A Method and System for Fire Source Localization and Extinguishing Control by Unmanned Aerial Vehicles Based on Multi-Source Sensing

Abstract

This application discloses a method and system for fire source location and extinguishing control using unmanned aerial vehicles (UAVs) based on multi-source sensing. The method includes: constructing an integrated ground-air multi-source sensing network and collecting multi-source data based on the terrain geographic information data of a photovoltaic power station; the construction of the integrated ground-air multi-source sensing network includes: deploying multimodal fire source sensing equipment on the ground and deploying inspection UAVs and fire-fighting UAVs in the air; fusing and processing the multi-source data and accurately locating the fire source; executing the scheduling and attitude control parameter configuration of the fire-fighting UAVs; implementing fire-fighting operations and conducting effect evaluation and network optimization to form a closed-loop control. This application can solve the inherent defects of traditional photovoltaic power station fire protection systems, effectively realize fire source location and fire-fighting control, and curb the spread of fire.

Description

Technical Field

[0001] This application relates to the field of fire monitoring and fire suppression technology, specifically to a method and system for fire source location and fire suppression control using unmanned aerial vehicles (UAVs) based on multi-source sensing. Background Technology

[0002] Photovoltaic power plants occupy an important position in the energy supply sector. As their scale continues to expand, they occupy increasingly larger areas and their equipment becomes more densely packed. At the same time, most photovoltaic power plants are located in complex outdoor environments, which exposes them to significant fire risks and makes fire prevention extremely difficult. Once a fire occurs, it not only severely damages the equipment of the photovoltaic power plant and affects power supply, but may also trigger a series of safety problems, threatening the surrounding environment and the safety of people.

[0003] In the past, fire protection systems for photovoltaic power plants primarily employed a passive prevention and control model. One common method relied on manual inspections, with staff periodically patrolling the power plant to identify potential fire hazards. However, manual inspections suffer from low efficiency and limited coverage, making it difficult to monitor the entire power plant in real time. Another common method involved monitoring using single sensors at fixed locations, such as thermal imaging sensors and smoke sensors. These sensors can detect fires to some extent, but they have many limitations. Moreover, due to their limited coverage, they cannot cover blind spots in large photovoltaic arrays, resulting in inaccurate fire source location. Regarding fire suppression control, most existing fire-fighting drones rely on remote manual control for bomb dropping. The operator's experience and communication delays affect the accuracy of bomb dropping, and they cannot adjust flight attitude and dropping strategies in a timely manner according to the dynamic changes of the fire source. Furthermore, data from inspection equipment, detection equipment, and fire suppression equipment are fragmented, lacking a unified multi-source data fusion and scheduling mechanism, making it difficult to form a closed-loop control system.

[0004] Existing fire suppression systems for photovoltaic power plants have significant shortcomings. Insufficient fusion of multi-source sensing data leads to low accuracy in fire source location. The limitations of single sensors necessitate manual verification even after coarse fire source location, undoubtedly delaying timely fire response. During firefighting, drones suffer from poor attitude control stability and insufficient bomb dropping accuracy. Influenced by factors such as operational experience and communication delays, they are unable to effectively extinguish fires. The lack of a closed-loop multi-device coordination system results in low emergency response efficiency, a high rate of missed fire reports, and an inability to effectively contain the initial spread of fires, posing a significant threat to the fire safety of photovoltaic power plants. Summary of the Invention

[0005] To address the inherent shortcomings of traditional photovoltaic power plant fire protection systems, effectively locate and control fire sources, and curb the spread of fire, this application provides a method and system for fire source location and fire suppression based on multi-source sensing using unmanned aerial vehicles (UAVs).

