Extra-high voltage main grid power transmission channel hidden danger identification method and system based on artificial intelligence large and small model cooperation

Abstract

The invention discloses an extra-high voltage main grid power transmission channel hidden danger identification method and system based on artificial intelligence large and small model cooperation. According to the method, images and environment metadata are collected at the edge end, and scene categories are judged and preprocessed based on the metadata; a lightweight first computer vision model is used for rapid preliminary screening, and a preliminary alarm with a scene label is generated and uploaded to a cloud end. And the cloud side dispatches a second computer vision large model to carry out fine review according to the alarm type and scene, selectively calls a time sequence analysis model or a multi-modal model to carry out trajectory analysis and cross-modal reasoning, realizes collaborative decision making and misjudgment correction, and finally outputs graded alarm. The system correspondingly comprises an edge end acquisition processing and preliminary screening module and a cloud end re-checking and outputting module. According to the invention, through cloud edge collaboration and multi-model division of labor, the real-time performance is ensured, and the recognition precision and the adaptive ability in a complex scene are greatly improved at the same time.

Description

Technical Field

[0001] This invention relates to the field of hazard identification in ultra-high voltage (UHV) main grid transmission channels, and in particular to a method and system for hazard identification in UHV main grid transmission channels based on the collaboration of artificial intelligence large and small models. Background Technology

[0002] With the continuous expansion of the ultra-high voltage power grid, the safe operation and maintenance of its transmission channels face severe challenges. Traditional manual inspection methods are inefficient and risky, making it difficult to meet the real-time monitoring needs of long-distance, large-scale lines. In recent years, intelligent inspection technology based on computer vision has been widely used, mainly by collecting images through cameras deployed on towers or drones and using artificial intelligence models to identify potential hazards.

[0003] Current technical solutions have the following limitations: First, deploying high-precision, large-scale visual models directly at the edge is limited by the computing power and power consumption of edge devices, making real-time response difficult. Furthermore, transmitting all raw images to the cloud consumes significant bandwidth. Second, while deploying lightweight models at the edge ensures real-time performance, the model's recognition accuracy is insufficient in the complex and variable environments of UHV transmission channels (such as windy areas, salt spray areas, and mountainous regions), resulting in high false alarm and false negative rates and generating numerous invalid alarms, significantly increasing the burden of manual review at the backend. In addition, existing solutions often employ single models or simple model concatenation, lacking refined and collaborative recognition mechanisms for different hazard types (such as static equipment defects, dynamic engineering machinery, and easily confused wildfire smoke) and different geographical and climatic scenarios. This makes it impossible to achieve high-precision, low-false-alarm hazard identification while ensuring real-time performance.

[0004] Therefore, how to design an intelligent hazard identification method that can balance identification efficiency and accuracy and adapt to complex ultra-high voltage scenarios has become an urgent technical problem to be solved in this field. Summary of the Invention

[0005] To address the problems existing in the prior art, the main objective of this invention is to provide a method and system for identifying hidden dangers in ultra-high voltage main grid transmission channels based on the collaboration of large and small models in artificial intelligence. By constructing a collaborative processing architecture consisting of preliminary screening by lightweight edge models and collaborative verification by large models in the cloud, and introducing a scene adaptation mechanism based on auxiliary metadata, this invention resolves the contradiction in the prior art where real-time identification at the edge side is insufficient in accuracy and has a high false alarm rate, while high-precision identification in the cloud side suffers from poor real-time performance and high resource consumption. Thus, under limited bandwidth and computing resources, this invention achieves efficient, accurate, and scene-adaptive identification of hidden dangers in transmission channels.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for identifying hidden dangers in ultra-high voltage main grid transmission channels based on the collaboration of artificial intelligence large and small models, applied to a hidden danger identification architecture composed of edge terminals and cloud. The method includes the following steps: Acquire monitoring image data and auxiliary metadata of the power transmission channel; based on the auxiliary metadata, determine the scene category to which the monitoring image belongs, and generate a structured data packet containing monitoring image data and scene category tags; The structured data packet is analyzed using a first computer vision model deployed at the edge, and the hazard image data, hazard type, and confidence level are output. A corresponding discrimination threshold is selected according to the scene category label. If the confidence level is not lower than the selected discrimination threshold, preliminary alarm data containing the hazard image data, hazard type label, and scene category label is generated and uploaded to the cloud. The preliminary alarm data is reviewed using a large model deployed in the cloud. The large model includes a second computer vision model and at least one of a time series analysis model or a multimodal model. The preliminary alarm data is classified according to the hazard type label and the scene category label. The second computer vision model corresponding to the hazard type and scene category is called according to the classification, and the hazard image data is reviewed. Based on the review result of the second computer vision model, the time series analysis model and / or multimodal model selected by the classification is used to perform a collaborative review of the preliminary alarm data. Based on the type of hazard confirmed during the review, an alarm message is output.

[0007] In the above method, preferably, the first computer vision model is a lightweight target detection model deployed at the edge; The first computer vision model is configured to: pre-set parameter templates associated with different scene categories, the parameter templates containing discrimination thresholds for different hazard types; load the corresponding parameter templates according to the scene category labels, and judge the confidence level based on the discrimination thresholds in the templates; and / or, The first computer vision model is further configured to: pre-set feature extraction modules associated with different scene categories, activate the feature extraction module corresponding to the scene category label, and perform feature extraction of monitoring image data based on the feature extraction module corresponding to the scene.

[0008] In the above method, preferably, the second computer vision model is a large visual model with more parameters than the first computer vision model, and is fine-tuned by a sample library of ultra-high voltage transmission channel scenarios. The second computer vision model is configured to: activate the corresponding scene-based feature library based on the scene category label, and perform multi-scale feature fusion and target recognition optimization on the hidden danger image data; The output of the second computer vision model includes: if the preliminary alarm data needs to be corrected, the reason for correction is marked; if the preliminary alarm data is confirmed, the hazard type label and the coordinate information corresponding to the label are retained.

[0009] In the above method, preferably, the time-series analysis model is based on a long short-term memory network architecture and is used to verify the preliminary alarm data generated in scenarios where the device is stationary or recurring based on the classification. The time-series analysis model is configured to: input the coordinates and time information of the current alarm, and perform trajectory analysis by combining historical alarm records within a preset time period, so as to filter duplicate alarms or identify high-risk alarms that are approaching the transmission line.

[0010] In the above method, preferably, the multimodal model is based on a vision-language pre-training architecture and is used to verify the potential hazards of environmental disasters or equipment defects determined based on the classification. The multimodal model includes: inputting the hazard image data and a preset text reasoning question, and performing cross-modal reasoning by matching image features with question semantics to confirm the authenticity of the hazard.

