Power transmission channel real-time dynamic target detection method and system based on edge computing

Abstract

The application discloses a power transmission channel real-time dynamic target detection method and system based on edge computing, and belongs to the technical field of image processing and computer vision, which comprises the following steps: acquiring real-time video stream of the power transmission channel, continuously storing and generating continuous video data; performing frame-by-frame analysis on the continuous video data at the edge, identifying dynamic targets by using an AI model, and generating dynamic target detection results; fusing the detection results and spatial information of the power transmission conductor, performing space-time behavior modeling, evaluating the motion trend and proximity of the dynamic targets, and generating threat evaluation vectors; combining power operation and maintenance rules to determine the risk level, generating graded alarm instructions, and driving local intervention devices to perform physical intervention actions. The closed-loop processing mode of the application, which directly drives the local intervention device, can realize the identification of dynamic threats, the pre-judgment of risks, and automatic on-site disposal by performing AI model analysis, space-time behavior modeling and graded decision at the edge.

Description

Technical Field

[0001] This invention relates to the fields of image processing and computer vision technology, and in particular to a method and system for real-time dynamic target detection in power transmission channels based on edge computing. Background Technology

[0002] Power transmission channels are the lifeline of a nation's energy supply, and their safe and stable operation is of paramount importance. Utilizing computer vision technology for real-time monitoring of power transmission channels and identifying anomalies such as external intrusions through image analysis is a crucial technological direction for ensuring power grid security. The core of this technology lies in accurately identifying and analyzing various dynamic targets and their behavioral patterns from continuous video images to achieve timely early warning of potential threats.

[0003] Currently, video surveillance of power transmission channels mainly employs two technical solutions. One relies on manual, 24 / 7 video patrol monitoring, using the naked eye of maintenance personnel to detect anomalies. The other uses a centralized intelligent video analysis system, transmitting all video streams collected by front-end cameras to a central server at the back end. The server cluster performs centralized, large-scale image processing and pattern recognition, and only pushes alarm information to maintenance personnel upon detecting anomalies.

[0004] Existing technical solutions have significant shortcomings in practical applications. Manual monitoring is not only labor-intensive but also prone to missed alarms due to fatigue or negligence, failing to achieve continuous and efficient monitoring. Centralized video analytics solutions require uploading massive amounts of high-definition video data from the scene to the cloud or data center in real time, placing enormous pressure on communication network bandwidth. Furthermore, inherent network latency during data transmission causes alarm information to lag significantly, making it difficult to meet the real-time response requirements for sudden high-risk events. In addition, traditional analysis models typically only perform simple target detection, lacking a comprehensive assessment of target dynamics and spatial threats, resulting in low alarm accuracy and a large number of invalid alarms. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a real-time dynamic target detection method and system for power transmission channels based on edge computing. It employs a closed-loop processing approach that performs AI model analysis, spatiotemporal behavior modeling, and hierarchical decision-making at the edge, and directly drives local intervention devices. This approach enables the identification of dynamic threats, the prediction of risks, and automated on-site handling.

[0006] The above objectives can be achieved through the following approach:

[0007] A real-time dynamic target detection method for power transmission channels based on edge computing includes the following steps:

[0008] The system acquires real-time video streams from power transmission channels, stores them continuously in a loop, and generates continuous video data.

[0009] At the edge, the continuous video data is analyzed frame by frame, and a pre-trained AI model is used to identify dynamic targets and generate dynamic target detection results.

[0010] By integrating the dynamic target detection results with the spatial information of the power transmission line, spatiotemporal behavior modeling is performed to assess the motion trend and proximity of the dynamic target and generate a threat assessment vector.

[0011] Based on the threat assessment vector, risk level is determined in conjunction with power transmission operation and maintenance rules, and graded alarm instructions are generated.

[0012] Based on the hierarchical alarm instructions, the local intervention device is driven to perform physical intervention actions.

[0013] Optionally, generating continuous video data specifically includes:

[0014] Real-time video streams are collected by industrial-grade high-definition monitoring equipment deployed on power transmission channels. The industrial-grade high-definition monitoring equipment automatically switches between day and night modes using a built-in light sensor unit, and actively cleans the lens in conjunction with an electric robotic arm and a superhydrophobic nano-shield architecture. The real-time video streams are sent to a storage unit configured for loop storage for full-time recording. Video clips containing the current moment are extracted from the storage unit to generate continuous video data.

[0015] Optionally, generating dynamic target detection results specifically includes:

[0016] The continuous video data is decoded using a hardware AI acceleration engine to generate video frames; the video frames are input into a pre-trained AI model for frame-by-frame analysis to identify dynamic targets and output bounding boxes and categories, forming dynamic target parameters; the dynamic target parameters are integrated to track targets and calculate motion states, generating target detection results.

[0017] Optionally, the generation of the threat assessment vector specifically includes:

[0018] The dynamic target detection results are analyzed over time to track the movement trajectory of the dynamic target and generate a trajectory sequence; the geographic information data of the power transmission line is obtained, and the spatial distance between the dynamic target and the power transmission line is calculated by combining the target location in the dynamic target detection results; the trajectory sequence and the spatial distance are fused and analyzed to quantify the potential threat and generate a threat assessment vector.

[0019] Optionally, the generation of the trajectory sequence specifically includes:

[0020] Based on the position changes of the dynamic target in consecutive frames, a velocity vector is calculated; based on the velocity vector, the possible position of the dynamic target at a future time point is predicted to form predicted position data; by combining the historical position, velocity vector and predicted position data of the dynamic target, a trajectory sequence is updated and generated.

[0021] Optionally, the generation of hierarchical alarm instructions specifically includes:

[0022] Based on the threat assessment vector and combined with power transmission operation and maintenance rules, the risk level is determined, and a preliminary alarm is generated. An alarm filtering mechanism is applied to screen the preliminary alarm, filtering out invalid alarms caused by background interference and duplicate alarms caused by the same event, and obtaining valid alarms. According to the risk level and target location information of the valid alarms, a graded alarm instruction is generated.

