A lightweight insulator defect detection method and system for power unmanned aerial vehicle inspection and a medium

Abstract

The application discloses a kind of lightweight insulator defect detection method and system for power unmanned aerial vehicle inspection, belong to low-altitude economic unmanned aerial vehicle target detection technical field.The method is based on YOLOv11n: implicit space-frequency decoupling unit is introduced in shallow layer of backbone network, high and low frequency characteristics are separated in spatial domain and dynamically calibrated, and small target high-frequency texture is retained;State space selection enhancement unit is deployed in middle layer of backbone network, and large-area background noise such as iron tower, forest and the like is removed by using differentiable sparse mask, and semantics is balanced by memory gate;Geometric perception deformation unit is deployed in neck network, geometric offset is predicted in a lightweight manner, so that convolution kernel is adaptively fitted to the edge of long string-shaped and other irregular defects.The application significantly improves the positioning accuracy of insulator micro and irregular defects in unmanned aerial vehicle aerial image under complex background while greatly reducing the parameter quantity and calculation quantity, and has great end-side edge real-time deployment value.

Description

Technical Field

[0001] This invention relates to the field of insulator surface defect detection, and specifically to a lightweight insulator defect detection method and system for power drone inspection. Background Technology

[0002] Insulators, as core insulating and supporting components in transmission lines, are exposed to the open environment for extended periods, enduring the combined effects of mechanical stress, strong electric fields, and harsh weather conditions. With increasing service life, insulators are highly susceptible to physical defects such as spontaneous explosion, breakage, flashover, or surface flashover. If these defects are not detected and addressed in a timely manner, they can directly lead to grounding short circuits or large-scale power outages, posing a serious threat to the safe and stable operation of the smart grid.

[0003] In recent years, using drones equipped with high-definition inspection pods to replace traditional manual tower climbing inspections has become a mainstream trend in modern power operation and maintenance. Drone inspections can acquire massive amounts of high-resolution images of transmission lines, but this also places extremely high demands on the automatic defect identification algorithms at the back-end or edge.

[0004] While traditional deep learning-based object detection algorithms demonstrate good detection accuracy on general datasets, they face the following technical bottlenecks when directly applied to power grid drone inspection scenarios:

[0005] High-frequency features of minute defects are severely lost in the spatial domain: Traditional networks use continuous stride convolution or pooling downsampling to crudely compress the spatial domain, which causes the high-frequency edge features of minute defects of insulators, which occupy only a few pixels in a wide aerial view, to be permanently lost in deep networks, resulting in serious missed detections.

[0006] The contradiction between redundant edge computing power and carpet search: The computing power of drone-borne edge computing devices is limited, while the traditional convolutional "carpet search" paradigm results in massive redundant computing power in invalid background areas such as the sky and forests, which cannot meet the requirements of high frame rate real-time detection.

[0007] Geometric mismatch between fixed receptive field and irregularly shaped targets: Insulators and defects often present irregular geometric shapes from an aerial perspective. The receptive field of the fixed square sampling grid of traditional convolution is rigid and cannot fit the physical edge of the defect. It is easy to introduce background noise, leading to false detections, and it is difficult to locate accurately.

[0008] Therefore, how to effectively filter out complex background noise and accurately capture the characteristics of irregular and irregular defects while strictly limiting the computational load of the model, and achieve the ultimate balance between lightweight and high precision, is a technical bottleneck that urgently needs to be solved in the field of intelligent power inspection. Summary of the Invention

[0009] To address the challenges of complex background interference, inaccurate detection of irregular and irregular defects, and the large parameter count and inability of traditional deep learning models to achieve high frame rate real-time inspection at edge devices in existing technologies, this invention provides a lightweight insulator defect detection method for power plant drone inspection. This method implicitly decouples and extracts high- and low-frequency features in the spatial domain, actively filters out complex background noise in the backbone network using a state-space sparse scanning mechanism and global memory gating, and combines geometric perception deformation units to fit irregular defects. This significantly improves the detection accuracy of small and irregular defects in insulators and greatly reduces the computational complexity of the algorithm model, making it easy to deploy on edge devices such as drones.

[0010] Includes the following steps:

[0011] Step 1: Use power drones to inspect transmission lines and acquire inspection images containing insulators. After preprocessing and scaling, construct an insulator defect dataset. Divide the dataset into training, validation, and test sets according to the proportions.

