Power transmission line fault detection method and device

Abstract

The invention provides a power transmission line fault detection method and device, and relates to the technical field of computer vision, and the method comprises the steps: firstly obtaining a target image of a to-be-detected power transmission line; inputting the target image into a target detection model, and performing target detection to obtain a target detection result output by the target detection model; wherein the target detection model comprises a first sub-model, a second sub-model and a fusion module; the first sub-model is used for performing global detection on the target image to generate a global detection result; the second sub-model is used for performing local detection on the target image to generate a local detection result; the fusion module is used for performing weighted fusion on the global detection result and the local detection result to generate a target detection result; and finally obtaining a power transmission line fault detection result based on the target detection result. The power transmission line fault detection method provided by the invention realizes accurate and efficient detection of the power transmission line fault.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and particularly relates to a power transmission line fault detection method and device. BACKGROUND

[0002] The power transmission line is a line for transmitting electric energy from a power plant to a power load center and interconnecting the power load centers, has large transmission capacity, long transmission distance, high line voltage level, is a backbone network frame of a power grid, and is an important carrier for electric energy transmission.

[0003] For the power transmission line in the high altitude, it is difficult for the detection personnel to perform fault detection, therefore, the existing detection method mainly adopts unmanned aerial vehicle autonomous inspection, and an image acquisition system and an edge computing platform are carried on the unmanned aerial vehicle to acquire power transmission line images and perform detection analysis.

[0004] However, since the power transmission line in the high altitude is not only long in distance but also has many natural interferences (such as smoke, fog or flying birds), the quantity of the acquired images is large and the quality is poor, therefore, when the detection analysis is performed by using the model in the edge computing platform, the calculation amount is large, the delay is high, and the accuracy of the detection result is low, and missing detection or false detection is prone to occur. SUMMARY

[0005] The present application provides a power transmission line fault detection method and device, which solves the technical problem of low efficiency and accuracy of power transmission line fault detection in the prior art, and realizes accurate and efficient detection of power transmission line faults.

[0006] The present application provides a power transmission line fault detection method, which comprises the following steps: acquiring a target image of a power transmission line to be detected; inputting the target image into a target detection model to perform target detection and obtaining a target detection result output by the target detection model; wherein the target detection model comprises a first sub-model, a second sub-model and a fusion module; the first sub-model is used for performing global detection on the target image to generate a global detection result; the second sub-model is used for performing local detection on the target image to generate a local detection result; and the fusion module is used for performing weighted fusion on the global detection result and the local detection result to generate a target detection result; obtaining a power transmission line fault detection result based on the target detection result.

[0007] According to the power transmission line fault detection method provided by the present application, the first sub-model comprises an encoder, a local enhancement module and a global detection module; the encoder is used for performing multi-scale global feature extraction on the target image to generate a multi-scale global feature map; The local enhancement module is configured to perform regional inference on the global feature map to generate a local mask of a region to be enhanced. The global detection module is configured to generate a global detection result according to the global feature map.

[0008] According to the power transmission line fault detection method provided by the application, the second sub-model comprises a cropping module, a feature reconstruction sub-model and a local detection module. The cropping module is configured to crop a local image of a corresponding position from the target image according to the local mask. The feature reconstruction sub-model is configured to generate a detailed local enlarged feature map according to the local image. The local detection module is configured to generate a local detection result according to the local enlarged feature map.

[0009] According to the power transmission line fault detection method provided by the application, the training step of the feature reconstruction sub-model comprises: Obtaining historical power transmission line images; Inputting the historical power transmission line images into a feature extraction model to perform deep feature extraction and obtain an original feature map; Inputting a down-sampled image of the historical power transmission line images into a feature reconstruction sub-model to perform deep feature diffusion generation and obtain an up-sampled feature map; Taking the up-sampled feature map as feature data and taking the original feature map as label data, the feature reconstruction sub-model is trained.

[0010] According to the power transmission line fault detection method provided by the application, the fusion module is specifically configured to: Determine a weight of the global detection result based on corresponding coordinates of the local detection result and corresponding coordinates of the global detection result. Merge the local detection result and the global detection result based on the weight of the global detection result to generate the target detection result.

[0011] According to the power transmission line fault detection method provided by the application, the determination of the weight of the global detection result based on the corresponding coordinates of the local detection result and the corresponding coordinates of the global detection result comprises: In the case that there is an overlapping area between the corresponding coordinates of the local detection result and the corresponding coordinates of the global detection result, the larger the overlapping area is, the smaller the weight of the global detection result is. In the case that there is no overlapping area between the corresponding coordinates of the local detection result and the corresponding coordinates of the global detection result, the weight of the global detection result is a preset value.

