A Multi-Source Fusion Intelligent Identification Method and System for Power Transmission Inspection Defects
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-09
- Publication Date
- 2026-08-14
AI Technical Summary
这些的存在使得在进行缺陷检测时回干扰缺陷特征的提取，导致虚假缺陷报警
[0009]利用知识蒸馏方式获取可见光异常分数图、红外异常分数图和紫外异常分数图，可以仅依靠正常样本进行训练，无需大量缺陷样本标注，且该模型轻量化、参数量小、推理速度快，可以适配无人机及边缘巡检设备。
Smart Images
Figure CN122574479A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent power line inspection technology, specifically to a multi-source integrated intelligent identification method and system for power transmission inspection defects. Background Technology
[0002] Transmission lines are a core component of the power system, and their safe and stable operation is directly related to the reliability of the power grid. With the continuous expansion of the transmission network, transmission lines are mostly distributed in complex terrains such as mountains and ravines, and are exposed to the outdoor natural environment all year round. They are prone to various defects such as broken conductor strands, damaged insulators, and corrosion of hardware, and require regular inspection.
[0003] With the development and advancement of science and technology, power transmission line inspection has gradually transformed from traditional manual inspection to intelligent inspection. The integration of artificial intelligence algorithms, such as deep learning, has enabled automatic defect identification, significantly improving inspection efficiency and reducing manual labor intensity and operational risks. However, misjudgments still exist in defect identification due to sources such as bird droppings, stains, shadows, and infrared noise. Specifically, bird droppings on the surfaces of insulators, conductors, and fittings of transmission lines can be easily confused with defects such as insulator damage and conductor corrosion due to their shape and color; dust, oil, and other stains adhering to equipment surfaces can also interfere with defect feature extraction during defect detection, leading to false defect alarms. Summary of the Invention
[0004] The purpose of this invention is to provide a multi-source fusion intelligent identification method and system for power transmission line inspection defects, which can effectively distinguish between power transmission line defects and interference items such as bird droppings, stains, shadows, and infrared noise, thereby improving the accuracy of defect identification.
[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A multi-source fusion intelligent identification method for power transmission inspection defects includes the following steps: Simultaneously acquire visible light, infrared, and ultraviolet images of power transmission line inspections and perform preprocessing; Based on preprocessed visible light, infrared, and ultraviolet images, knowledge distillation is used to obtain visible light anomaly score maps, infrared anomaly score maps, and ultraviolet anomaly score maps, respectively. The visible light anomaly score map, infrared anomaly score map, and ultraviolet anomaly score map are fused to generate a comprehensive anomaly score map, and Monte Carlo Dropout is used to output an uncertainty map associated with the comprehensive anomaly score map. The artifact source region was determined based on the comprehensive anomaly score map and uncertainty map; The artifact source area is re-inspected to identify any undetected abnormal areas, thereby determining the final power transmission inspection defects.
[0006] According to the above technical solution, the preprocessing includes pixel-level registration of visible light images, infrared images and ultraviolet images, and generating normalized grayscale images respectively.
[0007] During power transmission line inspections, visible light, infrared, and ultraviolet images are sequentially calibrated using camera intrinsic and extrinsic parameters, their corresponding feature points are extracted and matched, homography matrices are solved, projection transformations are performed, and sub-pixel interpolation resampling is applied to achieve pixel-level registration. Pixel-level registration ensures that the registered images correspond to the same physical inspection point on the power transmission line at the same coordinate position, resulting in spatial consistency among the visible light, infrared, and ultraviolet images.