[0006] In a first aspect, this application provides a method for fire source localization and fire suppression control using unmanned aerial vehicles (UAVs) based on multi-source sensing, including: Based on the topographical information data of photovoltaic power stations, an integrated ground-air multi-source sensing network is constructed and multi-source data is collected; the construction of the integrated ground-air multi-source sensing network includes: deploying multimodal fire source sensing equipment on the ground, and deploying inspection drones and fire-fighting drones in the air; Multi-source data is fused and processed for precise fire source localization. The data fusion and precise fire source localization include: extracting fire source features using an improved residual network based on embedded adaptive attention, fusing multi-scale feature maps through a feature pyramid network; identifying fire sources and outputting preliminary fire source localization using an improved anchorless detection algorithm based on an optimized loss function; and correcting fire source location deviations by combining extended Kalman filtering with data from ground sensing devices and data collected by inspection drones to obtain the fire source localization. The system executes the scheduling and attitude control parameter configuration for firefighting drones. The scheduling and attitude control of firefighting drones include: planning the flight path of the firefighting drone based on a pre-built digital twin model of a photovoltaic power station; automatically loading fire extinguishing bombs using the modular mounting mechanism of the hangar on the firefighting drone; and maintaining the attitude stability of the firefighting drone by dynamically adjusting the PID control parameters during flight according to the flight path using a quaternion attitude calculation architecture that integrates multi-sensor data collected by the attitude sensors on the firefighting drone. The system implements firefighting operations, conducts effectiveness evaluations, and optimizes the network to form a closed-loop control system. Real-time firefighting operations and effectiveness evaluation include: based on the fire source location and the flight path of the firefighting drone, initiating bomb dropping operations when the drone reaches the fire source location; after bomb dropping, continuously monitoring temperature and smoke data in the firefighting area using cameras onboard the drone, analyzing and evaluating the firefighting effect, determining the fire source location error, and providing feedback; network optimization in the closed-loop control includes: adjusting and optimizing the layout of ground sensing equipment and the drone's inspection routes based on the feedback on the firefighting effect and the fire source location error.

[0007] By adopting the above scheme, an integrated air-ground multi-source sensing network is constructed to achieve all-round data collection. The accuracy of fire source positioning is improved through multi-source data fusion processing. The digital twin model can plan the flight path and attitude control to ensure the stable operation of fire-fighting drones. Closed-loop control can optimize the sensing network, thereby improving fire-fighting efficiency and reducing fire losses.

[0008] Preferred options also include: Real-time collection of environmental data, electricity consumption data, and regional functional data of photovoltaic power plants; The environmental data, electricity consumption data and regional functional data of the photovoltaic power station collected in real time are input into the pre-built fire risk prediction model to predict the fire risk level of each area of ​​the photovoltaic power station in the future. Based on the predicted fire risk levels of various areas of the photovoltaic power station in the future, ground-based multimodal fire source sensing devices and aerial deployment devices are configured differently, and the constructed integrated ground-air multi-source sensing network is dynamically adjusted. The differentiated configuration of ground-based multimodal fire source sensing devices and aerial deployment devices includes: adapting the type, density, and operating mode of ground-based multimodal fire source sensing devices to the risk level; adapting the aerial deployment drone payload and drone inspection strategy to the risk level; and adapting the linkage mechanism between ground-based and aerial deployment devices to the risk level to execute preset early warning response linkage measures under corresponding risk levels.

[0009] By adopting the above scheme, the future fire risk level is predicted based on real-time collected data on the photovoltaic power station environment, electricity consumption, and regional functions. Differentiated configurations are made for ground-based multimodal fire source sensing equipment and aerial equipment, and the sensing network is dynamically adjusted to improve the sensing capability and response speed for different risk areas. This enhances the system's adaptability and flexibility, and enables more accurate responses to fire risks in different scenarios.

[0010] Preferably, it also includes: based on the predicted fire risk level of each area of ​​the photovoltaic power station in the future time period, performing feature fusion of multi-source data and fire source location in a differentiated manner; Differentiated feature fusion of multi-source data includes: based on features extracted by the improved residual network, the feature fusion weights are dynamically adjusted according to the risk level and the scenario type of each region of the photovoltaic power station; each combination of risk level and scenario type of each region of the photovoltaic power station is pre-adapted with preset feature fusion weights; the scenario type of each region is determined based on the environmental data, electricity consumption data and regional functional data of each region. Differentiated initial fire source location includes: based on the existing fire source location method using an improved anchorless detection algorithm based on an optimized loss function, a new fire source location method based on a pre-built rule-based judgment model and a YOLOv8 model is introduced, and the combination and weight of the fire source location methods are dynamically selected according to the risk level to achieve differentiated initial fire source location.