[0012] The application further provides a power transmission line fault detection device, comprising the following modules. An image acquisition module is configured to acquire a target image of the power transmission line to be detected. An image detection module is configured to input the target image into a target detection model to perform target detection and obtain a target detection result output by the target detection model. A fault detection module is configured to obtain a power transmission line fault detection result based on the target detection result.

[0013] The application further provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements the power transmission line fault detection method according to any one of the above when executing the computer program.

[0014] The application further provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the power transmission line fault detection method according to any one of the above.

[0015] The application further provides a computer program product comprising a computer program, wherein the computer program is executed by a processor to implement the power transmission line fault detection method according to any one of the above.

[0016] The power transmission line fault detection method provided by the application first acquires a target image of the power transmission line to be detected, then inputs the target image into a target detection model to perform target detection and obtain a target detection result output by the target detection model, wherein the target detection model comprises a first sub-model, a second sub-model, and a fusion module; the first sub-model is configured to perform global detection on the target image to generate a global detection result; the second sub-model is configured to perform local detection on the target image to generate a local detection result; and the fusion module is configured to perform weighted fusion on the global detection result and the local detection result to generate the target detection result, so that the model can take into account both large-scale structures and small defects of the power transmission line through parallel detection of global and local paths, and the double-path parallel processing mode reduces repeated calculation and improves calculation efficiency; the logical consistency and optimal integration of the double-path detection results are realized through weighted fusion, and the detection accuracy is improved; finally, a power transmission line fault detection result is obtained based on the target detection result, and accurate and efficient detection of the power transmission line fault is realized. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to make the technical solutions in the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort.

[0018] Figure 1 is a flowchart of a power line fault detection method provided by the present application.

[0019] Figure 2 is a flowchart of a target detection model in the training stage and the inference stage provided by the present application.

[0020] Figure 3 is a structural diagram of a power line fault detection device provided by the present application.

[0021] Figure 4 is a structural diagram of an electronic device provided by the present application. DETAILED DESCRIPTION

[0022] In order to make the technical solutions in the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort based on the embodiments in the present application.

[0023] Since there are both large-scale structures (such as tower) and micro-fault features (such as insulator crack) in the power line, and the imaging distance is far or the sensor resolution is limited when the unmanned aerial vehicle is shooting in the high altitude, the visual information (such as texture, structure) carried by the small target in the image itself is extremely sparse, which makes it difficult for the standard convolutional neural network detector to form a robust and discriminative representation in the feature extraction stage, resulting in serious missed detection and false detection, and thus the fault detection accuracy of the power line is low.

[0024] To cope with the above challenges, the prior art includes the following two technical paths: (1) Data domain enhancement path (Image-Domain Enhancement): This kind of method pre-processes the input data before detection. The typical scheme is to use a generative super-resolution model (such as SRGAN, ESRGAN or SR3 based on diffusion model, etc.) to perform pixel-level reconstruction on the image region containing small targets, in order to generate high-resolution (HR) images with high-frequency details, and then input them into the standard detector.

[0025] (2) Divide-and-Conquer with Zoom-In: This approach adopts a two-stage strategy. In the first stage, a small target aggregation region of interest (RoI) is identified by an auxiliary module (such as a density map prediction network, a clustering proposal network, or a detector that redefines classes). In the second stage, the RoI is cropped from the original high-resolution image and upsampled (usually bilinear interpolation), and then the enlarged image block is submitted to the detector for fine detection.

[0026] Although the above technical paths improve the detection performance of small targets under certain conditions, their inherent design flaws limit their application value in real-time scenarios: (1) Inevitable computational redundancy and inference delay: Both of the above two paths introduce a cascaded process of "pixel domain transformation -> feature domain recoding". Whether it is a complex generative reconstruction or a simple interpolation enlargement, the output pixel domain data must be re-fed into the backbone encoder of the detection model for full-path feature extraction from the bottom to the top. The backbone network is usually the most computationally intensive part of the model (such as ResNet, Swin Transformer), and repeated execution of this process leads to a multiple-level increase in total computational cost (FLOPs) and inference delay (Latency), which is contrary to the demand for real-time processing.