[0012] Step 2: Construct a lightweight insulator defect detection network based on YOLOv11n. The lightweight insulator defect detection network includes a backbone network, a neck network, and a detection head. The shallowest downsampling stage of the backbone network is equipped with an implicit frequency decoupling (LFD) unit. The shallow and middle layers of the backbone network are equipped with state space selection augmentation (SSA) units. The scale transformation downsampling nodes of the neck network are equipped with geometry-aware modulation (GAM) units. Iteratively train the lightweight insulator defect detection network based on the training set, and verify its performance using the validation set and the test set to obtain the trained lightweight insulator defect detection network. Iteratively train the lightweight insulator defect detection network based on the training set, and verify the network model for defect detection and performance using the validation set and the test set to obtain the trained lightweight insulator defect detection network.

[0013] Step 3: Obtain the inspection image of the insulators of the transmission line to be inspected, and input it into the trained lightweight insulator defect detection network to obtain the initial image features; in the shallowest downsampling stage of the backbone network, the initial image features are processed by the Latent Frequency Decoupling (LFD) unit to extract the low-frequency structural features and high-frequency damage edge features of the insulators to obtain the decoupling feature map;

[0014] The decoupled feature map is passed to the State Space Selection Augmentation (SSA) units in the shallow and middle layers of the backbone network to filter out complex background noise in the image and obtain a denoised feature map.

[0015] The denoised feature map is input into the Geometry-Aware Modulation (GAM) unit of the scale transformation downsampling node of the neck network to perform geometric alignment and feature fusion of the defect edges, and output a multi-scale fused feature map.

[0016] The multi-scale fused feature map is input into the detection head of the lightweight insulator defect detection network for target decoding, and the location bounding box and category prediction result of the insulator defect target are output, thereby completing the lightweight insulator defect detection of the transmission line.

[0017] Preferably, in the labeling process of step 1, the insulator defect dataset is specifically labeled according to the actual defect type of the power inspection as follows: insulator breakage, insulator micro-cracks, insulator edge missing, insulator spontaneous explosion, insulator flashover, and insulator surface pollution flashover.

[0018] Preferably, in step 1, the ratio of the training set, validation set, and test set is 8:1:1.

[0019] Preferably, in step 3, the specific processing steps of the LFD unit are as follows:

[0020] A1: The unit first applies a low-pass filter operator to the features of the initial input image in the spatial domain. Smoothing is performed to extract low-frequency components containing the macroscopic contours of large surfaces. ;

[0021] A2: By subtracting the low-frequency components from the features of the initial input image, a high-pass filter operator is simulated to remove background smoothing information, generating a high-frequency component that includes defect edges and texture details. ;

[0022] A3: Lightweight channel-wise grouped convolutions are used for low-frequency components. and high frequency components Independent feature extraction is performed, and learnable frequency band calibration weights are introduced to dynamically recalibrate the importance of the extracted low-frequency structural features and high-frequency damage edge features, resulting in recalibrated low-frequency structural features. High-frequency damage edge characteristics The calculation formula is as follows:

[0023] ;

[0024] ;

[0025] ;

[0026] ;

[0027] in This represents a two-dimensional average pooling operation. and These are the learnable low-frequency weights and high-frequency weights, respectively. This is a depthwise separable grouped convolution operation;

[0028] A4: Low-frequency structural characteristics after recalibration High-frequency damage edge characteristics Direct splicing is performed along the channel dimension, and through... Pointwise convolutional layers perform channel fusion and reconstruction, outputting decoupled feature maps. :

[0029] ;

[0030] ;

[0031] in, For channel splicing operations, SiLU is the activation function. For batch normalization layer.

[0032] Preferably, in step 3, the specific processing steps of the SSA unit are as follows:

[0033] B1: Decoupling feature map output from upstream unit Input feature map as the current state space selection enhancement unit The input feature map First, a spatial feature response map is generated using an explicit saliency scoring network. And flatten it along the spatial dimension into a one-dimensional sequence score The calculation formula is as follows:

[0034] ;

[0035] ;

[0036] in This indicates that the sequence can be flattened into a one-dimensional sequence.

