A method for identifying galloping of an edge-side power transmission line
By constructing a TLNet model and making lightweight improvements, and combining knowledge distillation and optical flow analysis, the real-time performance and accuracy issues of deep learning models on edge devices were solved, enabling efficient identification and real-time monitoring of power line galloping.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUHAN UNIV OF TECH
- Filing Date
- 2026-03-10
- Publication Date
- 2026-06-05
AI Technical Summary
Existing deep learning models are difficult to run in real time on edge devices, resulting in a large delay in the detection of power line galloping. Furthermore, the accuracy decreases after lightweighting, which cannot meet the needs of real-time monitoring.
A high-precision semantic segmentation model TLNet is constructed, which is combined with the line feature interaction parallel network structure LITB, detail recovery decoder DRD and line boundary perception auxiliary training head EPM. The model is lightweighted by neural architecture search NAS and general inverted bottleneck module UIB. Combined with knowledge distillation and dense optical flow analysis, the efficient identification of transmission line galloping is achieved.
High-precision, low-latency transmission line galloping detection has been achieved on edge devices, enhancing the intelligence and emergency response capabilities of the power grid.
Smart Images
Figure CN122156764A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision analysis technology, specifically relating to a method for determining the galloping of transmission lines on the edge side. Background Technology
[0002] With the increasing demand for electricity and the continuous expansion of the power grid, transmission line galloping has always been a significant factor affecting the safe and stable operation of the power grid. Transmission lines are prone to periodic vibrations under wind conditions. When wind speeds reach a certain range, conductors will exhibit low-frequency, large-amplitude periodic oscillations, the so-called "galloping" phenomenon. Galloping not only causes collisions, fatigue damage, and strand breakage between conductors, but can also trigger serious consequences such as short-circuit tripping, threatening the safe and stable operation of the power grid. To prevent galloping disasters from causing significant losses to social power safety and economic development, there is an urgent need for a transmission line galloping detection method that can operate efficiently on edge-side equipment.
[0003] Currently, the detection of power line galloping is typically performed at edge nodes, using methods including traditional sensors and surveillance video. Galloping detection relies on accurate, real-time monitoring of power line status changes, requiring the application of change thresholds. Video-based galloping detection involves feature extraction and motion analysis of consecutive frames, often relying on complex deep learning models. These models have large parameter counts and high computational costs, resulting in significant latency and failing to meet real-time monitoring requirements, making them difficult to run directly on edge devices.
[0004] Secondly, in practical applications, power transmission lines are widely laid in remote environments such as high mountains and deep valleys, spanning vast areas and complex environments. Traditional sensors are difficult to cover effectively, and they are susceptible to external interference such as wind, rain, and snow, resulting in insufficient detection range and a high false alarm rate. Furthermore, while pole cameras have been widely deployed to achieve full-area power transmission line monitoring, their utilization rate is insufficient, indicating significant potential for galloping detection. Making good use of the existing monitoring network is a crucial strategy for real-time and efficient galloping detection.
[0005] With the development of artificial intelligence and computer vision technologies, deep learning methods have been introduced into the detection of power transmission line galloping. Semantic segmentation networks can accurately classify each pixel in a video frame, thereby extracting the foreground region of the transmission line. Combined with optical flow analysis, the displacement changes of the transmission line between consecutive frames can be effectively calculated, achieving high-precision galloping recognition. This type of method significantly outperforms traditional image processing algorithms in terms of accuracy and robustness. However, in practical engineering applications, the complexity and resource consumption of deep learning models have become the main problems restricting their implementation. On the one hand, mainstream deep segmentation and detection networks have a large number of parameters and high computational cost, requiring high-performance GPUs or cloud servers to run, making direct deployment on edge computing devices at the transmission tower ends difficult. On the other hand, edge devices (such as embedded processors and smart cameras) are generally limited in terms of power consumption, memory, and computing power, unable to support large neural network models, resulting in insufficient real-time performance, large latency, and inability to meet the rapid response requirements for galloping detection.
[0006] Furthermore, the video surveillance scenes of power transmission lines are complex, with low cable pixel ratios, high background noise, and significant changes in lighting. While deep models can improve segmentation accuracy, without lightweight design and structural optimization, they can still create bottlenecks in inference speed and resource consumption. Therefore, relying solely on traditional deep networks is insufficient for achieving efficient dance detection on edge devices.
[0007] In summary, it is necessary to study a transmission line galloping discrimination method that combines deep learning and model lightweighting technology. This method can significantly reduce computational overhead while ensuring discrimination accuracy, enabling efficient deployment and real-time operation on edge devices, thereby improving the intelligence and response speed of power grid galloping monitoring. Summary of the Invention
[0008] The main objective of this invention is to provide a transmission line galloping identification method for edge-side applications, so as to achieve high-precision, low-latency, and deployable identification of transmission line galloping status.
[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a method for determining transmission line galloping on the edge side, the method comprising: A high-precision semantic segmentation model TLNet for power transmission line scenarios is constructed. TLNet is based on PIDNet and introduces the line feature interaction parallel network structure LITB, detail recovery decoder DRD and line boundary awareness auxiliary training head EPM to achieve semantic segmentation of power transmission lines. The constructed TLNet model is structurally lightweighted by using a neural architecture search NAS-based optimization method. Within a preset search space, the general inverted bottleneck module UIB is combined with multi-objective optimization search to obtain the lightweight model TLNet-Lite. Using the TLNet model as the teacher network, knowledge distillation is performed on the TLNet-Lite model to obtain the optimized lightweight segmentation model TLNet-S. The TLNet-S model is used to perform semantic segmentation on continuous video frames, extract the foreground mask of the transmission line, and calculate the dense optical flow field under the mask constraint to quantify the dancing features. The quantized galloping features are discriminated based on an adaptive threshold method, the galloping discrimination results are output, and the results are deployed on edge devices to realize the galloping discrimination of transmission lines.
[0010] Following the above technical solution, the TLNet model includes: The Line Feature Interaction Parallel Network Structure (LITB) simplifies the three-branch structure of PIDNet (proportional branch P, integral branch I, and differential branch D) into a two-branch architecture consisting of a detail mining branch (DB) and a context feature extraction branch (CB). The detail mining branch (DB) corresponds to the proportional branch (P) and is used to maintain high resolution to preserve transmission line details. The context feature extraction branch (CB) corresponds to the integral branch (I) and is used to capture global semantic information through downsampling. The detail mining branch (DB) and the context feature extraction branch (CB) interact with high-frequency features and context information through a bidirectional fusion block. The Detail Recovery Decoder (DRD) introduces cross-scale connections to fuse mid-level and deep features from the encoder. It also employs a Spatial Offset Alignment (SOA) module to adaptively learn and align the spatial offsets between high- and low-resolution feature maps in order to recover the complete outline and fine structure of the transmission line. The line boundary awareness auxiliary training head EPM uses wavelet convolution WTConv instead of ordinary convolution. By performing multi-level wavelet transform and inverse transform on the feature map, it separates and enhances the high-frequency components containing edge information. At the same time, it uses a boundary weighted loss function to assign higher weights to the pixels at the edge of the transmission line to strengthen the learning of the transmission line edge and improve the accuracy of semantic segmentation.