[0027] (2) Non-uniformity of optimization objectives and suboptimal solution problems: The data domain enhancement path has an objective mismatch problem. The training objective of the super-resolution model is to minimize the pixel-level or perceptual-level reconstruction error (such as L1 Loss, Perceptual Loss), aiming to improve the visual fidelity of the image. However, an image that is visually better does not necessarily contain more beneficial semantic features for the detection task. This separate optimization may lead to the convergence of the entire system to a suboptimal point, thereby reducing detection accuracy.

[0028] The present application provides a power line fault detection method, aiming to solve the inherent computational redundancy and optimization objective inconsistency problems in existing small target detection enhancement techniques for power line images. By constructing a framework that directly compensates information in the feature space, the present application avoids pixel domain reconstruction and repeated forward propagation of the backbone network, thereby effectively improving the detection accuracy of small targets in power line images while significantly reducing the computational complexity and inference delay of the system, improving computational efficiency, and enabling real-time deployment on edge computing platforms such as unmanned aerial vehicles.

[0029] The following will be combined Figures 1 to 4A power line fault detection method and device are described.

[0030] Figure 1 A flowchart of a power line fault detection method provided by the present application is shown in Figure 1 The method comprises the following steps: Step 101, obtaining a target image of a power line to be detected.

[0031] Step 102, inputting the target image into a target detection model to perform target detection and obtaining a target detection result output by the target detection model; wherein the target detection model comprises a first sub-model, a second sub-model and a fusion module; the first sub-model is used for global detection of the target image to generate a global detection result; the second sub-model is used for local detection of the target image to generate a local detection result; and the fusion module is used for weighted fusion of the global detection result and the local detection result to generate the target detection result.

[0032] Step 103, obtaining a power line fault detection result based on the target detection result.

[0033] Specifically, for the power line in the air, the high-definition camera carried by the unmanned aerial vehicle is used to collect a high-definition video stream or image of the power line as a target image in real time in the form of autonomous inspection of the unmanned aerial vehicle, and the power line is detected for faults. The high-definition video stream can be down-sampled to 1080p, and each frame of image is input into the model as a model input. The high-definition image can be pre-processed (noise reduction, enhancement, etc.) and then input into the model.

[0034] Since there are both large-scale structures (such as tower) and small fault features (such as insulator cracks) in the power line, and there may be many natural interferences (such as smoke, fog or flying birds, etc.) in the data, the existing detection model is difficult to balance the detection accuracy and the calculation efficiency.

[0035] Therefore, the present application designs a target detection model for high-precision detection of various small targets in the power line image. The first sub-model in the target detection model is used for global detection to detect large structures of the power line, the second sub-model in the target detection model is used for local detection to focus on small fault features of the power line, the double-path parallel detection is used to reduce repeated calculation, the calculation amount is reduced, the calculation speed is improved, and then the double-path detection results are weighted and fused to obtain the final target detection result. The target detection model can capture both large-scale structures and small defect features of the power line, cope with various power line images taken at different distances, angles or environments, and improve the adaptability of the model.

[0036] According to the bounding box and confidence data presented in the target detection result, a power line fault detection result is obtained. The detection result includes determining whether there is a fault in the power line, and in the case of a fault, determining the fault location according to the bounding box positioning, and analyzing the fault type according to the confidence data.

[0037] Figure 2 is a flowchart of the training and inference stages of a target detection model provided by the present application, as shown in Figure 2 The target detection model is an asymmetric teacher-student network model. The teacher network (first sub-model) provides high-quality features as supervision, and the student network (second sub-model) learns a shortcut from low-quality input to high-quality features.

[0038] The first sub-model can be constructed based on any advanced encoder-decoder structure target detector, such as DETR (Detection Transformer) or RT-DETR (Real-Time Detection Transformer). The first sub-model integrates a local enhancement module in parallel outside the detection head (encoder) of the target detector, which is used for regional inference on the global feature map to generate a local mask for the region to be enhanced, and is jointly constructed with the global detection module.

[0039] The second sub-model can be a lightweight generative network, i.e., a local feature enhancement network (LFEN), which is used to generate high-quality, high-dimensional semantic features corresponding to low-resolution image blocks as conditional input. The second sub-model can be constructed by adding necessary input and output modules to the Conditional Diffusion Model architecture. The Conditional Diffusion Model has advantages in learning complex data distribution and generating high-quality samples, and can better generate high-quality, high-dimensional semantic features.

