Intelligent round patrol and key area scheduling method for power plant work site

Abstract

The application discloses an intelligent round-patrol and key area scheduling method for power plant operation sites, comprising the following steps: step one, constructing a dynamic risk situation; step two, planning an adaptive round-patrol path; step three, distributing and executing a round-patrol task; step four, identifying and marking a key area; step five, implementing multi-level resource scheduling; step six, restoring the round-patrol state and archiving data. Based on real-time operation risk quantitative evaluation, the application dynamically generates an optimal round-patrol path and sampling frequency, and triggers a multi-level scheduling strategy after identifying a key area, thereby significantly improving the safety control level and emergency response efficiency of the power plant site; through the closed-loop feedback mechanism between the risk situation map and the round-patrol task, the application can realize intelligent inclination of monitoring resources to high-risk dynamic areas without increasing hardware costs, and effectively overcomes the defects of the fixed round-patrol mode.

Description

Technical Field

[0001] This invention relates to the field of power plant safety monitoring and intelligent inspection technology, specifically to a method for intelligent patrol and key area dispatching at power plant operation sites. Background Technology

[0002] Power plant operating environments are complex, containing hazardous sources such as high temperature, high pressure, and high voltage, requiring extremely high real-time performance and reliability for safety monitoring. Traditional video surveillance systems mostly use fixed cameras for fixed-point shooting or manual operation of pan-tilt units for inspection, making it difficult to achieve comprehensive coverage of the entire operating area without blind spots.

[0003] While existing technologies employ patrol schemes based on preset trajectories, their patrol paths and time intervals are fixed and cannot be dynamically adjusted according to real-time operational risks. This results in insufficient monitoring frequency in high-risk work areas, while potentially wasting monitoring resources in low-risk areas. Furthermore, when an anomaly occurs in a certain area (such as equipment alarms or unauthorized personnel entry), the dispatch system often responds slowly, failing to quickly locate and continuously track the event's evolution, thus missing the optimal time for response. Summary of the Invention

[0004] To address this issue, the present invention provides an intelligent patrol and key area dispatching method for power plant operation sites, in order to solve the problem that existing dispatching systems often have a delayed response, are unable to quickly locate and continuously track the evolution of events, and miss the best time to handle them.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] The intelligent patrol and key area dispatching method for power plant operations includes the following steps:

[0007] Step 1: Construct a dynamic risk situation by reading real-time personnel distribution, equipment status, and environmental sensor data at the power plant operation site, mapping multi-source data to grid cells of the electronic map, calculating the real-time risk coefficient of each grid cell, and generating a dynamic risk situation map that updates over time.

[0008] Step 2: Plan an adaptive polling path. Based on the dynamic risk situation map generated in Step 1, select grid cells whose real-time risk coefficient exceeds the first threshold as mandatory polling points. Use a path planning algorithm to generate a polling path that connects all mandatory polling points and passes through other grid cells under the constraints of monitoring resources. Assign a dwell time to each dwell point on the path that is proportional to the real-time risk coefficient of the grid cell in which it is located.

[0009] Step 3: Distribute and execute the polling task. Decompose the polling path planned in Step 2 into a sequence of instructions that includes the target camera identifier, pan-tilt angle, lens magnification and dwell time. Send the instructions to the corresponding camera or edge gateway through the communication network to drive it to perform image acquisition and bitstream transmission according to the instruction sequence.

[0010] Step 4: Identify and mark key areas. During or outside of the round-robin execution, perform real-time analysis on the returned video stream. When a preset abnormal event is detected or the average real-time risk coefficient of a continuous area in the dynamic risk situation map exceeds the second threshold and continues for a preset duration, the area is marked as a key area.

[0011] Step 5: Implement multi-level resource scheduling. For the key areas marked in Step 4, execute Level 1 scheduling: immediately command the main camera covering the area to pause its patrol task, lock the pan-tilt unit to the event center, and perform optical zoom; execute Level 2 scheduling: search for other nearby cameras, calculate the angle and path for the pan-tilt unit to rotate to face the key area, and issue instructions for multi-angle collaborative monitoring; execute Level 3 scheduling: if personnel are detected in the key area, capture facial features and compare them with the authorized database, triggering alarms and linkage devices for unauthorized personnel;

[0012] Step Six: Restore the patrol status and archive the data. Once the events in the key area are determined to have subsided, unlock the area, restore the camera to the pan-tilt angle and focal length before scheduling, encapsulate the video, alarms and scheduling records during the event handling process into a data packet and store it in the historical database. Then, re-execute Step Two to generate the patrol path for the next cycle.

