Electric power operation behavior intelligent identification and evaluation method in multi-machine cooperation mode

Abstract

The invention discloses a power operation behavior intelligent identification and evaluation method in a multi-machine cooperation mode. The method is realized through multiple steps: multi-source data collaborative acquisition and time-space consistency calibration are realized by combining Beidou / GPS dual-mode positioning with an inertial measurement unit to acquire data, and a unified time-space coordinate system is constructed through an improved synchronization algorithm and time-space calibration; edge end adaptive feature enhancement and fusion are carried out, image details are enhanced by means of a generative adversarial network and the like, and multiple types of features are extracted; the multi-machine collaborative decision and risk judgment realize accurate risk assessment through a three-level fusion mechanism and a dynamic weight distribution strategy; performing dynamic evaluation and closed-loop management and control, establishing a multi-dimensional index system, generating a real-time report in combination with an improved fuzzy analytic hierarchy process, and triggering intervention; and carrying out data security storage and model optimization, and ensuring security and optimizing the model by adopting hierarchical encryption and federal learning.

Description

Technical Field

[0001] The present invention relates to the field of intelligent power operation safety management and control, and in particular to a method for intelligently identifying and evaluating power operation behaviors in a multi-machine collaborative mode. Background Art

[0002] With the deepening of the intelligent transformation of the power industry, the need for safety risk management and control at power operation sites is becoming increasingly urgent. Traditional power operation models, which rely on manual inspections and post-event tracing, are no longer able to accurately identify violations (such as failure to wear protective equipment and illegal use of fire) in real time. Existing monitoring equipment suffers from multiple technical bottlenecks. For example, conventional recorders on the market lack standardized interfaces and rigid functionality, making them unable to meet the needs of multi-vendor device collaboration and real-time monitoring in complex environments. Differences in charging interfaces and data protocols between manufacturers lead to inefficient data collection, and manual export of historical data is time-consuming and labor-intensive. Furthermore, while computer vision and deep learning have made progress in behavioral recognition in common scenarios, power operation scenarios are unique, requiring identification of seven complex behaviors, such as helmet / safety belt wearing status and proper use of insulating tools. Existing algorithms lack accuracy for small target detection and are less robust in dynamic environments, such as those caused by obstructions during high-altitude operations and drastic changes in lighting, resulting in high rates of missed detection. Furthermore, power operations require strict step completeness and sequencing, and existing technologies lack the ability to analyze the temporal sequence of operation trajectories.

[0003] To meet these challenges, the industry urgently needs to break through technical bottlenecks such as device compatibility, complex scene recognition, and multi-source data collaboration. For example, the multi-device high-precision synchronization system launched by Chengdu Zhongke Four Point Zero Technology achieves nanosecond-level timing consistency between multiple devices through master-slave device initialization and clock synchronization alignment mechanisms, combined with FPGA clock generation technology, effectively solving the problem of behavioral trajectory correlation difficulties caused by time asynchrony among multiple devices at power operation sites. In addition, the decimeter-level positioning accuracy and timing capabilities of the Beidou satellite navigation system, combined with the real-life three-dimensional scenes constructed by three-dimensional lidar, provide precise spatial location information for power operations, enabling real-time tracking of personnel locations and dynamic alarms. In terms of anti-interference technology, the AIM+ anti-interference technology of the Septentrio Mosaic module, through RF chip design and baseband algorithm processing, can maintain stable positioning and navigation capabilities in strong electromagnetic interference environments, significantly improving the reliability of power equipment in complex environments.

[0004] In terms of data management and collaboration, the high bandwidth and low latency of 5G networks, combined with the localized processing capabilities of edge computing, provide efficient data transmission and real-time analysis solutions for power monitoring systems. For example, the Acrel-2000 power monitoring system, which utilizes a layered distributed architecture, enables real-time monitoring and fault warnings for substation equipment. Data is pre-processed and cached through edge computing nodes, reducing data transmission latency and improving system response speed. Furthermore, the National Energy Administration's newly revised "Regulations on the Security Protection of Power Monitoring Systems" emphasizes the protection principles of security zoning, dedicated networks, horizontal isolation, and vertical authentication. It also introduces cutting-edge technologies such as security immunity and situational awareness, requiring power monitoring system suppliers to ensure that their equipment is secure and controllable, strengthening supply chain security management.

[0005] Furthermore, the lack of industry-specific solutions is gradually improving. For example, collaborative safety monitoring applications based on 3D real-world scenarios, using Beidou's high-precision positioning module and inertial measurement units, enable digital scene construction of work sites and accurate analysis of personnel behavior. AI video analysis algorithms integrated into surveillance cameras can identify workers' clothing and work behavior in real time, and enable timely intervention in violations through voice prompts and platform warnings. These technological innovations not only improve the safety and efficiency of power operations but also provide strong support for the establishment of a closed-loop control system of "collection-identification-evaluation-intervention."