[0011] By adopting the above scheme, the weights of multi-source data feature fusion and the combination and weights of the initial fire source location methods are dynamically adjusted according to the fire risk level of each area of ​​the photovoltaic power station in the future, thereby improving the accuracy and adaptability of fire source location under different risk scenarios.

[0012] Preferably, it also includes: determining the location data and operation data of the ground-based fire extinguishing auxiliary devices within the photovoltaic power station; Based on the preliminary location data of the fire source and environmental data obtained, the range of fire spread in the future is predicted; According to the preset fire resource matching rules, the required number of firefighting drones and ground firefighting auxiliary devices are matched based on the predicted future fire spread range; based on the required number of firefighting drones and ground firefighting auxiliary devices... Based on the current operational data of firefighting drones and ground firefighting auxiliary devices, dispatch tasks for firefighting drones and ground firefighting auxiliary devices that meet the requirements are generated, and firefighting drones and ground firefighting auxiliary devices are dispatched according to the generated dispatch tasks. Based on the scheduling tasks of firefighting drones and ground firefighting auxiliary devices, during the planning of firefighting drone routes using a digital twin model of a photovoltaic power station, spatial and temporal conflicts are detected between the generated multiple firefighting drone routes. If spatial and temporal conflicts are detected among multiple firefighting drone routes, the flight altitude and flight time of the conflicting firefighting drones are adjusted to avoid conflicts. The system also monitors whether there are overlapping operation areas between bombing operations and ground firefighting auxiliary device operations. If overlapping operation areas exist, the operation time of the ground firefighting auxiliary device is adjusted to stagger the bombing operation time.

[0013] By adopting the above scheme, combining the initial location of the fire source and the prediction of the fire spread range based on environmental data, the required number of fire-fighting drones and ground fire-fighting auxiliary devices are matched according to the spread range and a scheduling task is generated to rationally allocate fire-fighting resources; when planning flight paths, conflicts in the flight paths of fire-fighting drones are detected and adjusted to avoid collisions during flight; when coordinating fire-fighting drones and ground fire-fighting auxiliary devices, the operation time is adjusted to avoid operational conflicts, thereby improving the safety and efficiency of fire-fighting operations.

[0014] Preferred options also include: An attitude prediction model is established based on the dynamic equations; The environmental data for future periods and the current operating status data of the inspection drone are obtained based on the pre-built environmental prediction model and input into the attitude prediction model to obtain the attitude for future periods as the actual attitude. Combined with the target attitude trajectory of the inspection drone, the goal is to minimize the deviation between the actual attitude and the target attitude and minimize the change in motor control quantity within the future prediction period. Under the constraints of attitude angle, angular velocity and torque, the optimal attitude adjustment value for future periods is obtained by solving. The system utilizes PID control to calculate motor control quantities and maintain UAV attitude stability based on the optimal attitude adjustment value for future time periods and the error of the actual attitude angle calculated by fusing multi-sensor data using a quaternion attitude calculation architecture. The dynamic adjustment of PID parameters for attitude control is divided into three stages: cruise, hovering, and bomb preparation. Each stage is configured with an independent set of proportional-integral-derivative parameters, and different stages correspond to future time periods of preset duration.

[0015] By adopting the above scheme, an attitude prediction model is established based on the dynamic equation and combined with environmental data, inspection UAV status data and target attitude trajectory to predict the attitude adjustment value and the error of the actual attitude angle in future periods. The optimal attitude adjustment value is output by combining multi-objective optimization solution, and the motor control quantity is calculated using the optimal attitude adjustment value to more accurately maintain the attitude stability of the UAV in different flight stages.

[0016] Preferred options also include: Before the fire-fighting drone reaches the fire source location and starts the bomb-dropping operation, the distance between the fire-fighting drone and the fire source is measured in real time using a laser rangefinder mounted on the fire-fighting drone. The optimal bomb-dropping angle and time are calculated in combination with the parabolic trajectory model of the fire-fighting bomb. The hovering attitude of the fire-fighting drone is adjusted according to the optimal bomb-dropping angle, and the bomb-dropping command is triggered at the optimal time.