[0013] Preferably, the real-time risk coefficient of each grid cell is calculated in step one as follows: real-time risk coefficient = (personnel density weight × real-time number of personnel in the unit grid) + (equipment anomaly weight × real-time equipment anomaly index) + (environmental exceedance weight × real-time environmental exceedance index), where the sum of personnel density weight, equipment anomaly weight, and environmental exceedance weight is 1, and the real-time equipment anomaly index is obtained by normalizing the percentage deviation between the measured values ​​of vibration, temperature, or electrical parameters and the preset threshold.

[0014] Preferably, in step two, when generating the round-robin path using a path planning algorithm, an ant colony algorithm or a genetic algorithm is used, and the sum of minimizing the total round-robin cycle time and maximizing the risk coverage of all dwelling points is taken as the joint optimization objective, wherein the risk coverage is equal to the sum of the real-time risk coefficients of all dwelling points on the path divided by the sum of the real-time risk coefficients of all grid cells in the dynamic risk situation map.

[0015] Preferably, the process of identifying and marking key areas in step four further includes maintaining an event priority queue: the system presets a mapping table between event types and priorities, marking equipment fires or personnel falls as first-priority events, high-risk area intrusions as second-priority events, and equipment temperature warnings as third-priority events; when multiple key areas are marked simultaneously, the highest-priority events in each area are compared, and the key areas corresponding to the first-priority events are given priority to perform all scheduling in step five, while the key areas corresponding to the second-priority and third-priority events are assigned idle cameras to perform fast round-robin, the frame rate of which is lower than the normal round-robin frame rate and the dwell time is shorter than the normal round-robin dwell time.

[0016] Preferably, in step five, when performing first-level scheduling for the locked key area, a progressively magnified observation strategy is adopted: In the first stage, the main camera is controlled to adjust to a wide-angle field of view covering the entire event area and maintain this for a first preset time to acquire the global context image; in the second stage, based on the event centroid coordinates provided by the video analysis algorithm, the main camera gimbal is driven to perform precise positioning and gradually move to the telephoto end to acquire event details; in the third stage, when the detected event type is abnormal personnel posture, the trajectory prediction function is activated, the predicted position within a second preset time period is calculated based on the personnel's historical movement position, and the gimbal is rotated in advance to track the target and prevent it from moving out of the frame.

[0017] Preferably, the multi-level resource scheduling in step five adopts a closed-loop confirmation mechanism: each time the edge device receives a scheduling instruction containing the target camera identifier, pan-tilt angle, and lens magnification, it replies to the central server with an execution status code after execution. The execution status code includes three types: success, angle not reached, or rotating. The central server decides whether to issue the next scheduling instruction or resend the current instruction based on the received execution status code.

[0018] Preferably, after allocating dwell time to dwell points in step two, the system also performs a dwell time correction step: obtaining the "last valid inspection time" of the grid cell corresponding to each dwell point, calculating the time elapsed since the current system time, and if the time elapsed exceeds a preset forgetting time threshold, multiplying the dwell time of the dwell point by a correction coefficient that is proportional to the uninspected time elapsed. The value of the correction coefficient is between 1.0 and 2.0, and the product result is used as the final dwell time of the dwell point.

[0019] Preferably, before planning the adaptive roving path in step two, a predictive risk correction step is also performed: a prediction model based on the simple displacement method of time series is called, the historical value of the real-time risk coefficient corresponding to the current time point in the previous three roving cycles is read, the average value of the risk coefficient change in adjacent cycles is calculated as the prediction increment, and the real-time risk coefficient calculated in step one is summed with the prediction increment to obtain the planning risk coefficient. The selection of the forced roving points and the allocation of the dwell time both use the planning risk coefficient.

[0020] Preferably, after the events in the key area are determined to have subsided in step six, a phased recovery strategy is executed: In the first phase, the lowest-level auxiliary camera called in step five is released, and the auxiliary camera is ordered to scan the area surrounding the original key area at a rate lower than the normal sampling frequency for a third preset time; In the second phase, if no new events are triggered within the third preset time, the main camera is controlled to switch from the locked state to a focused panoramic state that performs small-range reciprocating scans around the event center point for a fourth preset time; In the third phase, the main camera is controlled to return to its original position, and all cameras are ordered to take a snapshot of a representative preset position for global anomaly comparison. After confirming that there are no anomalies, step two is activated to generate the complete round-robin path for the next cycle.