[0037] B2: Set the sparsity filtering ratio =0.3, calculate the number of high-value tokens that need to be retained. Using a continuous Top-K relaxation selection strategy to obtain the response score The spatial index location of the maximum value is used to construct a hard threshold mask. ;

[0038] B3: To prevent the discrete masks generated by hard-threshold masks from causing gradient non-differentiability and gradient breakage problems during backpropagation, a straight-through estimator (STE) is introduced to reconstruct the soft gradient routes of the mask, resulting in a sparse state mask with soft gradients. :

[0039] ;

[0040] in This indicates the operation of truncating the backpropagation gradient;

[0041] B4: Utilizing the generated sparse state mask With input feature map Element-wise multiplication forces zeroing out large-area complex background noise sequences (such as grasslands and forests) that are not Top-K, yielding masking features. The masking features are fed into a state transition convolution, through... Local feature extraction is performed using a channel-wise grouped convolution operator to obtain updated local state features. At the same time, utilizing Standard pointwise convolution pairs input feature maps Perform global semantic mapping to obtain global contextual features : The global context features With the local state features Perform matrix addition, and through The convolutional layers complete the fusion, thereby generating preliminary selected features to characterize the new state. Finally, learnable global memory-gated residual parameters are introduced. And combined with spatial characteristic response map A global memory-gated residual compensation mechanism is constructed to adaptively adjust the fusion weights between the initially selected features and the decoupled feature map, and outputs a denoised feature map after background suppression. :

[0042] ;

[0043] ;

[0044] ;

[0045] ;

[0046] ;

[0047] in express Channel-by-channel grouped convolution, It can be represented as a memory gating scalar parameter for learning.

[0048] Preferably, in step 3, the specific processing steps of the GAM unit are as follows:

[0049] C1: For the elongated, string-like physical structure of insulators as seen from an aerial perspective, the GAM unit receives the denoised feature map transmitted from upstream. Input feature map as the current geometrically perceptive deformation unit A lightweight parameter prediction branch is constructed within the GAM unit, and the denoised feature map is processed using a cascaded operator based on depthwise separable convolution (DWConv). Perform adaptive feature mapping and generate the required 2D spatial offsets for deformable convolution in parallel. With sampling point weight mask :

[0050] ;

[0051] ;

[0052] in, and These represent lightweight, depthwise separable convolution operators used to predict the spatial bias tensor and the sampling weight tensor, respectively.

[0053] C2: An orthogonal direction-aware pooling decoupling mechanism is employed to extract the orientation information of the insulator string structure from the denoised feature map. This decoupling along horizontal and vertical orthogonal directions captures long-range dependencies, resulting in axial dependency features. Subsequently, through two layers of standards Convolution and nonlinear activation functions Regarding the axial dependence feature Perform channel dimension mapping and normalization, and generate spatial attention weights based on the directional information. :

[0054] ;

[0055] ;

[0056] in and These represent orthogonal strip-shaped adaptive average pooling operators along the height and width directions, respectively, which together constitute the aforementioned direction-aware pooling mechanism.

[0057] C3: Discarding the traditional fixed square convolution sampling window, and combining the spatial offset with the sampling point weight mask, a deformable convolution operator is used to adaptively sample and reconstruct edges of irregular and irregularly shaped defect regions (such as thin cracks and scattered flashes) on the denoised feature map, outputting deformation modulation features. Simultaneously, geometric regularization constraints are applied. Limiting the range of variation of the spatial offset to prevent it from diverging, and finally combining it with spatial attention weights. The fusion is completed, and a multi-scale fused feature map that achieves geometric alignment and feature wrapping of irregular defect edges is output. :

[0058] ;

[0059]

[0060] ;

[0061] in This is a standard deformable convolution operation. The semicolon (;) in the parentheses represents the unique subject feature being convolved, and the semicolon represents the auxiliary guiding parameters that assist the subject in performing adaptive edge deformation sampling and background noise removal.

[0062] Preferably, the present invention also provides a lightweight insulator defect detection system for power plant drone inspection scenarios, which physically includes an airborne edge computing terminal and a ground operation and maintenance control master station; specifically, it includes a data acquisition module, a preprocessing module, a defect detection core module, and a defect detection marking and positioning output module.

