Unmanned aerial vehicle-based high-voltage line area dangerous behavior identification method and system

Abstract

The invention discloses a high-voltage line area dangerous behavior identification method and system based on an unmanned aerial vehicle, and belongs to the technical field of electric power inspection safety monitoring, and the method comprises the steps: fusing historical texture features with historical temperature features through an attention mechanism, and obtaining historical fusion features; training an initial teacher model constructed based on YOLOv10-L based on the historical fusion features and historical subject types corresponding to the historical fusion features to obtain a final teacher model; knowledge of the final teacher model is migrated to a student model constructed based on YOLOv10-N through a knowledge distillation algorithm, and convolutional layers in the student model are screened through historical fusion features and historical subject types to obtain a lightweight student model; and inputting the real-time texture features and the real-time temperature features into a lightweight student model to obtain an identification result, obtaining position information of the unmanned aerial vehicle, and performing dangerous behavior identification according to the identification result and the position information. The technical problem that inspection efficiency and dangerous behavior recognition precision are difficult to improve at the same time in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of power line inspection and safety monitoring technology, specifically to a method and system for identifying dangerous behaviors in high-voltage line areas based on unmanned aerial vehicles (UAVs). Background Technology

[0002] Power transmission lines are the core channels for transmitting electrical energy, undertaking the crucial task of transmitting electricity generated by power plants to urban and rural end users. Their safe and stable operation is related to social production and the order of residents' lives. However, transmission lines are mostly located in urban-rural fringe areas and near rural waterways, with complex surrounding environments where dangerous activities such as illegal fishing are common. If conductive objects such as fishing rods and lines come into contact with high-voltage lines, they can easily cause short circuits or grounding faults, which may not only burn out line equipment but also directly threaten people's lives. Traditional line inspection methods mainly rely on manual inspection, but manual inspection is limited by personnel energy and environmental conditions, resulting in low inspection efficiency. With the widespread application of drones in power line inspections, inspection efficiency has improved to some extent. However, technical bottlenecks remain in the automatic identification of hazardous behaviors. Current technologies often employ YOLOv10-L to build identification models that assess the environment surrounding power transmission lines to ensure accuracy. Hazardous behavior is then judged based on the identified results. However, the YOLOv10-L-based identification models are computationally complex and have large parameter scales, while drones have limited computing power and storage resources, making it difficult to meet the deployment and real-time inference requirements of such models. Therefore, how to simultaneously improve inspection efficiency and hazardous behavior identification accuracy is a pressing technical challenge that needs to be addressed. Summary of the Invention

[0003] To address the technical challenge of simultaneously improving inspection efficiency and the accuracy of hazardous behavior recognition in existing technologies, this invention provides a method and system for hazardous behavior recognition in high-voltage power line areas based on unmanned aerial vehicles (UAVs). The method utilizes a knowledge distillation algorithm to transfer knowledge from a final teacher model built on YOLOv10-L to a student model built on YOLOv10-N. Furthermore, it filters the convolutional layers in the student model using historical fusion features and their corresponding historical subject types to obtain a lightweight student model. By employing real-time texture and temperature features, and utilizing this lightweight student model adapted to the computing power and storage resources of UAVs, environmental recognition is performed in high-voltage power line areas. This solution addresses the technical challenge of simultaneously improving inspection efficiency and the accuracy of hazardous behavior recognition in existing technologies.

[0004] To address the aforementioned technical problems, this invention provides a method for identifying hazardous behaviors in high-voltage power line areas based on unmanned aerial vehicles (UAVs), comprising the following steps: Historical fusion features are obtained by fusing historical texture features and historical temperature features through an attention mechanism; The initial teacher model built on YOLOv10-L is trained based on historical fusion features and the corresponding historical subject types to obtain the final teacher model. The knowledge of the final teacher model is transferred to the student model built on YOLOv10-N through the knowledge distillation algorithm, and the convolutional layers in the student model are filtered by the historical fusion features and the historical subject types corresponding to the historical fusion features to obtain a lightweight student model. Real-time texture features and real-time temperature features are input into a lightweight student model to obtain recognition results and the location information of the drone. Dangerous behaviors are then identified based on the recognition results and location information.