[0021] Preferably, in step four, when performing real-time analysis on the returned video stream, if the detected preset abnormal event type is a device spark or a temperature anomaly that reaches the third threshold, then in step five, when implementing multi-level resource scheduling, an additional linkage capture command is executed: the system calculates the precise coordinates of the spark or temperature anomaly in the electronic map, calls the non-main camera closest to the coordinate, drives its pan-tilt to rotate toward the coordinate and executes continuous capture mode, sets the capture frame rate to twice the normal frame rate, and the capture duration is 10 seconds. The captured image sequence is automatically stitched into a dynamic image segment, which is then attached to the encapsulated data packet and stored in the historical database.

[0022] The present invention has the following advantages: Based on real-time operational risk quantification assessment, the present invention dynamically generates the optimal patrol path and sampling frequency, and triggers multi-level scheduling strategies after identifying key areas, thereby significantly improving the safety management and emergency response efficiency of power plant sites; by establishing a closed-loop feedback mechanism between the risk situation map and the patrol task, the present invention can achieve intelligent tilting of monitoring resources towards high-risk dynamic areas without increasing hardware costs, effectively overcoming the defects of the fixed patrol mode. Attached Figure Description

[0023] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0024] Figure 1 A flowchart of an intelligent patrol and key area dispatching method for power plant operation sites provided in this application embodiment. Detailed Implementation

[0025] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Technical engineers in the field can make some non-essential improvements and adjustments to the present invention based on the above-described content. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figure 1 A method for intelligent patrol and key area dispatching at power plant operation sites includes the following steps:

[0027] Step 1: Construct a dynamic risk situation by reading real-time personnel distribution, equipment status, and environmental sensor data at the power plant operation site, mapping multi-source data to grid cells of the electronic map, calculating the real-time risk coefficient of each grid cell, and generating a dynamic risk situation map that updates over time.

[0028] Step 2: Plan an adaptive polling path. Based on the dynamic risk situation map generated in Step 1, select grid cells whose real-time risk coefficient exceeds the first threshold as mandatory polling points. Use a path planning algorithm to generate a polling path that connects all mandatory polling points and passes through other grid cells under the constraints of monitoring resources. Assign a dwell time to each dwell point on the path that is proportional to the real-time risk coefficient of the grid cell in which it is located.

[0029] Step 3: Distribute and execute the polling task. Decompose the polling path planned in Step 2 into a sequence of instructions that includes the target camera identifier, pan-tilt angle, lens magnification and dwell time. Send the instructions to the corresponding camera or edge gateway through the communication network to drive it to perform image acquisition and bitstream transmission according to the instruction sequence.

[0030] Step 4: Identify and mark key areas. During or outside of the round-robin execution, perform real-time analysis on the returned video stream. When a preset abnormal event is detected or the average real-time risk coefficient of a continuous area in the dynamic risk situation map exceeds the second threshold and continues for a preset duration, the area is marked as a key area.

[0031] Step 5: Implement multi-level resource scheduling. For the key areas marked in Step 4, execute Level 1 scheduling: immediately command the main camera covering the area to pause its patrol task, lock the pan-tilt unit to the event center, and perform optical zoom; execute Level 2 scheduling: search for other nearby cameras, calculate the angle and path for the pan-tilt unit to rotate to face the key area, and issue instructions for multi-angle collaborative monitoring; execute Level 3 scheduling: if personnel are detected in the key area, capture facial features and compare them with the authorized database, triggering alarms and linkage devices for unauthorized personnel;

[0032] Step Six: Restore the patrol status and archive the data. Once the events in the key area are determined to have subsided, unlock the area, restore the camera to the pan-tilt angle and focal length before scheduling, encapsulate the video, alarms and scheduling records during the event handling process into a data packet and store it in the historical database. Then, re-execute Step Two to generate the patrol path for the next cycle.

[0033] The following example illustrates a specific execution flow of this method.