[0011] In the above method, preferably, the auxiliary meta-structured data includes spatiotemporal data, environmental data, and device data; the step of determining the scene category based on the auxiliary metadata specifically includes: determining the scene category as a windy area, a salt fog area, or a mountainous area based on the wind speed, humidity, terrain type, or climate zone attributes in the environmental data; the scene category label is used to identify the determination result.

[0012] Preferably, the above method further includes: performing risk assessment and classification based on the confirmed hazard type and its scenario category label, generating graded alarm information and pushing it to the operation and maintenance terminal; at the same time, recording the alarm details, review results and handling status to the hazard management ledger.

[0013] Preferably, the above method further includes a model iteration step: collecting manual review feedback on alarm information and constructing a labeled sample library; using the labeled sample library to incrementally train and update the parameter templates of the first computer vision model, and / or the second computer vision model, the time series analysis model, and the multimodal model.

[0014] Preferably, the above method further includes a resource orchestration and scheduling step: configuring the collaborative workflow and triggering rules of the first computer vision model, the second computer vision model, the time series analysis model, and the multimodal model through a visual interface; and monitoring the computing load at the edge and the cloud, dynamically scheduling tasks and elastically scaling computing resources.

[0015] In a second aspect, the present invention provides a hidden danger identification system for ultra-high voltage main grid transmission channels based on the collaboration of artificial intelligence large and small models, applied in a hidden danger identification architecture composed of edge and cloud, characterized in that it includes: The edge data acquisition and processing module is used to acquire monitoring image data and auxiliary metadata of the power transmission channel; based on the auxiliary metadata, it determines the scene category to which the monitoring image belongs and generates a structured data packet containing monitoring image data and scene category labels; The edge-based initial screening module is equipped with a first computer vision model, which is used to analyze the structured data packet using the first computer vision model, output hazard image data, hazard type and confidence level; and to select the corresponding discrimination threshold according to the scene category label. If the confidence level is not lower than the selected discrimination threshold, preliminary alarm data containing the hazard image data, hazard type label and scene category label is generated and uploaded to the cloud. The cloud-based review module deploys a large model to review preliminary alarm data. This large model includes a second computer vision model and at least one of a time-series analysis model or a multimodal model. It is used to classify the preliminary alarm data based on hazard type labels and scene category labels, call the corresponding second computer vision model according to the classification, and perform review of the hazard image data. Based on the review results of the second computer vision model, it uses the time-series analysis model and / or multimodal model selected by the classification to perform collaborative review of the preliminary alarm data. The alarm output module is used to output alarm information based on the type of hidden danger confirmed by the review.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) In this invention, at the resource-constrained edge, a lightweight first computer vision model is used to quickly and initially screen massive monitoring images, filtering out most scenes without hidden dangers. Only a small amount of "preliminary alarm data" containing suspicious targets is uploaded to the cloud, which greatly reduces the amount of data to be uploaded, reduces network bandwidth pressure, and ensures the overall response efficiency of the system. Subsequently, in the resource-rich cloud, a second computer vision model with larger parameters and stronger capabilities is used, along with a combination of a time-series analysis model and / or a multimodal model, to conduct in-depth and detailed collaborative verification of the small number of high-value targets reported by the edge. This division of labor and cooperation mechanism decouples rapid filtering and accurate analysis at the system architecture level, allowing the edge to focus on efficiency and the cloud to focus on accuracy, thus achieving an organic unity of efficiency and accuracy at the system level. Overall, an effective balance between recognition efficiency and recognition accuracy is achieved.

[0017] (2) In this invention, a scene category determination and tagging mechanism based on auxiliary metadata is introduced. Thus, in the edge processing stage, the system not only analyzes the image content but also dynamically determines the scene category (e.g., windy areas, salt spray areas) based on auxiliary metadata such as wind speed, humidity, and geographical location. The scene category tag is then bound to the image data and applied throughout all subsequent processing stages. This allows the first computer vision model to select different discrimination thresholds based on the scene category tag (e.g., lowering the threshold in windy areas to increase sensitivity). It also enables the second computer vision model in the cloud to activate different recognition strategies or feature libraries based on the tag. This allows the entire system to dynamically adjust its behavior to adapt to different environmental interferences, significantly improving the recognition accuracy and robustness under various complex and harsh scenarios of ultra-high voltage, and effectively reducing misjudgments and missed judgments caused by environmental factors. Overall, it improves the recognition accuracy and adaptability in complex scenarios.

[0018] (3) In this invention, the collaborative review logic based on hazard type organically combines different types of models. The cloud first uses a second computer vision model to perform a preliminary review in a general and refined manner. Then, based on the "hazard type label" obtained from the preliminary review, it intelligently schedules its time series analysis model or (and) multimodal model for a second review. For example, for "engineering machinery" targets, the time series model is called to analyze its trajectory to filter duplicate alarms; for "wildfire smoke" targets, the multimodal model is called to perform image-text association reasoning to distinguish between true and false alarms. This type-based, serial-parallel combined model scheduling strategy fully leverages the professional advantages of different types of large models, forming a collaborative decision-making synergy, thereby achieving more targeted and accurate identification of various hazards. In particular, it effectively solves complex problems such as trajectory judgment and semantic understanding that are difficult to handle by traditional single vision models.

[0019] The invention will now be further described with reference to the accompanying drawings. Attached Figure Description

[0020] Figure 1 This is a collaborative flowchart of the method for identifying hidden dangers in ultra-high voltage main grid transmission channels based on the collaboration of artificial intelligence large and small models provided in this embodiment of the invention; Figure 2 This is a data model iterative dual-loop mechanism architecture diagram of the UHV main grid transmission channel hidden danger identification method based on artificial intelligence big-small model collaboration provided in the embodiments of the present invention; Figure 3 This is an alarm screening rule diagram of the method for identifying hidden dangers in ultra-high voltage main grid transmission channels based on the collaboration of artificial intelligence large and small models provided in this embodiment of the invention. Detailed Implementation

[0021] To better illustrate the objectives, technical solutions, and advantages of the present invention, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0022] like Figures 1 to 3 As shown, this embodiment of the invention provides a method for identifying potential hazards in ultra-high voltage (UHV) main grid transmission channels based on a collaborative approach using artificial intelligence-based large and small models. This method operates within a collaborative processing architecture comprised of edge servers (including edge servers deployed near transmission towers, smart cameras, or inspection drones) and cloud servers. Please refer to... Figure 1 It shows a schematic diagram of the overall process of one embodiment of the method, which specifically includes the following steps: S100: Acquire monitoring image data and auxiliary metadata of the power transmission channel; based on the auxiliary metadata, determine the scene category to which the monitoring image belongs, and generate a structured data packet containing monitoring image data and scene category labels; S200: Analyze the structured data packet using a first computer vision model deployed at the edge, and output the hidden danger image data, hidden danger type and confidence level; select the corresponding discrimination threshold according to the scene category label; if the confidence level is not lower than the selected discrimination threshold, generate preliminary alarm data containing the hidden danger image data, hidden danger type label and scene category label, and upload it to the cloud. S300: The preliminary alarm data is reviewed using a large model deployed in the cloud. The large model includes a second computer vision model and at least one of a time series analysis model or a multimodal model. The preliminary alarm data is classified according to the hazard type label and the scene category label. The second computer vision model corresponding to the hazard type and scene category is called according to the classification, and the hazard image data is reviewed. Based on the review result of the second computer vision model, the time series analysis model and / or multimodal model selected by the classification is used to perform a collaborative review of the preliminary alarm data. S400: Output alarm information based on the type of hazard confirmed by the review.

