Phased array radar and visible light sensor cooperative sensing power transmission channel external damage prevention method

Abstract

The invention discloses a phased array radar and visible light sensor cooperative sensing power transmission channel external damage prevention method, relates to the field of electric power, and aims to solve the problems that a background along a power transmission line is easily confused with target features, and the multi-scale target detection precision is insufficient. The method comprises the steps of obtaining a target image, and constructing a sample data set; performing pre-training by using the sample data set to obtain a weight model of the Yolov11s network model; dynamic snakelike convolution is introduced, an attention mechanism module and a detection head are added, and an improved Yolov11s network model is obtained; migrating the weight model to an improved Yolov11s network model, training to obtain a weight model suitable for external damage prevention hidden danger detection of the power transmission channel, and performing external damage prevention detection of the power transmission channel by using the model; the method improves the detection precision of various external damage targets of the power transmission channel, strengthens the attention of small targets, and is suitable for different target sizes and complex background scenes.

Description

Technical Field

[0001] This invention relates to the field of power technology, and in particular to a method for coordinating phased array radar and visible light sensors to detect and prevent external damage to power transmission channels. Background Technology

[0002] As the scale of the power system continues to expand, the safe operation of transmission channels, as the core carriers of power transmission, is directly related to the stability of the power grid. However, transmission channels typically cover a wide area and have complex terrain, making them susceptible to external damage. Traditional monitoring for external damage to transmission channels mainly relies on manual inspections and single-sensor monitoring, which is labor-intensive, costly, and extremely inefficient in complex terrains such as mountains and rivers, and prone to missed detections.

[0003] In recent years, with the rapid development of deep learning technology, many excellent machine vision object detection algorithms have emerged and have been widely applied in production and daily life. Among them, the YOLO algorithm primarily transforms the object detection task into a regression problem, using a single deep neural network model to simultaneously predict the position and category of all objects in an image during a single forward pass. YOLO does not require sliding window or candidate region detection on the image, thus significantly reducing computation time and improving detection efficiency. While it can be used for detecting external damage risks in power transmission channels, direct application still presents some problems, such as the background along the transmission line easily being confused with target features, and insufficient accuracy for multi-scale target detection. Therefore, it is necessary to propose a method for collaborative sensing of power transmission channel external damage risks using phased array radar and visible light sensors to address these issues. Summary of the Invention

[0004] The purpose of this invention is to provide a method for coordinating phased array radar and visible light sensors to detect and prevent external damage to power transmission channels, in order to solve the problems in the prior art where the background along the power transmission line is easily confused with the target features and the detection accuracy for multi-scale targets is insufficient.

[0005] This invention provides a method for coordinating phased array radar and visible light sensors to detect and prevent external damage to power transmission channels, comprising the following steps:

[0006] Step 1: Acquire target image; wherein, the transmission line tower is equipped with phased array radar and visible light sensor. When the target enters the warning area, the visible light sensor automatically adjusts the pan-tilt angle and focal length according to the coordinates provided by the phased array radar to acquire target image;

[0007] Step 2: Classify and label the target images and use them as labels; expand the dataset by rotating and scaling to construct a sample dataset.

[0008] Step 3: Use the sample dataset to perform pre-training to obtain the weight model of the Yolov11s network model;

[0009] Step 4: Improve the Yolov11s network model by introducing dynamic snake convolution, adding an attention mechanism module, and adding a detection head to obtain the improved Yolov11s network model.

[0010] Step 5: Transfer the weight model to the improved Yolov11s network model, and then train it using the sample dataset to obtain a weight model suitable for detecting external damage hazards in power transmission channels.

[0011] Step 6: Use the weighted model applicable to the detection of external damage hazards in power transmission channels to perform external damage detection of power transmission channels.

[0012] Furthermore, in step four, the introduction of dynamic serpentine convolution includes:

[0013] The original C3K2 module in the network is replaced by C3K2-DySnakeConv. The original features are received through the corresponding layers of the backbone network. The convolutional blocks and residual connections of the C3K2 module are used to ensure that the gradient does not vanish during the feature extraction process and to maintain the stability of network training. C3K2-DySnakeConv introduces dynamic snake convolution in each convolutional block. In the dynamic snake convolution, the standard convolutional kernel is linearized in both the x-axis and y-axis directions.