[0008] According to the above technical solution, the knowledge distillation execution steps include: The visible light grayscale image / infrared grayscale image / ultraviolet grayscale image of normal samples were selected as the training set, and the visible light grayscale image / infrared grayscale image / ultraviolet grayscale image of normal samples and various abnormal samples were selected as the test set. Construct a knowledge distillation network, which includes a teacher network and a student network that is structurally symmetrical to the teacher network. Normal visible light grayscale images / infrared grayscale images / ultraviolet grayscale images from the training set are input into the teacher network to extract normal transmission line features. The normal transmission line features are input into the student network to output reconstructed features. The student network parameters are updated with the residual loss between the normal transmission line features of the teacher network and the reconstructed features of the student network as the total loss. The test set of visible light grayscale images / infrared grayscale images / ultraviolet grayscale images to be tested is input and passed through the teacher network and student network respectively to obtain the normal features and reconstructed features of the transmission line; the difference between the normal features and reconstructed features of the transmission line is calculated pixel by pixel to obtain the anomaly score map, the anomaly score map is upsampled to the size of the visible light image to be tested and fused to obtain the single-channel anomaly score map; the single-channel anomaly score map is normalized and Gaussian smoothed to obtain the preprocessed visible light anomaly score map / infrared anomaly score map / ultraviolet anomaly score map.
[0009] The model utilizes knowledge distillation to obtain visible light anomaly score maps, infrared anomaly score maps, and ultraviolet anomaly score maps. It can be trained using only normal samples without the need for a large number of defect sample annotations. Furthermore, the model is lightweight, has a small number of parameters, and fast inference speed, making it compatible with drones and edge inspection equipment.
[0010] Knowledge distillation can achieve pixel-level anomaly localization by reconstructing differences in teacher-student network features, enabling accurate defect identification. At the same time, it can effectively suppress interference from changes in lighting, complex backgrounds, and textured clutter, improving detection robustness and generalization ability.
[0011] According to the above technical solution, the preprocessed visible light anomaly score map, infrared anomaly score map and ultraviolet anomaly score map are adaptively weighted and fused using an attention mechanism to obtain a comprehensive anomaly score map. According to the above technical solution, Monte Carlo Dropout is used to output visible light uncertainty, infrared uncertainty map, and ultraviolet uncertainty during the knowledge distillation reasoning stage; the visible light uncertainty, infrared uncertainty map, and ultraviolet uncertainty are then fused to obtain the final uncertainty map. The fusion method can be weighted using methods such as weighting or attention mechanisms.
[0012] In this process, Monte Carlo Dropout is used to infer multiple times on the same image, scoring each pixel multiple times. The average score is taken as the true anomaly score for that pixel, and the standard deviation or difference is used to measure the reliability of the model's judgment for that pixel. Then, the uncertainties of all pixels are combined into an uncertainty map that is perfectly aligned in size and position with the anomaly score map, achieving pixel-by-pixel matching between anomaly scores and uncertainties. The pixel anomaly scores in the anomaly score map are then correlated and matched pixel uncertainties output by Monte Carlo Dropout pixel by pixel, with uncertainty representing the confidence level of the anomaly detection result for each pixel.
[0013] According to the above technical solution, the regions in the comprehensive anomaly scoring map that are greater than the anomaly scoring threshold and the uncertainty map that are greater than the uncertainty threshold are identified as artifact source regions. If a region exhibits anomalies in only a single mode or any two modes in the visible light, infrared, and ultraviolet anomaly fraction maps, then that region is determined to be an artifact source region.
[0014] In power transmission line inspection and monitoring, actual defects will show anomalies simultaneously in visible light, infrared, and ultraviolet images; bird droppings, stains, shadows, and infrared noise will only show anomalies in one or two of these images. Furthermore, due to the large fluctuations and unstable judgments in the anomaly scores output by the model during multiple random inferences in Monte Carlo Dropout, the statistically obtained fusion uncertainty is synchronously high, implying that the anomaly scoring is inaccurate.
[0015] According to the above technical solution, the artifact source region is locally cropped and its size is normalized by adaptive pooling before being input into the defect re-inspection model for re-inspection. The output is a texture anomaly score in the range of 0 to 1. If the texture anomaly score is greater than the texture threshold, the artifact source region is determined to be an anomaly region that was missed. The defect re-examination model includes a two- to three-layer small convolutional network, global average pooling, and Sigmoid output; The defect re-inspection model is trained using real fine texture fractures and gradient abrupt defect samples of transmission lines as positive samples and artifact region samples as negative samples, with cross-entropy loss used in the training.