[0023] In step S100, specifically, acquiring the number of monitoring images of the power transmission channel is the foundation and starting point of the entire method. Its core objective is to obtain high-quality, full-dimensional raw data for hazard identification and analysis. In practice, the data acquisition operation simultaneously performs two main aspects: image information acquisition and auxiliary metadata acquisition, to ensure that subsequent analysis has sufficient contextual information.

[0024] In terms of image information acquisition, the system relies on a network of terminal monitoring equipment pre-deployed along the UHV transmission lines. This network typically consists of two types of equipment: one is fixed, high-definition cameras with night vision and optical zoom capabilities, used for continuous monitoring of key sections; the other is drones that perform inspection tasks periodically or on demand, which can cover blind spots of fixed cameras or provide more flexible inspection perspectives. Through these devices, the system acquires images of the transmission lines and their surrounding environment in real-time or near real-time using streaming or timed snapshot methods. The acquired images need to cover various potential hazard scenarios, including but not limited to: construction machinery (such as excavators, cranes, tower cranes, and other construction equipment), environmental disasters (such as wildfires and their associated smoke), foreign objects on the conductors (such as hanging plastic film, kite strings, etc.), and equipment defects (such as insulator damage, hardware corrosion, etc.).

[0025] To optimize resource utilization and increase monitoring density in high-risk areas, the frequency of image acquisition is not fixed but dynamically adjusted based on the environmental characteristics and risk levels of different sections of the channel. A typical adjustment strategy is as follows: in "windy areas" where there is a significant risk of wind damage and debris flow, the acquisition interval is shortened to once every 5 minutes to achieve more intensive monitoring; while in "normal areas" where the meteorological and geographical environment is relatively stable and the risk is low, the acquisition interval is set to once every 10 minutes, reducing data load and processing pressure while ensuring monitoring effectiveness. This differentiated frequency setting achieves intelligent matching between monitoring resources and risk levels, ensuring safety redundancy in high-risk areas while avoiding unnecessary computing and storage overhead in low-risk areas.

[0026] Regarding the acquisition of auxiliary metadata for power transmission channels, specifically, the system automatically associates and records various contextual information related to each frame of image capture. This metadata is crucial for understanding and accurately analyzing image content, and it mainly falls into three categories. The first category is spatiotemporal data, which precisely identifies "when and where" the image was captured. This includes the precise timestamp of image acquisition, the spatial coordinates of the acquisition equipment (usually expressed in latitude and longitude), the tower number associated with the equipment, and the name of the line to which it belongs. This data provides the foundation for subsequent hazard location, trajectory tracking, and spatiotemporal correlation analysis. The second category is environmental data, which describes the on-site physical conditions at the time the image was captured. This includes meteorological parameters such as temperature, humidity, wind speed, and rainfall, as well as macro-terrain type (e.g., mountainous, plain, or coastal) and long-term climate zone attributes (e.g., windy, salt spray, or high-altitude cold regions) determined by the equipment location. The technical purpose of collecting this type of data is to provide parameter basis for subsequent image preprocessing algorithms (such as defogging and image stabilization), and to provide key context for the hazard identification model to understand specific environmental interferences (such as salt spray corrosion and strong wind vibration), thereby improving the robustness and accuracy of the model in complex natural environments. The third type is equipment data, which records the technical status of the data source. Specifically, this includes the hardware model of the monitoring terminal, the angle and focal length parameters of the camera during shooting, and the historical maintenance records of the monitoring point (e.g., the types of hazards that frequently occurred at this location in the past). The effect of collecting equipment data is twofold: firstly, it can be used for data quality tracing and calibration; secondly, historical maintenance records can provide valuable prior knowledge for AI models, enabling them to pay more attention to historically high-incidence hazards when facing specific areas, achieving localized optimization of risk warnings.

[0027] In this embodiment of the invention, after data acquisition, the raw image data needs to undergo scene-specific preprocessing. The technical purpose of this step is to transform the raw images, which have different qualities and formats, into a standardized, high-quality set of input data that is highly adapted to the subsequent analysis scenario, thereby fundamentally improving the accuracy and reliability of subsequent AI model recognition. The preprocessing process includes three stages: basic preprocessing, scene adaptation processing, and data encapsulation.

[0028] Basic preprocessing primarily addresses the standardization of image data. Specifically, the system first denoises the acquired raw images to eliminate random noise interference introduced by the camera sensor or during transmission. Then, the resolution of all images is uniformly adjusted to the standard 1080P to ensure consistent input data size and meet the model's stringent input format requirements. Finally, the image files are uniformly converted to common compression formats such as JPEG. The technical benefit of this series of operations is the construction of a clean, uniformly sized image dataset, providing a stable and reliable foundation for subsequent feature extraction and model computation, and avoiding performance fluctuations caused by inconsistent data quality.

[0029] In terms of determining the scene category to which the monitoring image belongs, specifically, scene adaptation processing implements customized image enhancement and restoration operations to address the special and complex environmental challenges faced by ultra-high voltage transmission channels.

[0030] For example, in windy areas where the camera shakes slightly due to wind, causing motion blur, the system will invoke an image jitter correction algorithm. One possible implementation is a motion compensation technique based on optical flow. This algorithm can estimate and inversely compensate for the motion vectors between image frames, thereby effectively restoring image sharpness. The direct technical effect is that it makes previously blurry engineering machinery or small foreign objects captured in strong winds appear clearly defined, greatly reducing the risk of missed recognition due to image blur.

[0031] To address the occlusion effect of salt fog particles on camera lenses in coastal salt fog areas, the system employs a fog removal algorithm. A typical implementation is the dark channel prior algorithm, which, based on statistical prior knowledge of fog-free outdoor images, can effectively estimate and remove the fog layer from the image. The technical advantage of implementing this algorithm is that it significantly improves image clarity and contrast, removes the occlusion of equipment surface details (such as insulator damage and metal corrosion spots) by salt fog, and makes these key defect features stand out in the image, providing high-quality input for subsequent defect recognition models.