[0011] Following the above technical solution, the detail recovery decoder (DRD) includes a spatial offset alignment module (SOA), a stitching layer, and a context feature extraction (CB) module. The SOA connects the output of the Line Feature Interaction Parallel Network (LITB) and the feature map of the context feature extraction branch (CB), achieving cross-scale feature fusion and spatial alignment, and then reconstructing the feature map through feature upsampling. The stitching layer connects the reconstructed feature map with the downsampled feature map of the input image, and the stitched image is then input to the CB module. The CB module includes a convolutional layer (Conv) and a batch normalization (BN) layer, and the output of the CB module is connected to the segmentation head (SH). The line boundary awareness auxiliary training head EPM includes an EMP module and a boundary segmentation head BH; the input of the EMP module is connected to the feature map of the detail mining branch DB, and the output of the EMP module is connected to the boundary segmentation head BH; the boundary weighted loss function is described above. for: ; Where N, H, and W are the batch size, image height, and width, respectively; n is the sample number in the batch; and h and w are the pixel positions. This is the weight matrix; The labels are real binary boundaries, where 1 represents a boundary pixel and 0 represents a non-boundary pixel. This is the probability map of boundary pixels predicted by the TLNet model.
[0012] Following the above technical solution, structural lightweighting includes: The detail mining branch DB and the context feature extraction branch CB of the TLNet model are defined as optimization objects. The location of the standard residual block in the original network is defined as a replaceable candidate location. The set of operations available for each candidate location includes maintaining the original residual block structure or replacing it with a UIB module. The structural parameters of the UIB module are included in the search space to form a structure-parameter joint search space. The architecture search is performed using the ProxylessNAS method, which includes: constructing a weight-shared supernetwork containing all candidate operations; employing a path binarization mechanism during the search process; randomly activating only one candidate path during each forward and backward propagation; controlling the activation probability of each path through learnable architecture parameters; and optimizing the architecture parameters using gradient descent. The architecture search employs multi-objective optimization, with its objective function being... for: ; in, This is the loss for the semantic segmentation task, used to evaluate the semantic segmentation accuracy of the model; For the estimated inference latency of the target edge device; Penalty item for architectural parameter quantity; and This is a balancing factor used to adjust the trade-off between semantic segmentation accuracy and hardware efficiency.
[0013] Following the above technical solution, knowledge distillation includes: Using the TLNet model as the teacher network and the TLNet-Lite model as the student network, loss is distilled using soft labels. The soft-label distillation loss enables student networks to mimic the decision-making behavior of teacher networks. Defined as: ; in, and These are the outputs of the last layer of the teacher network and the student network, respectively. For the Softmax function; This is a temperature parameter used to smooth the probability distribution; Let KL divergence be a metric. Simultaneously employing hard label focus loss The student network is monitored to address the class imbalance problem caused by the extremely low pixel ratio in the power transmission line segmentation task, specifically the hard label focus loss. Defined as: ; Where N is the total number of pixels and C is the total number of categories; Predicting pixels for student networks Category The probability of; The corresponding real label is in one-hot format; For category The manual weighting factor is to increase the weight of the foreground class; This is an adjustment factor used to control the decay rate of the weights for easy and difficult samples; The optimized lightweight segmentation model TLNet-S is obtained by jointly optimizing the soft-label distillation loss and the hard-label focus loss. The total loss function is... for: ; in, The weighting coefficients are used to balance the two losses.
[0014] Following the above technical solution, the dance detection includes: Using the TLNet-S model to process two consecutive frames of images and Perform semantic segmentation to obtain the corresponding binary mask for the transmission line foreground. and In the mask, a pixel value of 1 represents the power line area, and 0 represents the background. The Farnebäck dense optical flow algorithm is used to focus the computation window or weights on the mask. The pixel region with a value of 1 and its neighborhood are calculated from the frame. arrive optical flow vector field Each valid position contains a horizontal displacement component. and vertical displacement components ; Based on optical flow vector field , calculate the mean value of the optical flow amplitude of all pixels within the foreground mask region : ; Among them, represents the foreground mask region, is the total number of pixels in the foreground mask region, and i is the pixel number; Obtain the video data in the non-dancing state and calculate the sequence of the mean value of the optical flow amplitude , and use the mean value of this sequence plus three times the standard deviation as the discrimination threshold , and adaptively update this threshold regularly during operation; If the current mean value of the optical flow amplitude satisfies , it is determined that the current transmission line is dancing; among them, is an adjustable sensitivity coefficient.
[0015] Continuing with the above technical solution, the dancing discrimination further includes: Calculate the standard deviation of the optical flow direction within the foreground region : ; Among them, is the optical flow direction of pixel i; is the average value of the optical flow direction; Distinguish disordered vibration and regular dancing through the standard deviation of the optical flow direction .
[0016] Continuing with the above technical solution, the method further includes the linkage response, regional collaborative response, and cloud collaborative management steps after the edge-side deployment; among them: The linkage response after the edge-side deployment includes: using the dancing discrimination result as the trigger signal for the anti-dancing device, automatically starting the damper to reduce the amplitude when regular dancing is detected; if the dancing amplitude continues to increase, linkage-adjust the line tension device; The regional collaborative response includes: when multiple adjacent transmission towers continuously detect dancing and form a regional risk, the dancing discrimination result is synchronized to the power grid dispatching center through the edge-cloud channel, and the dispatching system adjusts the load distribution of the regional power grid or temporarily cuts off the power supply of the high-risk line; The cloud collaborative management includes: uploading the dancing discrimination result and quantization data to the cloud, the cloud collects the historical discrimination data of all transmission towers, conducts an associated analysis of the dancing inducement in combination with meteorological data, counts the high-incidence periods and high-incidence line segments of dancing, and provides a basis for line renovation; at the same time, the cloud uses the accumulated historical data to reversely update the adaptive discrimination threshold of the edge-side device and dynamically adjust the sensitivity coefficient βTo improve the recognition sensitivity of low-amplitude dancing motions, and to fine-tune the TLNet-S model using new data, thereby reducing the false alarm rate in complex environments.