[0006] In summary, safety management and control at power operation sites requires the integration of advanced technologies such as high-precision timing, anti-interference algorithms, edge computing, and three-dimensional real-scene modeling. At the same time, combined with industry-specific standards and policy requirements, a set of solutions requiring hardware customization, intelligent algorithms, and distributed architecture collaboration should be built to address core challenges such as poor equipment compatibility, difficulty identifying complex environments, and low data management efficiency, thereby driving the power industry towards intelligent and automated development. Summary of the Invention

[0007] The technical problem to be solved by the present invention is a method for intelligent identification and evaluation of power operation behavior in a multi-machine collaborative mode. The method is based on multi-step implementation and solves a series of technical problems of intelligent identification and evaluation of power operation behavior in a multi-machine collaborative mode.

[0008] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows: A method for intelligent identification and evaluation of power operation behavior in a multi-machine collaborative mode is characterized by comprising the following steps: A1 multi-source data collaborative acquisition and spatiotemporal consistency calibration, A2 edge-end adaptive feature enhancement and fusion, A3 multi-machine collaborative decision-making and risk assessment, A4 dynamic evaluation and closed-loop control, and A5 data security storage and model optimization.

[0009] As a preferred technical solution of the present invention, the A1 multi-source data collaborative collection and precise spatiotemporal calibration: multiple intelligent collection terminals with multimodal perception capabilities are deployed at the power operation site, using Beidou and GPS dual-mode positioning technology combined with an inertial measurement unit to synchronously collect high-definition audio and video, personnel three-dimensional position and motion trajectory, and equipment operating status parameters including but not limited to voltage, current, and temperature data. Through an improved synchronization algorithm and a deep learning-based multi-perspective spatial calibration model, consistent calibration of multi-terminal data with nanosecond time accuracy and sub-meter spatial accuracy is achieved, and a high-precision unified scene spatiotemporal coordinate system is constructed to eliminate data deviations caused by differences in equipment installation position and angle. The improved synchronization algorithm is as follows: using high-precision timing signals such as Beidou satellites to ensure the initial time is accurate, and the device's own clock deviation is fine-tuned in real time through its own dynamic calibration mechanism; at the same time, it filters out abnormal errors caused by factors including but not limited to electromagnetic interference and network fluctuations.

[0010] As a preferred technical solution of the present invention, the A2 edge-end intelligent feature extraction and enhancement: the collected data is transmitted to the edge computing node, and the environmental parameters including but not limited to light, humidity, and electromagnetic interference are monitored in real time through the adaptive environmental perception module; based on these parameters, the filter kernel parameters are automatically adjusted according to the light intensity and noise type to perform noise reduction processing on the video data, and the generative adversarial network is combined with the attention mechanism to enhance the image details in low-light and high-noise environments, highlighting the key features of personnel and equipment; at the same time, the cross-modal feature fusion network is used to extract personnel behavior characteristics including but not limited to the wearing status of safety helmets, safety belts, and insulating gloves, tool operation action sequences and equipment interaction characteristics including but not limited to equipment connection, operation instruction execution status and environmental risk characteristics including but not limited to dust concentration exceeding the standard and abnormal electromagnetic radiation areas.

[0011] As a preferred technical solution of the present invention, the A3 multi-machine collaborative intelligent decision-making and accurate risk assessment: constructing a distributed hierarchical collaborative decision-making model, including a three-level fusion mechanism of data layer, feature layer and decision layer, specifically: At the data layer, A3-1 integrates multi-terminal perception data through timestamps and spatial coordinate transformations to generate point cloud maps and spatiotemporal sequence data of the global scene; At the feature layer, A3-2 dynamically allocates weights based on the geometric relationship between the device's viewpoint and the operating area, resolution, and environmental interference, and prioritizes integrating behavior and risk features from key viewpoints to generate a comprehensive feature vector. A3-3 At the decision-making level, using improved evidence theory and introducing prior knowledge and dynamic update mechanisms, the multi-terminal recognition results are integrated to output violation behavior judgments, including but not limited to failure to operate equipment according to specifications, illegal use of fire, and high-altitude throwing. Risk levels are assessed at high, medium, and low levels, and a confidence score of 0-1 is given. The improved evidence theory and the introduction of prior knowledge and a dynamic update mechanism are as follows: a dynamic update mechanism is implemented based on prior knowledge including but not limited to safety rules for power operations and historical violation cases; basic judgment rules are first set based on existing experience, and then the detection results of multiple devices are compared with these rules, and a comprehensive judgment is made through a method similar to "voting"; at the same time, the system continuously adjusts the judgment criteria based on new monitoring data.