[0023] Optionally, the alarm filtering mechanism specifically includes:

[0024] The preliminary alarm is compared with the records in the historical alarm cache to identify and filter out duplicate alarms; the visual features of the preliminary alarm image area are analyzed to identify and filter out background interference alarms caused by environmental factors.

[0025] Optionally, the driving local intervention device performs a physical intervention action, specifically:

[0026] The graded alarm commands are sent to the backend monitoring platform via the communication network. The platform parses the graded alarm commands to obtain the hazard level, target location information, and handling suggestions. When the hazard level is high-risk, the sound and light linkage device is triggered to emit sound waves and laser beams in a directional manner to the area indicated by the target location information. Through physiological stimulation, the workers are forced to stop the dangerous operation and evacuate.

[0027] Optionally, the method further includes:

[0028] Based on the timestamp in the hierarchical alarm instruction, the corresponding historical video data is retrieved from the storage unit of the circular storage; the operation and maintenance personnel review the historical video data and generate alarm verification tags; the alarm verification tags and the historical video data are used to form a feedback dataset for iterative optimization training of the AI ​​model.

[0029] Based on the same inventive concept, the present invention also provides a real-time dynamic target detection system for power transmission channels based on edge computing, the system comprising:

[0030] The video acquisition and storage module is used to acquire real-time video streams from the power transmission channel, store them continuously in a loop, and generate continuous video data.

[0031] The dynamic target recognition module is used to analyze the continuous video data frame by frame, identify dynamic targets using a pre-trained AI model, and generate dynamic target detection results.

[0032] The spatiotemporal behavior analysis module is used to fuse the dynamic target detection results with the spatial information of the power transmission line, perform spatiotemporal behavior modeling, evaluate the motion trend and proximity of the dynamic target, and generate a threat assessment vector.

[0033] The hierarchical decision-making module is used to determine the risk level based on the threat assessment vector and in combination with power transmission operation and maintenance rules, and generate hierarchical alarm instructions.

[0034] The active intervention module is used to drive the local intervention device to perform physical intervention actions according to the hierarchical alarm instructions.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] (1) By deploying a complete perception, analysis, decision-making and execution process at the edge, a rapid-response proactive defense closed loop is constructed. Since the core computing tasks are completed at the data source, the network latency and bandwidth pressure caused by the long-distance transmission of massive video data are avoided, enabling the entire process from discovering potential threats to executing physical intervention to be completed in a very short time, thus gaining a valuable time window to prevent dangerous behavior from occurring.

[0037] (2) Enhance the intelligence and predictive ability of threat identification. This method not only focuses on identifying objects, but also performs spatiotemporal behavior modeling by integrating the target's movement trajectory with the precise spatial information of the power transmission line. This enables in-depth analysis of the target's dynamic trends and proximity. This understanding and prediction of the target's behavioral intentions allows the system to identify high-risk events in advance, realizing the transformation from "post-event alarm" to "pre-event warning".

[0038] (3) Establish a continuous self-optimization learning mechanism to improve the long-term stability and adaptability of the system. By introducing the review and annotation of alarm events by operation and maintenance personnel, a high-quality feedback dataset is formed. The AI ​​model is periodically trained using this dataset, enabling the model to continuously learn and adapt to new situations and environmental changes in specific scenarios, continuously reduce false alarms and false negatives, and ensure the continuous improvement of detection performance. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart illustrating a real-time dynamic target detection method for power transmission channels based on edge computing, according to an embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram of a multi-dimensional threat assessment vector according to an embodiment of the present invention.

[0042] Figure 3 This is a schematic diagram of hierarchical alarm filtering according to an embodiment of the present invention.

[0043] Figure 4 This is a schematic diagram of the structure of a real-time dynamic target detection system for power transmission channels based on edge computing, according to an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0045] Reference Figure 1 One embodiment of the present invention proposes a real-time dynamic target detection method for power transmission channels based on edge computing. It adopts a closed-loop processing approach that performs AI model analysis, spatiotemporal behavior modeling and hierarchical decision-making at the edge and directly drives local intervention devices, which can realize the identification of dynamic threats, the prediction of risks and automated on-site handling.

[0046] The method described in this embodiment includes the following steps:

[0047] The system acquires real-time video streams from power transmission channels, stores them continuously in a loop, and generates continuous video data.

[0048] At the edge, continuous video data is analyzed frame by frame, and a pre-trained AI model is used to identify dynamic targets and generate dynamic target detection results.

[0049] By integrating dynamic target detection results with spatial information of power transmission lines, spatiotemporal behavior modeling is performed to assess the motion trend and proximity of dynamic targets and generate threat assessment vectors.

[0050] Based on threat assessment vectors and combined with power transmission operation and maintenance rules, risk levels are determined and graded alarm instructions are generated.

[0051] Based on the tiered alarm instructions, the local intervention device is driven to perform physical intervention actions.

[0052] Specifically, this method first ensures the continuity and traceability of the data source through uninterrupted video acquisition and cyclic storage. The core processing is completed at the edge, using an artificial intelligence model to analyze real-time video and achieve rapid identification of dynamic targets. Unlike simple target detection, this method deeply integrates the detected target information with pre-defined three-dimensional spatial information of the transmission line. Through spatiotemporal behavior modeling, it comprehensively evaluates the target's trajectory, speed, and orientation, thereby predicting its potential threat level to the transmission channel and quantifying it into a multi-dimensional threat assessment vector. Based on this vector, combined with existing operational safety rules, intelligent decision-making is performed to automatically determine the risk level and generate differentiated alarm commands. Finally, this command directly drives local physical intervention equipment to take action, completing the entire process from threat detection to proactive intervention. This invention shortens the response time from hazard detection to action by performing real-time computation and decision-making at the edge, solving the network latency problem inherent in traditional centralized processing models. This method not only identifies potential threat targets but also understands their intentions and predicts their trajectories through spatiotemporal behavioral analysis, enhancing the intelligence and foresight of threat identification and enabling the prediction and prevention of safety incidents. Ultimately, by directly linking tiered alarms with automated physical intervention devices, an unattended closed-loop control system is formed, enabling immediate and effective on-site handling of high-risk behaviors, enhancing the inherent safety level of power transmission channels, and improving the automation and efficiency of operation and maintenance management.