[0063] At the airborne edge computing terminal, an inspection image containing insulators is acquired through a data acquisition module. Subsequently, a preprocessing module performs preprocessing, scale normalization, and accelerated scaling on the inspection image to generate an inspection image of the transmission line insulators to be inspected. Next, the inspection image of the transmission line insulators to be inspected enters the defect detection core module. This module has a built-in lightweight insulator defect detection network that has been trained. It utilizes embedded LFD, SSA, and GAM units to collaboratively complete feature extraction and feature fusion, performing spatial high- and low-frequency separation, zeroing out complex background noise, and compressing branch parameters, significantly reducing the floating-point operation overhead at the airborne end. It outputs a multi-scale fused feature map and achieves real-time high-frame-rate defect search. The defect detection function cascaded to the backend... The measurement marking and positioning output module receives the multi-scale fused feature map and performs target decoding, outputting the location bounding box and category confidence of the insulator defect target, rendering a highlighted defect mark on the local image, and extracting the center coordinates of the target pixels; simultaneously, combining the current satellite 3D spatial geographic coordinates of the UAV, the yaw angle and pitch angle data of the airborne gimbal, the real-world absolute geographic coordinates of the defective insulator are calculated online through a geometric projection algorithm; finally, through data communication and the ground operation and maintenance control master station module, a standard message containing the defect mark image, defect confidence, category and absolute geographic coordinates is transmitted back to the ground operation and maintenance control master station in real time via a wireless network, triggering an alarm and automatically generating a hazard handling work order, realizing a closed-loop collaboration from autonomous inspection to ground handling. Attached Figure Description

[0064] The accompanying drawings are provided to further illustrate the technical solutions of the present invention and form part of the specification.

[0065] Figure 1 This invention provides an overall structural diagram of a lightweight insulator defect detection network for power plant drone inspection.

[0066] Figure 2 A control flowchart of an implicit space-frequency decoupling unit (LFD) provided in an embodiment of the present invention;

[0067] Figure 3 A control flowchart of a State Space Selection Enhancement Unit (SSA) provided in an embodiment of the present invention;

[0068] Figure 4 A control flowchart of a geometrically sensing deformation unit (GAM) provided for an embodiment of the present invention. Detailed Implementation

[0069] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. To make the objectives, technical solutions, and advantages of the invention clearer, the embodiments of the invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the invention, and not all of them. Based on the embodiments of the invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the invention.

[0070] like Figure 1-3 As shown, this invention provides a lightweight insulator defect detection method for power plant drone inspections, specifically including the following steps:

[0071] Step 1: Preprocessing and Dataset Construction

[0072] A drone is used to conduct aerial inspections of power transmission lines, acquiring raw inspection images including insulators. The raw inspection images are three-channel color pixel matrices. Since the input size of the target detection network is fixed (set to [value] in this embodiment), [further details are needed]. Image size normalization is required. For images with inconsistent aspect ratios, image padding is used to resize them, and a grid is filled around the edges using fixed pixels with the same color as the image edges to ensure that the image content is not distorted. For images that need to be scaled, a bilinear interpolation algorithm is used. Finally, data annotation software is used to annotate the bounding boxes of insulator defects in the images, constructing a standard insulator defect dataset, which is then divided into training, validation, and test sets in an 8:1:1 ratio.

[0073] Step 2: Construction of a lightweight insulator defect detection network

[0074] Level 0 (Stage-1 initial feature layer): Input after normalization The image matrix is ​​processed through a standard convolutional layer (Conv) for fast dimensionality reduction and shallow perception, with an output size of [size missing]. Shallow feature map;

[0075] Stage 1 (Stage-1 downsampling node): Replaces the original YOLOv11n downsampling convolution in situ, deploys LFD units, filters out macroscopic background at nodes that maintain high resolution, and outputs a size of [missing information]. Frequency enhancement feature map;

[0076] Stage 2 (Shallow Bottleneck Layer): Replace the original C3k2 cells in situ, deploy SSA cells, and the output size is... Spatial denoising feature map;

[0077] Stage 3 (Stage-3 conventional downsampling layer): uses a standard step size of 2. Convolutional layer (Conv), output size is ;

[0078] Level 4 (Stage-3 middle bottleneck layer): Replaces the original C3k2 unit in situ, deploys a second layer of SSA units, performs secondary active background noise suppression on the middle layer features, and the output size is... ;

[0079] Layers 5 to 10 (deep feature layers): These layers sequentially cascade standard convolutional downsampling layers, original C3k2 units, SPPF spatial pyramid pooling units, and C2PSA attention mechanism layers. Finally, layer 10 outputs a highly generalized, high-level abstract semantic feature map with a size of [size missing]. .

[0080] Step 3: Multi-scale feature fusion and alignment network construction

[0081] A neck network is constructed, which cross-fuses low-level spatial details and high-level semantic features from each layer of the backbone network with a bidirectional feature pyramid (FPN+PAN) from top to bottom and bottom to top. The key feature is that, at the scale-sampling nodes of the neck network, traditional strided convolutions are replaced in situ with GAM units.