[0016] By re-inspecting the artifact source area, the artifact areas identified in the initial screening can be further refined, effectively avoiding misjudging real weak defects such as microcracks and minor damage as artifacts such as bird droppings, stains, shadows, and infrared noise. This retains valid defect targets, eliminates false interference, and reduces the missed detection rate and false detection rate of power transmission line inspection. At the same time, the model structure is lightweight and has a small computational load, making it suitable for the real-time detection needs of inspection edge equipment.
[0017] This includes a technical solution: a multi-source fusion intelligent identification system for power transmission line inspection defects, comprising: The data acquisition module is used to simultaneously acquire visible light images, infrared images, and ultraviolet images of power transmission line inspections and perform preprocessing. The single-channel anomaly score map construction module uses knowledge distillation to obtain visible light anomaly score maps, infrared anomaly score maps, and ultraviolet anomaly score maps based on preprocessed visible light images, infrared images, and ultraviolet images, respectively. The comprehensive anomaly score map construction module is used to fuse the visible light anomaly score map, infrared anomaly score map, and ultraviolet anomaly score map to generate a comprehensive anomaly score map, and uses Monte Carlo Dropout to output an uncertainty map associated with the comprehensive anomaly score map; The artifact source region determination module determines the artifact source region based on the comprehensive anomaly score map and uncertainty map. The defect determination module is used to re-inspect the artifact source area to identify any missed abnormal areas in the artifact source area, and then determine the final power transmission inspection defects.
[0018] The invention includes a technical solution, an electronic device comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors execute a multi-source fusion intelligent identification method for power transmission inspection defects as described in any of the above technical solutions.
[0019] The invention includes a technical solution and a storage medium storing at least one instruction, which is loaded and executed by a processor to implement a multi-source fusion intelligent identification method for power transmission inspection defects as described in any of the above technical solutions.
[0020] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention identifies defects and draws anomaly score maps by analyzing visible light, infrared, and ultraviolet images from transmission line inspections. It then generates a comprehensive anomaly score map from these three anomaly score maps and uses Monte Carlo Dropout to output an uncertainty map associated with the comprehensive anomaly score map. Based on the comprehensive anomaly score map and the uncertainty map, it identifies artifact source regions and re-inspects these regions to determine the final defects. This invention quantifies the detection reliability of each anomaly region, solving the pain point of existing inspection methods that can only output defect identification results but cannot assess the reliability of the results. By jointly analyzing the comprehensive anomaly score map and the uncertainty map, artifact source regions are located, and re-inspection of these regions reduces the false detection rate in transmission line inspections, reduces ineffective maintenance costs, avoids unnecessary maintenance work caused by artifact misjudgments, and lowers power grid maintenance costs. This invention effectively avoids single-mode missed detections and false detections, significantly improving the accuracy and comprehensiveness of defect identification. Attached Figure Description
[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the steps of a multi-source fusion intelligent identification method for power transmission inspection defects according to the present invention; Figure 2 This is a flowchart for obtaining the comprehensive anomaly score chart. Detailed Implementation
[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0023] The example selected a 110kV high-voltage transmission line section with a length of 5km, including 30 towers, and covering key components such as conductors, insulators, vibration dampers, and spacers. The inspection environment included both sunny and cloudy scenarios, as well as interference factors such as partial light reflection and slight lens dirt, to simulate the complex environment in actual transmission line inspection.
[0024] A multimodal inspection drone is used, equipped with a high-definition visible light camera, an infrared thermal imager, and an ultraviolet imager. The drone flies at an altitude of 5-10m and a speed of 5m / s, flying parallel to the power transmission line conductor. It simultaneously collects visible light, infrared, and ultraviolet images. During the acquisition process, the drone's real-time positioning and attitude information is recorded to ensure accurate correspondence between the images and the power transmission line components.
