Low-voltage uninterruptible operation site visual monitoring optimization method based on edge computing architecture

Abstract

The invention discloses a low-voltage uninterruptible operation site visual monitoring optimization method based on an edge computing architecture, and the method comprises the following steps: constructing an operation scene graph analysis module, obtaining the distribution of operators, the operation type and the equipment voltage class information of each monitoring region, generating a risk class mapping graph, and carrying out the operation scene graph analysis module; positioning and initial identification of a high-risk area are realized; generating a risk perception matrix; constructing a priority queue of computing resources and risk levels; acquiring identification precision, processing delay and behavior capture integrity indexes, and establishing a task feedback mechanism; constructing a scheduling logic credibility evaluation module, analyzing a scheduling instruction, marking and recording an abnormal path, generating a scheduling behavior record library, and predicting and early warning a priority inversion risk; and iteratively optimizing scheduling model parameters, and dynamically updating risk scores and scheduling logic. According to the method, high-risk task resources are guaranteed preferentially, priority reversal is avoided, and the stability, robustness and safety response timeliness of the system in a complex operation environment are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent monitoring of power systems, and specifically relates to a low-voltage live working site visual monitoring optimization method based on an edge computing architecture. BACKGROUND

[0002] The "low-voltage live working site visual monitoring optimization based on an edge computing architecture" refers to introducing edge computing technology in the process of low-voltage live working, transferring complex computing tasks such as image recognition and behavior analysis of the working site from the traditional remote cloud to the edge computing nodes close to the site for local processing, thereby significantly improving the response speed and processing efficiency of the system. In combination with the material content, this method cooperates with AR glasses and on-site cameras to collect working pictures, analyzes the wearing condition of protective equipment, unsafe behaviors and dangerous source proximity of working personnel in real time by the edge node, automatically identifies risks and triggers the early warning mechanism, dynamically allocates computing resources to adapt to different risk levels, realizes synchronous processing and backtracking analysis of multiple video streams. Through this architecture optimization, the system not only reduces the dependence on network bandwidth and cloud platforms, but also enhances the stability and intelligent monitoring capability in complex communication and frequent working environments, and comprehensively improves the safety guarantee level of low-voltage live working.

[0003] The prior art has the following disadvantages: In the existing low-voltage live working visual monitoring system, although a task grading processing mechanism based on edge computing is introduced to alleviate the resource bottleneck problem caused by concurrent analysis of multiple video sources, the system may still face the potential risk of unbalanced computing resource scheduling in extremely high-load running scenarios. Specifically, when multiple working areas are in a high-concurrent monitoring state at the same time, the system needs to dynamically allocate the computing resource of the edge computing node according to the risk level of the working area. However, due to the logical deviation or abnormal decision path in the risk level identification, task priority allocation or behavior feature weight calculation process of the scheduling algorithm, the working area originally in a low-risk state may be misjudged as a high-priority task, thereby occupying a large amount of edge computing resources, causing the high-risk area to be marginalized in the allocation of computing resources, and further causing problems such as a sharp drop in video frame rate and a decrease in AI recognition frequency, ultimately leading to a serious decline in the detection capability of key dangerous behaviors, and a major safety hazard that high-risk behaviors such as violation of operation rules and approach to live areas cannot be identified and alarmed in time. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the application provides a low-voltage uninterrupted operation site visual monitoring optimization method based on an edge computing architecture, which realizes accurate identification and dynamic perception of high-risk areas by constructing an operation scene atlas and a multi-modal risk scoring mechanism, and introduces a multi-objective weighted scheduling and feedback mechanism to preferentially guarantee high-risk task resources and avoid priority inversion. At the same time, combined with scheduling credibility evaluation and time sequence supervised learning optimization, abnormal identification and adaptive evolution of scheduling strategies are realized, which significantly improves the stability, robustness and safety response timeliness of the system in a complex operation environment, to solve the problems in the above background technology.

[0005] The technical scheme adopted by the application to solve its technical problems is: a low-voltage uninterrupted operation site visual monitoring optimization method based on an edge computing architecture, comprising the following steps: An operation scene atlas analysis module is constructed to obtain operation personnel distribution, operation type and equipment voltage level information of each monitoring area, generate a risk level mapping diagram, and realize positioning and initial identification of high-risk areas; Based on image data, behavior feature data and environmental parameters, multi-modal features are extracted, a risk scoring model is constructed, and real-time risk scores of each monitoring area are calculated using a heterogeneous feature embedding method to generate a risk perception matrix; Based on the risk perception matrix, a multi-objective weighted scheduling algorithm is introduced to construct a priority queue of computing resources and risk levels, real-time adjust edge node resource allocation, and preferentially support identification tasks in high-risk areas; Collect identification accuracy, processing delay and behavior capture integrity indicators, establish a task feedback mechanism, dynamically adjust task execution parameters according to feedback, and correct resource scheduling strategies; A scheduling logic credibility evaluation module is constructed to analyze scheduling instructions, mark and record abnormal paths, generate a scheduling behavior record library, and predict and warn priority inversion risks; Based on a time sequence supervised learning strategy, historical operation scene data and current task data are fused to iteratively optimize scheduling model parameters, dynamically update risk scores and scheduling logic, and improve the environmental adaptability and robustness of the scheduling system.