[0053] In some embodiments, continuous video data is generated, specifically:

[0054] Real-time video streams are collected by industrial-grade high-definition monitoring equipment deployed in power transmission channels. The industrial-grade high-definition monitoring equipment automatically switches between day and night modes using built-in light sensors and works in conjunction with an electric robotic arm and a superhydrophobic nano-shield architecture to actively clean the lens.

[0055] The real-time video stream is sent to a storage unit configured for circular storage for continuous recording.

[0056] Extract the video clip containing the current moment from the storage unit to generate continuous video data.

[0057] Specifically, the first step is to deploy dedicated industrial-grade high-definition monitoring equipment at key locations along power transmission channels, such as the crossarms or tower bodies of transmission towers. This equipment boasts strong environmental adaptability, capable of withstanding harsh weather conditions such as wind, rain, high and low temperatures. The equipment integrates a built-in light sensor unit that monitors ambient light intensity in real time. When the light intensity falls below a preset threshold, such as at dusk or on cloudy days, the built-in light sensor unit automatically triggers the image sensor to switch to night vision mode. Infrared supplementary lighting or low-light enhancement technology ensures image clarity at night, enabling uninterrupted video capture around the clock. To address the issue of lenses being easily obstructed by rain, dust, bird droppings, or other contaminants in outdoor environments, the industrial-grade high-definition monitoring equipment employs an active cleaning design. Its lens surface is covered with a superhydrophobic nano-shield architecture. This architecture utilizes the surface properties of the material to prevent water droplets from adhering and to remove some dust. Simultaneously, the device is equipped with a miniature electric robotic arm. This arm automatically activates based on a preset time period or after an image quality diagnostic algorithm determines the degree of lens dirtiness, performing a physical scraping action to ensure the lens remains clean, thus guaranteeing the original quality of the acquired real-time video stream. The acquired high-definition real-time video stream is immediately transmitted to the storage unit built into the edge computing device. This storage unit is configured in a circular storage mode, meaning data is overwritten using a first-in, first-out (FIFO) principle. A fixed-size storage space is pre-allocated, for example, to store the most recent 72 hours of video recordings. When new video data is written, once the storage space is full, the oldest video data is automatically deleted to make room for the new data. This cycle repeats continuously, achieving full-time recording of the power transmission channel's status without causing storage overflow. When dynamic target analysis is required, the processing module retrieves all video frames from this storage unit, starting from the current processing time and working backward for a preset time length, such as 5 seconds. These consecutive, time-sequential video frames are then integrated into a single video segment, which serves as the continuous video data required for subsequent AI model analysis. This method solves the challenges of stability and clarity in video data acquisition under complex outdoor environments. The generated continuous video data not only contains rich spatial details but also retains the continuous motion sequence information of the target, providing a data foundation for subsequent dynamic target recognition, trajectory tracking, and threat assessment, thereby improving the reliability and accuracy of the entire dynamic target detection method.

[0058] In some embodiments, dynamic target detection results are generated, specifically as follows:

[0059] The hardware AI acceleration engine is used to decode continuous video data and generate video frames;

[0060] Video frames are input into a pre-trained AI model for frame-by-frame analysis to identify dynamic targets and output bounding boxes and categories, thus forming dynamic target parameters.

[0061] It integrates dynamic target parameters, tracks the target and calculates its motion state, and generates target detection results.

[0062] Specifically, the first step is to decode this continuous video data, restoring it into a series of independent images, i.e., video frames. Considering that edge computing devices typically have limited computing resources, and video decoding is a computationally intensive task, this method utilizes a hardware AI acceleration engine integrated into the edge device to perform this operation. A hardware AI acceleration engine is a dedicated processing unit, such as a Neural Processing Unit (NPU) or a Virtual Graphics Processing Unit (VGPU), capable of efficiently handling parallel computing tasks such as video encoding / decoding and AI model inference, thereby freeing up the main CPU to handle other logic. The decoded video frames are then fed one by one into a pre-trained AI model. This AI model is based on a deep learning-based object detection network, such as a YOLO or SSD architecture optimized for a specific dataset. This model has been trained to identify various dynamic targets commonly found in power transmission channel environments, such as construction vehicles, tower cranes, drones, and personnel. For each input frame, the AI ​​model performs a forward propagation calculation, outputting all the dynamic targets it identifies in the current frame. Each identified dynamic target is accompanied by structured dynamic target parameters, mainly including two core pieces of information: first, the target's bounding box, a set of coordinate values ​​used to accurately define the target's position and size on the two-dimensional image plane; second, the target's category, a label indicating which predefined object the target belongs to. Analyzing only a single frame is insufficient to understand the target's dynamic behavior. Therefore, it is necessary to integrate dynamic target parameters from multiple consecutive frames to track the target and calculate its motion state. Multi-target tracking algorithms, such as SORT or its improved versions, are employed. These algorithms compare the position, size, and even appearance features of detection boxes across consecutive frames, associating detection results belonging to the same physical entity and assigning a unique ID to each continuously tracked dynamic target. After establishing a stable tracking trajectory, its motion state can be quantified. For example, its instantaneous velocity can be estimated by calculating the displacement of the target's bounding box center point between two consecutive frames. Let the pixel coordinates of the target's center point in frame t be... The coordinates of the center point in the previous frame, i.e., frame t-1, are... The time interval between two frames is Then its velocity vector on the image plane It can be approximated as:

[0063] ;

[0064] in, and The center point coordinates are calculated using the bounding box output by the AI ​​model. This is the reciprocal of the video frame rate. This velocity vector describes the target's direction and speed of motion at the current moment. Finally, the identity ID of each tracked target, the dynamic target parameters of the current frame, and the calculated motion state are integrated to form a structured target detection result. By integrating frame-by-frame analysis and cross-frame tracking, this method can not only identify what the dynamic targets are and where they are in the power transmission channel, but also further understand their motion state. This provides rich and dynamic input data for subsequent deeper spatiotemporal behavior analysis and risk assessment, enhancing the system's situational awareness capabilities.