[0014] Furthermore, when dealing with irregularly shaped targets, the X-axis convolution kernel addresses elongated characteristics through horizontal extension, while the Y-axis adapts to shape changes through vertical deformation. Together, they capture the local curvature features of elongated targets. Equations 1 and 2 are the formulas for the changes in the X and Y axes, respectively:

[0015] Formula 1

[0016] Formula 2

[0017] in, This represents the coordinates of a point located at a horizontal distance c from the center grid along the i-direction. This represents the coordinates of a point located at a horizontal distance c from the center grid along the j-direction; c = 0, 1, 2, 3, 4, representing the horizontal distance from the center grid; ∆ represents the offset; x and y represent the X-axis and Y-axis coordinates, respectively; i represents the grid index in the Y-axis direction, and j represents the grid index in the X-axis direction.

[0018] Using bilinear interpolation, it can be expressed as:

[0019] Formula 3

[0020] Where K is the position to be interpolated, K' is the four nearest integer pixel coordinates around K, B is the interpolation kernel function, and B(K',K) is the weight value of position K' with respect to target position K.

[0021] Furthermore, in step four, an attention mechanism module is added, including:

[0022] The system filters feature channels specific to the external breach target. Channel attention uses the importance of feature channels to filter channels that are strongly correlated with the external breach target. For the input feature map, channel attention first uses global average pooling and global max pooling to capture the distribution pattern of channel features from two dimensions: global statistics and local extrema. Then, it generates weights for each channel through a shared fully connected layer and Sigmoid activation.

[0023] Furthermore, the channel attention weight matrix Mc is expressed as: R represents the set of real numbers, and C represents the number of channels;

[0024] In a multilayer perceptron, a dimensionality reduction coefficient r is used: The formula for calculating channel attention is shown in Equation 4:

[0025] Formula 4

[0026] in, This represents the channel attention weight matrix; This represents the Sigmoid activation function; Represents a multilayer perceptron; Indicates average pooling; Indicates the input feature map; This indicates max pooling; , This represents the weight parameters of the MLP; This represents the global average pooling feature. This represents the max pooling feature.

[0027] Furthermore, step four, which incorporates the attention mechanism module, also includes:

[0028] High-risk areas around the line are identified; spatial attention strengthens the features of high-risk areas through weight allocation in the feature space; for the feature map after channel attention processing, spatial attention first performs global average pooling and global max pooling on the channel dimension, and then generates weights for each spatial location through 1×1 convolution and Sigmoid activation.

[0029] Furthermore, the spatial attention weight matrix Ms is expressed as: R represents the set of real numbers, H represents the height of the feature map, and W represents the width of the feature map;

[0030] Two pooling methods are used in the channel dimension to generate 2D feature maps:

[0031]

[0032] The final formula for calculating spatial attention is shown in Equation 5:

[0033] Formula 5

[0034] in, Represents the spatial attention weight matrix. This represents the feature after global average pooling. This represents the feature after global max pooling. This represents a 7×7 convolution operation.

[0035] Furthermore, in step four, adding the detection head includes:

[0036] A new 160×160 small target detection head is added to the original detection head. The feature input of this detection head integrates the high-resolution output of the C3k2 module of the first layer of the backbone network and the shallow information sampled from the neck network layer. The small target features are enhanced and background interference is suppressed through the convolutional block attention module.

[0037] The beneficial effects of this invention are as follows: The phased array radar and visible light sensor collaborative sensing method for preventing external damage to power transmission channels replaces the backbone network C3K2 module with the C3K2-DySnakeConv module. While maintaining the residual structure to ensure training stability, it introduces dynamic serpentine convolution to adapt to the irregular and small-scale external damage target morphology in the power transmission channel, accurately capturing the local curvature features of the target and fusing them with global features, thus enhancing the feature extraction of small and irregular targets. A convolutional block attention module is added, which is a channel-space dual-dimensional attention module. Channel attention filters the feature channels specific to external damage targets, and spatial attention locks high-risk areas around the line. The progressive enhancement of the two provides the detection head with accurately focused feature input. A new 160×160 small target detection head is added, which fuses high-resolution and shallow upsampled features and enhances them through the convolutional block attention module, solving the problem of missed detection of small targets in the original detection head. These improvements, without significantly increasing computational costs, enhance the detection accuracy of various external damage targets in power transmission channels, strengthen the focus on small targets, adapt to different target sizes and complex background scenarios, and provide more reliable technical support for monitoring external damage to power transmission channels. Attached Figure Description