[0006] Preferably, the operation scene atlas analysis module comprises the following steps: Collect operation personnel distribution, operation type, equipment voltage level, operation task sheet and scheduling record of the monitoring area to form a regional semantic data pool; Based on operation personnel and operation type information, an association graph network is constructed, and behavior clustering mechanism is combined to identify operation task types and label risk levels; Graph neural network algorithm is applied to model node attributes and edge weights, calculate regional risk weights and generate a risk level mapping diagram; The time series analysis method is used to dynamically update the atlas structure, form a risk evolution prediction model, and link the scheduling module for response.

[0007] Preferably, the risk score model is constructed and includes the following steps: Collect image data, behavior characteristic data, and environmental parameter indicators in the monitoring area, and perform spatio-temporal alignment; Extract image, behavior, and environmental multi-modal features and encode them, and realize deep fusion expression through a heterogeneous feature embedding mechanism; A graph attention network model is constructed based on the fused features to output the risk score of each monitoring area.

[0008] Preferably, the risk perception matrix is generated by the following steps: structuring and integrating the risk scores according to the area number to generate a risk perception matrix as the input basis for resource scheduling.

[0009] Preferably, the multi-objective weighted scheduling algorithm includes the following steps: A risk score-driven task risk-resource matching model is constructed based on the risk perception matrix, and a scheduling objective function is established; A risk level and a corresponding priority queue of computing resources are constructed according to the risk score and the task attributes, and a resource application package is formed; According to the priority queue order, the computing resources are allocated, and in the case of resource bottleneck, the task rollback or migration strategy is executed; The scheduling parameters are dynamically adjusted based on the scheduling execution feedback indicators to improve the adaptability of the scheduling strategy to the job environment.

[0010] Preferably, the task execution feedback mechanism is established by the following steps: Collect model recognition accuracy, video processing delay, and behavior capture integrity indicators in each monitoring area to form a task running state table; According to the running state table, the task stability and resource scheduling matching degree are analyzed, and the task execution stability index is calculated; Based on the feedback results, the model frequency, video parameters, and concurrency level are dynamically adjusted to optimize the task execution parameters; According to the adjustment results, the scheduling strategy is corrected, the scheduling logic is optimized using reinforcement learning, and the area scheduling priority is updated.

[0011] Preferably, the scheduling logic credibility evaluation module is constructed by the following steps: Record the priority allocation instructions and their execution paths in each round of scheduling, and establish a scheduling path linked list; According to the preset rules and risk-resource ratio model, abnormal scheduling behaviors are identified and marked; Archive and cluster analyze normal and abnormal scheduling paths to construct a scheduling behavior record library and an abnormal feature template library; Template matching and risk prediction are performed on the current scheduling strategy, triggering scheduling early warning and generating a risk evolution curve.

[0012] Preferably, based on the time sequence supervision learning strategy, the historical operation scene data and the current task data are fused, and the scheduling model parameters are iteratively optimized, specifically including the following steps: Collect the input features, decision output and execution effect data of the scheduling task, and construct a task time sequence sample chain; A model combining long short-term memory network and attention mechanism is constructed, and the scheduling parameters are optimized according to the supervision signal; According to the model prediction result, the risk scoring mechanism and the scheduling logic are dynamically adjusted to realize the adaptive update of the scheduling strategy; A sliding window training and multi-model comparison mechanism is adopted to regularly update the scheduling model and deploy it to the edge system.

[0013] Compared with the prior art, the beneficial effects of the present application are: The present application realizes accurate identification and dynamic perception of high-risk areas by constructing an operation scene graph and a multi-modal risk scoring mechanism, and ensures the priority protection of computing resources in time delay sensitive and high risk tasks by introducing a multi-objective weighted scheduling and real-time feedback mechanism, avoiding the priority inversion problem in resource scheduling. At the same time, by constructing a scheduling credibility evaluation module and a continuous optimization framework based on time sequence supervision learning, the system has the ability to identify abnormal scheduling paths and strategy evolution, and can realize self-correction and scheduling logic evolution under complex environmental fluctuations, thereby significantly improving the stability, robustness and timeliness of safety warning response of the entire visual monitoring system in the actual power operation environment. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description.