[0034] Example 1:

[0035] Step S1: Construct a risk situation map of the work site

[0036] The system first acquires power plant work plans, personnel locations, equipment status, and environmental sensor data through interfaces. For example, it reads from the production management system that there are maintenance operations on the turbine platform from 10:00 to 12:00 on the current day, obtains the current personnel distribution density from the UWB positioning system, and receives real-time data from vibration and temperature sensors. The system maps this multi-source data onto a pre-built 3D or 2D electronic map of the power plant. Based on preset risk weights (e.g., personnel activity weight 0.4, equipment anomaly weight 0.35, environmental exceedance weight 0.25), it calculates the real-time risk coefficient for each grid cell on the map, forming a dynamically updated risk situation map over time.

[0037] Step S2: Perform dynamic round-robin path planning

[0038] Based on the risk situation map generated in step S1, the system runs a path planning algorithm. This algorithm has a joint objective of minimizing monitoring blind spots and maximizing the coverage of high-risk areas. Specifically, the algorithm decomposes the round-robin task into multiple dwell points (preset positions), each corresponding to a risk coefficient and an expected dwell time. The system first selects grid cells with risk coefficients higher than a first threshold (e.g., 0.7) as mandatory round-robin points. Then, using an ant colony algorithm or a genetic algorithm, under the constraint of remaining monitoring resources (available camera pan / tilt rotation capabilities), it plans a smooth round-robin path connecting all mandatory round-robin points and passing through other medium-risk areas. Simultaneously, the algorithm dynamically assigns a dwell time to each dwell point on the path, with higher-risk points having a longer dwell time (e.g., 30 seconds) than lower-risk points (e.g., 5 seconds).

[0039] Step S3: Distribute and execute round-robin tasks.

[0040] The planning engine converts the patrol path generated in the previous step into a specific instruction sequence. Each instruction includes the target camera ID, the pan / tilt angle (horizontal and vertical), the lens zoom ratio, and the execution duration. The system sends the instruction sequence to the corresponding network camera or edge gateway via a message queue. The camera executes the instructions, rotating sequentially to preset angles to acquire images or videos, and then sends the bitstream back to the central server for real-time analysis. The entire patrol cycle repeats continuously, and after each cycle, it returns to step S1 to reread the field data, thus achieving closed-loop adaptation to dynamic risks.

[0041] Step S4: Trigger adaptive scheduling for key areas

[0042] During both polling and non-polling periods, the system continuously performs intelligent analysis of the video stream. When a specific event is detected (such as a person falling, equipment sparks, or abnormal temperature reaching the second threshold) or when the average risk coefficient of a continuous area in the risk situation map exceeds the third threshold (e.g., 0.85) for more than 5 seconds, the system marks that area as a key area. The determination of the key area will be highlighted on the risk situation map and the scheduling module will be activated.

[0043] Step S5: Perform multi-level resource collaborative scheduling

[0044] For the key area marked in step S4, the system initiates a multi-level scheduling process. Level 1 Scheduling: The main camera covering the area is immediately commanded to pause its patrol task, lock its pan / tilt unit (PTZ) to the event center, and perform optical zoom to maximum magnification to capture detailed close-ups. Level 2 Scheduling: While locking the main camera, the system searches for other cameras adjacent to the key area to determine if there are any unobstructed viewpoints. If so, it calculates the optimal angle and shortest path for other cameras to rotate their PTZs to align with the key area and issues commands to provide auxiliary monitoring of the event area from different directions, offering panoramic and close-up multi-angle views. Level 3 Scheduling: If human activity is detected within the key area, the system automatically captures facial features and compares them with a pre-recorded database of authorized personnel; if unauthorized personnel are detected, an audible and visual alarm is issued on the scheduling interface, and the on-site broadcast system is activated to verbally urge them to leave.

[0045] Step S6: Round-robin recovery and historical data archiving

[0046] Once the incident in the key area is deemed resolved (e.g., hotspot temperature drops, personnel are evacuated, and video analysis detects no anomalies for 10 consecutive seconds), the system automatically unlocks the area. The called-in camera first performs a homing action, restoring its pan-tilt angle and focal length to the position recorded before scheduling. Then, the system encapsulates the video clips, alarm records, and scheduling logs from the entire incident handling process into a data packet and writes it to the historical database for post-event traceability and analysis. Finally, the system notifies the polling planning module to re-initiate step S2, generating a new polling path that includes the area.

[0047] Through the above steps, this embodiment realizes a fully automatic, intelligent, and adaptive method for monitoring power plant operations.

[0048] Example 2:

[0049] Based on Example 1, this embodiment further refines the determination and scheduling strategy for key areas, especially by introducing an event priority queue in steps S4 and S5.

