracking, which solves the problems disclosed in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A method for monitoring power line galloping based on virtual point tracking:

[0007] Acquire monitoring video streams of power transmission lines;

[0008] Decompose the surveillance video stream into an ordered sequence of frames;

[0009] An ordered sequence of frames is input into a pre-trained semantic segmentation network to generate a semantic mask;

[0010] Based on semantic masking, the pixel coordinate range of the power transmission line in the image is extracted, the length of the power transmission line is determined, and virtual points are evenly distributed according to the length of the power transmission line.

[0011] Centered on each virtual point, a ROI image block of preset pixels is extracted;

[0012] The ROI image patch is input into a pre-built CNN model, and the output is a feature map;

[0013] Based on the initial position of the virtual point and the corresponding feature map, bilinear interpolation sampling is performed to generate a feature template for tracking the virtual point;

[0014] Centered on the current virtual point prediction position, extract local search region features with a fixed radius on the corresponding feature map, and construct a correlation embedding vector based on the similarity between the feature template and the local search region features.

[0015] A token sequence is constructed using the virtual point's predicted location, the relevance embedding vector, and the visibility used to characterize whether the virtual point is occluded.

[0016] The token sequence is input into a pre-built Transformer architecture, which alternately performs temporal attention operations and spatial attention operations to perform multiple rounds of position and feature updates to achieve virtual point tracking and visibility prediction;

[0017] The instantaneous displacement during virtual point tracking is calculated to obtain the transmission line galloping amplitude. The FFT spectrum of the transmission line galloping is analyzed to obtain the dominant frequency of the transmission line. Based on the transmission line galloping amplitude and dominant frequency, it is determined whether the transmission line needs maintenance.

[0018] The type of power line galloping can be determined based on the characteristics of the power line's movement.

[0019] Furthermore, the method of extracting local search region features of a fixed radius on the corresponding feature map, centered on the current virtual point prediction position, and constructing a relevance embedding vector based on the similarity between the feature template and the local search region features, includes:

[0020] At time t, for the i-th virtual point, its current predicted position is... Using the center coordinates on the feature map, a square search region is defined with a preset fixed radius. All feature data within the square search region is extracted as the local search region features. ;

[0021] Construct a spatial weight matrix that follows a two-dimensional Gaussian distribution. :

[0022] ;

[0023] in, These are the relative spatial coordinates within the search area. The displacement component in the horizontal direction of the virtual point. This represents the displacement component in the vertical direction of the virtual point. To control the distribution variance of the weight decay rate;

[0024] To track feature templates As a convolution kernel, it is used to search for features in the local search region. Perform a deep local cross-correlation convolution operation to obtain the initial correlation response value, and then combine the initial correlation response value with the spatial weight matrix. Pixel-wise weighted multiplication is performed to complete weighted modulation and generate an optimized local correlation response map. :

[0025] ;

[0026] in, The cross-correlation convolution operation is repeated at multiple feature scales z to obtain a multi-scale correlation volume. This volume is then flattened and concatenated before being input into a multilayer perceptron for dimensionality reduction and feature fusion. The final output is a correlation embedding vector with a fixed dimension. :

[0027] .

[0028] Furthermore, the token sequence representation ;in, Visibility is used to characterize whether a virtual point is occluded.

[0029] Furthermore, the formula for calculating temporal attention is: ;

[0030] in, For time query vectors, For key vectors, For time-based attention weights.

[0031] Furthermore, the formula for calculating spatial attention is: ;

[0032] in, For spatial query vectors, For spatial key vectors, For spatial attention weights.

[0033] Furthermore, the Transformer architecture employs multiple rounds of position and feature updates to achieve virtual point tracking and visibility prediction, including:

[0034] The Transformer architecture outputs the position correction of the virtual point. and feature correction amount ;

[0035] Perform addition to update position and features:

[0036] ;

[0037] ;

[0038] Where m is any positive integer, representing the m-th update;

[0039] After repeating the above addition update M times, the predicted position is obtained. and characteristics ;

[0040] Visibility of virtual points ;

[0041] Where W is the learnable weight matrix, For bias terms, This is the Sigmoid activation function.