[0040] In the training stage, the first sub-model can automatically generate training targets for the local enhancement module in a self-supervised or weakly supervised manner. For example, the aggregated regions of real targets in the training set that are smaller than a certain threshold (such as 32x32 pixels) or have a detection confidence lower than a certain threshold in the initial training stage of the model can be generated by morphological operations (such as dilation and closure) to generate a binary mask as the supervision truth value for model training. The second sub-model can adopt a supervised learning manner, take a high-quality image block corresponding to a supervised true value area of the first sub-model as feature data, input any fixed and pre-trained teacher network (which can be a weight copy of the first sub-model or a more powerful external model), extract a feature map at a specific depth thereof as label data of the second sub-model, and perform supervised training.

[0041] The embodiment of the application integrates the loss functions of the first sub-model and the second sub-model into one unified loss function The first sub-model and the second sub-model are jointly trained in an end-to-end manner, and the loss function is which is a weighted sum of multiple task losses: In the formula, is a standard detection loss (such as a sum of a classification loss and a regression loss); is a mask segmentation loss (such as a Dice Loss or a Focal Loss) of a local enhancement module in the first sub-model; is a feature reconstruction loss of the second sub-model, which can be further refined as a combination of an L1 / L2 loss and an adversarial loss to ensure the accuracy and authenticity of the reconstructed features. represents a detection loss weight; represents a detection result of the first sub-model; represents a training label of the first sub-model; represents a mask segmentation loss weight; represents a predicted mask; represents a pseudo-label mask; represents a feature reconstruction loss weight; represents a reconstructed feature of the second sub-model; represents a generated feature of the first sub-model.

[0042] The embodiment of the application integrates the positioning of the difficult area and the final detection task in a single model and optimizes them through a joint loss function, ensures that the behavior of the local enhancement module in the first sub-model serves to improve the overall detection performance of the model, overcomes the suboptimal coupling problem between the modules in the separate scheme, and realizes a higher degree of task alignment.

[0043] The application provides a power transmission line fault detection method, which comprises the following steps: first, acquiring a target image of a power transmission line to be detected; then, inputting the target image into a target detection model to perform target detection and obtain a target detection result output by the target detection model; wherein the target detection model comprises a first sub-model, a second sub-model and a fusion module; the first sub-model is used for performing global detection on the target image to generate a global detection result; the second sub-model is used for performing local detection on the target image to generate a local detection result; and the fusion module is used for performing weighted fusion on the global detection result and the local detection result to generate the target detection result, so that the model can take into account both the large-scale structure and the small defects of the power transmission line through parallel detection of the global path and the local path, the double-path parallel processing mode is used to reduce repeated calculation and improve the calculation efficiency, the logical consistency and the optimized integration of the double-path detection results are realized through weighted fusion, and the detection accuracy is improved; finally, based on the target detection result, a power transmission line fault detection result is obtained to realize accurate and efficient detection of the power transmission line fault.

[0044] Further, the first sub-model comprises an encoder, a local enhancement module and a global detection module. The encoder is used for performing multi-scale global feature extraction on the target image to generate a multi-scale global feature map. The local enhancement module is used for performing regional reasoning on the global feature map to generate a local mask of the region to be enhanced. The global detection module is used for generating a global detection result according to the global feature map.

[0045] Specifically, the first sub-model comprises an encoder, a local enhancement module and a global detection module.

[0046] The encoder can be an encoder in DETR or RT-DETR, which is used for performing multi-scale feature extraction on the target image according to the pyramid feature representation theory of the convolutional neural network to generate a multi-scale global feature map. The local enhancement module can adopt a lightweight region for refinement proposal module (RPM) or a UNet model to perform regional reasoning on the multi-scale global feature map, identify whether there is a region to be enhanced, activate only when necessary, and generate one or more local masks or candidate boxes of the region to be enhanced. The global detection module integrates multi-scale features through a global detection head (such as an anchor-free detection head or a DETR detection head) according to the multi-scale global feature map to generate a set of global detection results containing a bounding box and a confidence.

[0047] The embodiment of the present application extracts multi-scale features of the target image by constructing a first sub-model, so that the model can process targets of different sizes at the same time, adapt to the complex environment in the power transmission line scene, and improve the robustness of model detection. At the same time, the local enhancement module in the first sub-model performs regional reasoning on the multi-scale global feature map to identify whether there is a region to be enhanced, and only activates when necessary, avoiding indiscriminate enhancement processing and feature recoding of the whole image. Specifically, the core computing unit (encoder) is only executed once in each reasoning, and only when a difficult region is detected, an additional but low-cost local enhancement module is activated. This "on-demand computing" mechanism reduces the average reasoning delay of the model by orders of magnitude compared to existing enhancement schemes, improves the response speed, and provides feasibility for real-time deployment on resource-limited platforms such as edge computing platforms deployed on unmanned aerial vehicles.