[0005] Preferably, before fusing historical texture features and historical temperature features through an attention mechanism to obtain historical fused features, the method further includes: Visible light and infrared images of the high-voltage line area are acquired, and noise reduction and image enhancement are performed on the visible light and infrared images respectively. Initial historical texture features and initial historical temperature features were obtained by extracting features from visible light images and infrared images respectively using convolutional neural networks. The initial historical texture features and initial historical temperature features were then normalized to obtain historical texture features and historical temperature features.

[0006] Preferably, the step of fusing historical texture features and historical temperature features through an attention mechanism to obtain historical fused features includes: Key features are extracted from historical texture features and historical temperature features using an attention mechanism, thereby obtaining the first weight of key features in historical texture features and the first weight of key features in historical temperature features. The second weights of historical texture features and historical temperature features are obtained based on the acquisition time of visible light images and infrared images, respectively; The final weight of the historical texture feature is obtained by combining the first weight of the key feature in the historical texture feature with the second weight of the historical texture feature. Similarly, the final weight of the historical temperature feature is obtained by combining the first weight of the key feature in the historical temperature feature with the second weight of the historical temperature feature. Historical fusion features are obtained by fusing historical texture features and historical temperature features based on the final weights of historical texture features and historical temperature features.

[0007] Preferably, the step of training the initial teacher model constructed based on YOLOv10-L to obtain the final teacher model based on historical fusion features and the corresponding historical subject types includes: The historical fusion features are input into the initial teacher model to obtain the first predicted subject type. The model parameters of the initial teacher model are then corrected based on the difference between the first predicted subject type and the historical subject type, thereby obtaining the final teacher model.

[0008] Preferably, the step of transferring knowledge from the final teacher model to the student model built on YOLOv10-N using a knowledge distillation algorithm includes: The historical fusion features are input into the final teacher model to obtain the second predicted subject type, and the historical fusion features are input into the student model to obtain the third predicted subject type. The model parameters of the student model are then corrected based on the difference between the second and third predicted subject types.

[0009] Preferably, the step of filtering convolutional layers in the student model based on historical fusion features and the corresponding historical subject types to obtain a lightweight student model includes: The convolutional layers in the student model are removed one by one to obtain the initial lightweight student model. The historical fusion features are input into the initial lightweight student model to obtain the fourth predicted subject type. If the fourth predicted subject type matches the historical subject type, the removed convolutional layers are marked as redundant convolutional layers. The redundant convolutional layers are removed from the student model to obtain the lightweight student model.

[0010] Preferably, the step of identifying dangerous behavior based on the identification result and location information includes: The danger range of the high-voltage line is defined according to safety requirements. When the target to be alerted and the dangerous interference source appear in the danger range at the same time in the identification results, it indicates that there is a serious dangerous behavior. When either the target to be alerted or the dangerous interference source appears in the danger range, it indicates that there is a minor dangerous behavior. If the target to be alerted and / or the dangerous interference source are still in the danger range within a preset time, it indicates that there is a continuous dangerous behavior.

[0011] By adopting the above technical solution, the present invention has the following advantages: The knowledge distillation algorithm is used to transfer the knowledge of the final teacher model built on YOLOv10-L to the student model built on YOLOv10-N. The convolutional layers in the student model are filtered by historical fusion features and the corresponding historical subject types to obtain a lightweight student model. Using real-time texture features and real-time temperature features, the lightweight student model that can adapt to the computing power and storage resources of drones is used to identify the environment in the high-voltage line area. This solves the technical problem that existing technologies cannot simultaneously improve inspection efficiency and the accuracy of dangerous behavior identification. Key features are screened for historical texture features and historical temperature features respectively, and weights are assigned to key features that are useful for identifying dangerous behaviors. This removes redundant and interfering information from texture and temperature features. A second weight is obtained by using the acquisition time of visible light and infrared images, which solves the problem of single features becoming invalid at different times, making it difficult to identify accurately. By generating high-quality historical fusion features that are both critical and dynamically adaptable to time changes, the problem of feature effectiveness fluctuation caused by day and night environmental differences in high-voltage line inspection is solved.