[0017] In a second aspect, the present invention provides a computer device / apparatus / system, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in the first aspect.
[0018] Thirdly, the present invention provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method described in the first aspect.
[0019] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: This invention proposes a transmission line galloping discrimination method for edge-side applications. It constructs a high-precision semantic segmentation model, TLNet, specifically designed for transmission line scenarios. Addressing the visual characteristics of transmission lines, such as their slender length, low pixel ratio, and sharp edges, it innovatively introduces a parallel network structure (LITB) for line feature interaction, a detail recovery decoder (DRD), and a line boundary awareness auxiliary training head (EPM). This enhances edge perception of small targets while maintaining segmentation accuracy. A multi-objective optimization strategy based on neural architecture search (NAS) and a general inverted bottleneck module (UIB) is employed to lightweight the TLNet model, maintaining high segmentation accuracy while significantly reducing parameter and computational costs. The accuracy of the lightweight model is effectively restored using the soft-hard label focus distillation method (SHD), ultimately yielding a lightweight segmentation model, TLNet-S, that balances accuracy and efficiency. Furthermore, by combining dense optical flow analysis with foreground mask constraints and an adaptive threshold discrimination method, it achieves quantitative identification and real-time discrimination of transmission line galloping states, meeting real-time inference requirements on edge-side devices. The galloping discrimination exhibits high accuracy and high recall. This invention effectively solves the problems of existing deep learning models, such as difficulty in deployment on edge devices, poor real-time performance, and decreased accuracy after lightweighting. It achieves high-precision, low-latency, and deployable real-time monitoring of transmission line galloping, significantly improving the intelligence level and emergency response capability of power grid operation and maintenance. Attached Figure Description
[0020] Figure 1 This is a flowchart of a transmission line galloping detection method oriented towards the edge side according to an embodiment of the present invention; Figure 2 This is an architecture diagram of a TLNet model according to an embodiment of the present invention; Figure 3 This is an architecture diagram of the Spatial Offset Alignment (SOA) module according to an embodiment of the present invention; Figure 4This is a schematic diagram illustrating the ProxylessNAS search UIB replacement location according to an embodiment of the present invention; Figure 5 This is a flowchart of the model optimization process based on soft and hard label focal distillation according to an embodiment of the present invention; Figure 6 This is an example result of the dance discrimination according to an embodiment of the present invention; Figure 7 This is a comparison diagram of a dancing example and a normal viewing example according to an embodiment of the present invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments provided by this invention without inventive effort are within the scope of protection of this invention.
[0022] Obviously, the accompanying drawings described below are merely some examples or embodiments of the present invention. Those skilled in the art can apply the present invention to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this invention, any changes to design, manufacturing, or production based on the technical content disclosed in this invention are merely conventional technical means and should not be construed as insufficient disclosure of the present invention.
[0023] In this invention, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this invention may be combined with other embodiments without conflict.
[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "a," "an," "an," "the," and similar words used in this invention do not indicate quantity limitation and may indicate singular or plural. The terms "comprising," "including," "having," and any variations thereof used in this invention are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms "connected," "linked," "coupled," and similar words used in this invention are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "A plurality" used in this invention refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships may exist; for example, "A and / or B" can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects have an "or" relationship. The terms "first," "second," and "third" used in this invention are merely to distinguish similar objects and do not represent a specific ordering of the objects.
[0025] To address the real-time galloping monitoring requirements of edge-side monitoring grid nodes, this invention proposes an edge-side-oriented transmission line galloping discrimination method, aiming to achieve high-precision, low-latency, and deployable identification of transmission line galloping states, thus resolving the following shortcomings in existing transmission line galloping monitoring technologies. The main technical problems addressed by this invention include: (1) Problem of insufficient real-time performance of dance detection: Although the existing deep learning-based video dance detection algorithm has achieved significant improvement in recognition accuracy, the model has a large number of parameters and high computational complexity, making it difficult to run in real time on edge devices at the end of transmission towers, resulting in an inability to respond to sudden dance phenomena in a timely manner.
[0026] (2) Problem of decreased accuracy after model lightweighting: During the process of model compression and parameter simplification, the network feature expression ability is easily reduced, resulting in a decrease in segmentation accuracy and edge perception ability, which cannot meet the engineering discrimination requirements.
[0027] (3) Limited computing power at the edge side prevents algorithm deployment: Transmission line monitoring terminals generally use low-power embedded computing platforms such as the Jetson series, which have limited memory, storage, and computing resources. Traditional deep neural networks cannot run directly on such devices, limiting the practical application of video dancing intelligent analysis functions.
[0028] In summary, there is an urgent need for a lightweight optimization method that can significantly reduce computational complexity while maintaining model accuracy. The purpose of this invention is to provide an edge-side transmission line galloping detection method. By introducing a lightweight deep learning model structure, knowledge distillation, and optical flow analysis strategy, it achieves real-time transmission line galloping identification, significantly improving the detection speed and accuracy at the edge, and meeting the real-time and deployable requirements of transmission line galloping monitoring.
[0029] This invention provides a method for identifying power transmission line galloping at the edge. By constructing a high-precision semantic segmentation model for power transmission line galloping and improving the structure of the semantic segmentation model to be lightweight, and combining optical flow analysis and knowledge distillation techniques, it achieves efficient identification and real-time monitoring of power transmission line galloping states. This method addresses the problems of high computational cost of traditional deep learning models, limited computing power of edge devices, and decreased accuracy after model lightweighting. It proposes a TLNet model based on PIDNet, and makes three key improvements targeting the unique visual characteristics of power transmission lines: fixed spatial shape, small pixel ratio, and clear edges. It proposes a lightweight structure scheme based on neural architecture search and designs a soft and hard label focus distillation mechanism to maintain segmentation accuracy after model compression. By introducing an optical flow offset calculation module, it quantitatively extracts the motion trajectory of the power transmission line in consecutive frames, achieving automatic identification of the galloping state. This invention effectively improves the model's running speed and recognition accuracy at the edge, and significantly enhances the real-time performance and stability of the system on low-power devices. Its main process is as follows: Figure 1 As shown.
[0030] like Figure 1 As shown, the transmission line galloping detection method facing the edge side of the present invention mainly includes the following steps: Step S110: Construction of a high-precision semantic segmentation model for power transmission line galloping detection.