[0012] As a preferred technical solution of the present invention, the A4 dynamic multi-dimensional evaluation and real-time closed-loop control: establish a dynamic multi-dimensional evaluation index system, including four dimensions: A4-1 safety compliance, A4-2 operational standardization, and A4-3 task completion quality and efficiency, specifically: A4-1 Safety compliance: A rule engine is built based on the power industry's safety regulations and standards to verify the compliance of operational processes in real time, including but not limited to power outages, electrical testing, and grounding operation sequences, and the implementation of safety measures, including but not limited to maintaining safe distances and wearing protective equipment. In terms of A4-2 operational standardization, the Transformer architecture is used to introduce position encoding and multi-head attention mechanisms to evaluate the completeness and correctness of operational steps, including but not limited to equipment maintenance procedures and switching operation procedures; In terms of A4-3 task completion quality and efficiency, the task completion time and operation accuracy are calculated by combining personnel and equipment trajectory data with the task plan, including but not limited to quantitative indicators such as equipment debugging accuracy and line connection accuracy.

[0013] As a preferred technical solution of the present invention, the A4 dynamic multi-dimensional evaluation and real-time closed-loop control is further characterized by: using an improved fuzzy hierarchical analysis method combined with grey correlation analysis to fuse the evaluation results, generating a real-time report containing risk level, violation traceability, and improvement suggestions, and pushing it to the operation terminal and management platform via a 5G encrypted communication link. The management platform automatically triggers audio and visual alarms, sends rectification notices, and intelligently blocks and intervenes in illegal operation processes based on the report results. The improved fuzzy hierarchical analysis method specifically breaks down the assessment objectives into multiple measurable small indicators, which are then scored layer by layer. The importance of each small indicator is first determined, and then a "fuzzy but reasonable" score is given for each indicator. The importance of each factor can be flexibly changed based on real-time operation scenarios and historical data.

[0014] As a preferred technical solution of the present invention, the A5 data security storage and efficient model optimization: a hierarchical encryption storage architecture is designed, and a distributed encryption NAS system based on blockchain is used at the edge to store original data and intermediate calculation results. The cloud uses a consortium chain combined with quantum encryption technology to store key evaluation results and decision data, making the data tamper-proof and highly secure. In terms of model optimization, each edge node trains a lightweight model based on local data, and the cloud collaboratively optimizes the global model through a transfer learning algorithm. At the same time, samples are selected based on model uncertainty and data diversity to dynamically update training data to avoid security risks caused by the external transmission of original data.

[0015] As a preferred technical solution of the present invention, the dynamic weight allocation strategy generates a weight matrix ranging from 0.1 to 1.0 by calculating the geometric correlation between the device perspective and the working area, the resolution contribution rate and the environmental interference coefficient. Specifically, the data of devices that are close to the working point, have clear shots and are less affected by interference are more dominant in judging violations; conversely, the data weight of devices in remote locations, blurred images and those affected by electromagnetic interference is reduced and is only used as an auxiliary reference; the system automatically generates a weight value of 0.1 to 1.0 for each device by real-time evaluation of the device's perspective effectiveness, data quality and degree of environmental interference, giving priority to integrating information from high-weight devices to reduce misjudgments in scenarios such as occlusion and low resolution.

[0016] As a preferred technical solution of the present invention, a system for intelligent identification and evaluation of electric power operation behavior in a multi-machine collaborative mode is characterized by comprising: B1 perception layer: It consists of multiple intelligent perception terminals, integrating environmental sensors, positioning modules and wide-angle lenses to achieve 360° data collection without blind spots; B2 edge layer: deploys multi-protocol adapters and edge computing units to support data preprocessing, feature extraction, and local decision-making; B3 cloud layer: includes a distributed decision engine, dynamic evaluation module, and secure storage cluster, enabling cross-regional collaboration through a microservices architecture; B4 application layer: provides a visual management and control interface, supports risk heat map generation, work report export, and API docking with third-party platforms.

[0017] As a preferred technical solution of the present invention, the cloud-layer integrated digital twin module builds a virtual mirror of the operation scene based on laser point cloud scanning technology, maps personnel behavior trajectories and equipment status in real time, and supports risk simulation and emergency strategy rehearsal.