[0082] The irregular geometric offsets predicted by GAM units are used to adaptively fit the physical topological edges of irregular defects. Multi-scale feature maps are concatenated through a channel concatenation (Concat) operation to construct a feature matrix with no redundancy and perfect scale alignment. Finally, three fused feature maps of different scales are fed to the head network, with the following dimensions: (Corresponding to small target detection) (Corresponding to target detection) and (Corresponds to large target detection).

[0083] Step 4: Iterative Model Training and Defect Labeling Output

[0084] The lightweight insulator defect detection network was trained end-to-end iteratively using the training set. During the test and inference phase, inspection images of the transmission line insulators to be detected were input into the trained network model. Three decoupled prediction heads decoded the multi-scale feature maps and directly output the bounding box coordinates of the insulator defects. The confidence probability of the defect category is used to achieve automated identification and highlighting during inspections.

[0085] In step 1, the constructed insulator defect dataset consists of high-resolution images collected by a drone pod under different weather conditions. Based on actual industrial defect standards for power line inspection, the dataset is categorized into various insulator defect types, including: insulator breakage, micro-cracks in insulators, missing edges in insulators, spontaneous explosion of insulators, flashover of insulators, and surface pollution flashover of insulators.

[0086] The data augmentation in step 1 is performed online during model training. Mosaic enhancement randomly crops and stitches together four different inspection images, forcibly enriching the background diversity of the images and increasing the number of samples of small targets in a single batch. Adaptive lighting adjustment simulates aerial photography under strong light and cloudy conditions by dynamically changing brightness and contrast in the color space.

[0087] In step 2, the specific processing steps and mathematical flow of the LFD unit are as follows:

[0088] The unit first receives the initial input image from the previous layer. (The dimensions in this layer are) This data is then fed into a two-dimensional average pooling layer. The average pooling layer acts as a low-pass filter in the spatial domain, smoothly filtering out details and noise from local abrupt changes in the image and extracting low-frequency components representing the macroscopic contour information of large surfaces. :

[0089] ;

[0090] in This represents a two-dimensional average pooling operation;

[0091] Using the initial input image With the extracted low-frequency components Perform element-wise residual subtraction. This operation is equivalent to a high-pass filter operator in the spatial domain, which can completely truncate smooth backgrounds and generate high-precision high-frequency components for edges, textures, and abrupt defects. :

[0092] ;

[0093] Two sets of lightweight depthwise separable convolutions are used respectively. For low-frequency components and high frequency components Independent feature extraction is performed. To enable the network to adaptively adjust the importance of frequency bands based on different shooting distances, two learnable frequency band calibration weight parameters are introduced. and The extracted low-frequency structural features and high-frequency damage edge characteristics The formula for dynamic recalibration is as follows:

[0094] ;

[0095] ;

[0096] Low-frequency structural features after frequency recalibration High-frequency damage edge characteristics Direct concatenation (Concat) is performed along the channel dimension to enable deep cross-channel interaction between the two sets of features and to achieve channel dimensionality reduction and downsampling alignment. This ensures the concatenated feature matrix passes through a step size of 2. The standard pointwise convolutional layer, after batch normalization (BN) and SiLU activation, outputs the final frequency-decoupled feature map. :

[0097] ;

[0098] ;

[0099] After this processing, the size of the output feature map is adjusted to... It perfectly achieves lossless preservation of high-frequency minute features.

[0100] In step 2, the specific processing steps and mathematical flow of the SSA unit are as follows:

[0101] Decoupling feature map of upstream unit output Input feature map as the current state space selection enhancement unit Firstly, through a The saliency scoring network, constructed using convolutional layers, calculates the correlation score between each pixel grid point and the insulator body in the spatial domain, outputting a single-channel spatial feature response map. Then, it is flattened along the spatial dimension into a one-dimensional sequence score. :

[0102] ;

[0103] ;

[0104] in This indicates that the sequence can be flattened into a one-dimensional sequence.