[0031] To achieve accurate extraction of power transmission line targets in surveillance videos and provide a reliable foreground mask for subsequent galloping detection, this step first constructs and optimizes a high-precision semantic segmentation model specifically for power transmission line scenarios, called TLNet (Transmission Line Network). Figure 2 As shown, this model is based on the real-time semantic segmentation network PIDNet and incorporates three key improvements tailored to the unique visual characteristics of power transmission lines: fixed spatial shape, small pixel ratio, and clear edges. The specific steps are as follows: S111. Benchmark Model Selection and Feature Analysis.
[0032] To achieve a balance between accuracy and speed, PIDNet was selected as the baseline model. The PIDNet model employs a three-branch structure: proportional (P), integral (I), and derivative (D), responsible for detail preservation, contextual semantic extraction, and edge perception, respectively. Its control theory-inspired design is suitable for handling targets with clear edges and structure. Considering the characteristics of transmission lines as appearing slender, radially distributed, and with complex backgrounds from the perspective of tower monitoring, targeted enhancements to the baseline model are necessary.
[0033] S112. Design of Line Feature Interactive Parallel Network Architecture (LITB).
[0034] To address the redundancy in function and structure between the P and D branches in PIDNet, a parallel network structure with line feature interaction is proposed. The original three branches are simplified into a two-branch architecture consisting of a detail mining branch (DB) and a context feature extraction branch (CB). The detail mining branch (DB) corresponds to the proportional branch (P), and the context feature extraction branch (CB) corresponds to the integral branch (I), while the original differential branch (D) is removed. The DB branch maintains high resolution to preserve transmission line details, while the CB branch captures global semantics through downsampling. The two branches interact with high-frequency features and contextual information through a bidirectional fusion block, reducing the number of parameters while enhancing the modeling capability of the transmission line's spatial structure.
[0035] S113. Detail Recovery Decoder (DRD) Design.
[0036] To address the issues of small pixel ratio in power transmission lines and the tendency for continuous downsampling to lead to breakage or loss of slender targets, a detail recovery decoder is proposed. This decoder introduces cross-scale connections to fuse features from the middle and deep layers of the encoder. The key innovation lies in the Spatial Offset Alignment (SOA) module, which adaptively learns and aligns the spatial positional offsets between high- and low-resolution feature maps, thereby more accurately recovering the complete outline and fine structure of the power transmission lines during feature upsampling reconstruction.
[0037] Specifically, such as Figure 2 As shown, the detail recovery decoder (DRD) includes a spatial offset alignment module (SOA), a stitching layer, and a CB module; the architecture of the spatial offset alignment module (SOA) is as follows: Figure 3 As shown, the spatial offset alignment module (SOA) connects the output of the parallel network structure LITB and the feature map of the context feature extraction branch (CB) to achieve cross-scale feature fusion and spatial alignment, and then reconstructs the feature map through feature upsampling. The stitching layer connects the reconstructed feature map with the downsampled feature map of the input image, and the stitched feature map is input to the CB module. The CB module includes a convolutional layer (Conv) and a batch normalization (BN). The output of the CB module is connected to the segmentation head (SH).
[0038] S114. Design of Line Boundary Awareness Assisted Training Head (EPM).
[0039] To enhance the model's sensitivity to clear transmission line boundaries, a line boundary awareness auxiliary training head is proposed. For example... Figure 2 As shown, the line boundary awareness auxiliary training head EPM includes an EMP module and a boundary segmentation head BH. The input of the EMP module is connected to the feature map of the detail mining branch DB, and the output of the EMP module is connected to the boundary segmentation head BH. This module uses wavelet convolution (WTConv) instead of ordinary convolution, and effectively separates and enhances high-frequency components containing edge information by performing multi-level wavelet transform and inverse transform on the feature map. Simultaneously, a boundary-weighted loss function is designed, defined as: ; Where N, H, and W are the batch size, image height, and width, respectively; n is the sample number in the batch; and h and w are the pixel positions. The core is the weight matrix; The labels are real binary boundaries, where 1 represents a boundary pixel and 0 represents a non-boundary pixel. This is the probability map of boundary pixels predicted by the TLNet model.
[0040] During training, higher weights are assigned to pixels at the edges of the transmission lines, thereby guiding the model to focus on the transition region between the foreground and background, significantly improving the sharpness and accuracy of the segmentation boundaries.
[0041] S115. Model Training and TLNet Acquisition.
[0042] The model integrating the three improvements was trained using a power transmission line semantic segmentation dataset. By jointly optimizing the main semantic segmentation loss and the boundary-aware auxiliary loss, a high-performance power transmission line-specific semantic segmentation model, TLNet, was finally obtained.
[0043] Step S120: Lightweight model structure based on neural architecture search and UIB module.
[0044] To address the issues of high computational cost and insufficient real-time performance of the TLNet model on edge devices, this paper proposes a lightweight model structure strategy based on Neural Architecture Search (NAS) and the Universal Inverted Bottleneck (UIB) module. The UIB module is derived from the MobileNetV4 network and significantly reduces computational overhead while maintaining feature representation capabilities through optional depthwise convolutions and efficient channel design.
[0045] The specific implementation process includes the following stages: S121. Define the lightweight search space: The backbone network and some contextual feature extraction branches of the TLNet model are used as the optimization objects. For example... Figure 4 As shown, a series of standard residual blocks in the original network are defined as replaceable candidate positions. The set of operations available for each position is: maintain the original residual block structure, or replace it with a UIB module. The structural parameters of the UIB module (expansion ratio, kernel size) are also included in the search space, forming a joint structure-parameter search space.
[0046] S122. A proxylessNAS approach is used for architecture search: a weight-shared supernet containing all candidate operations is constructed. During the search process, a path binarization mechanism is employed, randomly activating only one candidate path during each forward and backward propagation, thus significantly reducing the computational and memory overhead of the search process. This mechanism controls the activation probability of each path through learnable architecture parameters, and optimizes these parameters through gradient descent.
[0047] S123. Implement multi-objective optimization search: The optimization objective function of the search process comprehensively considers model accuracy and hardware efficiency, and is defined as: ; in, For semantic segmentation tasks, loss functions (such as cross-entropy loss) are used to evaluate model accuracy. Estimated inference latency for target edge devices (such as NVIDIA Jetson TXI); This is a penalty term for the number of model parameters, which is proportional to the number of parameters. and It is a balance coefficient used to adjust the trade-off between accuracy and efficiency.
[0048] S124. Exporting the Optimal Lightweight Model: After the search is complete, the optimal operation for each candidate position is selected based on the converged architecture parameters, and the final lightweight network structure, called TLNet-Lite, is exported. This stage only completes the structural lightweighting; the model accuracy needs to be restored and optimized in subsequent steps.
[0049] Step S130: Model optimization based on soft and hard label focus distillation.