[0050] In step S4, in addition to risk coefficient thresholds, the system also maintains an event type-priority mapping table. For example, "equipment fire" is defined as a P0 level event, "personnel fall injury" is P0 level, "high-risk area intrusion" is P1 level, and "equipment temperature warning" is P2 level. When multiple key areas are triggered simultaneously, the system does not immediately perform the first-level scheduling (lock and zoom) in S5 for all areas, but instead creates a scheduling priority queue. Specifically, the system first compares the event priorities of different areas: the area corresponding to the P0 level event will obtain absolute scheduling rights, and its related cameras will be locked immediately. If there are multiple P0 level events, they are then sorted from high to low according to their risk coefficient values ​​and scheduled sequentially. For P1 and P2 level events that are temporarily queued, the system will not completely ignore them, but will arrange a fast polling task: at least one idle adjacent camera is reserved, and these areas are cyclically scanned at a low frame rate (e.g., 2 frames / second) and a short dwell time (e.g., 2 seconds) to ensure that no escalation of the event is missed.

[0051] During the scheduling process, for the highest priority areas that are locked, the system employs a progressively magnified observation strategy. In the first phase, the main camera adjusts to a wide-angle view (approximately 60° horizontal angle) covering the entire event area for one second to acquire global context. In the second phase, based on the event centroid coordinates provided by video analysis algorithms (e.g., the center of the flame's base, the center of the fallen person's torso), the system drives the pan-tilt unit (PTU) for micrometer-level precise positioning and gradually pushes the lens to the telephoto end (approximately 15° horizontal angle) to acquire high-definition details of the target. In the third phase, if the event type is "abnormal personnel posture," the system automatically activates the trajectory prediction function, extrapolating the personnel's position two seconds into the future based on their historical movement location, and preemptively rotates the PTU to track the target, preventing it from moving out of the frame. Throughout all scheduling phases, the camera control commands generated by the multi-level scheduling employ a closed-loop confirmation mechanism: each time an edge device receives a command, it replies with an execution status code (e.g., "success," "angle reached," "rotating"). The main control system then determines whether to issue the next command or retry based on the status code, ensuring the reliable completion of the scheduling.

[0052] Example 3:

[0053] This embodiment mainly optimizes the dynamic round-robin path planning logic in step S2 by introducing predictive planning based on historical data.

[0054] The system establishes a historical risk record database, storing the average risk coefficient of each work area over the past 72 hours across different time slices (e.g., 15-minute time slices). During step S2, the planning module not only uses the real-time risk situation map but also invokes a prediction model. This model employs a simplified time-series forecasting method, estimating the risk coefficient increment within the first 5 minutes of the next cycle based on the risk coefficient change trend at the same relative time point in the previous three cycle periods. For example, if the risk coefficient of a certain equipment area increases from 0.5 to 0.6 in the previous two weeks, and the model predicts an increment of +0.1, then the base risk coefficient for the next cycle is the real-time value of 0.6 plus the predicted increment of 0.1, resulting in a planning risk value of 0.7. The system uses this planning risk value to recalculate the cycle path and dwell time weights.

[0055] In addition, to balance global coverage and key monitoring, a "historical forgetting factor" is introduced into the planning algorithm. The system calculates the "last valid inspection time" for each grid cell. If the "last valid inspection time" of a cell is more than a preset threshold (e.g., 5 minutes) from the current system time, the inspection priority of that cell is temporarily increased by one level. The specific calculation formula (the logic is only described in words here) is as follows:

[0056] The polling priority is calculated as: Current Risk Coefficient × (1 + Historical Forgotten Factor). The historical forgetting factor is directly proportional to the duration of no-inspection; the longer the duration, the larger the factor, with a maximum of 1.0. This ensures frequent polling of high-risk areas while preventing low-risk areas from becoming monitoring blind spots. The planned path is ultimately issued as a JSON-formatted instruction set, with each instruction including a verification field to prevent tampering during transmission.