[0018] The beneficial effects of adopting the above technical solutions are: real-time identification of key violations and dynamic verification of operational compliance combined with anti-interference hardware design and environmental adaptive algorithms to ensure stable operation of equipment in complex scenarios; through automated data processing, multi-machine collaborative decision-making and lightweight hardware adaptation, data collection efficiency is improved and cross-manufacturer equipment is compatible; at the same time, hierarchical encryption, blockchain evidence storage and federated learning technologies are used to promote continuous optimization of the model on the basis of ensuring data security and privacy, and ultimately build a closed-loop control system covering "people-equipment-environment", providing technical support for the intelligent transformation of the power industry. Figure 1 : Example image of helmet wearing recognition area; Figure 2 : Image processing flow chart; Figure 3 : Risk heat map; Figure 4 : Intelligent terminal parameter diagram; Figure 5 : Algorithm model deployment flow chart; Figure 6 : Multiple personnel gather real-time monitoring images; Figure 7 : Real-time operation monitoring diagram; Figure 8 : Schematic diagram of multi-stage collaboration; Figure 9 : Schematic diagram of feature fusion extraction; Attachment Figure 10 : Real-time snapshot of substation operation. DETAILED DESCRIPTION

[0019] The principles of the present disclosure will now be described with reference to several exemplary embodiments shown in the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that these embodiments are only described to facilitate those skilled in the art to better understand and implement the present disclosure, and are not intended to limit the scope of the present disclosure in any way. In the following description of the embodiments, specific details such as specific system structures and technologies are provided for illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, and methods are omitted to avoid obscuring the description of the present application with unnecessary details.

[0020] It should be understood that when used in this specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, components.

[0021] Example 1: Overview of overall steps This technical solution targets the safety management and control needs of power operations and builds a complete technical system that integrates multi-source data collection, edge intelligent processing, distributed decision-making, and closed-loop control. It achieves intelligent identification and evaluation of operational behaviors through five core steps.

[0022] 1. Collaborative Collection of Multi-Source Data and Temporal and Spatial Consistency Calibration Reference Attachment Figure 6 and 7 , deploying intelligent terminals with multimodal sensing capabilities at power operation sites, integrating Beidou / GPS dual-mode positioning modules and inertial measurement units, to simultaneously collect high-definition audio and video, three-dimensional personnel trajectories, and equipment operating parameters (voltage / current / temperature, etc.). Innovatively adopting improved space-time calibration technology: At the time synchronization level, relying on the high-precision timing signals of Beidou satellites to establish the initial time benchmark, combined with the equipment's dynamic calibration mechanism to correct clock deviations in real time, while filtering out abnormal errors such as electromagnetic interference and network fluctuations, achieving nanosecond-level time accuracy uniformity; at the spatial calibration level, using a deep learning-driven multi-perspective geometric matching model, sub-meter-level calibration is performed on the coordinate deviations caused by differences in installation position and angle of different devices, establishing a unified space-time coordinate system for the entire domain, and providing a precise space-time benchmark for subsequent data analysis.

[0023] 2. Edge-side Adaptive Feature Enhancement and Fusion Reference Attachment Figure 9 , the collected data is transmitted to the edge computing node in real time, and the environmental perception module dynamically monitors parameters such as light, humidity, and electromagnetic interference to drive the adaptive preprocessing algorithm: for video data, the filter kernel parameters are intelligently adjusted according to the light intensity and noise type, and the generative adversarial network (GAN) and attention mechanism are combined to enhance the image details in low-light and high-noise environments, and highlight key target features such as safety helmets and insulating gloves; in the feature extraction link, the cross-modal fusion network is used to integrate two-dimensional image texture and three-dimensional motion trajectory information to realize the extraction of three types of core features: personnel behavior characteristics (protective equipment wearing status, tool operation sequence), equipment interaction characteristics (connection status, instruction execution status), and environmental risk characteristics (dust concentration exceeding the standard area, electromagnetic radiation abnormal points), providing multi-dimensional feature vectors for upper-level decision-making.

[0024] 3. Multi-machine Collaborative Intelligent Decision-making and Risk Assessment Reference Attachment Figure 8, constructing a distributed hierarchical decision-making model, which includes a three-level fusion mechanism: the data layer integrates multi-terminal data into a global point cloud map and time-space sequence through timestamp alignment and spatial coordinate transformation, and constructs a visual scene model; the feature layer introduces a dynamic weight allocation strategy, comprehensively considers the geometric correlation of the equipment perspective (distance and angle with the operating area), resolution contribution rate (image clarity, target pixel ratio), and environmental interference coefficient (light / occlusion / electromagnetic interference influence), generates a 0.1-1.0 dynamic weight matrix, prioritizes the integration of key features of high-weight terminals, and improves the reliability of complex scene features; the decision-making layer is based on improved evidence theory, integrates prior knowledge such as power safety rules and historical violation cases, calibrates the judgment criteria in real time through a dynamic update mechanism, and integrates the multi-terminal recognition results in a "multi-source voting" manner to output 7 types of violation behavior judgments (non-standard operation of equipment, illegal fire, etc.), three-level risk levels (high / medium / low) and 0-1 confidence scores, to achieve accurate quantitative assessment of risks.