[0025] The present invention employs a multi-source fusion intelligent identification method for power transmission line defects, combining visible light, infrared, and ultraviolet images, to identify defects in power transmission lines. The specific steps include: S1. Simultaneously acquire visible light, infrared, and ultraviolet images of the power transmission line inspection and perform preprocessing. Preprocessing includes noise reduction, etc. But the most important step is to perform pixel-level registration of the visible light, infrared, and ultraviolet images to generate normalized grayscale images respectively.
[0026] S2. Based on the preprocessed visible light image, infrared image, and ultraviolet image, knowledge distillation is used to obtain the visible light anomaly score map, infrared anomaly score map, and ultraviolet anomaly score map, respectively.
[0027] In the knowledge distillation process, ResNet101 is used as the backbone network for the teacher network, and MobileNetV2 is used as the backbone network for the student network.
[0028] During the training process, visible light grayscale images / infrared grayscale images / ultraviolet grayscale images of normal samples were selected as the training set, and visible light grayscale images / infrared grayscale images / ultraviolet grayscale images of normal samples and various abnormal samples were selected as the test set.
[0029] Normal visible light grayscale images / infrared grayscale images / ultraviolet grayscale images from the training set are input into the teacher network to extract normal transmission line features. The normal transmission line features are input into the student network to output reconstructed features. The student network parameters are updated with the residual loss between the normal transmission line features of the teacher network and the reconstructed features of the student network as the total loss. The test set of visible light grayscale images / infrared grayscale images / ultraviolet grayscale images to be tested is input and passed through the teacher network and student network respectively to obtain the normal features and reconstructed features of the transmission line; the difference between the normal features and reconstructed features of the transmission line is calculated pixel by pixel to obtain the anomaly score map, the anomaly score map is upsampled to the size of the visible light image to be tested and fused to obtain the single-channel anomaly score map; the single-channel anomaly score map is normalized and Gaussian smoothed to obtain the preprocessed visible light anomaly score map / infrared anomaly score map / ultraviolet anomaly score map.
[0030] The student model uses the AdamW optimizer. During training, the teacher model parameters are completely frozen, and it only performs forward inference, outputting soft labels and multi-scale feature maps, without participating in parameter updates.
[0031] Simultaneously, during the knowledge distillation inference stage, Monte Carlo Dropout is used to output visible light uncertainty, infrared uncertainty maps, and ultraviolet uncertainty. Specifically, Monte Carlo Dropout is used to repeatedly infer multiple times for the same image to be tested, scoring each pixel multiple times. The average value is used as the true anomaly score for that pixel, and the standard deviation or difference is used to measure the reliability of the model's judgment for that pixel. Then, the uncertainties of all pixels are combined into an uncertainty map that is perfectly aligned in size and position with the anomaly scoring map, achieving pixel-by-pixel matching between anomaly scores and uncertainties. The pixel anomaly scores in the anomaly scoring map are then correlated and matched pixel uncertainties output by Monte Carlo Dropout pixel by pixel, with uncertainty representing the confidence level of the anomaly detection result for each pixel.
[0032] S3. The visible light anomaly score map, infrared anomaly score map, and ultraviolet anomaly score map are unified in size, normalized, and then concatenated to construct a three-modal input feature tensor. After extracting cross-modal shallow anomaly features through basic convolution with an attention mechanism, channel attention is used to adaptively allocate the anomaly weights for each modality. After spatial attention is used to enhance the local defect region, the final comprehensive anomaly score map is output through convolutional mapping. The basic convolutional layer of the attention mechanism controls two 3×3 convolutions, normalization, and activation functions.
[0033] Simultaneously, the visible light uncertainty map, infrared uncertainty map, and ultraviolet uncertainty map are weighted and fused to obtain the final uncertainty map. Specific steps are as follows: Visible light grayscale images, infrared grayscale images, and ultraviolet grayscale images were input into the Monte Carlo Dropout model above, and 100 independent forward propagations were performed. During each propagation, the Dropout layer randomly discarded some neurons, resulting in 100 different abnormality score results.