[0012] This invention also provides a drone-based system for identifying hazardous behaviors in high-voltage power line areas, applicable to the aforementioned drone-based method for identifying hazardous behaviors in high-voltage power line areas, comprising: The feature fusion module is used to fuse historical texture features and historical temperature features through an attention mechanism to obtain historical fused features; The final teacher model acquisition module is used to train the initial teacher model built on YOLOv10-L based on the historical fusion features and the historical subject types corresponding to the historical fusion features to obtain the final teacher model. The lightweight student model acquisition module is used to transfer the knowledge of the final teacher model to the student model built based on YOLOv10-N through the knowledge distillation algorithm, and to filter the convolutional layers in the student model by the historical fusion features and the historical subject types corresponding to the historical fusion features to obtain the lightweight student model. The hazardous behavior recognition module is used to input real-time texture features and real-time temperature features into a lightweight student model to obtain recognition results, acquire the drone's location information, and perform hazardous behavior recognition based on the recognition results and location information.

[0013] Preferred options also include: The alarm module is used to issue corresponding warning messages based on the type of dangerous behavior, and also to remind professionals to take corresponding emergency measures based on the type of dangerous behavior.

[0014] Preferably, the alarm module includes a high-voltage line area alarm module and a control center alarm module. The high-voltage line area alarm module is used to issue corresponding warning information according to the type of dangerous behavior, and the control center alarm module is used to remind professionals to take corresponding emergency measures according to the type of dangerous behavior.

[0015] By adopting the above technical solution, the present invention has the following advantages: The knowledge distillation algorithm is used to transfer the knowledge of the final teacher model built on YOLOv10-L to the student model built on YOLOv10-N. The convolutional layers in the student model are filtered by historical fusion features and the corresponding historical subject types to obtain a lightweight student model. Using real-time texture features and real-time temperature features, the lightweight student model that can adapt to the computing power and storage resources of drones is used to identify the environment in the high-voltage line area. This solves the technical problem that existing technologies cannot simultaneously improve inspection efficiency and the accuracy of dangerous behavior identification. By linking regional real-time alerts with professional handling by the control center, the problems of untimely intervention and unprofessional handling after the identification of dangerous behaviors on high-voltage lines have been solved, thereby improving the system's practicality and safety. Attached Figure Description

[0016] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0017] Figure 1 This is a flowchart illustrating the method for identifying hazardous behaviors in high-voltage power line areas based on unmanned aerial vehicles (UAVs) according to the present invention. Figure 2 This is a schematic diagram of the structure of the UAV-based hazardous behavior recognition system for high-voltage power line areas according to the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only one preferred embodiment of this invention and are only used to explain this invention. They do not limit the scope of protection of this invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0019] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures; the process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0020] Example 1: like Figure 1 As shown, the method for identifying hazardous behaviors in high-voltage power line areas based on drones includes the following steps: S1: Historical fusion features are obtained by fusing historical texture features and historical temperature features through an attention mechanism.

[0021] Before fusing historical texture features and historical temperature features through an attention mechanism to obtain historical fused features, the method further includes: Visible light and infrared images of the high-voltage line area are acquired, and noise reduction and image enhancement are performed on the visible light and infrared images respectively. Initial historical texture features and initial historical temperature features were obtained by extracting features from visible light images and infrared images respectively using convolutional neural networks. The initial historical texture features and initial historical temperature features were then normalized to obtain historical texture features and historical temperature features.