[0042] Furthermore, methods for calculating the instantaneous displacement of virtual points to obtain the amplitude of transmission line galloping include:

[0043] Record the vertical coordinates of the trajectories of N virtual points in a static state, which will serve as a static balance reference for the transmission line. ;

[0044] Using a sequence of N virtual point coordinates corrected for spatial attention as input, the least squares method is employed to fit the curve equation of the transmission line in real time. :

[0045] ;

[0046] Where (x, y) are the coordinates in the image coordinate system or physical coordinate system, and a, b, c are the fitting coefficients to be determined;

[0047] Using the coordinates of N virtual trajectory points ( , Construct the objective function and the sum of squared residuals S:

[0048] ;

[0049] To minimize S, we take the partial derivatives with respect to a, b, and c and set them to 0 to obtain a system of linear equations. We then convert these equations into matrix form and solve for the fitting coefficients a, b, and c using matrix inversion or Gaussian elimination to obtain the real-time fitting curve of the transmission line.

[0050] Compare the real-time fitting curve of the transmission line with the static equilibrium reference standard. ;

[0051] Constructing the relative displacement deviation function ;

[0052] Within the gear range Find the x-coordinate that maximizes the absolute value of the relative displacement deviation function. ,in Let the length of a single span of the transmission line be given; by differentiation... Locate the candidate positions of the antinodes and determine the global maximum deviation position by combining the boundary values;

[0053] locking Then, let the coordinates ( , () represents the instantaneous antinode;

[0054] in, The relative x-coordinate of the instantaneous antinode is The corresponding relative ordinate;

[0055] Calculate the instantaneous peak displacement h(t) of the instantaneous antinode relative to the baseline at time t:

[0056] ;

[0057] The instantaneous peak displacement of the instantaneous antinode relative to the baseline at the corresponding moment is calculated from T consecutive frames of images, resulting in a displacement time series reflecting the location of the most severe vibration of the transmission line: ;

[0058] The amplitude of the transmission line galloping is obtained by calculating the difference between the maximum and minimum values ​​in the displacement time series H. .

[0059] Furthermore, methods for obtaining the dominant frequency of transmission lines by analyzing the FFT spectrum of transmission line galloping include:

[0060] The frequency domain signal is obtained by processing the displacement time series H using FFT:

[0061] ;

[0062] Where X(k) is the frequency domain signal, h(l) is the l-th sample of the displacement time series, e is the natural constant, L is the signal length, k is the frequency index, and j is the imaginary unit;

[0063] ;

[0064] The modulus |X(k)| of the FFT result is calculated.

[0065] Calculate the power spectral density characterizing the energy distribution. ;

[0066] Map the discrete index K to the actual frequency , Indicates the sampling frequency;

[0067] Within the frequency range of dancing The frequency point corresponding to the maximum value of amplitude |X(k)| or power spectral density P(k) within the specified interval is used as the dominant frequency of transmission line galloping. .

[0068] Furthermore, methods for determining whether a transmission line needs maintenance based on its galloping amplitude and dominant frequency include: if the dominant frequency of the transmission line... Not in If so, the power transmission line needs repair;

[0069] If the amplitude of the power transmission line galloping If the safety threshold is exceeded, the power transmission line needs to be repaired.

[0070] Furthermore, methods for determining the type of power transmission line galloping based on the motion characteristics of the transmission line include:

[0071] Select the virtual point at the midpoint of the gear distance And the virtual point at 1 / 4 They are respectively denoted as and ;

[0072] Dance type parameters ;

[0073] Where r is the displacement response vector;

[0074] like A value approximately 1 indicates that the two points move in the same direction, and the transmission line bounces up and down as a whole; if The value is approximately -1, indicating that the two points move in opposite directions and the transmission line is undergoing torsional vibration.

[0075] The beneficial effects achieved by this invention are as follows:

[0076] 1. This invention proposes a targetless transmission line galloping monitoring scheme, which eliminates the need to deploy additional physical targets on the transmission line and directly monitors the galloping status through the analysis of the monitoring video stream. This avoids the installation and maintenance costs of targets and the potential impact on line operation, and is suitable for the actual application needs of complex field transmission environments.

[0077] 2. This invention integrates optical flow method and virtual point tracking technology. By uniformly distributing virtual points on the transmission line and capturing the motion trajectory of virtual points between adjacent frames using optical flow method, it achieves non-contact dynamic tracking of transmission line galloping, effectively solving the limitation of traditional contact monitoring methods that cannot cover the entire line.

[0078] 3. This invention constructs a dance trajectory optimization model based on the Transformer architecture, which organizes the position, correlation features and visibility information of virtual points into input tokens. By adding time and position embedding and sequence format reshaping, it provides an adaptive architecture support for the efficient processing of multi-dimensional dance information and improves the intelligence level of trajectory analysis.