[0034] The standard deviations of 100 anomaly scores were calculated for the visible light grayscale image, infrared grayscale image, and ultraviolet grayscale image, respectively. This standard deviation was used as the uncertainty of that pixel (between 0 and 0.5). A higher uncertainty indicates lower reliability of the detection result for that pixel and a greater susceptibility to interference factors. This resulted in visible light uncertainty, infrared uncertainty, and ultraviolet uncertainty images. The three images were then weighted to obtain the final uncertainty image. The weight for visible light uncertainty can be set to 0.4, for infrared uncertainty to 0.35, and for ultraviolet uncertainty to 0.23. The weights for these three uncertainty images can be adjusted according to the specific scenario.
[0035] S4. Regions in the comprehensive anomaly score map that are greater than the anomaly score threshold and in the uncertainty map that are greater than the uncertainty threshold are identified as artifact source regions.
[0036] The uncertainty of all pixels is visualized, with an uncertainty threshold of 0.3. Areas above this threshold are considered high-uncertainty areas, and areas below this threshold are considered low-uncertainty areas, forming a comprehensive anomaly score. Figure 1 A corresponding uncertainty diagram enables intuitive quantification of detection reliability.
[0037] If a region exhibits anomalies in only a single mode or any two modes in the visible light, infrared, and ultraviolet anomaly score maps, then that region is determined to be an artifact source region. For example, if region A has an anomaly score greater than the anomaly threshold B in the visible light anomaly score map, but is a normal region in the infrared and ultraviolet anomaly score maps, then region A is identified as an artifact source region.
[0038] S5. Construct a defect re-inspection model, which includes a two- to three-layer small convolutional network, global average pooling, and Sigmoid output. The defect re-inspection model is trained using real fine texture fractures and gradient abrupt defect samples of transmission lines as positive samples and artifact region samples as negative samples. The training adopts cross-entropy loss.
[0039] The artifact source region is locally cropped and its size is normalized by adaptive pooling before being input into the defect re-inspection model for re-inspection. The output is a texture anomaly score. If the texture anomaly score is greater than the texture threshold, the artifact source region is determined to be a missed anomaly region.
[0040] For any missed anomalies, the images and results from the re-inspection can be manually analyzed. By considering factors such as the actual structure of the transmission line components and the inspection environment, it can be determined whether the area represents a genuine defect. All non-artifact anomalies are then compiled into a final defect list, clearly identifying the location, type, and severity of each defect. This provides accurate data support for transmission line maintenance and completes the defect detection for this transmission line inspection.
[0041] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0042] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A multi-source fusion intelligent identification method for power transmission inspection defects, characterized in that, The steps include: Simultaneously acquire visible light, infrared, and ultraviolet images of power transmission line inspections and perform preprocessing; Based on preprocessed visible light, infrared, and ultraviolet images, knowledge distillation is used to obtain visible light anomaly score maps, infrared anomaly score maps, and ultraviolet anomaly score maps, respectively. The visible light anomaly score map, infrared anomaly score map, and ultraviolet anomaly score map are fused to generate a comprehensive anomaly score map, and Monte Carlo Dropout is used to output an uncertainty map associated with the comprehensive anomaly score map. The artifact source region was determined based on the comprehensive anomaly score map and uncertainty map; The artifact source area is re-inspected to identify any undetected abnormal areas, thereby determining the final power transmission inspection defects.
2. The intelligent identification method for power transmission inspection defects based on multi-source fusion as described in claim 1, characterized in that, The preprocessing includes pixel-level registration of the visible light image, infrared image, and ultraviolet image, and generating normalized grayscale images respectively.