[0048] Further, the second sub-model includes a cropping module, a feature reconstruction sub-model, and a local detection module. The cropping module is configured to crop a local image corresponding to a position from the target image according to the local mask. The feature reconstruction sub-model is configured to generate a detailed local enlarged feature map according to the local image. The local detection module is configured to generate a local detection result according to the local enlarged feature map.

[0049] Specifically, since direct detection on the whole image often cannot accurately detect small targets on the power transmission line, such as small defect features of insulator cracks or natural interference features such as smoke, fog, or flying birds, the second sub-model designed by the present application adopts a strategy of positioning first and then enlarging to improve detection performance In the case where the local enhancement module in the first sub-model is activated, first, the corresponding local image is accurately positioned and cropped according to the local mask output by the local enhancement module; then, through the feature reconstruction sub-model, a set of high-quality, semantically enhanced detailed local enlarged feature maps are generated according to the low-resolution local image using a generative method; finally, through the local detection module, a set of local detection results containing bounding boxes and confidence are generated according to the local enlarged feature map. The local detection module can be a dedicated small target detection head, such as the detection head in the YOLO or TinyDet series of models.

[0050] The embodiment of the present application optimizes the small target detection performance by designing a local detection path, and restores the details lost by downsampling or blurring through generative reconstruction, so that the model can capture the tiny defect features in the power transmission line, improve the detection accuracy, and at the same time, only activate the calculation when the small target region to be enhanced is identified, avoid indiscriminate enhancement processing and feature recoding on the whole image, reduce the calculation amount, reduce the delay, and improve the detection efficiency.

[0051] Further, the training step of the feature reconstruction sub-model comprises: acquiring historical power transmission line images; inputting the historical power transmission line images into a feature extraction model for deep feature extraction to obtain original feature maps; inputting the downsampled images of the historical power transmission line images into the feature reconstruction sub-model for deep feature diffusion generation to obtain upsampled feature maps; using the upsampled feature maps as feature data and the original feature maps as label data to train the feature reconstruction sub-model.

[0052] Specifically, the feature reconstruction sub-model is the core part of the second sub-model, and is preferably a conditional diffusion model.

[0053] The training of the model adopts a knowledge distillation framework. The theoretical basis is that the features extracted by a deep network contain richer semantic information and are suitable for being the reconstruction target. This way ensures that the reconstructed features are highly relevant to the detection task, rather than simply pursuing pixel-level similarity.

[0054] The training process specifically comprises: acquiring historical power transmission line images to construct a historical power transmission line image dataset; using a pre-trained ResNet-101 or a feature extraction model such as the first sub-model as a feature extraction model (teacher network) to extract deep features and obtain original feature maps; inputting the downsampled images of the historical power transmission line images into the feature reconstruction sub-model for deep feature diffusion generation to obtain upsampled feature maps; using the upsampled feature maps as feature data and the original feature maps extracted by the teacher network as label data to train the student network (conditional diffusion model).

[0055] The embodiment of the application ensures that the feature reconstruction process is task-driven by using the internal features of the detection model as a supervision target, and the reconstructed features are adapted to the downstream detection task, thereby maximizing the performance gain of the end-to-end, enabling the feature reconstruction sub-model to generate high-quality and task-related feature maps, thereby improving the local detection accuracy and realizing accurate extraction of the micro defect features in the transmission line image.

[0056] Further, the fusion module is specifically used for: determining a weight of the global detection result based on the corresponding coordinates of the local detection result and the corresponding coordinates of the global detection result; merging the local detection result and the global detection result based on the weight of the global detection result to generate the target detection result.

[0057] Further, the determination of the weight of the global detection result based on the corresponding coordinates of the local detection result and the corresponding coordinates of the global detection result comprises: in the case that there is an overlapping region between the corresponding coordinates of the local detection result and the corresponding coordinates of the global detection result, the larger the overlapping region is, the smaller the weight of the global detection result is; in the case that there is no overlapping region between the corresponding coordinates of the local detection result and the corresponding coordinates of the global detection result, the weight of the global detection result is a preset value.