[0105] Set sparse filtering ratio Calculate the number of high-value tokens that need to be retained at the current scale. A continuous Top-K relaxation selection strategy is used to filter out sequence scores. China ranked top The spatial index of the maximum value is assigned a value of 1.0 at the corresponding spatial grid point, while the value in the remaining non-candidate background regions is assigned a value of 0.0, thus constructing a spatial hard threshold mask. ;

[0106] To address the issues of gradient non-differentiability and gradient breakage during backpropagation caused by the discrete step characteristics of hard threshold masks, a straight-through estimator (STE) is introduced to reconstruct the computation graph, resulting in a sparse state mask with soft gradients. :

[0107] ;

[0108] in, This indicates direct locking during forward propagation. The 0 / 1 states are used to truncate background noise, and the hard thresholding operation is skipped during backpropagation, allowing the smooth gradient to flow fully towards differentiable values. Attention branching enables the selection of relaxed states;

[0109] The generated sparse state mask Input feature map of the state space selection enhancement unit Perform element-wise multiplication at the channel and spatial levels. At this point, large background features such as mountains, forests, and iron towers, whose scores are not Top-K, are forcibly cleared to zero due to multiplication by 0.0, resulting in masking features. Send the concealment features into Local feature extraction is performed using 3×3 channel-wise grouped convolutions, outputting local state features to characterize the morphology of defect edges. Meanwhile, to prevent the network from losing its global macroscopic view due to forcibly setting a large area of ​​background to zero, a parallel global context preservation channel was established. This utilizes a... The standard pointwise convolution operator selects the input feature map of the augmentation unit from the original state space. Perform global semantic mapping to extract global contextual features that contain the topological relationships of the entire graph. Then, the global context features are... With the local state features Perform matrix addition and feed the result into the cascaded array. The convolutional layer performs channel-dimensional interactions and reconstructions to generate preliminary filtered features for representing the new state. Finally, to prevent over-filtering from causing global context semantic breaks, a learnable global memory gating residual parameter is introduced. And combined with spatial characteristic response map We jointly construct an adaptive gating network to adaptively adjust the fusion ratio of suppressed background features and updated features, and output a denoised feature map after background suppression. :

[0110] ;

[0111] ;

[0112] ;

[0113] ;

[0114] ;

[0115] This processing completely blocks redundant background features that occupy a large area of ​​the image in the deep layers of the main structure, greatly reducing the computational overhead at the edges.

[0116] In step 3, the specific processing steps and mathematical flow of the GAM unit are as follows:

[0117] Denoising feature map output from upstream unit Input feature map as the current geometrically perceptive deformation unit The input feature map is fed into a lightweight parameter prediction branch reconstructed by a cascaded depthwise separable convolutional (DWConv) operator. Within this lightweight parameter prediction branch, the input feature map of the geometry-aware deformation unit is processed by two sets of parallel DWConv operators. Adaptive feature mapping at the channel and spatial dimensions is performed to predict the required two-dimensional spatial offset for subsequent deformable convolutions in parallel. With sampling point weight mask By replacing standard convolutions with DWConv, the number of parameters in the offset prediction branch is reduced by 85%.

[0118] ;

[0119] ;

[0120] in, and These represent lightweight, depthwise separable convolution operators used to predict the spatial bias tensor and the sampling weight tensor, respectively.

[0121] To address the elongated string-like physical structure of insulators under inspection perspective, an orthogonal direction-aware pooling decoupling mechanism is employed. Adaptive large-scale average pooling is performed in both the horizontal and vertical orthogonal directions to extract the axial large-scale long-range dependency features of the insulator string, thus obtaining the axial dependency features. Subsequently, through two layers of standards Convolution and nonlinear activation functions Regarding the axial dependence feature Channel dimension mapping and normalization are performed to generate spatial deformation attention weights. :

[0122] ;

[0123] ;

[0124] in, This indicates height-oriented compressed pooling. These represent width-direction compressed pooling, which together constitute the aforementioned direction-aware pooling mechanism.

[0125] Invoke the Deformable Convolution operator based on the predicted two-dimensional spatial offset. and sampling point weight mask Breaking away from the constraints of traditional square convolutional windows, this method performs multi-point adaptive sampling and reconstruction on the edges of irregularly shaped defects such as self-exploding edges and strip-shaped micro-cracks, outputting deformation modulation features. Simultaneously, a geometrical regularization constraint loss is added to the overall loss function. A sum-of-squares penalty constraint is applied to the spatial offset to limit its range of variation and prevent divergence. Finally, spatial deformation attention weights are incorporated. Feature fusion is performed to output a multi-scale fused feature map that achieves geometric alignment and feature wrapping of irregular defect edges. :

[0126] ;

[0127] ;

[0128] ;

[0129] This represents a deformable convolution operator.

[0130] During iterative training of the network model, the regularization loss generated by the built-in GAM units at each level is used. Multiply by the balance coefficient The results are directly added to the total loss function for synchronous backpropagation optimization.