[0079] 4. This invention innovatively introduces a decomposed attention mechanism, which alternately executes temporal attention and spatial attention operations within the Transformer architecture: temporal attention is based on the virtual point's full-time frame state to constrain the trajectory's temporal continuity, while spatial attention corrects local detection deviations through the interaction of multiple virtual points. The dual attention work together to improve the accuracy and stability of trajectory monitoring.

[0080] 5. This invention employs a multi-scale feature matching strategy, which samples features around the current predicted location and matches them with the tracking template. By repeatedly performing this operation on the multi-scale feature map, a correlation embedding vector is generated, which enhances the matching accuracy between "trajectory features" and "image features" and effectively addresses the recognition challenges caused by local detail changes in power transmission lines and environmental interference.

[0081] 6. This invention designs a visibility prediction mechanism based on Transformer output features and a learnable weight matrix. It calculates virtual point visibility parameters using the Sigmoid activation function, automatically identifies invalid data points obscured by towers, and removes them in post-processing. This avoids erroneous data interfering with the calculation of the galloping amplitude, ensuring the reliability of monitoring results.

[0082] 7. This invention constructs a virtual point tracking process, which initializes the tracking feature template with bilinear interpolation sampling, and combines the calculation of relative position offset between adjacent frames with sine and cosine position encoding to achieve accurate representation of virtual point motion information. With multiple rounds of position and feature update iterations, the robustness of virtual point tracking in complex dancing scenes is improved.

[0083] 8. This invention constructs a full-dimensional galloping feature monitoring system. By calculating the static equilibrium point, instantaneous displacement, apparent amplitude and dominant frequency, combined with trajectory morphology analysis and galloping type discrimination (based on the motion correlation of virtual points at different spans), it achieves comprehensive monitoring of the amplitude, frequency, type and stability of transmission line galloping, providing comprehensive data support for line safety assessment. Attached Figure Description

[0084] Figure 1This is a schematic diagram of the process of the present invention;

[0085] Figure 2 This is a schematic diagram of trajectory monitoring in the virtual point selection and tracking experiment of this invention;

[0086] Figure 3 This is a schematic diagram of the correlation calculation process in this invention;

[0087] Figure 4 This is a schematic diagram of the spatial and temporal attention mechanism in this invention;

[0088] Figure 5 This is a schematic diagram of the technical process for achieving stable tracking of virtual points in this invention. Detailed Implementation

[0089] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0090] like Figure 1 and Figure 5 As shown, this invention discloses a method for monitoring power line galloping based on virtual point tracking. It eliminates the need for physical targets and achieves comprehensive and accurate monitoring of power line galloping characteristics through non-contact technology. The method specifically includes the following steps:

[0091] Deploy image acquisition equipment and adjust it according to the span of the transmission line, terrain, and surrounding environment characteristics. Determine the placement points of industrial cameras or high-definition surveillance cameras to ensure that the imaging range of the equipment completely covers all sections of the monitored line without any blind spots. Select high-definition imaging equipment with a resolution of 384x512 and a frame rate of ≥25fps, and adjust the equipment's exposure time, white balance, focal length, and other parameters. Continuously acquire the transmission line monitoring video stream and store it in MP4 or AVI format.

[0092] The video framing algorithm in OpenCV is called, and the video path to be framing and the file path to save the images are passed in as two parameters. Based on the time order, the monitoring video stream is decomposed into an ordered sequence of frames. The resolution of each frame image is kept at the original resolution without scaling or compression, ensuring that key details such as the texture and edges of the power transmission lines are not lost.

[0093] A pre-trained semantic segmentation network (such as U-Net or Mask R-CNN) is used to generate a semantic mask, with the frame image at time t as input. Generate semantic masks The transmission line area is marked as the foreground and other areas as the background to separate the transmission line from the interference area. Simultaneously, the mask undergoes post-processing to remove isolated noise points and fill in minor gaps in the transmission line area, ensuring accurate mask positioning of the transmission line region.

[0094] ;

[0095] in, For a moment Semantic masking of moments For a moment The image, It is a semantic segmentation network.

[0096] like Figure 2 As shown, based on semantic masking Extract the pixel coordinate range of the transmission line in the image to determine the transmission line length (pixel level). Set the virtual point density (no less than 3 points per meter of transmission line), calculate the number of virtual points N based on the transmission line length, plan a uniform distribution scheme, avoid tower obstruction areas, and ensure that virtual points cover the entire effective area of ​​the transmission line. Following the principle of uniform distribution, deploy virtual points within the pixel range of the transmission line, ensuring consistent spacing between adjacent virtual points. Record the initial coordinates (x, y, y) of each virtual point. i , y i (i=1,2,…,N), establish a virtual point coordinate ledger to provide an initial position reference for subsequent tracking. For virtual point selection and tracking results, please refer to... Figure 2 "A schematic diagram of trajectory monitoring in the virtual point selection and tracking experiment."