[0050] To compensate for the performance loss caused by the lightweighting of the structure in step S120, this step proposes a soft-hard label alignment focus distillation method. Using the more accurate original TLNet model as the teacher network, knowledge transfer and performance enhancement are performed on the lightweight student network (TLNet-Lite), such as... Figure 5 As shown. The specific steps are as follows: S131. Construct the teacher-student distillation framework: Fix the pre-trained teacher network, i.e., the original TLNet model. Place the student network, i.e., the TLNet-Lite model obtained in step S120, into the distillation framework for learning. Training data flows through both the teacher network and the student network simultaneously.
[0051] S132. Soft-Label Knowledge Transfer: The predicted probability distribution (Softmax output) generated by the teacher network for the input image contains rich information about inter-class relationships and uncertainties, referred to as soft labels. By minimizing the Kullback-Leibler divergence between the student network output and the teacher network's soft labels, the student network is encouraged to imitate the teacher's decision-making behavior. Soft-Label Distillation Loss The calculation is as follows: ; in, and These are the logits outputs of the last layer of the teacher and student networks, respectively. For the Softmax function; Temperature parameter (usually) This is used to smooth probability distributions and highlight dark knowledge. Let KL divergence be denoted as KL divergence.
[0052] S133. Hard Label Focus Loss Supervision: Simultaneously, the student network must also receive supervision using ground truth hard labels. To address the severe class imbalance caused by the extremely low proportion of power line pixels in power line segmentation tasks, focus loss is used instead of standard cross-entropy loss. Focus loss automatically reduces the weights of easily classified background pixels, focusing training on hard-to-classify power line pixels and edge regions. Hard Label Loss The calculation is as follows: ; Where N is the total number of pixels and C is the total number of categories; Predicting pixels for student networks Category The probability of; The corresponding real label is in one-hot format; For category The manual weighting factor is to increase the weight of the foreground class; This is an adjustment factor used to control the decay rate of the weights of easy and difficult samples.
[0053] S134. Joint Optimization and Boundary-Aware Alignment: The overall training loss of the student network is a weighted sum of the soft-label distillation loss and the hard-label focus loss. ; in, To balance the weighting coefficients of the two losses (experiments have shown that setting it to 0.5 yields optimal results), joint optimization is employed. The student network can not only learn generalized knowledge from the teacher network, but also maintain the ability to distinguish small targets and clear boundaries of transmission lines under strong supervision of real labels.
[0054] S135. Obtain the optimized lightweight model: After sufficient training through the knowledge distillation process described above, the final performance-optimized lightweight segmentation model is obtained, named TLNet-S (S stands for Student). While maintaining a similar number of parameters and computational efficiency as TLNet-Lite, this model significantly restores and improves its segmentation accuracy (mIoU, mAcc, etc.), making it suitable for edge deployment.
[0055] Step S140: Determining the galloping state of transmission lines based on foreground mask optical flow analysis.
[0056] This step aims to quantify and discriminate the motion state of power transmission lines in consecutive video frames. Its core lies in using the lightweight, high-precision segmentation model (TLNet-S) obtained in step S130 to extract a stable power transmission line foreground, then calculating a high-confidence optical flow field on this binary mask, and finally achieving dance-like motion discrimination through statistical motion features. The specific steps are as follows: S141. Foreground mask extraction and alignment.
[0057] For the input video stream, extract two consecutive frames in sequence. and Using the TLNet-S model deployed at the edge, real-time semantic segmentation is performed on two frames of images to obtain the corresponding binary mask for the transmission line foreground. and In the mask, a pixel value of 1 represents the power line area, and 0 represents the background.
[0058] S142. Calculation of dense optical flow under mask constraints.
[0059] To reduce interference from complex background textures, the computation area is constrained to the foreground region of the transmission line. Specifically, the Farnebäck dense optical flow algorithm is employed, but the computation window or weights are focused on the mask. The pixel region with a value of 1 and its neighborhood are calculated from the frame. arrive optical flow vector field Each valid position contains a horizontal displacement component. and vertical displacement components .
[0060] S143. Quantification of dancing features.
[0061] Based on optical flow vector field The core features for calculating the degree of dancing are: (1) Average motion amplitude: Calculate the average optical flow amplitude of pixels in all foreground mask regions. .
[0062] ; in, Indicates the foreground mask area. This represents the total number of pixels in the region.
[0063] (2) Consistency of motion direction: Calculate the standard deviation of the optical flow direction in the foreground region. To distinguish between disordered vibrations and regular dancing movements: ; in, The direction of optical flow for pixel i; This represents the average value in the direction of optical flow.
[0064] When the light dances rhythmically, the direction of the light flow is concentrated near a certain main direction, resulting in... Smaller, below the threshold; during disordered vibration, the direction of motion is random and chaotic, and the optical flow direction is dispersed, leading to... The value is relatively large, exceeding the threshold. Therefore, a preset threshold can be used to distinguish between disordered vibration and regular galloping. In the transmission line galloping discrimination of this embodiment, this threshold is set to 26.3°. The threshold determination step is as follows: take a video segment of disordered vibration, and calculate the value of each frame. Values, then calculate all mean of values μ and standard deviation σ Finally, the threshold was set to , k The optical flow direction sensitivity coefficient is set to 2 in this embodiment.
[0065] S144. Adaptive threshold dance discrimination.
[0066] The dance detection does not rely on a single fixed threshold, but instead employs an adaptive thresholding method based on historical statistics: (1) Threshold initialization and update: During the initial run of the system, a video data segment under normal conditions without dancing is collected, and its optical flow amplitude mean sequence is calculated. and the mean of the sequence Plus three standard deviations As the initial discrimination threshold During operation, this threshold can be periodically updated with new, normal data.
[0067] (2) Decision-making: For continuous frames input in real time, calculate their... If the following formula is satisfied, then dancing is determined to have occurred during that time period: ; in, The currently used adaptive threshold, This is an adjustable sensitivity coefficient, usually set to 1.0.
[0068] S145. Results Output and Visualization.
[0069] The dance judgment results (normal / dancing), key quantitative indicators ( , The system outputs the timestamp and other data. It also generates visualizations for monitoring personnel, including examples and comparison charts. Figure 7 As shown.
[0070] Step S150: Edge-side deployment and real-time identification.
[0071] The lightweight and distilled optimized model is deployed on an edge computing device to achieve real-time operation of video input, foreground extraction and dance detection on the edge: First, the device takes a video stream as input and obtains the foreground mask of the transmission line through the lightweight segmentation model; then the system performs optical flow calculation and dance detection locally, and outputs the dance detection result and real-time alarm information.