[0015] Figure 1 The flowchart of the low-voltage uninterrupted operation site visual monitoring optimization method based on the edge computing architecture of the embodiment of the present application. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0017] In the description of the present application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, which are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0018] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connection" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integral connection; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through intermediate medium, or the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances. In addition, the technical features involved in the different embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0019] As shown in Figure 1 The low-voltage uninterrupted operation site visual monitoring optimization method based on edge computing architecture of the embodiment of the present application comprises the following steps: A work scene graph analysis module is constructed to obtain the work personnel distribution, work type and equipment voltage level information of each monitoring area in the low-voltage uninterrupted operation site, and a risk level mapping diagram is generated according to the obtained information to realize accurate positioning and initial risk identification of each high-risk operation area; A work scene graph analysis module for low-voltage uninterrupted operation site is constructed to realize comprehensive analysis of the work personnel distribution, work type and equipment voltage level and other multi-source structured and unstructured information of each monitoring area in the site, and to automatically generate a risk level mapping diagram based on this to provide high-precision input support for subsequent risk perception and resource scheduling strategy. The construction and application of this module can be divided into the following four steps: Firstly, the basic environmental information of the low-voltage uninterrupted operation site is collected, including monitoring camera coverage area, physical division of operation area, distribution equipment number and voltage level annotation, operation plan arrangement document, operation personnel identity information, operation task sheet and other original data. Through the intelligent terminal (such as AR glasses, wearable sensors, environmental perception devices, etc.) deployed on site, combined with the operation scheduling records in the power operation management system, a preliminary data pool of the site entity relationship is formed. The system accurately obtains the real-time distribution of each operation personnel in different monitoring areas by fusing GPS positioning, Wi-Fi / AP positioning and visual recognition and other multi-source location information, and labels the data in area granularity to provide semantic basis for graph construction.

[0020] Secondly, based on the job personnel and job type information in the data pool, a personnel-job behavior association network graph is constructed. The system extracts the risk level label corresponding to the job type (such as cable installation and removal, power supply connection, equipment detection, etc.) through the job task sheet and historical behavior record, and models the relationship between the job personnel and the task type through the graph database. In particular, to enhance the dynamic adaptability of the system, the application introduces a behavior clustering mechanism based on self-supervised learning, which is used to automatically identify new or variant job task types and their corresponding potential risk levels in the field, thereby avoiding the identification blind area of non-standard job scenes. At the same time, the voltage level of the field power distribution equipment is embedded in the graph as a key node attribute, so that the system can comprehensively consider the electrical danger level of the job object when judging the risk level.

[0021] Thirdly, the system combines the constructed job graph and spatial region division data, and applies a graph neural network algorithm to evaluate the comprehensive risk level of the job region. The graph neural network model takes the job personnel node, job task node, and power distribution equipment node as the basis of the graph structure, and fuses the node attributes and edge weights (such as personnel job experience, task complexity, voltage level, etc.) to calculate the risk weight value of each monitoring region. This model can effectively mine the potential high-risk behavior association patterns of each monitoring region, further enhancing the accuracy of risk level discrimination. At the same time, the system labels the heat level on the monitoring map according to the risk weight value, generates a risk level mapping graph, and displays the risk distribution state of each region in real time in the form of color coding, realizing intuitive risk positioning and dynamic early warning capability.

[0022] Finally, to ensure the real-time and robustness of the risk level mapping graph, the system introduces a dynamic updating mechanism. This mechanism automatically reconstructs the graph structure and node attributes by periodically collecting new human-machine environment data, and forms a prediction model of risk evolution path based on time series analysis method, modeling the fluctuation trend of regional risk level in a short time. Once the personnel distribution in the monitoring region mutates, the job type temporarily adjusts or the equipment state changes, the system can automatically trigger the graph reconstruction process to timely adjust the content of the risk level mapping graph, ensuring that the identified high-risk areas always reflect the true state of the current job site. This module also supports linkage with the subsequent resource scheduling module to realize real-time response of the scheduling strategy to the risk graph.

[0023] In summary, the job scene graph analysis module realizes the deep structured understanding of multi-dimensional information in the low-voltage uninterrupted power operation site through four steps of multi-source data collection, semantic modeling, graph structure risk evaluation, and dynamic updating mechanism, significantly improves the identification accuracy and update speed of high-risk operation areas, and provides a high-quality input basis for the subsequent dynamic scheduling of edge computing resources, which embodies significant technical creativity and application value.

[0024] Based on the image data, behavior feature data and environmental parameter indicators of each monitoring area, multi-modal features are extracted, a risk score model is constructed, heterogeneous feature embedding methods are used to process the extracted features, real-time risk scores of each monitoring area are calculated, and a risk perception matrix for subsequent scheduling is generated; In view of the complex scene characteristics of different monitoring areas in the low-voltage non-power operation site, in order to realize accurate quantitative evaluation of the potential risk level of each area, a multi-modal feature extraction and risk score modeling method is proposed, which integrates image data, behavior feature data and environmental parameter indicators, and on this basis, a risk perception matrix for resource scheduling is generated. The method includes the following four steps: Firstly, the synchronous acquisition and space-time alignment of multi-modal data are carried out. Through the deployment of high-definition cameras, wearable visual terminals, environmental perception sensors and edge computing nodes in each monitoring area of the operation site, image information, personnel behavior dynamics, temperature and humidity, noise, electromagnetic radiation and other environmental data in the operation area are collected. The system performs time series alignment and spatial registration on the collected heterogeneous data to ensure the corresponding relationship of all data in the same time axis and spatial coordinate system. In particular, the invention adopts a frame-level timestamp and video frame reconstruction synchronization mechanism to ensure that image behavior data and environmental parameters have millisecond-level pairing accuracy, laying a consistent foundation for subsequent feature fusion.