[0022] Because outdoor data acquisition is susceptible to haze, light fluctuations, and sensor noise, an adaptive median filtering algorithm was used to denoise the visible light images. The initial size of the filtering window was set to 3×3, and the maximum window size was dynamically adjusted to 7×7 based on the image noise intensity to remove Gaussian noise while preserving texture details. After denoising, the Retinex algorithm was used for image enhancement, which improved the contrast and texture clarity of the visible light images, making the texture features of people, fishing rods, and fishing lines in the high-voltage line area more prominent. Infrared images are prone to thermal noise and edge blurring. A wavelet thresholding denoising algorithm was used to effectively suppress thermal noise. Subsequently, a histogram equalization algorithm was used to enhance the denoised infrared images, stretching the temperature grayscale range to make the temperature differences of people, fishing rods, and fishing lines in the high-voltage line area more obvious.

[0023] In some preferred embodiments, the step of fusing historical texture features and historical temperature features through an attention mechanism to obtain historical fused features includes: Key features are extracted from historical texture features and historical temperature features using an attention mechanism, thereby obtaining the first weight of key features in historical texture features and the first weight of key features in historical temperature features. The second weights of historical texture features and historical temperature features are obtained based on the acquisition time of visible light images and infrared images, respectively; The final weight of the historical texture feature is obtained by combining the first weight of the key feature in the historical texture feature with the second weight of the historical texture feature. Similarly, the final weight of the historical temperature feature is obtained by combining the first weight of the key feature in the historical temperature feature with the second weight of the historical temperature feature. Historical fusion features are obtained by fusing historical texture features and historical temperature features based on the final weights of historical texture features and historical temperature features.

[0024] In this embodiment, if the historical texture features contain features related to the human body, fishing rod, fishing line, etc., then the key features in the historical texture features are those related to the human body, fishing rod, fishing line, etc. If the historical texture features do not contain features related to the human body, fishing rod, fishing line, etc., then the key features in the historical texture features are those related to the background of the high-voltage line area. The same applies to the key features in the historical temperature features. The first weight of features related to the human body, fishing rod, fishing line, etc., is greater than the first weight of features related to the background of the high-voltage line area. In some scenarios, such as during daytime hours (6:00, 10:00, 14:00, 18:00): with sufficient lighting, texture features are clear (high effectiveness), while temperature features are easily affected by environmental interference (low effectiveness). Therefore, the second weight of historical texture features during these times is greater than the second weight of historical temperature features. During nighttime hours (22:00, 02:00): with insufficient lighting, texture features are blurry (low effectiveness), while temperature features show significant differences (high effectiveness). Therefore, the second weight of historical texture features during these times is less than the second weight of historical temperature features. By performing key feature screening on historical texture and temperature features separately, and assigning higher weights to key features useful for identifying hazardous behaviors, redundant and interfering information in texture and temperature features is eliminated. A second weight is obtained by using the acquisition time of visible light and infrared images, thus solving the problem of single features becoming ineffective at different times, leading to inaccurate identification. By generating high-quality historical fusion features that are both critical and dynamically adaptable to time changes, the problem of fluctuating feature effectiveness due to day-night environmental differences in high-voltage line inspections is solved. In this embodiment, before fusing historical texture and temperature features, time synchronization and spatial registration of the historical texture and temperature features are required to obtain spatiotemporally aligned historical fusion features.

[0025] S2: Based on the historical fusion features and the corresponding historical subject types, train the initial teacher model built on YOLOv10-L to obtain the final teacher model.

[0026] The process of training the initial teacher model built on YOLOv10-L based on historical fusion features and the corresponding historical subject types to obtain the final teacher model includes: The historical fusion features are input into the initial teacher model to obtain the first predicted subject type. The model parameters of the initial teacher model are then corrected based on the difference between the first predicted subject type and the historical subject type, thereby obtaining the final teacher model.