[0025] 4. Dynamic Multi-dimensional Evaluation and Real-time Closed-Loop Control A dynamic evaluation system encompassing three key dimensions has been established. For safety compliance, a rules engine, based on power industry safety regulations, is built to verify the compliance of 27 operational processes, including "power outage, electrical testing, and grounding," as well as the implementation of safety measures (e.g., safety distances and the wearing of protective equipment) in real time. For operational compliance, a Transformer-based time series analysis model, incorporating position encoding and multi-head attention mechanisms, assesses the completeness and sequence correctness of steps in processes such as equipment maintenance and switching operations. For task effectiveness, efficiency and quality indicators, such as task completion time and equipment commissioning accuracy, are quantified by combining personnel / equipment trajectories with task plans. By integrating multi-dimensional evaluation results through an improved fuzzy analytic hierarchy process (FAHP), a real-time report is generated containing risk levels, violation traceability, and improvement suggestions. This report is then pushed to the operational terminal and management platform via a 5G encrypted link, triggering audible and visual alarms, rectification notices, and intelligent blocking of non-compliant processes, forming a closed-loop "identification-evaluation-intervention" management system.

[0026] 5. Secure Data Storage and Efficient Model Optimization A hierarchical encrypted storage architecture is designed to ensure data security: A distributed encrypted NAS system is built using blockchain technology at the edge, enabling local and secure storage of raw data and intermediate results. The cloud utilizes a consortium blockchain combined with quantum cryptography to provide tamper-proof evidence of key evaluation results and decision-making data, supporting secure cross-region sharing using CIFS / NFS protocols. For model optimization, each edge node trains a lightweight model (parameter size ≤ 10MB) based on local data. The cloud aggregates model parameters using a federated transfer learning algorithm and, combined with an active learning strategy, dynamically selects high-value samples to update training data. This ensures continuous optimization of the global model (cross-scenario generalization accuracy ≥ 92%) while protecting data privacy and minimizing the risk of raw data leakage.

[0027] Reference Attachment Figure 5 , system architecture design 1. Perception Layer It is composed of a three-dimensional perception network composed of multiple intelligent terminals, integrating environmental sensors (temperature, humidity, electromagnetic field), Beidou / GPS positioning modules and wide-angle lenses, supporting 360° data collection without blind spots, meeting the requirements of operating in a wide temperature environment of -40℃~70℃, and having IP68 protection level and anti-electromagnetic interference capabilities.

[0028] (2) Marginal layer Deploy multi-protocol adapters and edge computing units to support data access from multiple interfaces such as Type-C / Micro USB, implement data pre-processing (noise reduction / enhancement), feature extraction, and local primary decision-making (single-frame processing delay ≤ 50ms), and provide terminals with integrated automatic charging and data transmission services (single-port charging efficiency ≥ 3A, upload rate ≥ 100MB / s).

[0029] (3) Cloud layer It includes a distributed decision-making engine (three-level fusion model), a dynamic evaluation module (three-dimensional indicator system) and a secure storage cluster (blockchain evidence storage + federated learning platform), and realizes cross-regional collaboration through a microservice architecture; it integrates a digital twin module to build a virtual mirror of the operation scene based on laser point cloud scanning, mapping personnel trajectories and equipment status in real time, and supporting risk simulation and emergency rehearsals.

[0030] (4) Application layer It provides a visual management and control interface, supports real-time display of risk heat maps, automatic generation and export of work reports, and API interface development, and seamlessly connects to third-party platforms such as power ERP and production management to achieve centralized display of "people-equipment-environment" data and business collaboration.

[0031] Through technological innovation and architectural optimization, this technical solution breaks through the compatibility, real-time and robustness bottlenecks of traditional power operation monitoring, forming an industry-specific solution that integrates "precise perception - intelligent analysis - closed-loop control - continuous evolution", significantly improving operational safety and management efficiency.

[0032] Example 2: Safety Management and Control of Substation Equipment Maintenance Operations Reference Attachment Figure 10 ,Scenario: In June 2025, the No. 2 main transformer of a 500kV substation was overhauled. Six workers were divided into a high-altitude group (3 people), a ground group (2 people), and a monitoring group (1 person). They needed to perform the "power outage-electricity test-grounding-equipment replacement" process.