[0131] In step 4, the key hyperparameter settings for the lightweight network model during training are as follows: total number of epochs is set to 300, batch size is set to 32, the AdamW optimizer with weight decay is used, the initial learning rate is set to 0.001, and the momentum parameter is set to 0.937. After the training cycle is completed, the model evaluation metrics are calculated using an independent test set to assess the overall performance of the improved model of this invention. The evaluation metrics include precision, recall, and average precision (mAP).

[0132] This invention also provides a specific implementation of a lightweight insulator defect detection and location system for power line drone inspections: In physical deployment, the system consists of two parts: an airborne edge computing terminal and a ground-based operation and maintenance control main station. At the airborne edge computing terminal, the high-definition camera mounted on the drone in the data acquisition module acquires inspection images containing insulators through a video acquisition chip, and directly inputs them into the airborne embedded computing chip (such as the NVIDIA Jetson series edge computing platform). The preprocessing module residing within the embedded chip first performs hardware preprocessing, scale normalization, and accelerated scaling on the image to generate the inspection image of the transmission line insulators to be inspected. Subsequently, the inspection image of the transmission line insulators to be inspected enters the defect detection core module. Because this invention uses grouped convolution of LFD units instead of standard convolution in its algorithm design, it performs spatial high- and low-frequency separation, uses SSA units to block complex background noise such as large areas of mountains, forests, and sky, and uses DWConv branch parameter compression of GAM units. This significantly reduces the floating-point operation overhead (FLOPs) of the model on the airborne edge, enabling real-time, low-latency carpet search of defects at a high frame rate of greater than 30 FPS on edge devices with limited computing power, and outputs multi-scale fused feature maps. After the defect detection marking and positioning output module cascaded to the backend completes the target decoding, the system not only renders a highlighted red box mark on the local image frame, but also extracts the relative pixel center coordinates of the defect target. Combined with the GPS / BeiDou three-dimensional spatial geographic coordinates of the UAV's current flight, the gimbal yaw angle and pitch angle data, and the geographic mapping geometric projection formula, it calculates the real-world absolute geographic coordinates (longitude, latitude, and altitude) of the defective insulator (such as spontaneous explosion, micro-cracks, and surface flashover). Ultimately, the data communication and ground operation and maintenance control master station module, via low-latency wireless data transmission radio or a 5G / 6G integrated air-space-ground cellular network, packages the defect marking image, defect confidence level, category, and absolute geographic coordinates into a power operation and maintenance standard protocol message, and transmits it back to the ground operation and maintenance control master station in real time via wireless network. Upon receiving the message, the ground master station triggers an audible and visual alarm and automatically dispatches the location information to the defect and hazard work order system, guiding operation and maintenance personnel to accurately eliminate defects. This achieves automated positioning and coordination of the entire system, from autonomous UAV inspection to closed-loop ground handling.

[0133] While the embodiments disclosed in this invention are as described above, the content is merely for the purpose of facilitating understanding of the invention and is not intended to limit the invention. Any person skilled in the art to which this invention pertains may make any modifications and changes to the form and details of the implementation without departing from the spirit and scope disclosed herein; however, the scope of patent protection of this invention shall still be determined by the scope defined in the appended claims.

Claims

1. A lightweight insulator defect detection method for power plant unmanned aerial vehicle (UAV) inspection, characterized in that, Includes the following steps: Step 1: Obtain inspection images containing insulators, preprocess and normalize them to construct an insulator defect dataset, and divide the dataset into a training set, a validation set and a test set; Step 2: Construct a lightweight insulator defect detection network based on YOLOv11n; the lightweight insulator defect detection network includes a backbone network, a neck network, and a detection head; wherein, the shallowest downsampling stage of the backbone network is equipped with implicit space-frequency decoupling (LFD) units, the shallow and middle layers of the backbone network are equipped with state-space selection enhancement (SSA) units, and the scale transformation downsampling nodes of the neck network are equipped with geometry-aware deformation (GAM) units; the network is iteratively trained based on the training set, and the performance of the trained network is verified using the test set to obtain the trained lightweight insulator defect detection network; Step 3: Obtain the inspection image of the insulator of the transmission line to be inspected, and input it into the trained lightweight insulator defect detection network for preliminary feature extraction to obtain initial image features; in the shallowest downsampling stage of the backbone network, the LFD unit processes the initial image features to extract the low-frequency structural features and high-frequency damage edge features of the insulator to obtain the decoupled feature map. The decoupled feature map is passed to the SSA units in the shallow and middle layers of the backbone network to filter out complex background noise in the image and obtain a denoised feature map. The denoised feature map is input into the geometrically aware deformation (GAM) unit of the scale transformation downsampling node of the neck network to perform geometric alignment and feature fusion of the defect edges, and output a multi-scale fused feature map. The multi-scale fused feature map is input into the detection head of the lightweight insulator defect detection network for target decoding, and the location bounding box and category prediction result of the insulator defect target are output, thereby completing the lightweight insulator defect detection of the transmission line.