[0025] Secondly, the collected data are processed by multi-modal feature extraction and coding. In the image dimension, deep convolutional neural network is used to extract human key points, wearing equipment conditions, behavior posture trajectories and other semantic features of operation personnel; in the behavior data dimension, operation action recognition model and posture stability analysis model are combined to extract high-risk behavior features such as continuous bending, frequent approach to live area and personnel gathering anomaly; in the environmental dimension, time series parameter window is constructed according to the sampling frequency, and index change trend of abnormal environmental disturbance events such as temperature sudden rise and noise abnormal fluctuation is extracted. The above features are normalized and coded by a unified tensor representation method, and on this basis, a heterogeneous feature embedding mechanism is introduced, and a multi-channel attention fusion module is used to model the semantic relationship between different modalities, realizing deep fusion expression of image, behavior and environmental features.

[0026] Third, a risk score model is constructed to predict the risk value of the fused features. The system is based on a supervised learning framework, and the training set is composed of a large number of labeled job scene samples, each sample containing corresponding job task labels, violation risk level labels and key behavior trigger event labels. The model uses a graph attention network (GAT) structure, with fused features as node input, to capture the temporal behavior evolution pattern and potential correlation risk factors between regions, and to realize the scoring output of the current monitoring area risk level. The scoring result is represented as a floating point value between 0 and 1, and the higher the value, the greater the risk. To enhance the dynamic robustness of the model, the system introduces a short-term risk fluctuation smoothing mechanism to filter the sudden abnormal risk signals in time sequence, thereby reducing the misjudgment rate.

[0027] Finally, the risk score results of each monitoring area are structured and integrated to generate a risk perception matrix. The matrix uses region numbers as indexes and corresponding risk scores as elements to form a two-dimensional risk situation representation graph. The matrix update period is automatically adjusted according to the system running load, and in the resource abundant state, it can realize second-level refresh, and in the high load state, it automatically degenerates into a priority refresh mechanism centered on key areas, ensuring that the scoring results of high-risk areas are always up-to-date. The risk perception matrix will serve as an important input basis for edge computing node task scheduling, driving the construction and dynamic adjustment of the resource priority queue, and providing accurate and timely risk judgment support for subsequent high-reliability scheduling strategies.

[0028] In summary, through the above four steps, a multi-modal risk modeling process is realized that integrates image, behavior and environmental features, and ultimately forms a risk perception matrix that can be used for resource scheduling.

[0029] According to the risk perception matrix, a multi-objective weighted scheduling algorithm is introduced to construct a corresponding priority queue of computing resources and risk levels, and to adjust the resource allocation strategy of the edge computing node in real time according to the priority, and to preferentially allocate recognition computing resources to monitoring areas with high risk levels; To address the edge computing resource pressure caused by high-concurrency video analysis tasks in multiple monitoring areas in low-voltage uninterrupted job sites, a multi-objective weighted scheduling algorithm based on risk perception matrix is proposed to realize a risk-driven computing resource dynamic allocation mechanism, and a corresponding priority queue between risk levels and computing resources is constructed to ensure that recognition computing resources are preferentially served to high-risk monitoring areas. Specifically, the following four steps are included: First, a task risk-resource matching modeling system is constructed. The system takes the risk perception matrix generated in the previous stage as input, takes the risk score corresponding to each monitoring area as the main driving factor of the scheduling model, and considers the task characteristics of the area (such as the identification model complexity, video stream resolution, processing period), the current load state of the computing resource, the edge node communication bandwidth and delay and other system resource parameters, and establishes the objective function of the scheduling optimization problem. In the objective function, the risk score is used as the main weight variable to model the importance of different regional tasks, thereby forming a multi-objective optimization problem with "risk-driven + resource constraint" as the core. In particular, the invention introduces a hierarchical weight fusion mechanism to fuse the risk score and the time delay sensitivity of real-time identification tasks, enhancing the response ability of the scheduling model to the urgency of key event identification.

[0030] Second, the corresponding priority queue of risk level and computing resource is constructed according to the above model. The system sorts all the monitoring tasks to be processed according to the real-time risk score from high to low to form a preliminary priority task queue. In order to avoid scheduling jitter problems in areas with similar scores, the system sets a dynamic score difference threshold, and tasks with a risk score difference less than the threshold are classified into the same priority level, and the scheduling order is further subdivided according to the historical task stability and task complexity to form a stable and efficient hierarchical priority queue. Each task in the priority queue is bound to its required computing resource type (such as AI model inference power, cache bandwidth, temporary storage space), and a resource application package is formed to prepare for allocation.

[0031] Third, the priority-based resource allocation strategy adjustment is executed. The system monitors the computing capability state of each edge node (including CPU / GPU usage, memory occupation, task queue length, etc.), and establishes a global resource visualization state table. According to the priority queue order, the scheduling controller allocates resources from the global resource pool to high-priority tasks, and ensures that the image recognition tasks in high-risk areas run completely and in real time. If there is a local bottleneck in resources, the scheduling fallback mechanism is triggered, and through task rerouting or temporary task merging strategy, low-priority tasks are temporarily queued or transferred to other computing nodes for execution, ensuring that the computing resources in high-risk areas are always available. The embodiment also specially designs a cross-area collaborative scheduling strategy, which dynamically migrates computing tasks under the condition that multiple edge nodes have a shared computing power pool, to realize resource redundancy utilization and extreme guarantee of high-priority tasks.