[0027] YOLOv10-L is a large-scale model in the YOLOv10 object detection series, positioned for high-performance object detection tasks. It trades increased computational resource consumption for higher detection accuracy. YOLOv10-L employs a deeper network structure and richer parameter configurations. Its input size is 640×640, with approximately 24.4M-24.69M parameters, a model size of approximately 92.25M, and a computational cost of 120.3 GFLOPs. Its inference latency is 7.28ms. This model combines advanced anchor-free detection mechanisms, multi-scale feature fusion, and a decoupled head design, enabling more accurate detection of small targets and multi-category objects in complex backgrounds. It is suitable for edge devices and server environments with high computing power. In this embodiment, an initial teacher model is built using YOLOv10-L, and the initial teacher model is trained to obtain the final teacher model, thereby improving the final teacher model's accuracy in identifying subject types.

[0028] S3: The knowledge of the final teacher model is transferred to the student model built on YOLOv10-N through the knowledge distillation algorithm, and the convolutional layers in the student model are filtered by the historical fusion features and the historical subject types corresponding to the historical fusion features to obtain a lightweight student model.

[0029] The process of transferring knowledge from the final teacher model to the student model built on YOLOv10-N using a knowledge distillation algorithm includes: The historical fusion features are input into the final teacher model to obtain the second predicted subject type, and the historical fusion features are input into the student model to obtain the third predicted subject type. The model parameters of the student model are then corrected based on the difference between the second and third predicted subject types.

[0030] Knowledge distillation is a model compression and knowledge transfer technique. Its core idea is to transfer knowledge from a teacher model (typically a large, high-precision, but computationally complex model) to a student model (a small, lightweight, and efficient model). This allows the student model to maintain high performance while significantly reducing computational costs and inference latency, making it suitable for resource-constrained environments such as edge devices and real-time scenarios. YOLOv10-N is an ultra-lightweight version of the YOLOv10 series. By significantly reducing model parameters and computational load, YOLOv10-N achieves extremely fast inference speeds while maintaining basic detection capabilities. The model size of YOLOv10-N is much smaller than that of YOLOv10-L. In this embodiment, the knowledge from the final teacher model is directly transferred to the student model using the knowledge distillation algorithm, improving the recognition accuracy of the student model.

[0031] In some preferred embodiments, the step of filtering convolutional layers in the student model based on historical fusion features and the corresponding historical subject types to obtain a lightweight student model includes: The convolutional layers in the student model are removed one by one to obtain the initial lightweight student model. The historical fusion features are input into the initial lightweight student model to obtain the fourth predicted subject type. If the fourth predicted subject type matches the historical subject type, the removed convolutional layers are marked as redundant convolutional layers. The redundant convolutional layers are removed from the student model to obtain the lightweight student model.

[0032] In this embodiment, by detecting the importance of each convolutional layer, convolutional layers that do not significantly affect the identification of subject types are removed to obtain a lightweight student model. This lightweight student model not only has the high accuracy of the final teacher model, but can also adapt to the computing power and storage resources of the drone, thereby solving the technical problem that existing technologies cannot simultaneously improve inspection efficiency and dangerous behavior identification accuracy.

[0033] S4: Input real-time texture features and real-time temperature features into the lightweight student model to obtain recognition results, obtain the location information of the drone, and identify dangerous behaviors based on the recognition results and location information.

[0034] Specifically, the identification of dangerous behaviors based on the identification results and location information includes: The danger range of the high-voltage line is defined according to safety requirements. When the target to be alerted and the dangerous interference source appear in the danger range at the same time in the identification results, it indicates that there is a serious dangerous behavior. When either the target to be alerted or the dangerous interference source appears in the danger range, it indicates that there is a minor dangerous behavior. If the target to be alerted and / or the dangerous interference source are still in the danger range within a preset time, it indicates that there is a continuous dangerous behavior.