[0032] To address the issue of low contrast between targets and backgrounds caused by complex mountainous terrain and dense vegetation, the system employs image enhancement algorithms. For example, histogram equalization can be used to redistribute pixel intensity, thereby expanding the dynamic range of the image. The effect is that it significantly improves the visual contrast between potentially hazardous targets (such as small objects partially obscured by vegetation, or grounding wires that are similar in color to the soil) and complex backgrounds, making target features more prominent and helping the model accurately separate targets from cluttered backgrounds.

[0033] After optimizing the image content, the system enters the data encapsulation stage. In this stage, the pre-processed high-quality image is logically bound to its associated metadata (such as time, coordinates, and environmental parameters) from the acquisition phase, encapsulating it into a structured "image-attribute" data packet. Simultaneously, based on the environmental characteristics of the image source (e.g., wind speed indicating a windy area) and the preset monitoring task, a clear scene category label is assigned to this structured data packet, such as "Windy Area - Foreign Object Monitoring" or "Salt Spray Area - Equipment Corrosion." The encapsulated structured data packet is then transmitted to the designated edge computing node via the internal network. The technical effect of this approach is to achieve lossless association and efficient organization of data and contextual information, ensuring that edge nodes, upon receiving image data, also obtain its complete scene context. This allows them to intelligently load the most suitable AI model and parameters for processing the scene, providing ample data preparation for subsequent real-time edge analysis.

[0034] In step S300, specifically in this embodiment, the collaborative verification of large and small cloud-based models constitutes the core of accuracy assurance and decision-making for the entire hazard identification process. The technical objective of step S200 is to utilize the abundant computing resources and storage capacity of the cloud to deploy a series of specialized models with large parameter scales and deep understanding capabilities. This allows for multi-angle, in-depth cross-verification and comprehensive analysis of the initial alarms reported from the edge, ultimately eliminating false alarms, confirming missed alarms, and assessing the risk level, achieving a leap from "preliminary perception" to "precise understanding." Its implementation process is a logically coherent and clearly defined collaborative analysis chain, beginning with the intelligent routing of alarm data and proceeding through three types of large models, each with its own expertise, for targeted processing. Step S200 specifically includes: When the cloud-based collaborative hub receives the initial structured alarm data packet uploaded via encryption from the edge node, the first step is data reception and intelligent classification. The system parses the scene category label and "hazard type" carried in the structured data packet, using these two key metadata categories to automatically classify and route alarms from a dual-dimensional perspective. For example, based on the scene dimension, alarms are divided into different queues such as "high-wind zone alarms," ​​"salt fog zone alarms," ​​and "mountain area alarms"; based on the hazard type dimension, they are categorized into "construction machinery" (including excavators, cranes, etc.), "environmental disasters" (including wildfires, smoke), and "equipment defects" (including insulator damage, wire corrosion, etc.). The core technical effect of this classification operation is that it establishes a clear task context and pipeline for the subsequent review process, enabling models with different specializations to efficiently handle the alarm categories they are best at, achieving refined management of computing resources and optimization of task allocation.

[0035] After classification, the process enters the target recognition optimization stage of the large-scale computer vision model. This is the first step of collaborative review, aiming to "review" the alarm images using a more powerful and specialized visual model. The system dynamically calls the corresponding customized second computer vision model based on the alarm's classification label. This large model is typically based on a general-purpose visual foundation model (such as Vision Transformer or a large-scale convolutional network) and is deeply fine-tuned by incorporating over 300,000 scene-specific samples from ultra-high voltage transmission channels. Its training sample library specifically covers highly industry-specific data such as foreign objects in windy areas and corrosion in salt spray areas.

[0036] During inference, the model activates its internal "feature library" relevant to the current scene. For example, when handling alarms in windy areas, it focuses on the texture features of plastic films and kite strings, while when handling alarms in salt spray areas, it focuses on the morphological patterns of corrosion spots and seal failures. By applying advanced algorithms such as multi-scale feature fusion, this large model can perform more refined image analysis and use this to compare and verify the initial screening results of the smaller edge models. Its core function is to correct visual misjudgments caused by scene interference. A typical technical example is distinguishing between natural clouds and fog common in mountainous areas and real smoke: the model can accurately judge by analyzing the texture uniformity, diffusion pattern, and correlation with terrain heat sources (if present or absent); another example is distinguishing between normal aging and dangerous corrosion of equipment in salt spray areas: the model identifies this by recognizing the irregularity of spots, abnormal colors, and diffusion trends. After this round of optimization, the system outputs a corrected result. If the initial screening conclusion is overturned, the specific reason is marked; if confirmed, the hazard information is retained, thus significantly improving the accuracy of judgment based on a single visual dimension.

[0037] Next, for specific types of alarms, the system will initiate a time-series analysis model for repeated alarm filtering. This step is specifically applied to "construction machinery" hazards, aiming to solve the problem of alarm redundancy caused by such targets remaining in the monitoring field of view for extended periods or repeatedly appearing.

[0038] The system invokes a time-series analysis model, which is built on a time-series prediction architecture based on Long Short-Term Memory (LSTM) networks and trained using three years of historical power transmission line engineering machinery trajectory data. During analysis, the model receives the coordinates and acquisition time of the current alarm device and automatically queries historical alarm records for that coordinate point within the last 72 hours. Based on this, the model makes two key judgments: First, trajectory judgment: if the analysis finds that the spatial movement distance of the same target within the last 4 hours is extremely small (e.g., less than 10 meters under mountainous terrain constraints), or if historical records show that a work order has been generated for this target but has not yet been closed, the alarm will be judged as a "duplicate alarm" and marked as invalid. Second, trend judgment: if the model analysis finds that the target's coordinate sequence shows a trend of gradually approaching the power transmission line or tower, even if it is a new target, it will be marked as a "high-risk alarm" to increase its handling priority. The technical effect of this step is that by introducing time-dimensional analysis, it enables intelligent perception and alarm aggregation of static or continuous risks, greatly reducing the interference of invalid alarms to operation and maintenance personnel, while strengthening the early warning capability for progressive risks.