[0032] Finally, a dynamic feedback mechanism of scheduling execution process and a scheduling strategy fine-tuning model are established. After each scheduling is completed, the system automatically records the key indicators such as resource allocation efficiency, task execution success rate, and computing resource utilization, and correlates them with the delay and accuracy of the actual task identification results to generate a scheduling effect evaluation report. Based on the evaluation results, the system dynamically adjusts the priority sorting parameters, resource matching weights, and risk threshold settings to continuously fine-tune the scheduling logic and improve the adaptability of the scheduling algorithm to changes in complex job scenarios. In addition, the system tracks and monitors the changes in resources in key areas during two consecutive scheduling rounds. Once it finds that the resources are continuously decreasing or being occupied by low-risk tasks, it triggers a warning mechanism to prompt manual or system intervention to prevent potential priority inversion risks.

[0033] In summary, this step builds a multi-objective weighted scheduling mechanism that integrates risk perception, task characteristics, and resource status through the four steps of "modeling → sorting → allocation → feedback". This mechanism not only ensures the risk priority of computing resources, but also has good task stability and system response flexibility, effectively solving the resource allocation imbalance problem in high-concurrency running environments for edge computing systems and significantly enhancing the system's immediate response capability to high-risk behavior identification, providing a solid guarantee for the intrinsic safety of low-voltage uninterrupted power operation.

[0034] Collecting model identification accuracy, video processing delay, and behavior capture integrity parameters in each monitoring area, establishing a task execution feedback mechanism, dynamically adjusting task execution parameters of edge computing nodes according to the feedback indicators, and realizing real-time adaptive correction of resource scheduling strategy; To ensure that the identification tasks in different monitoring areas in the low-voltage uninterrupted power operation site can be optimized in real time according to the actual operation effect and improve the resource scheduling accuracy and adaptability of the edge computing system, a task execution feedback mechanism is proposed. This mechanism dynamically collects and analyzes operation indicators such as model identification accuracy, video processing delay, and behavior capture integrity to continuously optimize task execution parameters of edge computing nodes and realize real-time adaptive correction of scheduling strategy. This mechanism includes the following four steps: First, a task running state monitoring module is constructed to collect the core running indicators of model recognition tasks in each monitoring area in real time. In terms of recognition accuracy, the system dynamically calculates accuracy indicators such as accuracy and recall rate by comparing the model recognition results with manually annotated samples or high-credibility verification model outputs. In terms of video processing delay, the system records the complete time path from video frame collection, model loading, inference execution to result output to obtain a frame-level processing delay curve. In terms of behavior capture integrity, the system tracks the continuity of target recognition in a behavior action sequence to assess whether the model has phenomena such as recognition interruption and action loss. The above indicators are stored in real time in the form of time series, and form a multi-dimensional task running state table in units of areas to provide data basis for subsequent adjustment.

[0035] Second, a feedback analysis engine is established to evaluate the matching degree between the current scheduling resources and task execution according to the trend of the indicators in the task running state table. The system introduces a multi-factor scoring mechanism to give different weights to each indicator to form a comprehensive performance score. For example, in a high-risk operation area, the weight of recognition accuracy is increased; in a resource-constrained environment, the weight of processing delay is given priority. The system calculates a task execution stability index according to the score and identifies the bottleneck position in task execution, such as abnormal phenomena caused by a decrease in model inference frequency due to insufficient resource allocation or a sudden increase in video processing delay due to network congestion. By cross-analyzing the index with the risk level of the task, the system determines whether the current resource scheduling matches the actual task performance.

[0036] Third, the task execution parameters of the edge computing node are dynamically adjusted in real time according to the feedback results. The system can automatically adjust the following parameters according to the feedback indicators: 1) inference model running frequency, such as increasing the model calling frequency or switching to a lightweight backup model to reduce resource occupation in areas where the task stability is insufficient; 2) video stream resolution and frame rate configuration, compressing the video data stream within an acceptable recognition accuracy range to reduce processing delay; 3) task concurrency level limit, optimizing parallel scheduling of multiple recognition tasks to avoid conflicts of computing resources. All parameter adjustments are automatically completed by the scheduling control module without human intervention, and support hot switching and state rollback to ensure that the task is not interrupted during the adjustment process.

[0037] Finally, a feedback-driven scheduling strategy adaptive correction mechanism is established to realize the dynamic update of the global resource configuration strategy. After the system completes the adjustment, it continues to monitor the changes in the identification performance of the relevant area and calculates the adjustment effect based on the difference between the before and after states. If the task running stability index improves significantly, the current parameter configuration is retained; if the adjustment is ineffective or the performance decreases, the adjustment rollback mechanism is triggered. After multiple adjustments, the system optimizes the scheduling weight model through reinforcement learning method, automatically adjusts the influence factor of each feedback index in the scheduling logic, realizes the online self-learning and optimization of the resource scheduling logic. At the same time, the system updates the regional scheduling priority after completing each feedback closed-loop period, providing a more practical strategy reference for the next round of scheduling, realizing the close linkage and adaptive cooperation between the task execution state and the scheduling strategy.