[0097] With each virtual point P i Centered on a single point, a 64×64 pixel Region of Interest (ROI) image patch is extracted to generate an independent high-resolution image patch. The ROI image patch is then subjected to grayscale conversion and normalization to reduce the impact of illumination variations on feature extraction and improve its stability.

[0098] A CNN model (containing an input layer, convolutional layers, pooling layers, activation layers, and fully connected layers) is constructed to extract high-dimensional texture features. The convolutional layers use 3×3 convolutions to capture texture details and edge features within the Region of Interest (ROI). If higher resolution is required for feature extraction, bilinear interpolation upsampling is performed on the ROI image patches within the model. The preprocessed ROI image patches are input into the CNN model, and through forward propagation, a high-dimensional texture kernel corresponding to each ROI is output, using feature maps. .

[0099] ;

[0100] in, For the first Texture features of virtual points This is an image of the local region of interest around a virtual point.

[0101] Select the initial position of the virtual point at time t=0 (when the transmission line is stationary). Centered on this location, in the feature map The sampling range is determined in the middle.

[0102] With the initial position of the tracking point With feature map Based on this, bilinear interpolation sampling is performed. First, feature extraction is completed at the initial location, and then a local support region is constructed around it. This initializes the tracking feature Q, generating the feature template required for subsequent matching and providing adaptation conditions for the correlation calculation between image features and tracking features. The formula is as follows:

[0103] ;

[0104] in, For the first Initialization feature template for virtual points This is a bilinear interpolation operation. It is a texture feature.

[0105] like Figure 3 As shown, spatially distributed weight coefficients are introduced to optimize correlation calculation, construct dynamic correlation features, and generate input tokens for Transformer-based spatiotemporal joint tracking and iterative convergence updates. At time t, for the i-th virtual point, its current predicted position is used... Using the center coordinates on the feature map, a square search region is defined with a preset fixed radius. All feature data within the square search region is extracted as the local search region features. ;

[0106] To reduce the interference of complex backgrounds (such as foliage obscuring the image or similar tower structures) on the matching results, this invention introduces a spatial prior weighting mechanism in the correlation calculation. Considering the continuity of the physical displacement of transmission lines between adjacent frames, and that virtual points are likely to appear in the center of the search region, a spatial weight matrix following a two-dimensional Gaussian distribution is constructed. :

[0107] ;

[0108] in, These are the relative spatial coordinates within the search area. The displacement component in the horizontal direction of the virtual point. This represents the displacement component in the vertical direction of the virtual point. To control the distribution variance of the weight decay rate;

[0109] To track feature templates The kernel is used to search for features in the local search region. Perform a deep local cross-correlation convolution operation to obtain the initial correlation response value, and then combine the initial correlation response value with the spatial weight matrix. Pixel-by-pixel multiplication is performed to complete weighted modulation and generate an optimized local correlation response map. :

[0110] ;

[0111] in, This represents the cross-correlation convolution operation. By assigning high weights to the center of the search region and low weights to the edges, the false matching peaks caused by the surrounding background are significantly suppressed. The above weighted cross-correlation operation is repeated at multiple feature scales z to obtain the multi-scale correlation volume. After flattening and concatenating these volumes, they are input into a multilayer perceptron (MLP) for dimensionality reduction and feature fusion, ultimately outputting a correlation embedding vector with fixed dimensions. :

[0112] ;

[0113] The vector It represents the high reliability of visual matching confidence after spatial weighting constraints.

[0114] Construct an input grid and organize all tracking points into tokens: The input contains the following information:

[0115] Virtual point prediction location It contains the initial position and represents the relative position offset between adjacent frames through sine and cosine position coding.

[0116] Correlation characteristics That is, the one-dimensional matching confidence vector obtained after introducing spatial weight modulation is used to calculate the current transmission line point features and surrounding image features.

[0117] Visibility Initialized to 1, indicating whether it is occluded.

[0118] The above features are concatenated along the channel dimension and reshaped into a Transformer-compatible sequence format. At the same time, time position embedding is added to complete the preprocessing of the input data.