[0058] Specifically, the fusion module adopts a dynamic weighting strategy based on spatial relationship to weight and fuse the global and local detection results.

[0059] In the overlapping region, the local detection result obtained by local enhancement has higher accuracy, so the weight of the global detection result is inhibited; and in the non-overlapping region, the global detection is relied on, and the original prediction of the global detection head is reserved.

[0060] Specifically, the embodiment of the application adopts a Spatially-Prioritized Weighted Non-Maximum Suppression (SPW-NMS) algorithm based on spatial region priority to realize efficient and conflict-free merging of the global detection result and the local detection result. The core mechanism of the algorithm is that, instead of performing standard SPW-NMS processing on all detection boxes indiscriminately, the detection boxes are first grouped and prioritized according to their spatial source, i.e., whether they are located in the region to be enhanced defined by the local enhancement module. The algorithm flow is as follows: Firstly, all the bounding boxes from the local detection path (whose coordinates have been mapped back to the target image coordinate system) are marked as high priority; all the global detection boxes that do not overlap with any region to be enhanced in space are marked as regular priority (the weight remains the preset value); for the global detection boxes that overlap with the region to be enhanced, the confidence thereof is punished by a preset attenuation factor , the weight of the global detection result is dynamically adjusted according to the degree of overlap to reduce its competitiveness; finally, all the detection boxes that have been divided by priority and adjusted in confidence are merged into a unified list, and a standard non-maximum suppression (NMS) or its variant (such as Soft-NMS) processing is performed on the list.

[0061] The embodiment of the application ensures that, in a difficult area with small targets, the high-quality local detection result generated by the “small target expert” (i.e., the local detection head) has higher reservation priority through the SPW-NMS algorithm, effectively solving the potential conflict from the two paths; and in a regular area, the original prediction of the global detection head is reserved. Through the intelligent arbitration mechanism based on spatial priori, a fine weight allocation strategy is determined, which not only avoids the problem that high-quality detection results may be incorrectly suppressed due to simple merging, but also realizes the logical consistency and optimal integration of the dual-path detection results, improves the detection accuracy, and at the same time enables the model to adapt to the power transmission line fault detection requirements in different environments, improving the flexibility.

[0062] Based on the above embodiment, a typical embodiment of the application is applied to a UAV autonomous inspection system for high-voltage power transmission lines. In the system, the UAV is equipped with an edge computing platform embedded with an NVIDIA Jetson AGX Orin, and the target detection model provided by the application is deployed. The shared encoder is based on the RT-DETR-L architecture and is fine-tuned for the field adaptation of power transmission line components; the integrated local enhancement module is responsible for identifying small target aggregation areas; and the local feature enhancement network (LFEN) adopts a lightweight conditional diffusion model.

[0063] In the inspection task, the high-definition video stream collected by the UAV in real time is down-sampled to 1080p as the system input. For each image, the system first generates a global multi-scale feature map through the shared encoder. Based on the feature, the global detection head outputs a set of preliminary global detection results, and the local enhancement module identifies the region to be enhanced containing small components such as insulator strings and shock absorbers.

[0064] ​If the region to be enhanced is detected, the system activates the local enhancement path in parallel: the corresponding image block is cropped from the original image, and the high-quality semantic features thereof are reconstructed by the LFEN efficiently (through a 15-step accelerated reverse sampling); then a set of high-precision local detection results are output by the independent local detection head; finally, the system uses a spatial region priority-based weighted NMS algorithm to output the final target detection result by fusing the local detection results after coordinate system transformation and the original global detection results, and generates a fault detection report based on the target detection result.

[0065] The experimental results show that the above system can run in real time at a speed of about 32 FPS, and the recall rate of micro pin missing and other defects with a size of less than 32x32 pixels reaches 85.2%, which is more than 26 percentage points higher than the baseline scheme in the prior art using only the global path, and the average inference delay is about 31 milliseconds, which is significantly lower than the traditional two-stage scheme (usually more than 150 milliseconds) that needs to be repeatedly encoded, fully verifying that the present application has a significant advantage in computing efficiency while maintaining high accuracy, meeting the stringent requirements of industrial real-time and high-precision inspection.

[0066] Next, a power line fault detection device provided by the present application is described. The power line fault detection device described below can be correspondingly referred to the power line fault detection method described above.