[0038] In summary, this step builds a task execution feedback mechanism with real-time, adaptability and evolution ability through four steps of task state monitoring, feedback analysis, parameter adjustment and strategy correction. This mechanism not only can dynamically adjust the model running state and resource configuration, but also can feed back to the global optimization of the scheduling strategy, thereby effectively improving the rationality of multi-region edge computing resource allocation and the stability of system running in the low-voltage uninterrupted power operation site, providing continuous and efficient computing guarantee for high-risk behavior identification.

[0039] A scheduling logic credibility evaluation module is constructed to analyze the priority scheduling instructions generated during the scheduling execution process, mark and record abnormal scheduling path information, generate a scheduling behavior record library, and predict and warn potential priority inversion risks; To improve the scheduling decision transparency and risk prevention and control ability of the edge computing system in the low-voltage uninterrupted power operation site, a scheduling logic credibility evaluation module is proposed, which is used to continuously track and analyze the priority scheduling instructions generated during the scheduling process, automatically mark the scheduling abnormal path, establish a scheduling behavior record library, and realize the prediction and warning of potential priority inversion risks, thereby effectively preventing the risk of high-risk tasks being marginalized by resources. The module includes the following four steps: Firstly, a scheduling instruction tracking structure is constructed to collect and structure all priority allocation instructions and their execution path information generated during each scheduling process. The system records the priority decision sequence, target resource node number, scheduled task identification, corresponding risk score value, current resource occupation state and other information output by the scheduling logic module in each edge node task allocation period, forming a multi-field scheduling data packet. To ensure the reconfigurability of the scheduling process, the scheduling trajectory modeling method based on timestamp and task hash index is adopted to establish a unique index for each scheduling behavior, forming a scheduling path linked list, which is convenient for subsequent analysis and backtracking.

[0040] Secondly, the algorithm module is designed to identify and mark abnormal behaviors in the scheduling path. The system iterates the scheduling path linked list based on the set of scheduling credibility rules (such as: high-risk score tasks have not been scheduled for two consecutive periods, or low-risk tasks frequently occupy computing resources, etc.), to determine whether the scheduling behavior deviates from the normal strategy. In particular, the invention introduces a risk-resource ratio anomaly detection model to calculate the actual resource allocation value corresponding to each task unit risk score, and judges whether there is an abnormal scheduling tendency by comparing the deviation ratio with the historical average value; if the deviation exceeds the set threshold, the current scheduling instruction is marked as abnormal, and information such as trigger conditions, time points, and affected area numbers is recorded.

[0041] Thirdly, a scheduling behavior record library is constructed to archive, classify and extract features of historical scheduling behaviors. The system stores all normal and abnormal scheduling paths into the record library according to dimensions such as region, time, risk level, etc., and performs clustering analysis on the scheduling behavior patterns. To enhance the behavior learning ability of the system, the invention uses an unsupervised anomaly detection method to automatically extract typical error patterns from a large number of scheduling sequences, such as: priority inversion, scheduling continuous lag, resource repeated allocation, etc., thereby forming a scheduling anomaly feature template library that can be used for real-time comparison. The record library also supports manual review and labeling mechanisms, allowing engineers to manually confirm the rationality of the scheduling logic based on a graphical interface, assisting in model iteration training.

[0042] Finally, a scheduling credibility prediction and early warning mechanism is established to provide real-time prompts for potential priority inversion risks. The system calls the abnormal feature templates extracted from the record library before each round of scheduling, and performs real-time matching on the instruction sequence output by the current scheduling strategy. If the matching degree exceeds the risk threshold, the early warning process is triggered. The early warning mechanism includes: prompting the scheduling controller to roll back the current strategy and switch to a backup scheduling logic; notifying the resource allocation module to enter protection mode and prioritize critical area identification tasks; pushing scheduling anomaly risk information to the job safety supervision module to achieve a linked response. Further, the system analyzes the correlation between abnormal scheduling frequency and scheduling failure consequences through a time series prediction model, constructs a risk evolution curve, and predicts the stability trend of the system in continuous scheduling, providing decision-making basis for edge resource allocation.

[0043] In summary, this step builds a complete scheduling logic credibility evaluation mechanism through four steps: scheduling data collection, abnormal marking, behavior recording, and risk warning. This mechanism not only enables traceability, explainability, and controllability of scheduling behavior, but also can timely detect potential priority inversion risks, ensuring the stable, transparent, and secure operation of edge computing systems in the face of complex concurrent tasks. It is a key technical component for improving the reliability of intelligent scheduling in low-voltage uninterrupted power operation sites.