2. The method according to claim 1, characterized in that, In step 3, the specific process of constructing the decoupled feature map by the LFD unit includes: performing average pooling on the initial image features to generate the low-frequency components of the initial image features; The initial image features and the low-frequency components are subjected to residual operation to generate corresponding high-frequency components. Lightweight grouped convolution is used to extract features from the low-frequency and high-frequency components to obtain the low-frequency structural features and high-frequency damage edge features of the insulator. The obtained low-frequency structural features and high-frequency damage edge features are dynamically recalibrated using learnable frequency band weights. Finally, the recalibrated features are aggregated by channel fusion to output a decoupled feature map that has extracted the low-frequency structural features and high-frequency damage edge features of the insulator.

3. The method according to claim 1, characterized in that, In step 3, the specific process of the SSA unit constructing the denoised feature map is as follows: a saliency scoring network is constructed to process the decoupled feature map and generate a feature response map; the feature response map is evaluated using a continuous Top-K relaxation selection strategy to generate a sparse state mask to solve the gradient breakage problem caused by the hard threshold mask. The sparse state mask is applied to the input features for masking processing to obtain masked features, and the state of the masked features is updated using state transition convolution to obtain preliminary screening features. Simultaneously, a global memory-gated residual compensation mechanism is introduced to adaptively learn the fusion weights between the initially screened features and the decoupled feature map; finally, a denoised feature map after background suppression is output.

4. The method according to claim 1, characterized in that, In step 3, the specific process of constructing a multi-scale fusion feature map by the GAM unit is as follows: a lightweight parameter prediction branch is constructed inside the GAM to generate the spatial offset and sampling point weight mask required for deformable convolution; orientation-aware pooling is used to extract the orientation information of the insulator string structure from the denoised feature map, and long-range dependencies are captured by decoupling along the horizontal and vertical orthogonal directions; spatial attention weights are generated according to the orientation information, and combined with the spatial offset and the sampling point weight mask, deformable convolution is used to complete the adaptive sampling of irregular defect regions in the denoised feature map, and the variation range of the spatial offset is limited by geometric regularization constraints; finally, a multi-scale fusion feature map that achieves geometric alignment and feature wrapping of irregular defect edges is output.

5. The method according to claim 1, characterized in that, The specific categories of insulator defects include: insulator breakage, microcracks in insulators, missing edges in insulators, spontaneous explosion of insulators, flashover of insulators, and surface pollution flashover of insulators, among other types of insulator defects.

6. A lightweight insulator defect detection system for power plant unmanned aerial vehicle (UAV) inspection, characterized in that, include: Data acquisition module: used to acquire inspection images of insulators on power transmission lines to be inspected using power drones; Preprocessing module: used to preprocess and scale normalize the inspection images acquired by the data acquisition module to generate images to be detected; The core module for defect detection includes a built-in lightweight insulator defect detection network, which is used to extract and fuse features from the image to be detected. Specifically, the network deploys LFD units in the shallowest downsampling stage of the backbone network to extract low-frequency structural and high-frequency edge features of the insulator; SSA units are deployed in the shallow and middle layers of the backbone network to filter background noise; and GAM units are deployed at the downsampling scale conversion node of the neck network to drive deformable convolutions to adaptively fit the defect geometry, ultimately outputting a multi-scale fused feature map. Defect detection marking and localization output module: cascaded at the back end of the defect detection core module, used to receive the multi-scale fused feature map and perform target decoding, output the location bounding box and category confidence of the insulator defect target, generate a bright defect mark and extract the center coordinates of the target pixel; at the same time, combined with the UAV spatial geographic coordinates, gimbal yaw angle and pitch angle data, the real-world absolute geographic coordinates of the defective insulator are calculated by geometric projection algorithm. Data communication and ground operation and maintenance control master station module: used to transmit standard messages containing defect marker images, defect confidence levels, categories and absolute geographic coordinates back to the ground operation and maintenance control master station via wireless network, trigger alarms and automatically generate hazard handling work orders.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the method according to any one of claims 1 to 5.

Images