[0033] Operational procedures and data support Equipment deployment and initialization (08:00-08:10) Deploy four head-mounted recorders (A / B / C / D): A / B are fixed to the aerial work vehicle operating console (facing the equipment area, 5m away, 4K resolution), and C / D are worn by ground personnel (with a side viewing angle of 45°, a distance of 10m, and a resolution of 1080P). Time and space calibration: Beidou timing + Kalman filtering is used to achieve multi-device time synchronization (clock deviation ≤ 50ns), and spatial calibration is completed based on the Transformer feature matching model (the coordinate system error between devices A and B is 8cm).

[0034] Reference Attachment Figure 1 and 2 , Data Acquisition and Edge Preprocessing (08:15-09:30) The high-altitude crew wears a safety belt to enter the work area. Device A collects the "seat belt buckled state" feature (confidence level 0.95). Device B collects the feature confidence level 0.82 due to slight obstruction (obstruction rate 15%). The edge computing unit performs real-time processing: CycleGAN enhancement is applied to low-light images of device C (20 lux), increasing the helmet recognition accuracy from 75% to 92%. The tool operation trajectory is also extracted (the torque value of the insulated wrench is 120 N·m, and the compliance range is 100-150 N·m).

[0035] Multi-machine collaborative decision-making and risk assessment (09:40) Reference Attachment Figure 3 ,Dynamic weight allocation: Device A has a weight of 0.95 because it faces the work point (geometric correlation 0.9), has high resolution (resolution contribution rate 1.0), and has no interference (environmental coefficient 0.95); Device D has a side view and is far away (geometric correlation 0.6), so it has a weight of 0.6; The decision-making layer fusion result: The high-altitude group personnel "did not wear the double-hook safety belt according to the specifications" (equipment A / B both detected single hook hanging, confidence level 0.92+0.85, risk level 4 after evidence theory synthesis), triggering the terminal vibration alarm (response within 0.3 seconds).

[0036] Closed-loop control and data storage (10:00) The management platform generates a rectification work order: mark the violation time (09:40:12), responsible person (work number 003), and evidence image (snapshot of device A); Data storage: Edge-side encrypted storage of original video (10GB), cloud-based blockchain storage of evaluation results (hash value generation takes 200ms), and federated learning to update the "seat belt wearing" model parameters (local training takes 15 minutes).

[0037] Technical effect: The seat belt wearing recognition accuracy rate in complex occlusion scenarios is 91% (traditional single device is 73%), and the process compliance verification takes ≤ 2 seconds / step, which is 400% more efficient than manual inspection.

[0038] Example 3: Risk Prevention and Control of High-Above-Head Transmission Line Operations Scenario: In December 2025, a 220kV transmission line was installed in a mountainous area. The ambient temperature was -15°C, the wind speed was level 6, and there was electromagnetic interference (electromagnetic field strength 0.8μT).

[0039] Operational procedures and data support Reference Attachment Figure 4 , Extreme Environment Equipment Deployment (07:30-08:00) Two fixed monitoring devices (E / F) were installed at the tower base (15 meters from the operation point, IP68 protection, and resistant to -40°C low temperature). One drone was equipped with a mobile terminal (G, with a flight time of 15 hours and support for 50x zoom). Time and space calibration: When the Beidou signal is blocked by mountains, it switches to inertial navigation compensation (the positioning accuracy is temporarily maintained at 1.2m and calibrated to 0.5m within 5 seconds after the signal is restored).

[0040] Multimodal Data Acquisition and Enhancement (08:30-10:00) Drone terminal G captured images of overhead power lines (target pixels accounted for 0.3%). After being enhanced with the attention mechanism, the accuracy of wire connector bolt recognition increased from 80% to 95%. Device E detects an electromagnetic field anomaly (0.8μT, threshold 0.5μT), triggering anti-interference mode: the edge computing unit switches to a low-noise filtering algorithm, and the data processing delay is reduced from 80ms to 60ms.

[0041] Dynamic Weight Allocation and Collaborative Decision-Making (10:15) Weight calculation: Drone G has an unobstructed view from high altitude (geometric correlation 0.85), but is at a long distance (18m, geometric weight 0.7), a resolution contribution rate of 0.8, an environmental factor of 0.8 (unobstructed at low temperatures), and a comprehensive weight of 0.78; Identification results: The ground personnel were found to be "touching the wire without wearing insulating gloves" (Equipment E / F detected the exposed skin area of ​​the hand, confidence level 0.91, UAV G-assisted positioning coordinates (X=123.456, Y=78.901), risk level 3).