[0065] In some embodiments, a threat assessment vector is generated, specifically as follows:

[0066] Perform time-series analysis on the dynamic target detection results, track the motion trajectory of the dynamic target, and generate trajectory sequences;

[0067] Geographic information data of power transmission lines is obtained, and the spatial distance between the dynamic target and the power transmission line is calculated by combining the target location in the dynamic target detection results.

[0068] By fusing trajectory sequences with spatial distances, potential threats are quantified, and threat assessment vectors are generated.

[0069] Specifically, to generate a threat assessment vector, dynamic targets are first continuously tracked through temporal analysis to determine their motion paths in consecutive video frames, thus generating a trajectory sequence. This trajectory sequence is a structured representation of the target's history and current state, forming the basis for assessing its dynamic behavior. Next, to accurately assess the threat, the target's position in the 2D image must be mapped to 3D physical space, and its actual spatial distance to the power transmission line must be calculated. This step relies on geographic information data of the power transmission line pre-configured in edge computing devices. This data is based on a high-precision 3D model generated from LiDAR scanning or BIM modeling, describing the spatial orientation and shape of the power transmission line in a specific coordinate system. Simultaneously, extrinsic parameters such as the installation location and orientation of the monitoring equipment, as well as intrinsic parameters such as lens focal length, also need to be pre-calibrated. Using these camera calibration parameters, a mapping relationship from image pixel coordinates to real-world 3D coordinates is established. When a dynamic target is detected, its position in the image is extracted, such as the midpoint of the bottom edge of the bounding box. Combined with the assumption of a flat ground surface, a coordinate transformation algorithm is used to calculate the target's approximate coordinates in 3D physical space. Once the three-dimensional coordinates of the target and the transmission line are obtained, the shortest spatial distance between them can be calculated, which is the Euclidean distance between the point on the target closest to the transmission line and the point on the transmission line closest to that point. Finally, the information obtained in the first two steps is fused and analyzed to quantify the potential threat and encapsulate it into a structured threat assessment vector. This vector is a multi-dimensional data structure designed to comprehensively describe the nature and urgency of the threat. Its structure is typically as follows:

[0070] ;

[0071] in, This represents the final generated threat assessment vector; the first component of the vector. The first component is the minimum spatial distance between the target and the guide wire calculated at the current moment, which directly reflects the static degree of danger; the second component... This is the rate of change of spatial distance with respect to time, i.e., the target's approach rate. It is calculated by differencing the spatial distances at consecutive moments in the trajectory sequence. It reveals dynamic threat trends; a negative value indicates the target is approaching the guide wire. The third component... The first component is the target category information, which is an enumerated value used to distinguish the inherent risk levels of different targets, such as large machinery, vehicles, and personnel; the fourth component... The duration of a target's stay within the warning area, obtained through trajectory sequence analysis, is used to measure the sustained risk of potential operations. This method not only solves the problem of correlating two-dimensional video information with three-dimensional spatial risks but also captures the target's dynamic behavioral characteristics through temporal analysis. The generated threat assessment vector integrates spatial proximity, temporal movement trends, target attributes, and behavioral persistence, forming a three-dimensional, multi-dimensional description of potential threats to the power transmission channel. This allows subsequent risk assessments to move beyond simple threshold triggering to a comprehensive evaluation of the entire situation, thereby improving the accuracy and foresight of alarms and avoiding false alarms and missed alarms.

[0072] In some embodiments, the trajectory sequence is generated as follows:

[0073] Calculate the velocity vector based on the positional changes of a dynamic target in consecutive frames;

[0074] Based on the velocity vector, the possible position of a dynamic target at a future time point is predicted, thus forming predicted position data;

[0075] By combining historical position, velocity vector, and predicted position data of dynamic targets, the trajectory sequence is updated and generated.

[0076] Specifically, to achieve the generation and updating of trajectory sequences, this method introduces a dynamic tracking and prediction mechanism based on a state-space model, typically implemented using a Kalman filter or its variant. This process begins by acquiring the positional changes of the same dynamic target identified in consecutive frames. Let the target position detected by the AI ​​model at time point t be... Based on this location information and the location information of the previous time t-1. It is possible to calculate the target within a time interval. The instantaneous velocity vector within the image plane reflects the target's direction and speed of motion. However, prediction based solely on instantaneous velocity is susceptible to detection noise. Therefore, this method establishes a more complete state model to describe the dynamic target. The target's state vector... It can be defined as a combination containing its position and velocity information, for example ,in These are the target's location coordinates. This represents the target's velocity component. Based on this state, a kinematic model can be used to predict the target's possible position at future time points; this is how predicted position data is generated. A commonly used linear motion model is represented as follows:

[0077] ;