[0044] Based on the time sequence supervised learning strategy, the historical operation scene data and the current task execution data are collected, the scheduling model parameters are continuously iterated and optimized, the risk score and the scheduling logic are dynamically updated, and the adaptation of the scheduling strategy to the change of the operation environment and the robustness of the scheduling system are realized. To enhance the adaptive ability of the scheduling strategy of the edge computing architecture in the low-voltage uninterrupted operation site to the environmental changes and the robustness in the long-term operation process, a scheduling model optimization method based on the time sequence supervised learning strategy is proposed. The method continuously collects historical operation scene data and current task execution data, dynamically trains and iteratively updates the core parameters of the scheduling model by using time sequence modeling and supervised learning technology, so as to realize the self-evolution of the risk score mechanism and the scheduling logic. The method includes the following four steps: Firstly, the data sample set for training is collected and constructed. The system records the input features, decision outputs and task execution effects of each round of scheduling tasks during the deployment of the edge node, forming a complete scheduling behavior sample. The input features include risk score matrix, resource available state, task complexity, environmental parameters, etc.; the decision output includes priority queue order, resource allocation instruction, etc.; the execution effect includes task completion rate, recognition accuracy, response time delay and feedback index change. All data are organized in time sequence to form a task time sequence sample chain, and combined with the environmental factors of historical operation scene (such as load level, operation type distribution, weather condition, etc.), a multi-dimensional time sequence training set is formed to provide basic corpus for model learning.

[0045] Secondly, the structure of the time sequence supervised learning model is designed and the initial training is carried out. The invention adopts a deep learning model combining long short-term memory network (LSTM) and attention mechanism as the core architecture, in which the LSTM module is used to capture the potential time dependence and task execution trend in the scheduling process, and the attention mechanism is used to dynamically focus on the key features affecting the scheduling effect (such as sudden risk score, frequent failure task, etc.). In the model training stage, the difference between the scheduling output and the actual task execution effect is used as the supervision signal to guide the model to learn and correct the scheduling parameters (such as risk weight coefficient, resource allocation threshold, priority rule factor). By minimizing the error between the predicted scheduling effect and the real feedback, the internal parameters of the model are optimized, and the prediction ability and adaptability of the model are improved.

[0046] Third, dynamically update the risk scoring mechanism and scheduling logic structure based on model prediction results. Before each round of task scheduling, the system calls the trained time series learning model to simulate the possible scheduling path, predicts the expected performance of each scheduling scheme, and dynamically fine-tunes the current scheduling logic accordingly. For example, for areas with a risk of declining recognition accuracy, the system can recommend increasing the risk score weight to enter the scheduling priority early; for tasks predicted to have delay backlog, the system can actively reduce the amount of allocated resources to prevent node overload. The risk scoring mechanism and scheduling priority logic are no longer driven by static parameter tables, but are adjusted in real time according to the evolution of the scene and historical performance, forming a "prediction-adjustment-execution-feedback" closed loop.

[0047] Finally, build a model continuous iteration optimization and deployment mechanism to realize online evolution of scheduling strategy. The system uses a sliding window update strategy to periodically train samples from multiple scheduling cycles in the recent period and integrate the latest task performance in a timely manner. At the same time, the model version is managed and rolled back to protect against performance fluctuations caused by new model deployment. Model training tasks are automatically started when the system is under low load, and after training is complete, the system is updated with hot deployment in the next scheduling cycle. To improve robustness, the system also designs a multi-model comparison mechanism, with different learning models (such as time convolution network TCN and gated recurrent unit GRU) running simultaneously, and the best performer is selected as the main scheduling predictor to enhance the stability and fault tolerance of the system in complex environments.

[0048] In summary, this step builds a time series supervised learning scheduling optimization mechanism for low-voltage uninterrupted power operation scenarios through four steps: sample collection, model training, dynamic prediction, and online optimization. This mechanism has the ability to understand task context, adaptively adjust strategies, and optimize online in real time, significantly improving the ability of edge scheduling systems to respond to high-dynamic, multi-source task environment changes. It is one of the key technologies to realize intelligent, self-evolving resource scheduling systems.

[0049] Through the low-voltage uninterrupted operation site visual monitoring optimization method based on the edge computing architecture, high coupling and adaptive linkage between the dispatching logic and the operation risk state are effectively realized, and the intelligent resource allocation ability and risk perception accuracy of the system in a high-concurrency task scenario are significantly improved. The method not only realizes accurate identification and dynamic perception of high-risk areas through the construction of an operation scene atlas and a multi-modal risk scoring mechanism, but also ensures the priority protection of computing resources in time delay sensitive and high risk tasks through the introduction of a multi-objective weighted scheduling and real-time feedback mechanism, avoiding the priority inversion problem in resource scheduling. At the same time, by constructing a scheduling credibility evaluation module and a continuous optimization framework based on time sequence supervised learning, the system has the ability to identify abnormal scheduling paths and the ability to evolve strategies, and can realize self-correction and scheduling logic evolution under complex environmental fluctuations, thereby significantly improving the stability, robustness and timeliness of safety warning response of the entire visual monitoring system in the actual power operation environment.