[0039] Finally, for complex hazards that are visually easily confused, the system employs a multimodal model for reasoning in complex scenarios. This process primarily targets alarms susceptible to environmental interference, such as those related to "environmental disasters" (e.g., smoke, wildfires) and "equipment defects" (e.g., corrosion in salt spray areas). The multimodal model invoked by the system is based on a vision-language pre-training architecture (e.g., CLIP), and incorporates professional terminology and scene descriptions from the power industry during training. Its operation differs from traditional visual models; it employs a cross-modal reasoning path: the system inputs the alarm image, scene category label, and a predefined reasoning question expressed in natural language, such as "Is the target area in a windy zone a plastic film drifting in the wind?" or "Are the spots on the equipment surface in a salt spray area corrosion defects?". The core mechanism of the model is to perform "image feature-language semantics" matching and alignment. It deeply correlates and calculates probabilities between visual features extracted from the image (such as the light transmittance and floating shape of the plastic film, and the irregular outline and color of corrosion spots) and semantic concepts in the question text (such as "floating in the wind" and "corrosion defects"), ultimately generating an inference conclusion and a corresponding confidence level. For example, the model may infer that the white clumps in the image are "natural clouds and fog, without heat source features," rather than smoke. The system is set to only confirm the alarm as valid when the inference conclusion is "a hidden danger" and the confidence level reaches or exceeds 85%; otherwise, it is marked as an "invalid alarm," and the model's reasoning basis is recorded (such as "the clouds and fog blend with the surrounding terrain, with no trace of artificial creation"). The significant technical effect of this step is that by introducing prior knowledge of language semantics and an understanding of common sense, it breaks through the limitations of pure visual models in the feature similarity dilemma, providing a new and more reliable solution to solve complex recognition problems that rely on context and logical judgment (such as distinguishing between smoke and clouds and fog with similar shapes).

[0040] In step S400, the hierarchical output and ledger recording of alarm results are the value realization and closed-loop management links of the hazard identification process. Its technical purpose is to transform the accurate analysis results generated by cloud verification into structured instructions with clear priorities and complete contextual information that can guide on-site operation and maintenance actions, and to form traceable and auditable data assets, thereby completing the final transformation from "intelligent perception" to "scientific decision-making" and "effective control." The implementation of this link ensures that the identification results can be efficiently and orderly integrated into the existing production management system, improving the overall intelligence and refinement of operation and maintenance response.

[0041] Specifically, in step S400, during the alarm classification phase, after the system completes the review in the cloud, it immediately performs automated classification of alarms based on a set of composite assessment rules. The core of these rules is a comprehensive decision based on two dimensions: the inherent risk level of the hazard type and the severity of the hazard's occurrence scenario. Based on this, the reviewed alarms are systematically divided into three levels. The first level is "emergency alarms," ​​which are defined as hazards that, if they occur, could directly cause power grid operation accidents or serious equipment damage. Typical examples include: ongoing wildfires, high-risk construction machinery (such as crane booms too close to conductors, within dangerous operating distances), and severe equipment corrosion in salt spray areas (such as insulators showing visible cracks). These alarms represent the highest level of immediate risk, and the system sets extremely strict response time limits for them, such as requiring the emergency response process to be initiated within 15 minutes. The second level is "routine alarms," ​​which refer to hazards that have been confirmed to exist but will not immediately cause an accident and have a certain buffer period for handling. Examples include low-risk, small-volume floating objects (such as a piece of plastic strap hanging in a non-critical location) or minor signs of aging on equipment (such as paint peeling without structural corrosion). For these types of alarms, the system sets a relatively long processing window, such as requiring verification or handling within 24 hours. The third level is "invalid alarms," ​​specifically referring to alarms that have been explicitly rejected after multi-model verification in the cloud, such as false alarms confirmed as natural fog or duplicate construction machinery alarms that have been filtered out by the time-series model. The system filters these alarms directly at the front end, without pushing them to the operations and maintenance personnel interface, thus avoiding information interference. The technical effect of this hierarchical mechanism is that it achieves precise matching of risk resources, ensuring that limited human and emergency resources can be prioritized for the most urgent and dangerous threats, while shielding a large amount of low-value interfering information from the operations and maintenance personnel's view, greatly improving the decision-making efficiency of operations and maintenance command and the scientific nature of emergency response.

[0042] Specifically, in step S400, during the result output phase, the system pushes alarms to relevant systems and personnel in real time, based on the graded results, in different forms and with different content. The main recipients of the push include the backend "Power Transmission Production Operation Support System" and the mobile terminals of frontline maintenance personnel (such as through the enterprise-level instant messaging application elink). The system employs a differentiated design in organizing the output content. For emergency alarms, the push information must be highly action-oriented, therefore it usually includes the most critical decision support elements, namely "on-site images of the hazard," "precise geographical coordinates," and specific "preliminary handling suggestions." For example, a push message for a wildfire alarm might be: "[Wildfire Alarm] An open flame has been discovered 50 meters southeast of XX tower, coordinates (X,Y). It is recommended to immediately dispatch nearby drones to the site for verification and notify the fire department." For routine alarms, the push information focuses more on providing background knowledge to assist in the assessment, usually including "detailed description of the hazard" and "reference to similar historical cases," helping maintenance personnel evaluate its prevalence and handling methods. In addition to targeted push notifications, the system also provides global visualizations on large screens such as the "Cloud View Platform" in the central control center. For example, it generates dynamically updated "alarm heat maps" to intuitively display the alarm density of each area using different color depths; or it periodically generates "scenario adaptation analysis reports" to statistically analyze the distribution patterns of various hazards in different scenarios (such as "Foreign object alarms in windy areas accounted for 35% of the total alarms this week"). The technical effect of this three-dimensional output method is that it ensures that key information reaches the front line directly and drives rapid action, while also providing management with tools for macro-situational awareness and trend analysis, achieving a unity of tactical command and strategic control.

[0043] Specifically, in step S400, during the record-keeping phase, while outputting alarms, the system simultaneously records the key information of each valid alarm (including emergency and routine alarms) in a structured manner into the "Hazard Identification Ledger" database. Each ledger record constitutes a complete event file, whose fields typically include: a unique alarm ID, the precise hazard type and subclass, the associated scene category tag, the original time and coordinates of data collection, the final conclusion of multi-model collaborative review in the cloud, and the entire lifecycle handling status of the alarm from generation to closure (such as "not responded," "in progress," or "closed"). This ledger data is usually stored in a high-performance unstructured or distributed storage platform to facilitate rapid writing and complex querying of massive historical data. The core technical effect of establishing this ledger is that it forms an auditable and traceable data closed loop for the entire intelligent inspection system. On the one hand, it supports full-process tracing of any historical alarm, facilitating post-event analysis and responsibility determination; on the other hand, the accumulated massive ledger data provides a valuable data foundation for subsequent in-depth data mining, model effect evaluation, hidden danger pattern analysis, and even the optimization of power grid risk assessment models, and is a key carrier for realizing continuous self-optimization of the system and the accumulation of operation and maintenance knowledge.

[0044] According to the present invention, a method for identifying hidden dangers in ultra-high voltage main grid transmission channels based on artificial intelligence size model collaboration further includes the following steps: S500, Model Iteration Steps: Collect manual review feedback on alarm information and construct a labeled sample library; use the labeled sample library to incrementally train and update the parameter templates of the first computer vision model and / or the second computer vision model, the time series analysis model, and the multimodal model.