[0035] In this embodiment, the target to be alerted is specifically a human body, and the dangerous interference source is specifically a dangerous object such as a fishing rod, fishing line, or kite string that could threaten the safe and stable operation of the high-voltage line.

[0036] Example 2: This embodiment also provides a drone-based system for identifying hazardous behaviors in high-voltage power line areas, applicable to the aforementioned drone-based method for identifying hazardous behaviors in high-voltage power line areas, including: The feature fusion module is used to fuse historical texture features and historical temperature features through an attention mechanism to obtain historical fused features; The final teacher model acquisition module is used to train the initial teacher model built on YOLOv10-L based on the historical fusion features and the historical subject types corresponding to the historical fusion features to obtain the final teacher model. The lightweight student model acquisition module is used to transfer the knowledge of the final teacher model to the student model built based on YOLOv10-N through the knowledge distillation algorithm, and to filter the convolutional layers in the student model by the historical fusion features and the historical subject types corresponding to the historical fusion features to obtain the lightweight student model. The hazardous behavior recognition module is used to input real-time texture features and real-time temperature features into a lightweight student model to obtain recognition results, acquire the drone's location information, and perform hazardous behavior recognition based on the recognition results and location information.

[0037] Also includes: The alarm module is used to issue corresponding warning messages based on the type of dangerous behavior, and also to remind professionals to take corresponding emergency measures based on the type of dangerous behavior.

[0038] The alarm module includes a high-voltage line area alarm module and a control center alarm module. The high-voltage line area alarm module is used to issue corresponding warning information according to the type of dangerous behavior, and the control center alarm module is used to remind professionals to take corresponding emergency measures according to the type of dangerous behavior.

[0039] like Figure 2As shown, in this embodiment, the system also includes a dual-modal camera, an image preprocessing module, and a feature extraction module. The dual-modal camera has a dual-light source system with white light and infrared light, used to simultaneously acquire visible light and infrared images of the high-voltage line area. It has a built-in automatic lighting mode that activates the white light when personnel move, maintaining it for 20 seconds before switching to infrared night vision mode. It also features built-in 6KV-level lightning protection, electrostatic discharge protection, and a wide voltage adaptation function of 10-17V. It is equipped with a 1.5A constant current drive circuit and multiple protection mechanisms. The visible light camera has a resolution of no less than 1920×1080, used to capture high-resolution images in daytime environments to clearly identify personnel and dangerous objects that threaten the safe and stable operation of high-voltage lines. The infrared camera has a resolution of no less than 640×512, used to identify human thermal signals and dangerous objects that threaten the safe and stable operation of high-voltage lines through thermal imaging in nighttime or low-light environments, achieving all-weather detection. The image preprocessing module is used to perform noise reduction, image enhancement, and normalization processing on the visible light and infrared images. The feature extraction module is used to extract features from the acquired visible light and infrared images using a convolutional neural network. The lightweight detection module comprises a lightweight student model acquisition module, an image preprocessing module, and a feature extraction module. The edge computing unit consists of a feature fusion module, a hazardous behavior recognition module, and an alarm module. The high-voltage line area alarm module includes an audible and visual alarm and a voice broadcast. The audible and visual alarm simultaneously emits a warning sound and flashing light; the voice broadcast plays alarm prompts, reminding on-site personnel to immediately move away from the high-voltage line area. Additionally, the high-voltage line area alarm module sends the alarm level, hazardous behavior category, target location coordinates, and alarm time to the control center alarm module, thereby alerting professionals to take appropriate emergency measures in a timely manner. In this embodiment, the system also includes a drone platform and a host computer. The drone flies along a preset power transmission line inspection route. The drone platform is a quadcopter drone equipped with a flight control module, which has automatic obstacle avoidance, autonomous cruise, fixed-point hovering and path planning functions. The inspection altitude, speed and trajectory can be set according to task requirements. The host computer adopts a Python+SQL architecture to work together. Python is used to build the host computer monitoring interface and is responsible for data visualization and behavior analysis. The SQL database is used to store historical detection results, alarm records and inspection logs. The host computer stores detection data and alarm status once per minute and can retain at least 10 minutes of data cache for subsequent traceability analysis.