[0038] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart of the method for coordinating phased array radar and visible light sensor to detect and prevent external damage to power transmission channels according to the present invention;

[0040] Figure 2 To improve the YOLOv11S architecture;

[0041] Figure 3 This is the original C3K2 module structure diagram;

[0042] Figure 4 To improve the C3K2 module structure diagram;

[0043] Figure 5 A network framework diagram for the convolutional block attention module;

[0044] Figure 6 Diagram of the channel attention network framework;

[0045] Figure 7 A spatial attention network framework diagram. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The technical solutions provided by various embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0047] Please see Figures 1 to 7 This invention provides a method for coordinating phased array radar and visible light sensors to detect and prevent external damage to power transmission channels, comprising the following steps:

[0048] Step 1: Acquire target image; wherein, the transmission line tower is equipped with phased array radar and visible light sensor. When the target enters the warning area, the visible light sensor automatically adjusts the pan-tilt angle and focal length according to the coordinates provided by the phased array radar to acquire target image.

[0049] Specifically, this invention utilizes a collaborative sensing method combining phased array radar and visible light sensors. Through deep fusion of multiple sensors and algorithm optimization, it achieves all-weather detection and high-precision identification of external damage hazards. A high-frequency phased array radar is installed on the transmission line towers. When the phased array radar detects a target entering the warning area, a high-definition visible light camera is triggered to automatically adjust the pan-tilt angle and focal length based on the target coordinates provided by the radar, acquiring a high-definition image of the target. An improved YOLOv11s algorithm enables accurate identification of the types of external damage hazards in the transmission channel and allows for grading.

[0050] Step two: Classify and label the target images and use them as labels; expand the dataset by rotating and scaling to construct a sample dataset.

[0051] Step 3: Use the sample dataset to perform pre-training to obtain the weight model of the Yolov11s network model.

[0052] Step four: Improve the Yolov11s network model by introducing dynamic snake convolution, adding an attention mechanism module, and adding a detection head to obtain the improved Yolov11s network model.

[0053] Specifically, this invention improves and optimizes the existing YOLOv11s network model in the following three aspects:

[0054] First, replacing the existing C3K2 module with C3K2-DySnakeConv can effectively enable the network to focus on targets such as hanging objects and tilted construction machinery in the power transmission channel.

[0055] In power transmission channel image acquisition scenarios, due to limitations in monitoring distance and equipment resolution, the pixel area of ​​externally damaged targets in the image is relatively small. They contain very little feature information such as edges and textures, and are easily interfered with by environmental factors such as tree branches and shadows from power lines. This causes the native C3K2 module, due to its fixed grid sampling, to be unable to effectively capture target features, resulting in a high false negative rate. C3K2-DySnakeConv, a dynamic snake convolution enhanced C3K2 module, retains the original residual structure of C3K2 but replaces the core convolution with dynamic snake convolution (DySnakeConv). By dynamically sampling to adapt to the target shape and combining spatial attention to focus on key areas, it enhances the feature extraction capability for irregular, small-scale externally damaged targets while suppressing complex background interference, thereby improving the model's detection accuracy and robustness for small targets in power transmission channels. Its working principle is as follows:

[0056] Preserving the residual feature framework: Initially, the original features are received through the corresponding layers of the backbone network, using the convolutional blocks and residual connection basic structure of the C3K2 module to ensure that the gradient does not vanish during the feature extraction process and to maintain the stability of network training.