[0045] In step S500, the data and model iterative optimization closed loop is the core self-optimization mechanism that ensures the continuous evolution of the entire hazard identification system and maintains high accuracy and strong adaptability over the long term. The technical purpose of this step is to systematically transform real feedback generated in actual operation and maintenance scenarios (especially boundary cases that the system cannot perfectly handle) into fuel driving continuous improvement in model performance, thereby forming a virtuous cycle of "application-feedback-learning-optimization," enabling the system to cope with new hazards and changing environments. Logically, its implementation constitutes a closed loop that connects end to end, beginning with manual quality verification and knowledge solidification of the system's output results, and ending with targeted enhancement training of the models on the cloud and edge using the newly solidified knowledge.

[0046] Specifically, in step S500, after the alarm information, verified in the cloud, is pushed to the maintenance personnel, the system does not consider this the end of the process. Instead, it sets a quality feedback window, typically completed within 24 hours. During this period, professionally trained maintenance personnel will conduct a secondary review of the alarm results pushed by the system based on their on-site experience and professional knowledge. The core objective of this review is to identify two types of errors in the system's automatic identification: first, "missed identification," meaning that actual potential hazards are not detected by any model, such as a small plastic foreign object not being captured by a large model in a complex mountainous background; second, "misjudgment," meaning that normal equipment components or environmental phenomena are incorrectly identified as potential hazards, such as misjudging normal glaze reflection on an insulator as a crack under a specific lighting angle. When such errors are found, the maintenance personnel will use the annotation tools provided by the system to mark the corresponding original images as "samples to be annotated." Subsequently, these samples need to be finely annotated. The annotations not only include the correct type of hazard that should actually be present in the image and its precise coordinates (for missed samples, the omitted target must be marked; for misjudged samples, they must be corrected to the correct category or labeled "no hazard"), but also must include a textual analysis of the possible causes of the system error, such as "Reason for missed detection: the foreign object is too small and its color is similar to the background shrubs" or "Reason for misjudgment: the evening sidelight forms a bright band on the insulator surface, resembling a crack." These images and annotation pairs rich in diagnostic information are uniformly collected into a dedicated "Ultra-High Voltage Transmission Channel Scenario-Based Sample Library." The technical effect of this step is that it precipitates the judgment and experience of human experts in the form of structured data, transforming one-off maintenance actions into "teaching cases" that can be digested and absorbed by machine learning systems, providing the most direct material for solving the current cognitive blind spots and shortcomings of the system.

[0047] Specifically, in step S500, as new samples accumulate in the scenario-based sample library, the system periodically triggers a retraining phase of the model to absorb new knowledge. Typically, the system sets a trigger threshold; for example, when the cumulative number of new samples reaches 10,000, a training task is automatically initiated. The training process employs differentiated optimization strategies tailored to the different architectures and deployment characteristics of the edge and cloud-based models.

[0048] Specifically, in step S500, for lightweight CV models deployed on resource-constrained edge devices, the optimization focus is on adjusting their dynamically loadable "scenario-specific parameter templates," rather than directly altering the model's massive base network weights. During training, the system extracts new positive and negative samples from the sample library for specific scenarios (such as windy areas) and re-optimizes the key parameters in the scenario template. For example, if a large number of new samples indicate that the current "windy area template" has a high false negative rate for small-volume floating objects, the training algorithm may automatically lower the recognition confidence threshold for this type of target; if specific background interference is found to increase false positives, the feature weights may be adjusted to strengthen background suppression. The optimized new parameter template is deployed remotely and in batches to the corresponding edge computing nodes via secure OTA (Over-The-Air) technology, thereby achieving silent upgrades of the edge model's recognition strategy without replacing the model itself, resulting in high update efficiency and minimal impact on business operations.

[0049] Specifically, step S500 employs an incremental training strategy for various large models deployed in the cloud (including the second computer vision model, the temporal analysis model, and the multimodal model). The system uses the full set of samples (or a subset divided by scene) from the sample library for training. The advantage of incremental training is that it focuses primarily on learning new features and patterns from newly added samples, without requiring large-scale relearning of the knowledge learned by the model from the existing massive dataset, thus significantly reducing the consumption of computing resources. For example, through incremental training, the second computer vision model can learn the appearance features of a new type of engineering machinery and add it to the feature library; the temporal analysis model can learn new equipment movement patterns; and the multimodal model can enhance its understanding of the language-visual associations of certain easily confused concepts. After training is complete, the updated model parameters and feature library are deployed to the cloud inference service. The fundamental technical effect of this series of retraining operations is that it enables the AI ​​model of the entire system to keep pace with the times, continuously learn from actual errors and new forms of hidden dangers, thereby continuously improving the accuracy of adaptation to different scenarios, expanding the scope of identifiable hidden dangers, and ultimately enabling the system to have the vitality of dynamic evolution and self-improvement, effectively responding to the new challenges that constantly emerge in the power grid operation and maintenance environment.

[0050] According to the present invention, a method for identifying hidden dangers in ultra-high voltage main grid transmission channels based on the collaboration of artificial intelligence large and small models further includes the following steps: S600, resource orchestration and scheduling steps: configuring the collaborative workflow and triggering rules of the first computer vision model, the second computer vision model, the time series analysis model and the multimodal model through a visual interface; and monitoring the computing load of the edge and the cloud, dynamically scheduling tasks and elastically scaling computing resources.

[0051] In step S500, cloud-edge-device model orchestration and dynamic computing power scheduling are key support layers providing flexibility and resilience for the entire collaborative recognition system. The technical objective is to achieve flexible assembly of workflows for various AI models distributed across the cloud and edge through software-defined methods, and intelligent perception and dynamic allocation of underlying heterogeneous computing resources. This ensures efficient, stable, and low-latency execution of recognition tasks even under changing business needs, continuous model iteration, and drastic load fluctuations. Logically, this mechanism can be divided into two complementary layers: model visualization orchestration oriented towards business logic, and dynamic computing power scheduling oriented towards infrastructure.