[0042] Real-time Intervention and Model Optimization (10:20) The operating terminal receives an alarm: "Insulation protection is missing, stop operation immediately!" (voice broadcast + red dot flashing on the screen), and the operator completes the correction within 10 seconds; Cloud records: This scenario added 32 new "low-temperature environment glove recognition" samples. By updating the model through federated learning, the recognition accuracy of similar scenarios increased from 85% to 92%.

[0043] Technical effect: The device can run continuously for 14 hours in a low-temperature environment of -15°C (nominal endurance is 12 hours, dynamic power adjustment saves 30% power), the false alarm rate in strong electromagnetic interference scenarios is reduced by 60%, and the risk event response time is ≤1 second.

[0044] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for intelligent identification and evaluation of power operation behavior in a multi-machine collaborative mode, characterized by: The following steps are involved: A1 Collaborative collection of multi-source data and temporal and spatial consistency calibration, A2 Edge-end adaptive feature enhancement and fusion, A3 Multi-machine collaborative decision-making and risk assessment, A4 Dynamic evaluation and closed-loop control, A5 Data security storage and model optimization.

2. The method for intelligently identifying and evaluating electric power operation behaviors in a multi-machine collaborative mode according to claim 1 is characterized by: The A1 multi-source data collaborative collection and precise spatiotemporal calibration: Multiple intelligent collection terminals with multimodal perception capabilities are deployed at the power operation site. They utilize Beidou and GPS dual-mode positioning technology combined with an inertial measurement unit to synchronously collect high-definition audio and video, personnel's three-dimensional position and motion trajectory, and equipment operating status parameters including but not limited to voltage, current, and temperature data. Through an improved synchronization algorithm and a deep learning-based multi-perspective spatial calibration model, consistent calibration of multi-terminal data with nanosecond temporal accuracy and sub-meter spatial accuracy is achieved, establishing a high-precision unified scene spatiotemporal coordinate system to eliminate data deviations caused by differences in equipment installation positions and angles. The improved synchronization algorithm is as follows: using high-precision timing signals such as Beidou satellites to ensure the initial time is accurate, and the device's own clock deviation is fine-tuned in real time through its own dynamic calibration mechanism; at the same time, it filters out abnormal errors caused by factors including but not limited to electromagnetic interference and network fluctuations.

3. The method for intelligent identification and evaluation of electric power operation behaviors in a multi-machine collaborative mode according to claim 1 is characterized by: A2 edge-end intelligent feature extraction and enhancement: The collected data is transmitted to the edge computing node, and the adaptive environment perception module monitors environmental parameters including but not limited to light, humidity, and electromagnetic interference in real time; Based on these parameters, the filter kernel parameters are automatically adjusted according to the light intensity and noise type to perform denoising on the video data, and a generative adversarial network combined with an attention mechanism is used to enhance image details in low-light and high-noise environments, highlighting the key features of people and equipment. At the same time, a cross-modal feature fusion network is used to extract personnel behavior features, including but not limited to the wearing status of safety helmets, safety belts, and insulating gloves, tool operation action sequences, and equipment interaction features, including but not limited to equipment connection, operation instruction execution status, and environmental risk features, including but not limited to excessive dust concentration and abnormal electromagnetic radiation areas.

4. The method for intelligently identifying and evaluating electric power operation behaviors in a multi-machine collaborative mode according to claim 1 is characterized by: The A3 multi-machine collaborative intelligent decision-making and accurate risk assessment: build a distributed hierarchical collaborative decision-making model, including a three-level fusion mechanism of data layer, feature layer and decision layer, specifically: At the data layer, A3-1 integrates multi-terminal perception data through timestamps and spatial coordinate transformations to generate point cloud maps and spatiotemporal sequence data of the global scene; At the feature layer, A3-2 dynamically allocates weights based on the geometric relationship between the device's viewpoint and the operating area, resolution, and environmental interference, and prioritizes integrating behavior and risk features from key viewpoints to generate a comprehensive feature vector. At the decision-making level, A3-3 utilizes improved evidence theory, introduces prior knowledge, and employs a dynamic update mechanism. It integrates multi-terminal recognition results to output violation determinations, including but not limited to failure to operate equipment according to specifications, illegal use of fire, and throwing objects from heights. It assesses risk levels at high, medium, and low levels, and assigns a confidence score of 0 to 1. The improved evidence theory and the introduction of prior knowledge and a dynamic update mechanism are as follows: a dynamic update mechanism is implemented based on prior knowledge including but not limited to safety rules for power operations and historical violation cases; basic judgment rules are first set based on existing experience, and then the detection results of multiple devices are compared with these rules, and a comprehensive judgment is made through simulated "voting"; at the same time, the system continuously adjusts the judgment criteria based on new monitoring data.