[0057] Dynamic serpentine sampling adapts to target shape: Unlike the fixed standard convolution in the traditional C3K2 module, C3K2-DySnakeConv introduces dynamic serpentine convolution in each convolutional block. In dynamic serpentine convolution, the standard convolution kernel is linearized in both the x-axis and y-axis directions. Taking the x-axis direction as an example, the specific position of each grid in K is represented by c = 0, 1, 2, 3, 4, which represents the horizontal distance from the center grid. When dealing with irregularly shaped targets, the convolution kernel in the x-axis extends horizontally to address elongated characteristics, while the y-axis undergoes vertical deformation to adapt to shape changes. The two work together to capture the local curvature features of elongated targets. Equations 1 and 2 are the formulas for the changes in the x and y axes, respectively.

[0058] Formula 1

[0059] Formula 2

[0060] in, This represents the coordinates of a point located at a horizontal distance c from the center grid along the i-direction. This represents the coordinates of a point located at a horizontal distance c from the center grid along the j-direction; c = 0, 1, 2, 3, 4, representing the horizontal distance from the center grid; ∆ represents the offset; x and y represent the X-axis and Y-axis coordinates, respectively; i represents the grid index in the Y-axis direction, and j represents the grid index in the X-axis direction.

[0061] Using bilinear interpolation, it can be expressed as:

[0062] Formula 3

[0063] Where K is the position to be interpolated, K' is the four nearest integer pixel coordinates around K, B is the interpolation kernel function, and B(K',K) is the weight value of position K' with respect to target position K.

[0064] The refined local features extracted by dynamic serpentine convolution are fused with the global features of the C3K2 residual branch. For example, when a hanging object is detected entangled on the line, the dynamic serpentine convolution captures the details of the curved edge of the hanging object, while the residual branch conveys the spatial position and direction of the line. After the two are fused, the model can accurately locate the hanging object and clearly define its spatial relationship with the line.

[0065] Second, a Convolutional Block Attention Module (CBAM) is added to the network. CBAM is a channel-space dual-dimensional attention module. Addressing the pain points of scarce target features and concentrated risk areas in power transmission channel external damage targets, it achieves dual-dimensional synergy of accurate target feature selection and targeted focusing on high-risk areas.

[0066] Channel Attention Mechanism (CAM): This mechanism filters feature channels specific to external damage targets. Transmission line images contain complex elements such as lines, towers, and trees, and the features of external damage targets, such as construction machinery and hanging objects, are often obscured by the background. Channel Attention precisely filters channels strongly correlated with external damage targets by weighting the importance of feature channels. For the input feature map, CAM first uses Global Average Pooling (GAP) and Global Max Pooling (GMP) to capture the distribution patterns of channel features from both global statistical and local extremum dimensions; then, it generates weights for each channel through a shared fully connected layer and sigmoid activation.

[0067] The channel attention weight matrix Mc is expressed as follows: R represents the set of real numbers, and C represents the number of channels;

[0068] In a multilayer perceptron, a dimensionality reduction coefficient r is used: The formula for calculating channel attention is shown in Equation 4:

[0069] Formula 4

[0070] in, This represents the channel attention weight matrix; This represents the Sigmoid activation function; Represents a multilayer perceptron; Indicates average pooling; Indicates the input feature map; This indicates max pooling; , This represents the weight parameters of the MLP; This represents the global average pooling feature. This represents the max pooling feature.

[0071] Spatial Attention Mechanism (SAM): This mechanism identifies high-risk areas around power transmission lines. External damage risks are not uniformly distributed but concentrated within a 5-50 meter radius around the line. SAM uses weight allocation in the feature space to specifically enhance features in high-risk areas. For the feature map processed by channel attention, the SAM first performs global average pooling and global max pooling on the channel dimension, then generates weights for each spatial location using 1×1 convolution and sigmoid activation. The spatial attention weight matrix Ms is represented as follows: R represents the set of real numbers, H represents the height of the feature map, and W represents the width of the feature map;

[0072] Two pooling methods are used in the channel dimension to generate 2D feature maps:

[0073]

[0074] The final formula for calculating spatial attention is shown in Equation 5:

[0075] Formula 5

[0076] in, Represents the spatial attention weight matrix. This represents the feature after global average pooling. This represents the feature after global max pooling. This represents a 7×7 convolution operation.