[0052] At the model visualization and orchestration level, the system empowers operations and maintenance personnel with a high degree of process customization capabilities through a core software module called the "Model Orchestration Center." This center provides an intuitive graphical user interface, allowing users to easily drag and drop pre-packaged functional components such as "CV mini-models," "secondary computer vision models," "time series analysis models," and "multimodal models" to create customized analysis pipelines suitable for different inspection scenarios, much like building blocks. A typical application example is that, for windy areas where the focus is on identifying lightweight floating objects, operations and maintenance personnel can orchestrate a simplified process consisting of "initial screening by edge CV mini-models" and "final material and intent reasoning by cloud-based multimodal models." Conversely, for equipment corrosion detection needs in salt spray areas, a process can be orchestrated where "initial screening by edge CV mini-models" is followed directly by "high-precision surface defect verification by cloud-based secondary computer vision models." After the workflow is built, users can configure granular triggering and execution rules for each step. For example, they can set "only trigger cloud-based large model review when the confidence level of the edge small model output is below 70%", thereby enabling on-demand resource access. The final orchestrated workflow will be saved in a structured data format (such as JSON) and can be deployed to relevant cloud services and edge nodes with one click. The core technical effect of this orchestration mechanism is that it transforms the fixed recognition process into a flexibly definable business strategy, enabling the system to quickly respond to differentiated operational priorities in different regions and at different times. This significantly improves the adaptability and maintainability of the entire technical solution at the application layer, allowing for upgrades and adjustments to the recognition logic without modifying the code.

[0053] At the dynamic scheduling level of computing power, the system has built an intelligent resource management system that runs through the cloud-edge infrastructure, aiming to ensure optimal utilization of computing resources and continuity of task execution in real time throughout the entire process. This scheduling plays a role in three key locations. First, at the edge, the system relies on the underlying Dianhong IoT operating system to continuously monitor the real-time hardware load indicators of each edge node, including CPU and GPU utilization, memory usage, etc. When the system detects that the load of a node exceeds a safe threshold (e.g., CPU utilization is consistently higher than 80%), it will automatically and intelligently schedule some of the image screening tasks to be processed to nearby idle edge nodes with lighter loads, thereby achieving load balancing within the edge computing cluster and preventing processing delays or interruptions caused by overload of a single node. Second, in the cloud, the scheduling mechanism is deeply integrated with the cloud platform infrastructure, such as utilizing the elastic scaling service of China Southern Power Grid Cloud. When extreme weather causes a surge in the number of alarms or a sudden doubling of the concurrent calls to large model services in the cloud, the system will automatically trigger an elastic scaling strategy to quickly increase the number of container instances carrying large model inference to cope with peak pressure; when the call volume drops, it will automatically reduce the number of instances to release resources. The technical advantage of this automated scaling capability lies in ensuring that the cloud-based review service always maintains high availability with elastic scalability, while avoiding the cost waste caused by maintaining excessive computing resources during low-load periods. Finally, the system also possesses cross-level task offloading capabilities. For edge nodes with particularly limited computing power, if they cannot independently complete the required scenario-based image preprocessing tasks, the system can temporarily offload these computationally intensive preprocessing operations to dedicated "lightweight processing nodes" in the cloud. After processing, the results are then sent back to the edge node for subsequent model screening. This flexible offloading strategy ensures that even in cases of heterogeneous or insufficient computing power at edge nodes, the initial stage of the entire recognition process remains uninterrupted, guaranteeing the integrity and real-time nature of monitoring coverage. In summary, this dynamic computing power scheduling system, through multi-level and automated resource management, achieves efficient aggregation and on-demand allocation of distributed computing power, powerfully supporting the stable, efficient, and low-cost operation of the aforementioned complex AI collaborative processes from an infrastructure perspective.

[0054] This invention provides a hidden danger identification system for ultra-high voltage (UHV) main grid transmission channels based on the collaborative operation of artificial intelligence-based large and small models. The system is deployed in a collaborative architecture consisting of edge devices (such as edge computing devices and smart cameras deployed on the transmission towers) and cloud servers. Through the collaborative work of each module, it achieves efficient and accurate hidden danger identification. Figure 2 A schematic diagram of one embodiment of the system is shown.

[0055] The edge data acquisition and processing module is physically deployed within various monitoring terminals along the power transmission corridor. This module is responsible for data acquisition and initial organization. It acquires real-time monitoring image data of the power transmission corridor and synchronized auxiliary metadata (such as time, location, wind speed, and equipment model) through integrated sensors (such as high-definition cameras, meteorological sensors, and positioning units). The module integrates scene analysis logic, which can automatically analyze and determine the scene category of the current monitoring image based on the auxiliary metadata (e.g., classifying it as a "windy area," "salt fog area," or "mountainous area"). Subsequently, the module encapsulates the optimized monitoring image data and the generated scene category tags, outputting standardized structured data packets for subsequent module processing.

[0056] The edge-based initial screening module is typically deployed on edge computing nodes (such as edge servers on the side of power poles) close to the data source. The core of this module is a pre-integrated and running first computer vision model, a lightweight target detection model designed to accommodate limited computing resources at the edge. This module receives structured data packets from the data acquisition and processing module and uses the first computer vision model to quickly analyze the image data, outputting hazard image data containing potential targets, hazard types, and corresponding confidence levels. The module has pre-configured or dynamically loaded parameter configurations associated with different scene categories, allowing it to select corresponding discrimination thresholds based on scene category labels in the input data packets. The module logic compares the confidence level output by the model with this threshold; if the condition is met (e.g., the confidence level is not lower than the threshold), a preliminary alarm is generated. This data packet integrates key information such as hazard image data, hazard type labels, and scene category labels, and is encrypted and uploaded to the cloud via a communication unit.

[0057] The cloud-based verification module is deployed in a powerful cloud server cluster. This module is the core of the system's deep analysis and decision-making, and it internally deploys multiple dedicated large models, including a second computer vision model, a time-series analysis model, and a multimodal model. This module first receives and parses the preliminary alarm data uploaded from the edge. Its internal scheduling and management unit intelligently classifies and distributes alarms based on the hazard type and scene category tags in the alarm data. First, the scheduling unit calls the second computer vision model (whose model size and capabilities are larger than the first model at the edge) that matches the hazard type and scene to perform refined verification and validation of the hazard image data. Next, based on the verification results of the second computer vision model and the hazard type, the scheduling unit further decides whether to selectively call the time-series analysis model (suitable for trajectory analysis of engineering machinery hazards) or the multimodal model (suitable for cross-modal reasoning of environmental disasters and equipment defects hazards) for deeper collaborative analysis and verification, ultimately forming a comprehensive and highly reliable hazard confirmation conclusion.

[0058] The alarm output module, also deployed in the cloud, interfaces with the business system. This module receives the final hazard confirmation conclusion from the cloud-based review module. It integrates risk grading logic and information formatting functions, generating alarm messages of different levels and formats based on the confirmed hazard type and its severity. This module is responsible for pushing the final alarm information to designated operation and maintenance management systems (such as power transmission production operation support systems) and the mobile terminals of operation and maintenance personnel, effectively delivering the identification results.

[0059] Through the coordinated operation of the above four modules, this system realizes a complete automated process from data acquisition, rapid edge filtering, in-depth collaborative analysis in the cloud to result output, forming an efficient, accurate and scalable intelligent hidden danger identification system for UHV transmission channels.