[0040] The drone platform is used to carry a dual-modal camera and an edge computing unit. A lightweight detection model is deployed on the edge computing unit. The host computer and the edge computing unit communicate wirelessly. The edge computing unit generates alarm information (alarm level, dangerous behavior category, target location coordinates, and alarm time) and transmits this information to the host computer so that professionals can take timely emergency measures. Audible and visual alarms and voice broadcasts are also connected to the edge computing unit. In this embodiment, the final teacher model acquisition module is not mounted on the drone platform; it only serves as a process model for acquiring the lightweight student model acquisition module. By mounting the lightweight student model acquisition module on the drone platform, the lightweight student model not only possesses the high accuracy of the final teacher model but also adapts to the drone's computing power and storage resources, thus solving the technical problem of simultaneously improving inspection efficiency and dangerous behavior identification accuracy in existing technologies. The system provided by this invention is applicable to safety inspections of power transmission lines, water crossings, urban-rural fringe areas, and other high-risk areas, enabling rapid identification and real-time early warning of illegal fishing, kite flying, and other dangerous behaviors.

[0041] The specific embodiments described above are preferred embodiments of the method and system for identifying dangerous behaviors in high-voltage line areas based on unmanned aerial vehicles (UAVs) of the present invention, and are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A method for identifying dangerous behavior in a high-voltage line area based on a UAV, characterized in that, The method comprises the following steps: fusing the historical texture features and the historical temperature features through an attention mechanism to obtain historical fusion features; training an initial teacher model based on YOLOv10-L based on the historical fusion features and historical subject types corresponding to the historical fusion features to obtain a final teacher model; migrating the knowledge of the final teacher model to a student model based on YOLOv10-N through a knowledge distillation algorithm, and screening the convolutional layers in the student model through the historical fusion features and the historical subject types corresponding to the historical fusion features to obtain a lightweight student model; inputting real-time texture features and real-time temperature features into the lightweight student model to obtain a recognition result, obtaining position information of the unmanned aerial vehicle, and performing dangerous behavior recognition according to the recognition result and the position information. 2.The unmanned aerial vehicle-based high-voltage line area dangerous behavior recognition method according to claim 1, characterized in that, Before the historical fusion features are obtained by fusing the historical texture features and the historical temperature features through the attention mechanism, the method further comprises the following steps: obtaining a visible light image and an infrared image of the high-voltage line region, and respectively performing denoising and image enhancement processing on the visible light image and the infrared image; respectively extracting initial historical texture features and initial historical temperature features from the visible light image and the infrared image through a convolutional neural network, and performing normalization processing on the initial historical texture features and the initial historical temperature features to obtain the historical texture features and the historical temperature features. 3.The unmanned aerial vehicle based high-voltage line area dangerous behavior recognition method according to claim 2, characterized in that, The historical fusion features are obtained by fusing the historical texture features and the historical temperature features through the attention mechanism, comprising the following steps: respectively extracting key features from the historical texture features and the historical temperature features through the attention mechanism to obtain a first weight of the key features in the historical texture features and a first weight of the key features in the historical temperature features; respectively obtaining a second weight of the historical texture features and a second weight of the historical temperature features according to the collection time of the visible light image and the infrared image; obtaining a maximum weight of the historical texture features according to the first weight of the key features in the historical texture features and the second weight of the historical texture features, and obtaining a maximum weight of the historical temperature features according to the first weight of the key features in the historical temperature features and the second weight of the historical temperature features; fusing the historical texture features and the historical temperature features according to the maximum weight of the historical texture features and the maximum weight of the historical temperature features to obtain the historical fusion features. 4.The UAV-based high-voltage line area dangerous behavior recognition method according to claim 1, characterized in that, The final teacher model is obtained by training an initial teacher model based on YOLOv10-L based on the historical fusion features and historical subject types corresponding to the historical fusion features, comprising the following steps: inputting the historical fusion features into the initial teacher model to obtain a first predicted subject type, modifying the model parameters of the initial teacher model according to the difference between the first predicted subject type and the historical subject type, and thereby obtaining the final teacher model. 5.The unmanned aerial vehicle based high-voltage line area dangerous behavior recognition method according to claim 1, characterized in that, The knowledge of the final teacher model is migrated to a student model based on YOLOv10-N through a knowledge distillation algorithm, comprising the following steps: inputting the historical fusion features into the final teacher model to obtain a second predicted subject type, inputting the historical fusion features into the student model to obtain a third predicted subject type, and modifying the model parameters of the student model according to the difference between the second predicted subject type and the third predicted subject type. 6.The unmanned aerial vehicle based high-voltage line area dangerous behavior recognition method according to claim 1, characterized in that, The student model is screened by the historical fusion feature and the historical subject type corresponding to the historical fusion feature, so as to obtain a lightweight student model, and the method comprises the steps of: The convolutional layer in the student model is removed from the student model in sequence to obtain an initial lightweight student model, the historical fusion feature is input into the initial lightweight student model to obtain a fourth predicted subject type, if the fourth predicted subject type matches the historical subject type, the removed convolutional layer is marked as a redundant convolutional layer, and the redundant convolutional layer is removed from the student model to obtain a lightweight student model. 7.The UAV-based high-voltage line area dangerous behavior recognition method according to claim 1, characterized in that, The dangerous behavior is identified according to the recognition result and the position information, and the method comprises the steps of: The dangerous range of the high-voltage line is defined according to the safety requirement, when the to-be-alarmed target and the dangerous interference source appear in the dangerous range at the same time in the recognition result, it indicates that there is a serious dangerous behavior, when the to-be-alarmed target or the dangerous interference source appears in the dangerous range, it indicates that there is a slight dangerous behavior, and when the to-be-alarmed target or / and the dangerous interference source is still in the dangerous range within a preset time, it indicates that there is a continuous dangerous behavior.