[0017] By adopting the above scheme, the accuracy of fire extinguishing bombs can be improved by using a laser rangefinder to measure the distance in real time and combining it with the parabolic trajectory model of the fire extinguishing bomb to calculate the optimal throwing angle and time.

[0018] Preferred options also include: When the fire extinguishing effect assessment result is determined to be partially extinguished or there is a risk of reignition, a second bomb drop is triggered. When the second bomb drop is triggered, multi-source data acquisition, multi-source data fusion processing and precise fire source location are carried out again, fire extinguishing drone scheduling and attitude control parameter configuration are executed, fire extinguishing operations are carried out and effect assessment is conducted until the fire extinguishing effect assessment result is that there is no risk of reignition.

[0019] By adopting the above scheme, when the fire extinguishing effect assessment results indicate a risk of reignition, a second bombing is triggered, and multi-source data acquisition, fusion processing, and precise fire source location are carried out again, along with the scheduling and attitude control parameter configuration of fire-fighting drones, fire extinguishing operations, and effect assessment, until there is no risk of reignition. This effectively eliminates the hidden danger of reignition of photovoltaic power station fires and curbs the spread of the initial fire.

[0020] Secondly, this application provides a UAV fire source location and extinguishing control system based on multi-source sensing, comprising: The data acquisition module is used to construct an integrated ground-air multi-source sensing network and collect multi-source data based on the terrain and geographic information data of the photovoltaic power station. The construction of the integrated ground-air multi-source sensing network includes: deploying multimodal fire source sensing equipment on the ground and deploying inspection drones and fire-fighting drones in the air. The fire source localization module is used to fuse multi-source data and accurately locate fire sources. The data fusion processing and accurate fire source localization include: extracting fire source features using an improved residual network based on embedded adaptive attention, fusing multi-scale feature maps through a feature pyramid network; identifying fire sources and outputting preliminary fire source localization using an improved anchor-free detection algorithm based on an optimized loss function; and correcting fire source position deviations by using extended Kalman filtering combined with data from ground sensing devices and data collected by inspection drones to obtain the fire source localization. The firefighting dispatch module is used to execute the dispatching and attitude control parameter configuration of firefighting drones. The firefighting drone dispatching and attitude control includes: planning the flight path of the firefighting drone based on a pre-built digital twin model of a photovoltaic power station, and automatically loading fire extinguishing bombs using the modular mounting mechanism of the hangar on the firefighting drone; during flight according to the flight path, a quaternion attitude calculation architecture is used to integrate multi-sensor data collected by the attitude sensors on the firefighting drone, and the attitude stability of the firefighting drone is maintained by dynamically adjusting the PID control parameters. The fire suppression execution module is used to carry out fire suppression operations and perform effect evaluation and network optimization to form a closed-loop control. The real-time fire suppression operation and effect evaluation include: based on the fire source location and the flight path of the fire suppression drone, initiating bomb dropping operations when the drone reaches the fire source location; after bomb dropping, continuously monitoring the temperature and smoke data of the fire suppression area using the camera device on the fire suppression drone, analyzing and evaluating the fire suppression effect, determining the fire source location error, and providing feedback; the network optimization in the closed-loop control includes: adjusting and optimizing the layout of ground sensing equipment and the drone inspection route based on the feedback of the fire suppression effect and the fire source location error.

[0021] By adopting the above scheme, an integrated ground-air multi-source sensing network can be constructed to collect data comprehensively. Multi-source data fusion processing can accurately locate fire sources. Automatic scheduling and attitude stabilization control of firefighting drones can efficiently execute firefighting tasks. Combined with a closed-loop control mechanism of effect evaluation and network optimization, firefighting efficiency can be improved, fire false alarm rate can be reduced, and emergency response capability can be enhanced.

[0022] Thirdly, this application provides a computer-readable storage medium including a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to perform the method described above.

[0023] Fourthly, this application provides a computer device, the computer device including a memory, a processor and a program stored in the memory and executable thereon, the program being executed by the processor to implement the steps of the method described above.