[0057] Example 4:

[0058] This embodiment enhances the polling recovery mechanism in step S6, implementing a smooth transition strategy. When the system determines that the event in the key area has subsided, it does not immediately return all cameras to their positions and initiate a full replanning; instead, it executes a phased recovery process. In the first phase, the lowest-level auxiliary camera (i.e., the non-main camera called in step S5) is released, allowing it to slowly scan the area surrounding the original event region at a sampling frequency of once every 5 seconds for 30 seconds to observe for any signs of recurrence. In the second phase, if no new events are triggered within 30 seconds, the main camera switches from a close-up locked state to a "focused panning" state, i.e., it performs a small-range reciprocating scan around the event center point at ±20°, maintaining a medium-long focal length for 15 seconds. In the third phase, the main camera returns to its position and performs a rapid full polling. The system only performs a quick image capture and comparison on a representative preset position (not all preset positions) of all cameras in the entire plant. Only after confirming that there are no other abnormalities globally is step S2 activated, generating a complete polling path for the next cycle. This phased recovery strategy effectively avoids frequent interruptions to the rotation plan due to the brief calm of the event, and reduces unnecessary wear and tear on the gimbal motor.

[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent patrol and key area dispatching at power plant operation sites, characterized in that, Includes the following steps: Step 1: Construct a dynamic risk situation by reading real-time personnel distribution, equipment status, and environmental sensor data at the power plant operation site, mapping multi-source data to grid cells of the electronic map, calculating the real-time risk coefficient of each grid cell, and generating a dynamic risk situation map that updates over time. Step 2: Plan an adaptive polling path. Based on the dynamic risk situation map generated in Step 1, select grid cells whose real-time risk coefficient exceeds the first threshold as mandatory polling points. Use a path planning algorithm to generate a polling path that connects all mandatory polling points and passes through other grid cells under the constraints of monitoring resources. Assign a dwell time to each dwell point on the path that is proportional to the real-time risk coefficient of the grid cell in which it is located. Step 3: Distribute and execute the polling task. Decompose the polling path planned in Step 2 into a sequence of instructions that includes the target camera identifier, pan-tilt angle, lens magnification and dwell time. Send the instructions to the corresponding camera or edge gateway through the communication network to drive it to perform image acquisition and bitstream transmission according to the instruction sequence. Step 4: Identify and mark key areas. During or outside of the round-robin execution, perform real-time analysis on the returned video stream. When a preset abnormal event is detected or the average real-time risk coefficient of a continuous area in the dynamic risk situation map exceeds the second threshold and continues for a preset duration, the area is marked as a key area. Step 5: Implement multi-level resource scheduling. For the key areas marked in Step 4, execute Level 1 scheduling: immediately command the main camera covering the area to pause its patrol task, lock the pan-tilt unit to the event center, and perform optical zoom; execute Level 2 scheduling: search for other nearby cameras, calculate the angle and path for the pan-tilt unit to rotate to face the key area, and issue instructions for multi-angle collaborative monitoring; execute Level 3 scheduling: if personnel are detected in the key area, capture facial features and compare them with the authorized database, triggering alarms and linkage devices for unauthorized personnel; Step Six: Restore the patrol status and archive the data. Once the events in the key area are determined to have subsided, unlock the area, restore the camera to the pan-tilt angle and focal length before scheduling, encapsulate the video, alarms and scheduling records during the event handling process into a data packet and store it in the historical database. Then, re-execute Step Two to generate the patrol path for the next cycle.

2. The intelligent patrol and key area dispatching method for power plant operation sites according to claim 1, characterized in that, The calculation of the real-time risk coefficient for each grid cell in step one is as follows: Real-time risk coefficient = (Personnel density weight × Real-time number of personnel per unit grid) + (Equipment anomaly weight × Real-time equipment anomaly index) + (Environmental exceedance weight × Real-time environmental exceedance index), The sum of the weights for personnel density, equipment anomaly, and environmental exceedance is 1. The real-time equipment anomaly index is obtained by normalizing the percentage deviation between the measured values ​​of vibration, temperature, or electrical parameters and the preset threshold.

3. The intelligent patrol and key area dispatching method for power plant operation sites according to claim 2, characterized in that, In step two, when generating the round-robin path using a path planning algorithm, an ant colony algorithm or a genetic algorithm is used, and the sum of minimizing the total round-robin cycle time and maximizing the risk coverage of all dwelling points is taken as the joint optimization objective. The risk coverage is equal to the sum of the real-time risk coefficients of all dwelling points on the path divided by the sum of the real-time risk coefficients of all grid cells in the dynamic risk situation map.