[0067] Based on any of the above embodiments, Figure 3 is a structural schematic diagram of a power line fault detection device provided by the present application, as Figure 3 shown. The embodiment of the present application provides a power line fault detection device, which comprises an image acquisition module 301, an image detection module 302 and a fault detection module 303, wherein: The image acquisition module 301 is configured to acquire a target image of a power line to be detected; the image detection module 302 is configured to input the target image into a target detection model to perform target detection and obtain a target detection result output by the target detection model; wherein the target detection model comprises a first sub-model, a second sub-model and a fusion module; the first sub-model is configured to perform global detection on the target image to generate a global detection result; the second sub-model is configured to perform local detection on the target image to generate a local detection result; and the fusion module is configured to perform weighted fusion on the global detection result and the local detection result to generate a target detection result; and the fault detection module 303 is configured to obtain a power line fault detection result based on the target detection result.

[0068] The application provides a power transmission line fault detection device. First, a target image of a power transmission line to be detected is acquired. Then, the target image is input into a target detection model for target detection to obtain a target detection result output by the target detection model. The target detection model comprises a first sub-model, a second sub-model and a fusion module. The first sub-model is used for global detection of the target image to generate a global detection result. The second sub-model is used for local detection of the target image to generate a local detection result. The fusion module is used for weighted fusion of the global detection result and the local detection result to generate the target detection result. Thus, the model can take into account large-scale structures and small defects of the power transmission line through parallel detection of global and local paths, and the double-path parallel processing mode reduces repeated calculation, improves calculation efficiency, realizes logical consistency and optimized integration of the double-path detection results through weighted fusion, and improves detection accuracy. Finally, a power transmission line fault detection result is obtained based on the target detection result, and accurate and efficient detection of the power transmission line fault is realized.

[0069] Figure 4 An example of a schematic diagram of a physical structure of an electronic device is shown in Figure 4 The electronic device can include a processor 410, a communications interface 420, a memory 430 and a communications bus 440, wherein the processor 410, the communications interface 420 and the memory 430 communicate with each other through the communications bus 440. The processor 410 can invoke logical instructions in the memory 430 to execute a power transmission line fault detection method, which comprises: Acquiring a target image of a power transmission line to be detected; Inputting the target image into a target detection model for target detection to obtain a target detection result output by the target detection model. The target detection model comprises a first sub-model, a second sub-model and a fusion module. The first sub-model is used for global detection of the target image to generate a global detection result. The second sub-model is used for local detection of the target image to generate a local detection result. The fusion module is used for weighted fusion of the global detection result and the local detection result to generate the target detection result. Obtaining a power transmission line fault detection result based on the target detection result.

[0070] In addition, the logic instructions in the memory 430 described above can be implemented in the form of software functional units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0071] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the power line fault detection method provided by the above-mentioned methods, the method comprising: obtaining a target image of a power line to be detected; inputting the target image into a target detection model to perform target detection and obtain a target detection result output by the target detection model; wherein the target detection model comprises a first sub-model, a second sub-model and a fusion module; the first sub-model is used for global detection of the target image to generate a global detection result; the second sub-model is used for local detection of the target image to generate a local detection result; and the fusion module is used for weighted fusion of the global detection result and the local detection result to generate a target detection result; obtaining a power line fault detection result based on the target detection result.

[0072] In another aspect, the present application also provides a non-transitory computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the power line fault detection method provided by the above-mentioned methods, the method comprising: obtaining a target image of a power line to be detected; input the target image into a target detection model to perform target detection, and obtain a target detection result output by the target detection model; wherein the target detection model comprises a first sub-model, a second sub-model, and a fusion module; the first sub-model is configured to perform global detection on the target image to generate a global detection result; the second sub-model is configured to perform local detection on the target image to generate a local detection result; and the fusion module is configured to perform weighted fusion on the global detection result and the local detection result to generate the target detection result. Based on the target detection result, a power transmission line fault detection result is obtained.

[0073] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment scheme according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0074] Through the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software and the necessary general hardware platform, and of course, can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0075] It should be noted that, as used in this document, the terms "includes," "including," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements is not required to only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises a... " does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element. Additionally, it should be noted that the scope of the methods and apparatus of the present embodiments are not limited by the order of the steps or the order of the sequence, and can include performing the steps in different order than described, and can also include adding, omitting, or combining various steps, and features described with respect to some examples can be combined in other examples.