[0050] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which shall be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A low-voltage uninterrupted operation site visual monitoring optimization method based on an edge computing architecture, characterized in that, The method comprises the following steps: Constructing a work scene graph analysis module to obtain work personnel distribution, work type, and equipment voltage level information of each monitoring area, generating a risk level mapping diagram, and realizing positioning and initial identification of high-risk areas; Based on image data, behavior feature data, and environmental parameters, multi-modal features are extracted, a risk scoring model is constructed, and a real-time risk score of each monitoring area is calculated using a heterogeneous feature embedding method to generate a risk perception matrix; Based on the risk perception matrix, a multi-objective weighted scheduling algorithm is introduced, a priority queue of computing resources and risk levels is constructed, and the resource allocation of the edge node is adjusted in real time to preferentially support the identification tasks of high-risk areas; Collecting identification accuracy, processing delay, and behavior capture integrity indicators, establishing a task feedback mechanism, dynamically adjusting task execution parameters according to feedback, and correcting resource scheduling strategies; Constructing a scheduling logic credibility evaluation module, analyzing scheduling instructions, marking and recording abnormal paths, generating a scheduling behavior record library, and predicting and warning priority inversion risks; Based on the time sequence supervised learning strategy, historical work scene data and current task data are fused, the scheduling model parameters are iteratively optimized, the risk score and scheduling logic are dynamically updated, and the environmental adaptability and robustness of the scheduling system are improved.

2. The method of claim 1, wherein the method is performed by an edge computing architecture. The construction of the work scene graph analysis module comprises the following steps: Collecting work personnel distribution, work type, equipment voltage level, work task list, and scheduling records of the monitoring area to form a regional semantic data pool; Based on the work personnel and work type information, an association graph network is constructed, the work task type is identified by combining the behavior clustering mechanism, and the risk level is labeled; The graph neural network algorithm is applied to model the node attributes and edge weights, calculate the regional risk weight, and generate a risk level mapping diagram; Based on the time series analysis method, the graph structure is dynamically updated, a risk evolution prediction model is formed, and the scheduling module is linked to respond. 3.The low-voltage live-line work site visual monitoring optimization method based on the edge computing architecture according to claim 1, characterized in that, The construction of the risk scoring model comprises the following steps: Collecting image data, behavior feature data, and environmental parameter indicators of the monitoring area and performing spatio-temporal alignment; Extracting image, behavior, and environmental multi-modal features and encoding processing, and realizing deep fusion expression through a heterogeneous feature embedding mechanism; Based on the fusion features, a graph attention network model is constructed to output the risk score of each monitoring area.

4. The method of claim 1, wherein the method is performed by an edge computing architecture. The generation of the risk perception matrix comprises the following steps: structuring and integrating the risk scores according to the area number to generate a risk perception matrix as the input basis for resource scheduling.

5. The method of claim 1, wherein the method further comprises: The introduction of the multi-objective weighted scheduling algorithm comprises the following steps: Based on the risk perception matrix, a risk score driven task risk-resource matching model is constructed, and a scheduling objective function is established; According to the risk score and task attributes, a corresponding priority queue of risk levels and computing resources is constructed to form a resource application package; According to the priority queue order, the computing resources are allocated, and in the case of resource bottleneck, the task rollback or migration strategy is executed; Collecting scheduling execution feedback indicators and dynamically adjusting scheduling parameters to improve the adaptability of the scheduling strategy to the work environment.

6. The method of claim 1, wherein the method further comprises: The establishment of the task execution feedback mechanism comprises the following steps: Collecting model identification accuracy, video processing delay, and behavior capture integrity indicators of each monitoring area to form a task running state table; According to the running state table, the stability of the task and the matching degree of the resource scheduling are analyzed, and the task execution stability index is calculated; Based on the feedback results, the model frequency, video parameters and concurrency level are dynamically adjusted to optimize the task execution parameters; According to the adjustment results, the scheduling strategy is corrected, the scheduling logic is optimized by reinforcement learning, and the regional scheduling priority is updated.

7. The method of claim 1, wherein the method further comprises: The steps for building the scheduling logic credibility evaluation module include: Record the priority allocation instructions and their execution paths in each round of scheduling, and establish a scheduling path linked list; According to the preset rules and risk-resource ratio model, identify and mark abnormal scheduling behavior; Archive and cluster analyze normal and abnormal scheduling paths to build a scheduling behavior record library and an abnormal feature template library; Template matching and risk prediction are performed on the current scheduling strategy to trigger scheduling warning and generate risk evolution curve.

8. The method of claim 1, wherein the method further comprises: Based on the time sequence supervised learning strategy, the historical job scene data and the current task data are fused to iteratively optimize the scheduling model parameters, including the following steps: Collect the input features, decision outputs and execution effect data of the scheduling task to build a task time sequence sample chain; Build a model that integrates long short-term memory network and attention mechanism, and optimize the scheduling parameters according to the supervision signal; According to the model prediction results, dynamically adjust the risk scoring mechanism and scheduling logic to realize adaptive update of the scheduling strategy; Use sliding window training and multi-model comparison mechanism to regularly update the scheduling model and deploy it to the edge system.

Images