4. The intelligent patrol and key area dispatching method for power plant operation sites according to claim 3, characterized in that, The process of identifying and marking key areas in step four also includes maintaining an event priority queue: the system presets a mapping table between event types and priorities, marking equipment fires or personnel falls as first-priority events, high-risk area intrusions as second-priority events, and equipment temperature warnings as third-priority events; when multiple key areas are marked simultaneously, the highest-priority events in each area are compared, and the key areas corresponding to the first-priority events are given priority to perform all scheduling in step five, while the key areas corresponding to the second-priority and third-priority events are assigned idle cameras to perform fast round-robin, the frame rate of which is lower than the normal round-robin frame rate and the dwell time is shorter than the normal round-robin dwell time.

5. The intelligent patrol and key area dispatching method for power plant operation sites according to claim 4, characterized in that, In step five, when performing first-level scheduling for the locked key area, a progressively magnified observation strategy is adopted: In the first stage, the main camera is controlled to adjust to a wide-angle field of view covering the entire event area and continue for a first preset time to obtain the global context image; In the second stage, based on the event centroid coordinates provided by the video analysis algorithm, the main camera pan-tilt unit is driven to perform precise positioning and gradually move to the telephoto end to obtain event details. In the third stage, when the detected event type is abnormal human posture, the trajectory prediction function is activated. Based on the human's historical movement position, the predicted position within the next preset time period is calculated, and the gimbal is rotated in advance to track the target and prevent it from moving out of the frame.

6. The intelligent patrol and key area dispatching method for power plant operation sites according to claim 5, characterized in that, The multi-level resource scheduling in step five adopts a closed-loop confirmation mechanism: each time the edge device receives a scheduling instruction containing the target camera identifier, pan-tilt angle, and lens magnification, it replies to the central server with an execution status code after execution. The execution status code includes three types: success, angle not reached, or rotating. The central server decides whether to issue the next scheduling instruction or resend the current instruction based on the received execution status code.

7. The intelligent patrol and key area dispatching method for power plant operation sites according to claim 1, characterized in that, After allocating dwell time to dwell points in step two, the system also performs a dwell time correction step: obtain the "last valid inspection time" of the grid cell corresponding to each dwell point, calculate the time elapsed since the current system time, and if the time elapsed exceeds a preset forgetting time threshold, multiply the dwell time of the dwell point by a correction coefficient that is proportional to the uninspected time elapsed. The value of the correction coefficient is between 1.0 and 2.0, and the product result is used as the final dwell time of the dwell point.

8. The intelligent patrol and key area dispatching method for power plant operation sites according to claim 7, characterized in that, Before planning the adaptive round-robin path in step two, a predictive risk correction step is also performed: a prediction model based on the simple displacement method of time series is called, the historical value of the real-time risk coefficient corresponding to the current time point in the previous three round-robin cycles is read, the average value of the risk coefficient change in adjacent cycles is calculated as the prediction increment, and the real-time risk coefficient calculated in step one is summed with the prediction increment to obtain the planning risk coefficient. The selection of the forced round-robin points and the allocation of the dwell time both use the planning risk coefficient.

9. The intelligent patrol and key area dispatching method for power plant operation sites according to claim 1, characterized in that, After the events in the key area are determined to have subsided in step six, a phased recovery strategy is executed: In the first phase, the lowest-level auxiliary camera called in step five is released, and the auxiliary camera is ordered to scan the area around the original key area at a rate lower than the normal sampling frequency for a third preset time. In the second stage, if no new event is triggered within the third preset time, the main camera is controlled to switch from the locked state to the focus panning state, which performs small-range reciprocating scanning around the center point of the event, and continues for the fourth preset time. In the third stage, the main camera is controlled to return to its original position and all cameras are ordered to take a snapshot of a representative preset position for global anomaly comparison. After confirming that there are no anomalies, step two is activated to generate the complete round-robin path for the next cycle.

10. The intelligent patrol and key area dispatching method for power plant operation sites according to claim 9, characterized in that, In step four, when performing real-time analysis of the returned video stream, if the detected preset abnormal event type is a device spark or a temperature anomaly that reaches the third threshold, then in step five, when implementing multi-level resource scheduling, an additional linkage capture command is executed: the system calculates the precise coordinates of the spark or temperature anomaly in the electronic map, calls the non-main camera closest to the coordinate, drives its pan-tilt to rotate toward the coordinate and executes continuous capture mode, sets the capture frame rate to twice the normal frame rate, and the capture duration is 10 seconds. The captured image sequence is automatically stitched into a dynamic image segment, which is then attached to the encapsulated data packet and stored in the historical database.

Images