[0119] like Figure 4 As shown, a decomposed attention mechanism is introduced, causing the Transformer module to alternately execute two differentiated attention operations, specifically processing temporal and spatial information respectively, to achieve precise optimization of the transmission line galloping trajectory. Temporal attention targets the same virtual trajectory point on the same transmission line, and by calling this virtual trajectory point across all time frames... to The state information within (T being the total number of time frames) enables time-domain smoothing of the transmission line galloping trajectory. This time attention mechanism can constrain the trajectory change trend based on the temporal correlation of the trajectory. For example, if the virtual trajectory point corresponding to the transmission line is in an upward motion state in frames 10 and 12, the motion state of the virtual trajectory point in frame 11 can be constrained to prevent a sudden drop. The formula is as follows:

[0120] ;

[0121] in, For time query vectors, For key vectors, For time-based attention weights.

[0122] Establishing a spatial attention mechanism:

[0123] At the same time t, the state information of all virtual trajectory points at that time is extracted synchronously, and the interaction between virtual trajectory points is used to enhance the inter-point collaboration and correct local detection deviations.

[0124] In specific application scenarios, when external environmental interference (such as strong light causing reflection) affects the weighted correlation embedding vector of a certain section of the transmission line, If low confidence is still observed, it indicates that the visual features of that point are no longer reliable. At this point, the spatial attention mechanism automatically reduces its reliance on the features of the blurred point itself, and instead assigns high confidence to adjacent points based on the consistency of group movement (i.e.,...). Virtual points with strong responses receive higher attention weights, thus playing a role in group collaborative correction.

[0125] For example, if a virtual trajectory point with high confidence on the left side of a transmission line is moving upwards, the information transmission and interaction through this spatial attention mechanism can drive the adjacent virtual trajectory point on the right side, which is in a low-confidence state, to synchronously generate an upward movement trend. The formula is as follows:

[0126] ;

[0127] in, For spatial query vectors, For spatial key vectors, For spatial attention weights.

[0128] The dual safeguards of the aforementioned "Gaussian weighted cross-correlation + spatial attention mechanism" can effectively compensate for the detection error caused by external interference of a single virtual trajectory point, ensuring extremely high reliability of trajectory detection results.

[0129] The collective synergy effect formed by the aforementioned spatial attention mechanism can effectively compensate for the detection error of a single virtual trajectory point caused by external interference, significantly improving the detection accuracy and stability of the overall power line galloping trajectory and ensuring the reliability of the trajectory detection results. After spatiotemporal attention operations and feature fusion processing, the Transformer architecture outputs the position correction of the virtual point. and feature correction amount ; Perform addition to update position and features::

[0130] ;

[0131] ;

[0132] Where m is any positive integer, representing the m-th update; by repeating the above additive update M times, the predicted position can be obtained. and characteristics Additionally, this program supports visibility prediction, calculating the result after M repeated operations. The value:

[0133] ;

[0134] Where W is the learnable weight matrix, For bias terms, Sigmoid activation function

[0135] If a section of the transmission line is completely blocked by a tower, its corresponding value will be close to 0. Such data points can be directly removed in the post-processing stage to avoid erroneous data interfering with the calculation results of the galloping amplitude.

[0136] Through the above operations, this invention can achieve stable tracking of virtual points. The tracking process is as follows: Figure 5 As shown.

[0137] A static equilibrium benchmark was established, and the spatial state of the transmission line was fitted using a multi-point coordinate flow, with the maximum amplitude-frequency characteristic at a single point extracted. In the initial stage of the analysis, the vertical coordinates of N virtual trajectory points in a static state were recorded as the static equilibrium reference benchmark for the transmission line. .

[0138] Using a sequence of N point coordinates corrected for spatial attention as input, the least squares method is employed to fit the curve equation of the transmission line in real time. Set the equation for the fitted curve Construct the objective function of residual sum of squares:

[0139] ;

[0140] Taking the partial derivatives of a, b, and c and setting them to zero yields a system of linear equations. The coefficients are then solved using matrix inversion or Gaussian elimination to obtain the real-time fitting curve f. t (x) accurately represents the real-time spatial position of the transmission line.

[0141] Real-time comparison of fitted curves With static equilibrium reference benchmark The system dynamically searches for the geometric location with the largest deviation from the baseline across the entire span of the curve, identifying it as the instantaneous antinode. The vertical displacement time series of this dynamic antinode is then extracted. And its peak-to-peak value is calculated as the apparent gobling amplitude. The calculation method is as follows:

[0142] First, the real-time curve equation of the transmission line obtained by fitting is used. With static equilibrium reference benchmark Constructing the relative displacement deviation function In the gear range Find the x-coordinate that maximizes the absolute value of the deviation function. By differentiation Quickly locate candidate antinode positions and combine them with boundary values ​​to determine the global maximum deviation position. Lock the x-axis. Then, calculate the instantaneous peak displacement h(t) of the point relative to the baseline at that moment.