3. The intelligent identification method for multi-source fusion defects in power transmission inspection according to claim 1, characterized in that, The distillation process using the aforementioned knowledge includes the following steps: The visible light grayscale image / infrared grayscale image / ultraviolet grayscale image of normal samples were selected as the training set, and the visible light grayscale image / infrared grayscale image / ultraviolet grayscale image of normal samples and various abnormal samples were selected as the test set. Construct a knowledge distillation network, which includes a teacher network and a student network that is structurally symmetrical to the teacher network. Normal visible light grayscale images / infrared grayscale images / ultraviolet grayscale images from the training set are input into the teacher network to extract normal transmission line features. The normal transmission line features are input into the student network to output reconstructed features. The student network parameters are updated with the residual loss between the normal transmission line features of the teacher network and the reconstructed features of the student network as the total loss. The test set of visible light grayscale images / infrared grayscale images / ultraviolet grayscale images to be tested is input and passed through the teacher network and student network respectively to obtain the normal features and reconstructed features of the transmission line; the difference between the normal features and reconstructed features of the transmission line is calculated pixel by pixel to obtain the anomaly score map, the anomaly score map is upsampled to the size of the visible light image to be tested and fused to obtain the single-channel anomaly score map; the single-channel anomaly score map is normalized and Gaussian smoothed to obtain the preprocessed visible light anomaly score map / infrared anomaly score map / ultraviolet anomaly score map.
4. The intelligent identification method for multi-source fusion defects in power transmission inspection according to claim 3, characterized in that, The preprocessed visible light anomaly score map, infrared anomaly score map, and ultraviolet anomaly score map are adaptively weighted and fused using an attention mechanism to obtain a comprehensive anomaly score map.
5. The intelligent identification method for multi-source fusion defects in power transmission inspection according to claim 1, characterized in that, During the knowledge distillation reasoning stage, Monte Carlo Dropout is enabled to output visible light uncertainty, infrared uncertainty map, and ultraviolet uncertainty; the visible light uncertainty, infrared uncertainty map, and ultraviolet uncertainty are fused to obtain the final uncertainty map.
6. The intelligent identification method for multi-source fusion defects in power transmission inspection according to claim 1, characterized in that, Regions in the comprehensive anomaly score map that are greater than the anomaly score threshold and in the uncertainty map that are greater than the uncertainty threshold are identified as artifact source regions. If a region exhibits anomalies in only a single mode or any two modes in the visible light, infrared, and ultraviolet anomaly fraction maps, then that region is determined to be an artifact source region.
7. The intelligent identification method for power transmission inspection defects based on multi-source fusion as described in claim 1, characterized in that, The artifact source region is locally cropped and its size is normalized by adaptive pooling before being input into the defect re-inspection model for re-inspection. The output is a texture anomaly score. If the texture anomaly score is greater than the texture threshold, the artifact source region is determined to be a missed anomaly region.
8. A multi-source fusion intelligent identification system for power transmission inspection defects, characterized in that, include: The data acquisition module is used to simultaneously acquire visible light images, infrared images, and ultraviolet images of power transmission line inspections and perform preprocessing. The single-channel anomaly score map construction module uses knowledge distillation to obtain visible light anomaly score maps, infrared anomaly score maps, and ultraviolet anomaly score maps based on preprocessed visible light images, infrared images, and ultraviolet images, respectively. The comprehensive anomaly score map construction module is used to fuse the visible light anomaly score map, infrared anomaly score map, and ultraviolet anomaly score map to generate a comprehensive anomaly score map, and uses Monte Carlo Dropout to output an uncertainty map associated with the comprehensive anomaly score map; The artifact source region determination module determines the artifact source region based on the comprehensive anomaly score map and uncertainty map. The defect determination module is used to re-inspect the artifact source area to identify any missed abnormal areas in the artifact source area, and then determine the final power transmission inspection defects.
9. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the method according to any one of claims 1-7.
10. A storage medium, characterized in that, The storage medium stores at least one instruction, which is loaded and executed by a processor to implement the method as described in any one of claims 1-7.