[0076] It should also be noted that the terms "first", "second", and the like, are used herein to distinguish one element from another, and do not necessarily have an ordinal or chronological significance. It is to be understood that such terms are used interchangeably, and that the embodiments of the present application can operate in other sequences than those described or illustrated herein, and that the embodiments of the present application can be adapted, to operate in other sequences than those described or illustrated herein, and that the objects distinguished by "first", "second", etc. are generally a class, and the number of objects is not limited, for example, the first object can be one or more.

[0077] The "determining B based on A" in the embodiments of the present application means that A is considered as a factor when determining B. It is not limited to "determining B based on A only", but also includes "determining B based on A and C", "determining B based on A, C and E", "determining C based on A, and determining B based on C further", etc. In addition, it can also include A as a condition for determining B, for example, "when A meets the first condition, determining B using the first method"; for example, "when A meets the second condition, determining B"; for example, "when A meets the third condition, determining B based on the first parameter"; etc. Of course, A can also be a condition for determining B as a factor, for example, "when A meets the first condition, determining C using the first method, and further determining B based on C", etc.

[0078] The term "a plurality of" in the present application means two or more, and other quantifiers are similar.

[0079] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit the same; and although the present application has been described in detail with reference to the foregoing embodiments, it should be appreciated by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features thereof can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting faults in power transmission lines, characterized in that, include: Acquire the target image of the power transmission line to be tested; The target image is input into a target detection model for target detection, and the target detection result output by the target detection model is obtained. The target detection model includes a first sub-model, a second sub-model, and a fusion module. The first sub-model performs global detection on the target image and generates a global detection result. The second sub-model performs local detection on the target image and generates a local detection result. The fusion module performs weighted fusion of the global detection result and the local detection result to generate the target detection result. Based on the target detection results, the power transmission line fault detection results are obtained.

2. The method for detecting transmission line faults according to claim 1, characterized in that, The first sub-model includes an encoder, a local enhancement module, and a global detection module; The encoder is used to extract global features at multiple scales from the target image and generate a global feature map at multiple scales. The local enhancement module is used to perform region reasoning on the global feature map and generate a local mask for the region to be enhanced. The global detection module is used to generate global detection results based on the global feature map.

3. The method for detecting transmission line faults according to claim 2, characterized in that, The second sub-model includes a cropping module, a feature reconstruction sub-model, and a local detection module; The cropping module is used to crop out a local image at a corresponding position from the target image based on the local mask; The feature reconstruction sub-model is used to generate a detailed magnified local feature map based on the local image; The local detection module is used to generate local detection results based on the magnified local feature map.

4. The method for detecting transmission line faults according to claim 3, characterized in that, The training steps of the feature reconstruction sub-model include: Obtain historical images of power transmission lines; The historical transmission line images are input into a feature extraction model for deep feature extraction to obtain the original feature map. The downsampled image of the historical transmission line image is input into the feature reconstruction sub-model to generate deep feature diffusion and obtain an upsampled feature map. The feature reconstruction sub-model is trained using the upsampled feature map as feature data and the original feature map as label data.

5. The method for detecting transmission line faults according to claim 3, characterized in that, The fusion module is specifically used for: The weight of the global detection result is determined based on the corresponding coordinates of the local detection result and the corresponding coordinates of the global detection result. Based on the weights of the global detection results, the local detection results are merged with the global detection results to generate the target detection results.

6. The method for detecting transmission line faults according to claim 5, characterized in that, The step of determining the weight of the global detection result based on the corresponding coordinates of the local detection result and the corresponding coordinates of the global detection result includes: When there is an overlap between the coordinates of the local detection result and the coordinates of the global detection result, the larger the overlap area, the smaller the weight of the global detection result. If there is no overlap between the coordinates of the local detection result and the coordinates of the global detection result, the weight of the global detection result is a preset value.

7. A power transmission line fault detection device, characterized in that, include: The image acquisition module is used to acquire the target image of the power transmission line under test. An image detection module is used to input the target image into a target detection model for target detection and obtain the target detection result output by the target detection model. The target detection model includes a first sub-model, a second sub-model, and a fusion module. The first sub-model performs global detection on the target image and generates a global detection result. The second sub-model performs local detection on the target image and generates a local detection result. The fusion module performs weighted fusion of the global detection result and the local detection result to generate the target detection result. The fault detection module is used to obtain the fault detection result of the transmission line based on the target detection result.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the power transmission line fault detection method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the power transmission line fault detection method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the power transmission line fault detection method as described in any one of claims 1 to 6.

Images