[0078] in, It is the predicted value of the target state at time t at time t-1, which includes the predicted information of the future location; It is the optimal state estimate after correction at time t-1; This is the state transition matrix, which describes how the state evolves over time. For a uniform motion model, The current position can be predicted by adding the previous position to the velocity and multiplying by the time interval, while keeping the velocity constant. This prediction process essentially infers the target's future movement based on the existing velocity vector, thus forming predicted position data. Finally, when the actual observed position at time t—the latest position data in the dynamic target detection results—arrives, historical information, predicted information, and current observation information are comprehensively utilized to update and generate the final trajectory sequence. This process will determine the predicted state... The optimal state estimate at time t is generated by weighted fusion with the actual observations of the current frame. This fusion process comprehensively considers the uncertainties of both prediction and observation, providing a more accurate and smoother state estimate than either observation alone or prediction alone. This updated state... It includes not only a precise description of the current position but also corrections to the current velocity vector. The trajectory sequence is the optimal state estimate obtained from this series of continuous predictions and corrections. This method endows the system with the ability to make forward-looking predictions, inferring the potential future location of a target based on current movement trends. This predictive capability is crucial for risk assessment, enabling the system to move from "detecting danger" to "predicting danger," thus gaining a valuable time window for taking preventative measures and enhancing the initiative and effectiveness of the entire power transmission channel safety monitoring system.

[0079] In some embodiments, a hierarchical alarm instruction is generated, specifically as follows:

[0080] Based on the threat assessment vector and combined with power transmission operation and maintenance rules, the risk level is determined and an initial alarm is generated.

[0081] An alarm filtering mechanism is applied to screen preliminary alarms, filtering out invalid alarms caused by background interference and duplicate alarms caused by the same event, thus obtaining valid alarms.

[0082] Based on the risk level and target location information of valid alarms, generate graded alarm commands.

[0083] Specifically, the process begins by determining the risk level based on a threat assessment vector, combined with pre-embedded power transmission maintenance rules embedded in edge computing devices, to generate an initial alert. These power transmission maintenance rules are digitized into a multi-dimensional decision-making logic that compares various components of the threat assessment vector, such as spatial distance, approach rate, target category, and dwell time, with a series of preset thresholds. For example, the rule defines a "high" risk level if a target of the "tower crane" category is less than 20 meters away and has a negative approach rate; while a "personnel" target loitering at a distance of 50 meters might be classified as "low" risk. Figure 2 As shown in the figure, this diagram visually illustrates the composition of the threat assessment vector. The four dimensions in the diagram correspond to spatial distance, approach rate, target category risk, and dwell time, respectively. By comparing "high-risk events," "medium-risk events," and "low-risk events," it can be understood how this method combines power transmission operation and maintenance rules to determine the risk level. Through this refined logical matching, a preliminary alarm containing the risk level, target information, and timestamp is generated for each event considered to have a potential threat. However, the preliminary alarm may contain false alarms due to environmental complexity. Therefore, an alarm filtering mechanism is applied to filter the preliminary alarm stream to extract effective alarms. This mechanism consists of two core parts. The first part is duplicate alarm suppression. When a new preliminary alarm is generated, it checks whether there are existing alarms in the buffer targeting the same target ID with very close timestamps. If so, and the risk level has not significantly escalated, the new alarm is considered a continuation of the same event and is filtered out, avoiding continuous bombardment of alarms for the same threat event. The second part is background interference filtering. It retrieves the image region corresponding to the preliminary alarm and performs a rapid secondary analysis of its visual features. For example, by analyzing the target's texture, morphological stability, and other characteristics, invalid alarms caused by non-physical environmental factors such as swaying tree shadows and water reflections are identified and filtered out. After filtering, the remaining alarms are considered high-confidence and valid. Finally, based on the risk level determined in each valid alarm and the precise target location information, standardized hierarchical alarm commands are generated. Figure 3As shown, the flowchart clearly depicts the transformation process from a large number of "preliminary alarms" to the final "tiered alarm results." The instruction is a structured data packet that not only contains the risk level and location coordinates but may also encapsulate suggested contingency plan code. For example, for a valid alarm at the "high" risk level, the generated tiered alarm instruction will include a command to immediately activate local intervention devices; for the "medium" risk level, the instruction might notify the backend monitoring platform for manual review; and "low" risk alarms might only be logged. This method, by introducing a risk assessment mechanism combining operational rules and a two-stage alarm filtering mechanism, achieves intelligent decision-making from initial threat perception to precise, executable instructions, solving the common problems of false alarms and information redundancy in dynamic target detection, and ensuring the quality and relevance of alarms. By classifying threats into different levels and generating corresponding tiered alarm instructions, the system can achieve optimized resource allocation and differentiated responses, i.e., immediate and powerful intervention for high-risk events and monitoring and recording of low-risk events, improving the automation level and operational efficiency of the entire power transmission channel monitoring system, making security protection more targeted and effective.

[0084] In some embodiments, the alarm filtering mechanism is specifically as follows:

[0085] The initial alarm is compared with the records in the historical alarm cache to identify and filter out duplicate alarms;

[0086] Analyze the visual features of the initial alarm image area to identify and filter out background interference alarms caused by environmental factors.

[0087] Specifically, the alarm filtering mechanism consists of two parallel sub-processes designed to filter out invalid and redundant information from the initial alarm stream. The first sub-process is the identification and filtering of duplicate alarms. Whenever a new initial alarm is generated, its core elements are extracted, including the unique ID of the tracked target, the risk level, and the timestamp of the event. Then, all records in the historical alarm cache are traversed. The comparison logic is as follows: if a record already exists in the cache, and its target ID is the same as the ID of the current initial alarm, and the difference between the timestamps of the two alarms is less than a preset "silent time," and the risk level has not increased, then the current initial alarm is determined to be a duplicate report of the same ongoing threat event. In this case, the initial alarm will be marked as "duplicate" and discarded directly, not entering the subsequent processing flow, thus suppressing the frequency of alarms for the same event. The second sub-process is the identification and filtering of background interference alarms. This process mainly addresses false alarms caused by environmental factors, such as changes in light and shadow, and swaying vegetation. Once an initial alarm is generated, the image region corresponding to the alarm is cropped from the original video frame based on the target bounding box provided in the alarm information. Next, a series of rapid visual feature analyses are performed on this image region. The analysis includes: first, calculating the motion optical flow field of the pixels within the region; if the optical flow field exhibits large-area, patterned collective motion rather than the motion of independent entities with clear outlines, it is determined to be background interference. Second, the texture features and color histogram of the region are extracted and compared with a pre-established non-threatening background sample library. If its features are highly similar to typical background interference patterns in the library, it is also determined to be a background interference alarm. Initial alarms determined to be background interference are also filtered out. Repeated alarm filtering ensures that the monitoring system does not continuously generate interfering alarms due to a persistent threat, allowing maintenance personnel to focus on new or escalating events. Background interference filtering improves the signal-to-noise ratio of alarms; it utilizes deeper image analysis methods to compensate for potential misjudgments by the primary target detection model in complex dynamic backgrounds, reducing false alarms caused by environmental changes.