[0143] ;

[0144] ;

[0145] The amplitude of the galloping is calculated by performing the above operation on a continuous T-frame image, resulting in a displacement time series reflecting the location of the most severe vibration of the transmission line. Subsequently, the difference between the maximum and minimum values ​​in the displacement sequence H is calculated to obtain the maximum swing range, i.e., the galloping amplitude, within this analysis period. ;

[0146] ;

[0147] Single-sided peak ;

[0148] The dominant frequency is calculated, and then a Fast Fourier Transform (FFT) is performed on the single-point displacement sequence h(t). The time-domain signal H is converted into a frequency-domain signal X(k), and the magnitude of the FFT result is calculated to obtain the amplitude distribution of each frequency component.

[0149] The frequency domain signal is obtained by processing the displacement time series H using FFT:

[0150] ;

[0151] Where X(k) is the frequency domain signal, h(l) is the l-th sample of the displacement time series, e is the natural constant, L is the signal length, k is the frequency index, and j is the imaginary unit;

[0152] ;

[0153] Calculate the power spectral density (PSD) to characterize the energy distribution:

[0154] ;

[0155] ;

[0156] The discrete index K is mapped to the actual frequency. , Indicates the sampling frequency;

[0157] Then in the typical dancing frequency range (usually The search is performed within this interval to find the frequency point corresponding to the maximum value of the amplitude |X(k)| or power spectrum P(k). :

[0158] ;

[0159] Should This is the dominant frequency as defined in this invention; determining the dominant frequency. Is it in Range; amplitude of dancing If the safety threshold is exceeded, appropriate measures must be taken immediately.

[0160] To determine the type of galloping, whether it is overall vertical jumping or torsional vibration, a virtual point at the midpoint of the span is selected. And the virtual point at 1 / 4 They are respectively denoted as and ; Calculate relevant parameters ,like A value approximately 1 indicates that the two points move in the same direction, and the transmission line bounces up and down as a whole. The value is approximately -1, indicating that the two points move in opposite directions and the transmission line is undergoing torsional vibration.

[0161] ;

[0162] Where r is the displacement response vector;

[0163] To determine stability, then, using the horizontal coordinate... The X-axis represents the dominant frequency of the signal, and the vertical coordinate is [missing information]. To plot the Y-axis, use the original... Plot the trajectory using coordinates. If the trajectory shows a clear circular or elliptical shape, it indicates that the power line is galloping smoothly and there is no safety risk; if the trajectory is chaotic, it indicates that the power line is experiencing complex multi-frequency vibrations, and timely measures should be taken.

[0164] A safety threshold is set, and the galloping amplitude threshold is determined based on the line voltage level and transmission line type. If the amplitude exceeds the threshold, the frequency exceeds the range, or the stability is poor, the system automatically triggers an early warning signal, records the warning time, characteristic parameters, and other information, and prompts maintenance personnel to take timely measures such as inspection and reinforcement. For line galloping parameters: Galloping amplitude ( ), dominant frequency ( Dance type () The weights for trajectory stability and data validity are set as W1, W2, W3, W4, and W5, respectively. The total weight is 1.0, emphasizing the impact of galloping amplitude and dominant frequency on line safety.

[0165] See Table 1 for specific weight settings:

[0166]

[0167] Table 1

[0168] Determine whether the measured values ​​of each parameter are within the safe range, and set the judgment value as B. i =1; when out of range, B i =0. In the data validity parameters, B is defined as follows: valid data percentage ≥ 90%. i =1, otherwise B i =0. The structural scoring formula is as follows:

[0169] ;

[0170] Among them, G t The structure is rated (0-100 points), with higher scores indicating safer lines; W i For the weights of each parameter; A i The measured value of parameter A; i,0.9 It is 90% of the parameter safety threshold (e.g., amplitude A of a 220kV line). i,0.9 =1.8m); A i,max For parameter limit values ​​(such as amplitude A) i,max =3.0m, frequency A i,max =5.0Hz)

[0171] This monitoring method uses a seven-day monitoring cycle, measuring, processing, and statistically analyzing the data within the cycle. Three predictions are made on day i, and structural scores are calculated for each of the three measurements, resulting in a vector on day i containing the structural scores from the three measurements. , , ], assuming Let i = 1, 2, 3…7. We select the representative structural score value for day i. As the true score value of the structure on day i, the representative score value of the structure on day i. The calculation process is as follows:

[0172] 1. and respectively with Perform the difference. ,when When the condition is constant, select the cubic structural score. , , The arithmetic mean is used as the representative value of the structural score on day i. .