[0060] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0061] The above embodiments mainly describe the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the present invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.

Claims

1. A method for identifying hidden dangers in ultra-high voltage main grid transmission channels based on the collaborative modeling of artificial intelligence, applied to a hidden danger identification architecture consisting of edge terminals and cloud computing, characterized in that... The method includes the following steps: Acquire monitoring image data and auxiliary metadata of the power transmission channel; based on the auxiliary metadata, determine the scene category to which the monitoring image belongs, and generate a structured data packet containing monitoring image data and scene category tags; The structured data packets are analyzed using a first computer vision model deployed at the edge, and the hazard image data, hazard type, and confidence level are output. Select the corresponding discrimination threshold based on the scene category label. If the confidence level is not lower than the selected discrimination threshold, generate preliminary alarm data containing the hazard image data, hazard type label and scene category label, and upload it to the cloud. The preliminary alarm data is reviewed using a large model deployed in the cloud. The large model includes a second computer vision model and at least one of a time series analysis model or a multimodal model. The preliminary alarm data is classified according to the hazard type label and the scene category label. The second computer vision model corresponding to the hazard type and scene category is called according to the classification, and the hazard image data is reviewed. Based on the review result of the second computer vision model, the time series analysis model and / or multimodal model selected by the classification is used to perform a collaborative review of the preliminary alarm data. Based on the type of hazard confirmed during the review, an alarm message is output.

2. The method for identifying hidden dangers in UHV main grid transmission channels based on the collaborative modeling of artificial intelligence as described in claim 1, characterized in that, The first computer vision model is a lightweight object detection model deployed at the edge; The first computer vision model is configured to: pre-set parameter templates associated with different scene categories, the parameter templates containing discrimination thresholds for different hazard types; load the corresponding parameter templates according to the scene category labels, and judge the confidence level based on the discrimination thresholds in the templates; and / or, The first computer vision model is further configured to: pre-set feature extraction modules associated with different scene categories, activate the feature extraction module corresponding to the scene category label, and perform feature extraction of monitoring image data based on the feature extraction module corresponding to the scene.

3. The method for identifying hidden dangers in UHV main grid transmission channels based on the collaborative modeling of artificial intelligence as described in claim 1, characterized in that, The second computer vision model is a large visual model with more parameters than the first computer vision model, and it has been fine-tuned using a sample library of ultra-high voltage transmission channel scenarios. The second computer vision model is configured to: activate the corresponding scene-based feature library based on the scene category label, and perform multi-scale feature fusion and target recognition optimization on the hidden danger image data; The output of the second computer vision model includes: if the preliminary alarm data needs to be corrected, the reason for correction is marked; if the preliminary alarm data is confirmed, the hazard type label and the coordinate information corresponding to the label are retained.

4. The method for identifying hidden dangers in UHV main grid transmission channels based on the collaborative modeling of artificial intelligence as described in claim 1, characterized in that, The time-series analysis model is based on a long short-term memory network architecture and is used to verify the preliminary alarm data generated in scenarios where the device is stationary or recurring based on the classification. The time-series analysis model is configured to: input the coordinates and time information of the current alarm, and perform trajectory analysis by combining historical alarm records within a preset time period, so as to filter duplicate alarms or identify high-risk alarms that are approaching the transmission line.

5. The method for identifying hidden dangers in UHV main grid transmission channels based on the collaborative modeling of artificial intelligence as described in claim 1, characterized in that, The multimodal model is based on a vision-language pre-training architecture and is used to verify the potential hazards of environmental disasters or equipment defects determined based on the classification. The multimodal model includes: inputting the hazard image data and a preset text reasoning question, and performing cross-modal reasoning by matching image features with question semantics to confirm the authenticity of the hazard.

6. The method for identifying hidden dangers in UHV main grid transmission channels based on the collaborative modeling of artificial intelligence as described in claim 1, characterized in that, The auxiliary metadata structured data includes spatiotemporal data, environmental data, and equipment data; the determination of scene category based on auxiliary metadata specifically includes: determining the scene category as a windy area, salt fog area, or mountainous area based on wind speed, humidity, terrain type, or climate zone attributes in the environmental data; the scene category label is used to identify the determination result.

7. The method for identifying hidden dangers in UHV main grid transmission channels based on the collaborative modeling of artificial intelligence as described in claim 1, characterized in that, The output alarm information also includes: risk assessment and classification based on the confirmed hazard type and its scenario category label, generating graded alarm information and pushing it to the operation and maintenance terminal; at the same time, alarm details, review results and handling status are recorded in the hazard management ledger.

8. The method for identifying hidden dangers in UHV main grid transmission channels based on the collaborative modeling of artificial intelligence as described in claim 1, characterized in that, The method further includes a model iteration step: collecting manual review feedback on alarm information and constructing a labeled sample library; using the labeled sample library to incrementally train and update the parameter templates of the first computer vision model and / or the second computer vision model, the time series analysis model, and the multimodal model.

9. The method for identifying hidden dangers in UHV main grid transmission channels based on the collaborative modeling of artificial intelligence as described in claim 1, characterized in that, The method also includes a resource orchestration and scheduling step: configuring the collaborative workflow and triggering rules of the first computer vision model, the second computer vision model, the time series analysis model, and the multimodal model through a visual interface; and monitoring the computing load at the edge and the cloud, dynamically scheduling tasks and elastically scaling computing resources.

10. A hidden danger identification system for ultra-high voltage main grid transmission channels based on the collaborative model of artificial intelligence, applied in a hidden danger identification architecture consisting of edge and cloud, characterized in that, include: The edge data acquisition and processing module is used to acquire monitoring image data and auxiliary metadata of the power transmission channel; Based on auxiliary metadata, the scene category to which the monitored image belongs is determined, and a structured data packet containing the monitored image data and scene category label is generated; The edge-based initial screening module is equipped with a first computer vision model, which is used to analyze the structured data packet using the first computer vision model, output hazard image data, hazard type and confidence level; and to select the corresponding discrimination threshold according to the scene category label. If the confidence level is not lower than the selected discrimination threshold, preliminary alarm data containing the hazard image data, hazard type label and scene category label is generated and uploaded to the cloud. The cloud-based verification module is equipped with a large model. It uses the large model deployed in the cloud to verify the preliminary alarm data. The large model includes a second computer vision model and at least one of a time series analysis model or a multimodal model. It is used to classify the preliminary alarm data according to the hazard type label and the scene category label, call the second computer vision model corresponding to the hazard type and scene category according to the classification, and perform verification of the hazard image data. Based on the verification results of the second computer vision model, the system uses the time-series analysis model and / or multimodal model selected by the classification to perform collaborative verification of the preliminary alarm data; The alarm output module is used to output alarm information based on the type of hidden danger confirmed by the review.

Images