[0088] In some embodiments, driving the local intervention device to perform a physical intervention action specifically includes:

[0089] The tiered alarm commands are sent to the backend monitoring platform via the communication network. The tiered alarm commands are parsed to obtain the hazard level, target location information, and handling suggestions.

[0090] When the hazard level is high-risk, the sound and light linkage device is triggered to emit sound waves and laser beams in a directional manner to the area indicated by the target location information, forcing the workers to stop the dangerous operation and evacuate through physiological stimulation.

[0091] Specifically, the edge computing device first uses its built-in communication module to send structured, tiered alarm commands to the backend monitoring platform in real time via 4G / 5G, fiber optic, or other reliable communication networks. Upon receiving the data packet, the monitoring platform immediately parses it. This parsing process primarily extracts key information fields encapsulated in the command, including but not limited to the hazard level, the target location information triggering the alarm, and the handling suggestion code automatically generated based on the operation and maintenance rule base. Subsequent actions are differentiated based on the parsed hazard level. When the hazard level is determined to be the highest level, "high-risk," it indicates that the current situation poses an immediate and serious threat to the safety of the power transmission channel, requiring immediate physical intervention. At this point, the monitoring platform or the edge device itself triggers an automated execution sequence. The goal of this sequence is to activate local intervention devices deployed on-site, particularly sound-and-light linkage devices. This device typically consists of a high-decibel sound wave transmitter and a high-intensity laser beam transmitter, mounted on a pan-tilt unit capable of precise angle adjustment. After the trigger command is issued, the target location information is extracted from the tiered alarm command. Using this coordinate data, the control system precisely drives the pan-tilt unit of the audio-visual linkage device, aligning it with the physical area where the target is located. Once the device is aligned with the target, it activates immediately. The sound wave emitter emits a strong warning sound wave, such as a pre-recorded voice warning or a piercing alarm, directed towards the target area. Simultaneously, the laser beam emitter projects a conspicuous, typically green or red, laser beam, which is highly visible even during the day, illuminating the hazardous work area or directly near the workers, creating a strong visual warning. This combined audio-visual intervention method, by directly stimulating the hearing and vision of on-site workers, aims to forcibly interrupt their attention, making them aware of the danger, thereby compelling them to stop ongoing hazardous work and quickly evacuate the danger zone. This method, by directly linking high-risk alarms with local physical intervention devices, constructs a rapid response closed loop from intelligent identification to automatic handling. It changes the traditional monitoring system's passive mode of merely "seeing" and "reporting," giving the system the ability to proactively intervene and stop dangerous behaviors. Especially for workers who approach power transmission lines due to negligence or violation of regulations, this non-contact, strong physiological stimulation method can prevent the situation from deteriorating before human intervention arrives, reducing the probability of power safety accidents caused by external factors.

[0092] In some embodiments, the method further includes:

[0093] Based on the timestamp in the hierarchical alarm command, retrieve the corresponding historical video data from the circular storage unit;

[0094] The system receives and verifies historical video data from maintenance personnel, and generates alarm verification tags.

[0095] The AI ​​model is iteratively optimized and trained by using alarm verification tags and historical video data to form a feedback dataset.

[0096] Specifically, this method constructs a closed-loop self-optimizing learning mechanism, the process of which begins with the generation of a hierarchical alarm command. After any alarm event occurs, key metadata, especially the timestamp recording the precise moment of the event, is extracted from the hierarchical alarm command. This timestamp is then used as an index to initiate a data retrieval request to the storage unit of the circular storage deployed at the edge. The goal of this request is to obtain historical video data related to the alarm event, typically a small video segment centered on the timestamp, such as 10 seconds before and 10 seconds after the event, to ensure a complete retracing of the entire event process and provide sufficient contextual information for manual review. This historical video data containing the alarm scenario is transmitted to the backend monitoring platform and presented in a visual manner to power transmission maintenance personnel with professional knowledge. The maintenance personnel's task is to review this alarm, that is, to judge the accuracy of the AI ​​model's alarm by watching the video. The review results are recorded in a structured manner, forming an alarm verification label. If the alarm is accurate, it is labeled as a "positive sample" and may include fine-tuning confirmation of the target category and location bounding box. If the alarm is incorrect, i.e., a false alarm, it is labeled as a "negative sample" and the reason for the false alarm is attached, such as "tree shadow interference" or "animal misidentification." This manual labeling, generated by the professional judgment of operations personnel, is a precise evaluation of the AI ​​model's behavior. Over time, each alarm review generates a set of data pairs, namely a piece of historical video data and its corresponding alarm verification label. These data pairs are continuously collected and accumulated to build a feedback dataset specifically for the actual application scenario of this power transmission channel. When this dataset reaches a certain scale, it is used to iteratively optimize and train the AI ​​model deployed at the edge. This process is usually done on the server side, using techniques such as transfer learning or incremental learning to fine-tune the existing AI model using this feedback dataset containing a large number of real-world positive and negative samples. After the new version of the AI ​​model is trained, it undergoes rigorous testing before being redeployed to the edge computing device to replace the old model. This iterative optimization process can reduce the long-term false alarm rate and false negative rate of the system, ensuring that the performance of the entire dynamic target detection method continuously approaches the optimal level in practical applications, and achieving truly intelligent and highly reliable operation and maintenance.