[0024] In summary, this application has the following beneficial effects: This invention overcomes the limitations of limited sensing range and single data dimension of traditional photovoltaic power plant fire protection systems by constructing an integrated ground-air multi-source sensing network. Building upon this, a fusion processing mechanism combining deep learning and filtering algorithms significantly improves the accuracy and reliability of fire source location, resolving the shortcomings of traditional solutions that suffer from low fire source location accuracy due to insufficient data fusion and require secondary manual confirmation, thus delaying response. Furthermore, this invention utilizes a digital twin model for flight path planning, combined with quaternion attitude calculation and adaptive PID control to achieve precise and stable control of UAV attitude, overcoming the problems of existing UAVs relying on remote manual operation and exhibiting poor flight stability, laying a solid foundation for precise bombing. Finally, this invention forms a complete closed-loop control system through post-fire suppression effect evaluation and optimization of the sensing network configuration based on the evaluation results, solving the fundamental problem of fragmented and uncoordinated data between inspection, detection, and fire suppression equipment in existing technologies. In summary, this invention, by systematically integrating perception, decision-making, execution, and evaluation, achieves intelligent closed-loop control throughout the entire process, from precise fire source location to efficient fire suppression and system self-optimization, effectively improving the emergency response efficiency and fire suppression success rate of photovoltaic power plants in the face of initial fires. Specific advantages are as follows: 1. By constructing an integrated ground-air multi-source sensing network and collecting multi-source data, we can obtain more comprehensive and accurate information related to photovoltaic power station fires; by fusing and processing multi-source data and accurately locating fire sources, we can improve the accuracy and reliability of fire source location; by executing the scheduling and attitude control parameter configuration of fire-fighting drones, we can improve the rationality of fire-fighting drone scheduling and the stability of flight attitude; by implementing fire-fighting operations and conducting effect evaluation and network optimization to form a closed-loop control, we can ensure the fire-fighting effect, and dynamically optimize the sensing network based on feedback, thereby improving the efficiency and capability of responding to fires. 2. Based on real-time collected data, predict the fire risk level of each area of ​​the photovoltaic power station in the future, and accordingly configure ground and aerial equipment, dynamically adjust the sensing network, so that the deployment and operation of the sensing equipment are more in line with the actual risk situation, and further improve the accuracy and timeliness of fire source location; 3. Dynamically adjust the feature fusion weights and the methods and weights for preliminary fire source location based on different risk levels and scenario types, to more accurately adapt to the actual conditions of different areas of the photovoltaic power station and enhance the ability to prevent and control fires in different areas. Attached Figure Description

[0025] Figure 1 This is a flowchart of the UAV fire source localization and fire extinguishing control method based on multi-source perception described in a specific embodiment; Figure 2 This is a schematic diagram of the structure of the UAV fire source location and extinguishing control system based on multi-source perception described in a specific embodiment. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] like Figure 1 As shown in the figure, this application discloses a method for fire source location and fire suppression control of unmanned aerial vehicles (UAVs) based on multi-source perception. The method includes: constructing a perception network to collect data, fusing the data to locate the fire source, scheduling the UAV and controlling its attitude, extinguishing the fire, and evaluating and optimizing the results. These steps aim to improve the accuracy of fire source location, the precision of fire suppression, and the efficiency of emergency response. Each step will be described in detail below.

[0028] S1. Based on the terrain and geographic information data of photovoltaic power stations, construct an integrated ground-air multi-source sensing network and collect multi-source data.

[0029] This embodiment targets photovoltaic power plants, focusing on locating and extinguishing fires that may occur at these plants. Specifically, a situational awareness map is constructed based on the terrain and geographic information data of the photovoltaic power plant. The optimal deployment locations of sensing devices are determined through simulation, completing the construction of an integrated ground-air multi-source sensing network. This integrated ground-air multi-source sensing network includes: deploying multimodal fire source sensing devices on the ground and deploying inspection drones and firefighting drones in the air.

[0030] The ground-based multimodal fire source detection equipment includes various types of fire source sensors, such as temperature sensors and smoke sensors, as well as visible light or infrared cameras. These basic sensors and cameras are evenly distributed throughout the photovoltaic power station. In addition, thermal imaging observation-type photoelectric turntables and binocular intelligent monitoring gimbals are deployed. The photoelectric turntables are deployed at the high-altitude edges of the array to achieve large-scale initial fire detection, while the binocular cameras are deployed in key equipment clusters such as transformer substations and inverters to achieve precise mesoscale monitoring. Both types of equipment integrate dual-path detection modules for visible light and infrared, have built-in intelligent fire point and smoke recognition algorithms, and integrate satellite positioning modules for synchronized location information. Aerial deployment includes inspection drones and firefighting drones. The inspection drones are equipped with dual-light cameras and depth cameras, while the firefighting drones are modified firefighting models equipped with infrared thermal imagers, laser rangefinders, and attitude sensors, possessing heavy-load capabilities.