[0072] The superior performance of TLNet-S was verified by comparing it with several mainstream lightweight backbone networks. The experimental results are shown in Table 1.
[0073] Table 1. Experimental results comparing TLNet-S with mainstream lightweight solutions
[0074] The table shows a comparison between the NAS-UIB and SHD hybrid lightweight strategy and other lightweight solutions. Experiments show that although the introduction of NAS-UIB slightly reduces mIoU by 0.35%, SHD partially restores the model's accuracy, thus maintaining a level comparable to TLNet. In terms of inference speed, TLNet-S's frame rate increases from 65.09 FPS to 102.28 FPS, an increase of 56.9%.
[0075] Experimental results show that the proposed algorithm significantly enhances real-time inference capabilities while maintaining essentially the same accuracy, and has better potential for edge applications.
[0076] In addition, experiments were conducted to demonstrate the effectiveness of the lightweight improvement of the TLNet structure based on the NAS-UIB module. The experimental results, validating the effectiveness of the lightweight improvement of the TLNet structure using the NAS-UIB module, are shown in Table 2.
[0077] Table 2. Experimental results comparing NAS-UIB before and after replacing residual convolution.
[0078] As shown in the table, the introduction of the NAS-UIB module reduced the number of model parameters by 48.6% and the computational cost from 41.37G to 18.38G, a reduction of 55.6%. In terms of real-time inference speed, the speed improved by 57.1%. This demonstrates that UIB position insertion under NAS search makes the model more lightweight. However, the reduction in the number of lightweight modules and channels resulted in a slight decrease in mIoU and mDice, from 87.21% to 86.25% and from 92.73% to 92.10%, respectively.
[0079] Experimental results show that the NAS-UIB lightweight strategy enables the model to improve inference speed without sacrificing too much accuracy.
[0080] The effectiveness of the lightweight soft and hard label focused distillation (SHD) model was verified, and the experimental results are shown in Table 3.
[0081] Table 3. Verification of the effectiveness of SHD distillation
[0082] As shown in the table, the SHD method significantly improved the model's accuracy, increasing the lightweight model's mIoU from 86.25% to 86.86% (an improvement of 0.61%) and mDice from 92.10% to 92.50% (an increase of 0.4%). This indicates that distilling students through the teacher network can achieve considerable accuracy recovery. Compared to CWD distillation, introducing SHD improved mIoU by 0.31%, mAcc by 1.5%, and mDice by 0.18%. This demonstrates that through soft and hard label focus distillation, the student model can better learn the knowledge from the teacher model, especially on more difficult samples, thus enhancing the model's performance.
[0083] Experimental results show that the student model can learn the knowledge of the teacher model better through soft and hard label focus distillation, especially on more difficult samples, thereby enhancing the model's performance.
[0084] The final transmission line galloping discrimination results were verified. The experimental results of galloping discrimination combining TLNet-S semantic segmentation and optical flow method are shown in Table 4.
[0085] Table 4. Effect of transmission line galloping detection
[0086] As shown in the table, the accuracy, recall, and F1 score of the dance discrimination method combining TLNet-S semantic segmentation and optical flow are 84.25%, 90.31%, and 87.17%, respectively. Figure 6 The results of optical flow-based galloping detection in three long video clips are presented, with solid origins indicating frames identified as galloping. Experimental accuracy results and real-world examples demonstrate that the proposed edge-side transmission line galloping detection system can accurately and quantitatively identify galloping phenomena. The HSV color space is used to map the amplitude and direction of optical flow onto the image, generating an intuitive RGB image representation. Test results are shown below. Figure 7 As shown, the arrows represent the direction of optical flow, and the size of the arrows represents the amplitude of the dance. Comparing the dance instance and the normal instance, it can be found that the dance amplitude between two consecutive frames in the three frames of the dance instance is significantly greater than that in the normal instance.
[0087] The transmission line galloping detection method for the edge side of the present invention can be used for: (1) Real-time emergency response.
[0088] Automatic linkage adjustment of on-site equipment: The judgment result can be used as a trigger signal for anti-galloping devices such as dampers and phase spacers. When regular galloping is detected (low standard deviation of directional consistency), the damper is automatically activated to reduce the amplitude. If the galloping amplitude continues to increase, the line tension device can be adjusted in linkage to reduce the risk of conductor collision and strand breakage.
[0089] Regional power grid emergency dispatch: When multiple adjacent towers are continuously detected to be galloping (forming regional risks), the judgment results are synchronized to the power grid dispatch center through the edge-cloud channel. The dispatch system can quickly adjust the load distribution of the power grid in the region, or temporarily cut off the power supply to high-risk lines (to avoid the spread of short circuit trips), and restore power supply after the galloping subsides.
[0090] Tiered alarms and precise push notifications: The judgment results are linked to local audible and visual alarms on edge devices (such as tower end warning lights and buzzers). When multiple adjacent towers are continuously detected to be gyrating, the alarm information is pushed to the maintenance personnel's APP and the regional monitoring center via 5G / IoT. The information includes "gyrating tower ID, gyrating intensity (average motion amplitude), and duration", enabling precise positioning and tiered response (such as triggering an emergency alarm when the amplitude is greater than 1.5 times the threshold, and prioritizing dispatching of orders).
[0091] (2) Cloud-based collaborative management.
[0092] The judgment results are not only used for local response. After the quantitative data is uploaded to the cloud, it can support macro-management and system optimization at the network level, making up for the limitations of local monitoring of edge devices.
[0093] Deep Dive into the Patterns of Pole Movement: Historical data (including timestamps and environmental parameters) of all poles are collected in the cloud, and core patterns are uncovered through big data analysis.
[0094] Trigger correlation: By combining local meteorological data (wind speed, wind direction, temperature), we analyze the correlation between "specific wind speed range + terrain (high mountains and steep ridges)" and gossamer pulsars, and clarify the gossamer pulsar trigger thresholds in different regions.
[0095] High-incidence characteristics: Statistics on high-incidence periods of troop dancing (such as the spring windy season) and high-incidence line sections (such as cross-valley conductors) provide a basis for subsequent line renovation.
[0096] Full network situation visualization and resource allocation: The cloud will map the real-time judgment results into a heat map of transmission line galloping risk, marking the galloping intensity and duration in different areas. The dispatch center can optimize the allocation of operation and maintenance resources based on the heat map (such as prepositioning operation and maintenance teams in high-incidence areas and stockpiling emergency supplies).