[0097] Based on the same inventive concept, such as Figure 4 As shown, the present invention also provides a real-time dynamic target detection system for power transmission channels based on edge computing, the system comprising:

[0098] The video acquisition and storage module is used to acquire real-time video streams from the power transmission channel, store them continuously in a loop, and generate continuous video data.

[0099] The dynamic target recognition module is used to analyze continuous video data frame by frame, identify dynamic targets using a pre-trained AI model, and generate dynamic target detection results.

[0100] The spatiotemporal behavior analysis module is used to fuse dynamic target detection results with the spatial information of power transmission lines, perform spatiotemporal behavior modeling, assess the motion trend and proximity of dynamic targets, and generate threat assessment vectors.

[0101] The hierarchical decision-making module is used to determine the risk level based on the threat assessment vector and combined with the power transmission operation and maintenance rules, and generate hierarchical alarm instructions.

[0102] The active intervention module is used to drive the local intervention device to perform physical intervention actions based on the hierarchical alarm instructions.

[0103] To verify the feasibility of this invention in practice, it was applied to the real-time safety monitoring of a 220kV high-voltage transmission corridor. This transmission corridor traverses an industrial park under development, where large construction machinery operates frequently. Traditional manual inspections are insufficient for full-time coverage, posing a significant risk of line safety accidents due to unauthorized operations. The purpose of this invention is to deploy an edge computing-based intelligent monitoring and proactive intervention system to achieve real-time detection, risk prediction, and automated handling of dynamic targets within this area.

[0104] In this embodiment, the industrial-grade high-definition monitoring equipment of this invention is deployed on transmission towers No. 35 and No. 36 along the transmission channel. This equipment integrates a built-in light sensor unit, which can automatically switch between day and night modes according to ambient light. It also works in conjunction with an electric robotic arm through a superhydrophobic nano-shield to ensure the lens remains clean even in rainy or dusty weather. The acquired high-definition video stream is sent to an edge computing device installed near the tower base, whose built-in storage unit is configured to cyclically store the video data of the most recent 72 hours.

[0105] To verify the effectiveness of the present invention, the system underwent continuous operation testing from August to October 2024, during which multiple potential threat events were recorded. Typical cases are selected below for illustration.

[0106] At 10:30 AM on August 15, 2024, the system's dynamic target recognition module detected a large tower crane entering the monitoring area near tower number 35 in the real-time video stream acquired by the video acquisition and storage module. The system used a hardware AI acceleration engine to decode the video in real time and fed it into a pre-trained AI model for analysis. The AI ​​model successfully identified the target as a "tower crane," output its bounding box, assigned it a unique ID through a multi-target tracking algorithm, and began continuously tracking its movement.

[0107] In the subsequent spatiotemporal behavior analysis, the system integrated the dynamic detection results of the tower crane with pre-stored 3D geographic information data of the power transmission line. Initially, the system calculated the minimum spatial distance between the end of the tower crane boom and the power transmission line to be 45 meters. As the tower crane rotated, the system continuously updated the threat assessment vector. By 10:32, the system detected that the boom was approaching the power transmission line at a speed of approximately -1.5 m / s, and the real-time spatial distance had decreased to 22 meters. At this point, the system's threat assessment vector H was updated, where the spatial distance... meters, approach speed m / s, target category C = "large machinery", time spent in the danger zone More than 2 minutes have passed. Meanwhile, the trajectory prediction algorithm based on the Kalman filter determines that if the current trend continues, the target will enter the 20-meter absolute safety red line within 10 seconds.

[0108] Based on this threat assessment vector, the hierarchical decision-making module, combined with built-in power transmission operation and maintenance rules, determines the event to be at a "high-risk" level and generates a preliminary alarm. The alarm filtering mechanism then activates, comparing the alarms with the historical alarm cache to confirm that this is a new event rather than a duplicate alarm; simultaneously, by analyzing the visual features of the image region, the possibility of background interference is eliminated, generating a valid alarm. Finally, the system generates a hierarchical alarm instruction containing the "high-risk" level, the target's precise coordinates, and handling suggestions.

[0109] At 10:32:05, the command triggered the active intervention module to execute a physical intervention action. The audio-visual linkage device located on tower No. 35 was activated. Based on the coordinate data provided by the command, its pan-tilt unit precisely oriented towards the tower crane cab, emitting a high-decibel voice warning: "Warning, high voltage danger, please stop work immediately and evacuate," while simultaneously projecting a bright red laser beam onto the boom. Upon receiving the strong audio-visual warning, the on-site personnel immediately stopped the tower crane's rotation at 10:32:20 and moved the boom away from the power lines. By 10:33:15, the system confirmed that the target had retreated to a safe distance of 30 meters, the risk level was reduced to "low," and the alarm was automatically deactivated.

[0110] Furthermore, the closed-loop self-optimization mechanism of this invention has been verified. During a strong wind event in early July 2024, the system generated a "medium-risk" false alarm due to the violent swaying of a large tree's crown. After the operations and maintenance personnel reviewed the historical video of the event retrieved by the system on the backend platform, they generated an alarm verification label of "negative sample - tree shadow interference." This data, along with the video clip, was added to the feedback dataset. After a month of data accumulation and one iteration of model optimization training, the new model was redeployed to the edge in early August. Data shows that the optimized model has enhanced ability to identify background interference such as swaying tree shadows, and the false alarm rate decreased by approximately 85% under subsequent similar weather conditions.