5. The method for intelligent identification and evaluation of electric power operation behavior in a multi-machine collaborative mode according to claim 1 is characterized by: The A4 dynamic multi-dimensional evaluation and real-time closed-loop control: establish a dynamic multi-dimensional evaluation indicator system, including four dimensions: A4-1 safety compliance, A4-2 operational standardization, and A4-3 task completion quality and efficiency. Specifically: A4-1 Safety compliance: A rule engine is built based on the power industry's safety regulations and standards to verify the compliance of operational processes in real time, including but not limited to power outages, electrical testing, and grounding operation sequences, and the implementation of safety measures, including but not limited to maintaining safe distances and wearing protective equipment. In terms of A4-2 operational standardization, the Transformer architecture is used to introduce position encoding and multi-head attention mechanisms to evaluate the completeness and correctness of operational steps, including but not limited to equipment maintenance procedures and switching operation procedures; In terms of A4-3 task completion quality and efficiency, the task completion time and operation accuracy are calculated by combining personnel and equipment trajectory data with the task plan, including but not limited to quantitative indicators such as equipment debugging accuracy and line connection accuracy.

6. According to the method for intelligent identification and evaluation of electric power operation behavior in a multi-machine collaborative mode according to claim 5, the A4 dynamic multi-dimensional evaluation and real-time closed-loop control is further characterized by: An improved fuzzy hierarchical analysis method combined with grey correlation analysis integrates evaluation results to generate real-time reports containing risk levels, violation traceability, and improvement suggestions. These reports are then pushed to the operation terminal and management platform via a 5G encrypted communication link. Based on the report results, the management platform automatically triggers audio and visual alarms, sends rectification notices, and intelligently blocks and intervenes in illegal operation processes. The improved fuzzy hierarchical analysis method specifically breaks down the assessment objectives into multiple measurable small indicators, which are then scored layer by layer. The importance of each small indicator is first determined, and then a "fuzzy but reasonable" score is given to each indicator. The importance of each factor can be flexibly changed based on real-time operation scenarios and historical data.

7. The method for intelligent identification and evaluation of electric power operation behaviors in a multi-machine collaborative mode according to claim 1 is characterized by: The A5 data security storage and efficient model optimization: a hierarchical encrypted storage architecture is designed, and a distributed encrypted NAS system based on blockchain is used at the edge to store original data and intermediate calculation results. The cloud uses a consortium chain combined with quantum encryption technology to store key evaluation results and decision data, making the data tamper-proof and highly secure. In terms of model optimization, each edge node trains a lightweight model based on local data, and the cloud collaboratively optimizes the global model through a transfer learning algorithm. At the same time, samples are selected based on model uncertainty and data diversity to dynamically update training data to avoid security risks caused by the external transmission of original data.

8. The method for intelligent identification and evaluation of electric power operation behaviors in a multi-machine collaborative mode according to claim 5 is characterized by: The dynamic weight allocation strategy generates a weight matrix ranging from 0.1 to 1.0 by calculating the geometric correlation between the device's perspective and the work area, the resolution contribution rate, and the environmental interference coefficient. Specifically, the data of devices that are close to the work point, have clear images, and are less affected by interference are more dominant in judging violations; conversely, the data weight of devices in remote locations, with blurred images, and those affected by electromagnetic interference is reduced and only used as an auxiliary reference; the system automatically generates a weight value of 0.1 to 1.0 for each device by real-time evaluation of the device's perspective effectiveness, data quality, and degree of environmental interference, and prioritizes the integration of information from high-weight devices to reduce misjudgments in scenarios such as occlusion and low resolution.

9. An intelligent identification and evaluation system for electric power operation behavior in a multi-machine collaborative mode, characterized by: include: B1 perception layer: It consists of multiple intelligent perception terminals, integrating environmental sensors, positioning modules and wide-angle lenses to achieve 360° data collection without blind spots; B2 edge layer: deploys multi-protocol adapters and edge computing units to support data preprocessing, feature extraction, and local decision-making; B3 cloud layer: includes a distributed decision engine, dynamic evaluation module, and secure storage cluster, enabling cross-regional collaboration through a microservices architecture; B4 application layer: provides a visual management and control interface, supports risk heat map generation, work report export, and API docking with third-party platforms.

10. The intelligent identification and evaluation system for electric power operation behavior in a multi-machine collaborative mode according to claim 9 is characterized by: The cloud-layer integrated digital twin module builds a virtual mirror of the operation scene based on laser point cloud scanning technology, maps personnel behavior trajectories and equipment status in real time, and supports risk simulation and emergency strategy rehearsal.

Images