[0097] Model iteration and threshold optimization: The cloud uses accumulated historical data to update the adaptive threshold (Tcurrent) of edge devices in reverse—for example, if the amplitude of the dance is found to be generally low in a certain area during winter, the threshold coefficient can be dynamically adjusted. β (e.g., from 1.0 to 0.8) to improve the recognition sensitivity of low-amplitude dancing motions; at the same time, the TLNet-S model is fine-tuned with new data to further reduce the false alarm rate in complex environments (such as blizzards and strong backlight).
[0098] (3) Precise optimization of operation and maintenance: reduce costs and increase efficiency, and improve the level of intelligent operation and maintenance.
[0099] Traditional power transmission line operation and maintenance faces challenges such as wide coverage, environmental risks, and high costs. The assessment results can transform blind inspections into precise operation and maintenance, with the core being to improve the efficiency and targeting of operation and maintenance. ① Targeted and precise inspection: Once a pole is identified as galloping, the system automatically marks it as a key inspection target. Maintenance personnel do not need to conduct a full-area inspection, but only need to perform detailed inspections on this area (such as checking whether the conductors are broken or the insulators are damaged). Based on the galloping amplitude data, the system prioritizes the inspection (the larger the amplitude, the higher the inspection priority), thus avoiding the waste of manpower.
[0100] ② Full life cycle management of equipment health status: Based on the quantitative data of "frequency + intensity" of galloping, establish a conductor fatigue damage model - for example, if a line has an average annual galloping frequency of >50 times and an average amplitude of >3cm, it can be predicted that its service life will be shortened by 20%, and a conductor replacement plan can be formulated in advance; at the same time, track the operation effect of anti-galloping device (such as whether the galloping amplitude decreases after activation), evaluate the reliability of the device and maintain it in a timely manner.
[0101] ③ Dynamic adjustment of operation and maintenance plan: Based on the high-incidence season / time period of galloping analyzed by the cloud, adjust the operation and maintenance plan in advance: for example, before the spring windy season, carry out pre-reinforcement of lines in high-incidence areas (such as tightening wire joints); during the low-incidence period of galloping (such as summer), carry out concentrated equipment maintenance to avoid conflicts between operation and maintenance and the high-incidence period of galloping.
[0102] (4) Cross-scenario intelligent upgrade: build a full-link smart grid.
[0103] The judgment results, as a core component of the dynamic status data of transmission lines, can be integrated with other systems to support the power grid's upgrade from passive monitoring to active prediction and intelligent control. ① Meteorological-Galloping Linkage Prediction: The judgment results are integrated with meteorological forecast data (wind speed, wind direction, air pressure) to train the "meteorological-galloping prediction model". For example, when the wind speed in a certain area is predicted to reach 10-15 m / s (the critical wind speed for galloping), the model combines the historical galloping data of the area to push out a potential galloping risk warning 6-12 hours in advance, so that maintenance personnel can be on site in advance to prepare.
[0104] ② Multi-dimensional line status fusion assessment: The galloping judgment results are linked with other monitoring data (such as conductor ice thickness, temperature, and insulator leakage current) to form a comprehensive line health score. For example, when there is a dual risk of "galloping + icing", the score triggers the highest level warning and initiates more stringent emergency measures (such as shutdown for maintenance). If it is only low-amplitude galloping and the conductor temperature and icing are normal, operation can be maintained and continuous monitoring can continue.
[0105] ③ In-depth application of digital twin scenarios: The judgment results are synchronized to the power grid digital twin model in real time, and the conductor galloping trajectory and amplitude are restored in the virtual scene to simulate the stress effect of different galloping intensities on line towers, joints and insulators; through simulation analysis, the installation position of anti-galloping devices is optimized (such as adding dampers to the conductor section with the largest galloping amplitude), or the effectiveness of line modification schemes (such as adjusting conductor spacing) is verified.
[0106] Furthermore, the present invention also provides a computer device / apparatus / system, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method.
[0107] The present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.
[0108] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, systems, and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, as well as 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, a special-purpose computer, an 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, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or block diagrams.
[0109] These computer program instructions may also be stored in a computer-readable storage medium capable of directing 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 that implement the functions specified in one or more flowcharts and / or one or more blocks in a block diagram.
[0110] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, thereby providing steps for implementing the functions specified in one or more flowcharts and / or one or more blocks in a block diagram.
[0111] It should be noted that, depending on the implementation needs, the various steps / components described in this invention can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.
[0112] Those skilled in the art will readily understand that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for detecting galloping of transmission lines oriented towards the edge side, characterized in that, The method includes: A high-precision semantic segmentation model TLNet for power transmission line scenarios is constructed. TLNet is based on PIDNet and introduces the line feature interaction parallel network structure LITB, detail recovery decoder DRD and line boundary awareness auxiliary training head EPM to achieve semantic segmentation of power transmission lines. The constructed TLNet model is structurally lightweighted by using a neural architecture search NAS-based optimization method. Within a preset search space, the general inverted bottleneck module UIB is combined with multi-objective optimization search to obtain the lightweight model TLNet-Lite. Using the TLNet model as the teacher network, knowledge distillation is performed on the TLNet-Lite model to obtain the optimized lightweight segmentation model TLNet-S. The TLNet-S model is used to perform semantic segmentation on continuous video frames, extract the foreground mask of the transmission line, and calculate the dense optical flow field under the mask constraint to quantify the dancing features. The quantized galloping features are discriminated based on an adaptive threshold method, the galloping discrimination results are output, and the results are deployed on edge devices to realize the galloping discrimination of transmission lines.
2. The transmission line galloping detection method for the edge side according to claim 1, characterized in that, The TLNet model includes: The Line Feature Interaction Parallel Network Structure (LITB) simplifies the three-branch structure of PIDNet (proportional branch P, integral branch I, and differential branch D) into a two-branch architecture consisting of a detail mining branch (DB) and a context feature extraction branch (CB). The detail mining branch (DB) corresponds to the proportional branch (P) and is used to maintain high resolution to preserve transmission line details. The context feature extraction branch (CB) corresponds to the integral branch (I) and is used to capture global semantic information through downsampling. The detail mining branch (DB) and the context feature extraction branch (CB) interact with high-frequency features and context information through a bidirectional fusion block. The Detail Recovery Decoder (DRD) introduces cross-scale connections to fuse mid-level and deep features from the encoder. It also employs a Spatial Offset Alignment (SOA) module to adaptively learn and align the spatial offsets between high- and low-resolution feature maps in order to recover the complete outline and fine structure of the transmission line. The line boundary awareness auxiliary training head EPM uses wavelet convolution WTConv instead of ordinary convolution. By performing multi-level wavelet transform and inverse transform on the feature map, it separates and enhances the high-frequency components containing edge information. At the same time, it uses a boundary weighted loss function to assign higher weights to the pixels at the edge of the transmission line to strengthen the learning of the transmission line edge and improve the accuracy of semantic segmentation.