[0111] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0112] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A real-time dynamic target detection method for power transmission channels based on edge computing, characterized in that, The method includes the following steps: The system acquires real-time video streams from power transmission channels, stores them continuously in a loop, and generates continuous video data. At the edge, the continuous video data is analyzed frame by frame, and a pre-trained AI model is used to identify dynamic targets and generate dynamic target detection results. By integrating the dynamic target detection results with the spatial information of the power transmission line, spatiotemporal behavior modeling is performed to assess the movement trend and proximity of the dynamic target and generate a threat assessment vector. Specifically, this involves: performing time-series analysis on the dynamic target detection results to track the movement trajectory of the dynamic target and generate a trajectory sequence; acquiring the geographic information data of the power transmission line and calculating the spatial distance between the dynamic target and the power transmission line based on the target location in the dynamic target detection results; and fusing the trajectory sequence with the spatial distance to quantify the potential threat and generate a threat assessment vector. Based on the threat assessment vector, risk level is determined in conjunction with power transmission operation and maintenance rules, and graded alarm instructions are generated. Based on the hierarchical alarm instructions, the local intervention device is driven to perform physical intervention actions.

2. The real-time dynamic target detection method for power transmission channels based on edge computing according to claim 1, characterized in that, The generation of continuous video data specifically involves: Real-time video streams are collected by industrial-grade high-definition monitoring equipment deployed in power transmission channels. The industrial-grade high-definition monitoring equipment automatically switches between day and night modes using a built-in light sensor unit, and actively cleans the lens in conjunction with an electric robotic arm and a superhydrophobic nano-shield architecture. The real-time video stream is sent to a storage unit configured for circular storage for full-time recording; The video clip containing the current moment is extracted from the storage unit to generate continuous video data.

3. The real-time dynamic target detection method for power transmission channels based on edge computing according to claim 2, characterized in that, The generation of dynamic target detection results specifically includes: The continuous video data is decoded using a hardware AI acceleration engine to generate video frames; The video frames are input into a pre-trained AI model for frame-by-frame analysis to identify dynamic targets and output bounding boxes and categories, thus forming dynamic target parameters. By integrating the dynamic target parameters, the target is tracked and its motion state is calculated to generate dynamic target detection results.

4. The real-time dynamic target detection method for power transmission channels based on edge computing according to claim 3, characterized in that, The generated trajectory sequence is specifically as follows: Calculate the velocity vector based on the positional changes of the dynamic target in consecutive frames; Based on the velocity vector, the possible position of the dynamic target at a future time point is predicted, thus forming predicted position data; By combining the historical position, velocity vector, and predicted position data of the dynamic target, a trajectory sequence is updated and generated.

5. The real-time dynamic target detection method for power transmission channels based on edge computing according to claim 4, characterized in that, The specific method for generating hierarchical alarm instructions is as follows: Based on the threat assessment vector and combined with power transmission operation and maintenance rules, the risk level is determined and a preliminary alarm is generated. An alarm filtering mechanism is applied to screen the preliminary alarms, filtering out invalid alarms caused by background interference and duplicate alarms caused by the same event, so as to obtain valid alarms. Based on the risk level and target location information of the valid alarms, a graded alarm command is generated.

6. The real-time dynamic target detection method for power transmission channels based on edge computing according to claim 5, characterized in that, The alarm filtering mechanism is as follows: The initial alarm is compared with the records in the historical alarm cache to identify and filter out duplicate alarms; Analyze the visual features of the preliminary alarm image area to identify and filter out background interference alarms caused by environmental factors.

7. The real-time dynamic target detection method for power transmission channels based on edge computing according to claim 6, characterized in that, The driving local intervention device performs physical intervention actions, specifically as follows: The graded alarm commands are sent to the backend monitoring platform through the communication network, the graded alarm commands are parsed, and the hazard level, target location information and handling suggestions are obtained; When the hazard level is high-risk, the sound and light linkage device is triggered to emit sound waves and laser beams in a directional manner to the area indicated by the target location information, forcing the workers to stop the dangerous operation and evacuate through physiological stimulation.

8. The real-time dynamic target detection method for power transmission channels based on edge computing according to claim 7, characterized in that, The method further includes: Based on the timestamp in the hierarchical alarm instruction, retrieve the corresponding historical video data from the circular storage unit; The operation and maintenance personnel review the historical video data and generate alarm verification tags. The AI ​​model is iteratively optimized and trained using the alarm verification tags and the historical video data to form a feedback dataset.

9. A real-time dynamic target detection system for power transmission channels based on edge computing, applied to the real-time dynamic target detection method for power transmission channels based on edge computing as described in any one of claims 1-8, characterized in that, The system includes: The video acquisition and storage module is used to acquire real-time video streams from the power transmission channel, store them continuously in a loop, and generate continuous video data. The dynamic target recognition module is used to analyze the continuous video data frame by frame, identify dynamic targets using a pre-trained AI model, and generate dynamic target detection results. The spatiotemporal behavior analysis module is used to fuse the dynamic target detection results with the spatial information of the power transmission line, perform spatiotemporal behavior modeling, evaluate the movement trend and proximity of the dynamic target, and generate a threat assessment vector. Specifically, it performs time-series analysis on the dynamic target detection results, tracks the movement trajectory of the dynamic target, and generates a trajectory sequence; acquires the geographic information data of the power transmission line, and calculates the spatial distance between the dynamic target and the power transmission line by combining the target location in the dynamic target detection results; and fuses and analyzes the trajectory sequence and the spatial distance to quantify the potential threat and generate a threat assessment vector. The hierarchical decision-making module is used to determine the risk level based on the threat assessment vector and in combination with power transmission operation and maintenance rules, and generate hierarchical alarm instructions. The active intervention module is used to drive the local intervention device to perform physical intervention actions according to the hierarchical alarm instructions.

Images