[0077] The convolutional block attention module employs a progressive enhancement approach, first filtering features and then defining regions. The first step involves channel attention filtering for target-related features and removing background interference, resulting in a set of channels containing only the potential features of the externally damaged target. The second step involves spatial attention focusing on the risk areas surrounding the line within the filtered channels, further enhancing the target's feature response within these risk zones. This collaborative approach allows the feature map to retain the core features of the externally damaged target while focusing on high-risk spatial areas, providing accurate and relevant input to the detection head. The detection head only needs to identify targets within high-weight channels and regions, significantly reducing the probability of false positives and false negatives.

[0078] Third, the original YOLOv11 has three detection heads, but small targets such as hanging objects and small construction tools in power transmission channels often have a pixel size of less than 30×30px in the image. Existing detection heads are prone to missed detections due to insufficient feature resolution and low anchor box matching. Therefore, a new 160×160 small target detection head is added to the original detection head. The feature input of this detection head integrates the high-resolution output of the C3k2 module of the first layer of the backbone network and the shallow information sampled from the neck network layer. The small target features are enhanced and background interference is suppressed through the convolutional block attention module, thereby improving the detection accuracy of small targets such as hanging objects and small construction tools.

[0079] Step 5: Transfer the weight model to the improved Yolov11s network model, and then train it using the sample dataset to obtain a weight model suitable for detecting external damage hazards in power transmission channels.

[0080] Step 6: Use the weighted model applicable to the detection of external damage hazards in power transmission channels to perform external damage detection of power transmission channels.

[0081] As can be seen from the above embodiments, the phased array radar and visible light sensor collaborative sensing method for preventing external damage to power transmission channels of the present invention replaces the backbone network C3K2 module with the C3K2-DySnakeConv module. While retaining the residual structure to maintain training stability, it introduces dynamic serpentine convolution to adapt to the irregular and small-scale external damage target morphology in the power transmission channel, accurately captures the local curvature features of the target and fuses them with global features, and strengthens the feature extraction of small and irregular targets. A convolutional block attention module is added. This module is a channel-space dual-dimensional attention module. Channel attention filters the exclusive feature channels of external damage targets, and spatial attention locks the high-risk areas around the line. The two progressively enhance each other to provide the detection head with accurately focused feature input. A new 160×160 small target detection head is added, which integrates high-resolution and shallow upsampling features and is enhanced by the convolutional block attention module to solve the problem of missed detection of small targets in the original detection head. These improvements, without significantly increasing computational costs, enhance the detection accuracy of various external damage targets in power transmission channels, strengthen the focus on small targets, adapt to different target sizes and complex background scenarios, and provide more reliable technical support for monitoring external damage to power transmission channels.

[0082] The embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention.

Claims

1. A method for coordinating phased array radar and visible light sensors to detect and prevent external damage to power transmission channels, characterized in that, Includes the following steps: Step 1: Acquire target image; wherein, the transmission line tower is equipped with phased array radar and visible light sensor. When the target enters the warning area, the visible light sensor automatically adjusts the pan-tilt angle and focal length according to the coordinates provided by the phased array radar to acquire target image; Step 2: Classify and label the target images and use them as labels; expand the dataset by rotating and scaling to construct a sample dataset. Step 3: Use the sample dataset to perform pre-training to obtain the weight model of the Yolov11s network model; Step 4: Improve the Yolov11s network model by introducing dynamic snake convolution, adding an attention mechanism module, and adding a detection head to obtain the improved Yolov11s network model. Step 5: Transfer the weight model to the improved Yolov11s network model, and then train it using the sample dataset to obtain a weight model suitable for detecting external damage hazards in power transmission channels. Step 6: Use the weighted model applicable to the detection of external damage hazards in power transmission channels to perform external damage detection of power transmission channels.