[0173] 2. When there exists a set j such that The other set of j satisfies When, take the median value. As the representative value of the structural score on day i .

[0174] 3. When If the condition is not met, the vector of the measurement structure score should not be used as the basis for evaluation, and three more measurements need to be performed.

[0175] The above method can be used to obtain seven representative structural scores for a monitoring period. ], for the 7-day representative value [G t1 G t2 , …, G t7 After performing a stationarity test (Spearman correlation coefficient method) and first-order differencing, a regression moving average algorithm was used to predict the score G on day 8. t8 .

[0176] Eight sets of scoring data (7 days of actual measurement + 1 day of prediction) were obtained using the above method. From these eight sets of data, monitoring indicators A and B can be derived. The calculation methods for monitoring indicators A and B are as follows:

[0177] Eight sets of data were filtered, and representative values ​​of the structured score were calculated. The percentage of scores greater than 90 is recorded as monitoring indicator A; the representative value of the structural score is calculated. The percentage of scores below 80 is designated as monitoring indicator B. The higher the hazard level, the more dangerous the structure. Maintenance and line control plans are implemented using a tiered control approach, selecting the appropriate plan based on the structure's hazard level.

[0178] The relationship between structural hazard level, judgment criteria and response plan is shown in Table 2.

[0179]

[0180] Table 2

[0181] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0182] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0183] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0184] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0185] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A method for monitoring power line galloping based on virtual point tracking, characterized in that: Acquire monitoring video streams of power transmission lines; Decompose the surveillance video stream into an ordered sequence of frames; An ordered sequence of frames is input into a pre-trained semantic segmentation network to generate a semantic mask; Based on semantic masking, the pixel coordinate range of the power transmission line in the image is extracted, the length of the power transmission line is determined, and virtual points are evenly distributed according to the length of the power transmission line. Centered on each virtual point, a ROI image block of preset pixels is extracted; The ROI image patch is input into a pre-built CNN model, and the output is a feature map; Based on the initial position of the virtual point and the corresponding feature map, bilinear interpolation sampling is performed to generate a feature template for tracking the virtual point; Centered on the current virtual point prediction position, extract local search region features with a fixed radius on the corresponding feature map, and construct a correlation embedding vector based on the similarity between the feature template and the local search region features. A token sequence is constructed using the virtual point's predicted location, the relevance embedding vector, and the visibility used to characterize whether the virtual point is occluded. The token sequence is input into a pre-built Transformer architecture, which alternately performs temporal attention operations and spatial attention operations to perform multiple rounds of position and feature updates to achieve virtual point tracking and visibility prediction; The instantaneous displacement during virtual point tracking is calculated to obtain the amplitude of the transmission line galloping. The FFT spectrum of the transmission line galloping is analyzed to obtain the dominant frequency of the transmission line. Based on the amplitude and dominant frequency of the transmission line galloping, it is determined whether the transmission line needs maintenance. Determine the type of power transmission line galloping based on the characteristics of power transmission line movement; Methods for calculating the amplitude of transmission line galloping during virtual point tracking include: record the vertical coordinates of the N virtual points in the stationary state as the static equilibrium reference datum of the power transmission line ; The N virtual point coordinate sequences corrected by the spatial attention are used as input quantities, and a least square method is used to fit the curve equation of the power transmission line in real time : ； Where (x, y) are the coordinates in the image coordinate system or physical coordinate system, and a, b, c are the fitting coefficients to be determined; Using the coordinates of N virtual trajectory points ( , Construct the objective function and the sum of squared residuals S: ； To minimize S, we take the partial derivatives with respect to a, b, and c and set them to 0 to obtain a system of linear equations. We then convert these equations into matrix form and solve for the fitting coefficients a, b, and c using matrix inversion or Gaussian elimination to obtain the real-time fitting curve of the transmission line. Comparing a real-time fitted curve of a power transmission line to a static equilibrium reference ; Constructing relative displacement deviation function ; In the span interval find the abscissa that makes the absolute value of the relative displacement deviation function maximum where is the length of a single span of the transmission line; by taking the derivative lock the antinode candidate position and determine the global maximum deviation position in combination with the boundary value locking Then, let the coordinates ( , () represents the instantaneous antinode; in, The relative x-coordinate of the instantaneous antinode is The corresponding relative ordinate; Calculate the instantaneous peak displacement h(t) of the instantaneous antinode relative to the baseline at time t: ； The instantaneous peak displacement of the instantaneous antinode relative to the baseline at the corresponding moment is calculated from T consecutive frames of images, resulting in a displacement time series reflecting the location of the most severe vibration of the transmission line: ; The amplitude of the transmission line galloping is obtained by calculating the difference between the maximum and minimum values ​​in the displacement time series H. ; Methods for determining the type of power transmission line galloping based on the motion characteristics of power transmission lines include: Select the virtual point at the midpoint of the gear distance And the virtual point at 1 / 4 They are respectively denoted as and ; Dance type parameters ; Where r is the displacement response vector; like A value approximately 1 indicates that the two points move in the same direction, and the transmission line bounces up and down as a whole; if The value is approximately -1, indicating that the two points move in opposite directions and the transmission line is undergoing torsional vibration.