3. The transmission line galloping detection method for the edge side according to claim 2, characterized in that, The Detail Recovery Decoder (DRD) comprises a Spatial Offset Alignment (SOA) module, a stitching layer, and a Context Extraction (CB) module. The SOA connects the output of the Line Feature Interaction Parallel Network (LITB) and the feature map of the Context Feature Extraction (CB) branch, achieving cross-scale feature fusion and spatial alignment, and then reconstructs the feature map through feature upsampling. The stitching layer connects the reconstructed feature map with the downsampled feature map of the input image, and the stitched map is then input to the CB module. The CB module includes a convolutional layer (Conv) and a batch normalization (BN) layer; the output of the CB module is connected to the segmentation head (SH). The line boundary awareness auxiliary training head EPM includes an EMP module and a boundary segmentation head BH; the input of the EMP module is connected to the feature map of the detail mining branch DB, and the output of the EMP module is connected to the boundary segmentation head BH; the boundary weighted loss function is described above. for: ； Where N, H, and W are the batch size, image height, and width, respectively; n is the sample number in the batch; and h and w are the pixel positions. This is the weight matrix; The labels are real binary boundaries, where 1 represents a boundary pixel and 0 represents a non-boundary pixel. This is the probability map of boundary pixels predicted by the TLNet model.
4. The transmission line galloping detection method for the edge side according to claim 2, characterized in that, Structural lightweighting includes: The detail mining branch DB and the context feature extraction branch CB of the TLNet model are defined as optimization objects. The location of the standard residual block in the original network is defined as a replaceable candidate location. The set of operations available for each candidate location includes maintaining the original residual block structure or replacing it with a UIB module. The structural parameters of the UIB module are included in the search space to form a structure-parameter joint search space. The architecture search is performed using the ProxylessNAS method, which includes: constructing a weight-shared supernetwork containing all candidate operations; employing a path binarization mechanism during the search process; randomly activating only one candidate path during each forward and backward propagation; controlling the activation probability of each path through learnable architecture parameters; and optimizing the architecture parameters using gradient descent. The architecture search employs multi-objective optimization, with its objective function being... for: ； in, This is the loss for the semantic segmentation task, used to evaluate the semantic segmentation accuracy of the model; For the estimated inference latency of the target edge device; Penalty item for architectural parameter quantity; and This is a balancing factor used to adjust the trade-off between semantic segmentation accuracy and hardware efficiency.
5. The transmission line galloping detection method for the edge side according to claim 1, characterized in that, Knowledge distillation includes: Using the TLNet model as the teacher network and the TLNet-Lite model as the student network, loss is distilled using soft labels. The soft-label distillation loss enables student networks to mimic the decision-making behavior of teacher networks. Defined as: ； in, and These are the outputs of the last layer of the teacher network and the student network, respectively. For the Softmax function; This is a temperature parameter used to smooth the probability distribution; Let KL divergence be a metric. Simultaneously employing hard label focus loss The student network is monitored to address the class imbalance problem caused by the extremely low pixel ratio in the power transmission line segmentation task, specifically the hard label focus loss. Defined as: ； Where N is the total number of pixels and C is the total number of categories; Predicting pixels for student networks Category The probability of; The corresponding real label is in one-hot format; For category The manual weighting factor is to increase the weight of the foreground class; This is an adjustment factor used to control the decay rate of the weights for easy and difficult samples; The optimized lightweight segmentation model TLNet-S is obtained by jointly optimizing the soft-label distillation loss and the hard-label focus loss. The total loss function is... for: ； in, The weighting coefficients are used to balance the two losses.
6. The transmission line galloping detection method for the edge side according to claim 1, characterized in that, Dance discrimination includes: Using the TLNet-S model to process two consecutive frames of images and Perform semantic segmentation to obtain the corresponding binary mask for the transmission line foreground. and In the mask, a pixel value of 1 represents the power line area, and 0 represents the background. The Farnebäck dense optical flow algorithm is used to focus the computation window or weights on the mask. The pixel region with a value of 1 and its neighborhood are calculated from the frame. arrive optical flow vector field Each valid position contains a horizontal displacement component. and vertical displacement components ; Based on optical flow vector field Calculate the mean optical flow amplitude of all pixels within the foreground mask region. : ； in, Indicates the foreground mask area. is the total number of pixels in the foreground mask region, and i is the pixel number; Acquire video data in a non-dancing state and calculate the mean sequence of optical flow amplitude. The mean of the sequence plus three standard deviations is used as the discrimination threshold. And during operation, the threshold is updated adaptively and periodically. If the current average optical flow amplitude is satisfy If this is the case, then it is determined that the current power transmission line is galloping; among them, This is an adjustable sensitivity coefficient.
7. The transmission line galloping detection method for the edge side according to claim 6, characterized in that, Dance discrimination also includes: Calculate the standard deviation of the optical flow direction within the foreground region. : ； in, The direction of optical flow for pixel i; This represents the average value along the optical flow direction. Standard deviation of optical flow direction Distinguish between disordered vibrations and regular dancing movements.
8. The transmission line galloping detection method for the edge side according to claim 7, characterized in that, The method also includes steps for coordinated response after edge deployment, regional collaborative response, and cloud-based collaborative management; among which: The linkage response after deployment on the edge side includes: using the galloping judgment result as the trigger signal for the anti-galloping device; automatically activating the damper to reduce the amplitude when regular galloping is detected; and adjusting the line tension device in conjunction with the galloping amplitude if the galloping amplitude continues to increase. Regional coordinated response includes: when multiple adjacent towers are continuously detected to gallop, forming a regional risk, the galloping judgment results are synchronized to the power grid dispatch center through the edge-cloud channel, and the dispatch system adjusts the regional power grid load distribution or temporarily cuts off the power supply to high-risk lines; Cloud-based collaborative management includes: uploading galloping detection results and quantitative data to the cloud; the cloud aggregating historical detection data from all towers; combining this data with meteorological data to conduct correlation analysis of galloping causes; and statistically analyzing high-incidence periods and sections of lines to provide a basis for line upgrades. Simultaneously, the cloud uses accumulated historical data to update the adaptive detection thresholds of edge-side devices, dynamically adjusting the sensitivity coefficients. β To improve the recognition sensitivity of low-amplitude dancing motions, and to fine-tune the TLNet-S model using new data, thereby reducing the false alarm rate in complex environments.
9. A computer device / equipment / system, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 8.