2. The method for coordinating phased array radar and visible light sensor to detect and prevent external damage to power transmission channels according to claim 1, characterized in that, Step four involves introducing dynamic serpentine convolution, including: The original C3K2 module in the network is replaced by C3K2-DySnakeConv. The original features are received through the corresponding layers of the backbone network. The convolutional blocks and residual connections of the C3K2 module are used to ensure that the gradient does not vanish during the feature extraction process and to maintain the stability of network training. C3K2-DySnakeConv introduces dynamic snake convolution in each convolutional block. In the dynamic snake convolution, the standard convolutional kernel is linearized in both the x-axis and y-axis directions.

3. The method for coordinating phased array radar and visible light sensor to detect and prevent external damage to power transmission channels according to claim 2, characterized in that, When dealing with irregularly shaped targets, the X-axis convolution kernel addresses elongated characteristics through horizontal extension, while the Y-axis adapts to shape changes through vertical deformation. Together, they capture the local curvature features of elongated targets. Equations 1 and 2 are the formulas for the changes in the X and Y axes, respectively: Formula 1 Formula 2 in, This represents the coordinates of a point located at a horizontal distance c from the center grid along the i-direction. This represents the coordinates of a point located at a horizontal distance c from the center grid along the j-direction; c = 0, 1, 2, 3, 4, representing the horizontal distance from the center grid; ∆ represents the offset; x and y represent the X-axis and Y-axis coordinates, respectively; i represents the grid index in the Y-axis direction, and j represents the grid index in the X-axis direction. Using bilinear interpolation, it can be expressed as: Formula 3 Where K is the position to be interpolated, K' is the four nearest integer pixel coordinates around K, B is the interpolation kernel function, and B(K',K) is the weight value of position K' with respect to target position K.

4. The method for coordinating phased array radar and visible light sensor to detect and prevent external damage to power transmission channels according to claim 2, characterized in that, In step four, an attention mechanism module is added, including: The system filters feature channels specific to the external breach target. Channel attention uses the importance of feature channels to filter channels that are strongly correlated with the external breach target. For the input feature map, channel attention first uses global average pooling and global max pooling to capture the distribution pattern of channel features from two dimensions: global statistics and local extrema. Then, it generates weights for each channel through a shared fully connected layer and Sigmoid activation.

5. The method for coordinating phased array radar and visible light sensor to detect and prevent external damage to power transmission channels according to claim 4, characterized in that, The channel attention weight matrix Mc is expressed as follows: R represents the set of real numbers, and C represents the number of channels; In a multilayer perceptron, a dimensionality reduction coefficient r is used: The formula for calculating channel attention is shown in Equation 4: Formula 4 in, This represents the channel attention weight matrix; This represents the Sigmoid activation function; Represents a multilayer perceptron; Indicates average pooling; Indicates the input feature map; This indicates max pooling; , This represents the weight parameters of the MLP; This represents the global average pooling feature. This represents the max pooling feature.

6. The method for coordinating phased array radar and visible light sensor to detect and prevent external damage to power transmission channels according to claim 5, characterized in that, Step four, which includes adding an attention mechanism module, also includes: High-risk areas around the line are identified; spatial attention strengthens the features of high-risk areas through weight allocation in the feature space; for the feature map after channel attention processing, spatial attention first performs global average pooling and global max pooling on the channel dimension, and then generates weights for each spatial location through 1×1 convolution and Sigmoid activation.

7. The method for coordinating phased array radar and visible light sensor to detect and prevent external damage to power transmission channels according to claim 6, characterized in that, The spatial attention weight matrix Ms is expressed as: R represents the set of real numbers, H represents the height of the feature map, and W represents the width of the feature map; Two pooling methods are used in the channel dimension to generate 2D feature maps: The final formula for calculating spatial attention is shown in Equation 5: Formula 5 in, Represents the spatial attention weight matrix. This represents the feature after global average pooling. This represents the feature after global max pooling. This represents a 7×7 convolution operation.

8. The method for coordinated sensing of power transmission channels by phased array radar and visible light sensors to prevent external damage, as described in claim 2, is characterized in that... Step four, adding the detection head includes: A new 160×160 small target detection head is added to the original detection head. The feature input of this detection head integrates the high-resolution output of the C3k2 module of the first layer of the backbone network and the shallow information sampled from the neck network layer. The small target features are enhanced and background interference is suppressed through the convolutional block attention module.

Images