2. The method for monitoring power line galloping based on virtual point tracking according to claim 1, characterized in that: The method of extracting local search region features with a fixed radius from the corresponding feature map, centered on the current virtual point prediction location, and constructing a relevance embedding vector based on the similarity between the feature template and the local search region features includes: At time t, for the i-th virtual point, its current predicted position is... Using the center coordinates on the feature map, a square search region is defined with a preset fixed radius. All feature data within the square search region is extracted as the local search region features. ; Construct a spatial weight matrix that follows a two-dimensional Gaussian distribution. : ； in, These are the relative spatial coordinates within the search area. This represents the horizontal displacement component of the virtual point. This represents the displacement component in the vertical direction of the virtual point. To control the distribution variance of the weight decay rate; To track feature templates The kernel is used to search for features in the local search region. Perform a deep local cross-correlation convolution operation to obtain the initial correlation response value, and then combine the initial correlation response value with the spatial weight matrix. Pixel-by-pixel multiplication is performed to complete weighted modulation and generate an optimized local correlation response map. : ； in, The cross-correlation convolution operation is repeated at multiple feature scales z to obtain a multi-scale correlation volume. This volume is then flattened and concatenated before being input into a multilayer perceptron for dimensionality reduction and feature fusion. The final output is a correlation embedding vector with a fixed dimension. : 。 3. The method for monitoring power line galloping based on virtual point tracking according to claim 2, characterized in that, Token sequence representation ;in, Visibility is used to characterize whether a virtual point is occluded.

4. The method for monitoring power line galloping based on virtual point tracking according to claim 1, characterized in that, The formula for calculating time attention is: ; in, For time query vectors, For key vectors, For time-based attention weights.

5. The method for monitoring power line galloping based on virtual point tracking according to claim 1, characterized in that, The formula for calculating spatial attention is: ; in, For spatial query vectors, For spatial key vectors, For spatial attention weights.

6. The method for monitoring power line galloping based on virtual point tracking according to claim 1, characterized in that, Methods for implementing virtual point tracking and visibility prediction using the Transformer architecture through multiple rounds of position and feature updates include: The Transformer architecture outputs the position correction of the virtual point. and feature correction amount ; Perform addition to update position and features: ； ； Where m is any positive integer, representing the m-th update; After repeating the above addition update M times, the predicted position is obtained. and characteristics ; Visibility of virtual points ; Where W is the learnable weight matrix, For bias terms, This is the Sigmoid activation function.

7. The method for monitoring power line galloping based on virtual point tracking according to claim 1, characterized in that, Methods for analyzing the FFT spectrum of power transmission line galloping to obtain the dominant frequency of the power transmission line include: The frequency domain signal is obtained by processing the displacement time series H using FFT: ； Where X(k) is the frequency domain signal, h(l) is the l-th sample of the displacement time series, e is the natural constant, L is the signal length, k is the frequency index, and j is the imaginary unit; ； The modulus |X(k)| of the FFT result is calculated. Calculate the power spectral density characterizing the energy distribution. ; The discrete index K is mapped to the actual frequency. , Indicates the sampling frequency; Within the frequency range of dancing The frequency point corresponding to the maximum value of amplitude |X(k)| or power spectral density P(k) within the specified interval is used as the dominant frequency of transmission line galloping. .

8. The method for monitoring power line galloping based on virtual point tracking according to claim 7, characterized in that, Methods for determining whether a transmission line needs maintenance based on its galloping amplitude and dominant frequency include: if the dominant frequency of the transmission line... Not in If so, the power transmission line needs repair; If the amplitude of the power transmission line galloping If the safety threshold is exceeded, the power transmission line needs to be repaired.

Images