8. The UAV-based high-voltage line area dangerous behavior recognition system, applicable to the UAV-based high-voltage line area dangerous behavior recognition method of any one of claims 1-7, characterized in that, The method comprises the steps of: The feature fusion module is used for fusing the historical texture feature and the historical temperature feature by the attention mechanism to obtain a historical fusion feature; The final teacher model acquisition module is used for training an initial teacher model based on YOLOv10-L to obtain a final teacher model based on the historical fusion feature and the historical subject type corresponding to the historical fusion feature; The lightweight student model acquisition module is used for migrating the knowledge of the final teacher model to a student model based on YOLOv10-N by a knowledge distillation algorithm, and screening the convolutional layer in the student model by the historical fusion feature and the historical subject type corresponding to the historical fusion feature, so as to obtain a lightweight student model; The dangerous behavior identification module is used for inputting the real-time texture feature and the real-time temperature feature into the lightweight student model to obtain a recognition result, obtaining the position information of the unmanned aerial vehicle, and identifying the dangerous behavior according to the recognition result and the position information. 9.The UAV-based high-voltage line area dangerous behavior recognition system of claim 8, wherein, The method further comprises the steps of: The alarm module is used for issuing corresponding warning information according to the dangerous behavior type, and is also used for reminding professional personnel to take corresponding emergency measures according to the dangerous behavior type. 10.The UAV-based high-voltage line area dangerous behavior recognition system of claim 9, wherein, The alarm module comprises a high-voltage line area alarm module and a control center alarm module, the high-voltage line area alarm module is used for issuing corresponding warning information according to the dangerous behavior type, and the control center alarm module is used for reminding professional personnel to take corresponding emergency measures according to the dangerous behavior type.

Images