[0031] After the deployment of the integrated ground-air multi-source sensing network, the data acquisition and preprocessing process is initiated: ground sensing devices collect video streams, temperature data, and positioning information in real time and transmit them to the data platform; inspection drones cruise along preset routes, collecting images, fire source depth data, and their own trajectory data, and transmit them back to the data platform in real time. The data platform preprocesses the received multi-source data, performing noise reduction and image enhancement on the video streams, and outlier removal and timestamp alignment on the sensor data. This ensures data consistency and availability, laying the foundation for subsequent fusion processing. In addition, the data platform can also perform early warning analysis on the data collected by ground sensing devices and aerial devices, and synchronize this analysis to the corresponding devices, controlling them to execute preset early warning response measures to achieve ground-air collaboration.

[0032] S2. Perform fusion processing on the preprocessed multi-source data and complete the precise location of the fire source.

[0033] Specifically, the preprocessed multi-source data is fused to extract feature data in advance. To ensure the accuracy of fire source location, an improved residual network backbone feature extraction network is used to extract features from image data of ground equipment and inspection drones. The improved residual module is constructed with a multi-branch structure and incorporates an adaptive attention module to enhance the extraction of key features such as flame texture and temperature gradient.

[0034] For the extracted feature data, a feature pyramid network is used to fuse multi-scale feature maps to obtain multi-scale fused features. In addition, an efficient global context network is introduced to model the remote dependency relationship between the fire source and the background, reducing interference from fire-like targets such as strong light and high-temperature equipment.

[0035] Based on multi-scale fusion features, an improved anchorless frame detection algorithm is used for fire source identification and outputs preliminary fire source location. The improved anchorless frame detection algorithm adjusts the bounding box regression accuracy through an optimized loss function, such as introducing a dynamic IoU (Intersection over Union) weighted loss function to optimize the bounding box regression accuracy.

[0036] By combining the data collected by the inspection drone (depth data) with the data from the ground sensing device (satellite positioning data of the ground device), the extended Kalman filter algorithm is used to verify the location of the fire source, correct the positioning deviation caused by terrain undulation, and finally output the accurate three-dimensional coordinates of the fire source and fire data (such as: burning range, temperature distribution characteristics), and synchronize them to the visualization platform to trigger alarms, providing core basis for subsequent fire fighting and dispatch.

[0037] The fire source location verification using the Extended Kalman Filter (EKF) algorithm specifically includes: establishing an EKF system model, including selecting the three-dimensional coordinates of the fire source and incorporating the attitude error of the UAV; defining a state vector; constructing a state equation, with a constant-velocity model for state transition, where the current state equals the previous state plus process noise; constructing an observation equation, where the observation vector includes the fire source coordinates retrieved from UAV depth data and the fire source coordinates calibrated from ground satellite positioning reference points; and iteratively merging and refining the state estimate through two steps: prediction (predicting the current state based on the previous state estimate and state equation) and updating (calculating the Jacobian matrix, calculating the Kalman gain, and correcting the predicted state based on observation data to obtain the optimal state estimate for the current moment), fusing multi-source data, refining the state estimate, and iteratively converging to output the location for verification. Correcting the location deviation caused by terrain undulations involves: using preprocessed UAV depth data, generating a DEM of the area surrounding the fire source using an interpolation algorithm to obtain the actual elevation of each location on the ground; calculating the terrain deviation by comparing the measured elevation of the ground satellite positioning reference point with the terrain elevation of the corresponding point in the DEM; and using the obtained elevation deviation to correct the output elevation of the fire source. Based on the corrected three-dimensional coordinates of the fire source, combined with images and temperature data collected by the inspection drone, the combustion range and temperature distribution characteristics are calculated.

[0038] S3. Configure the scheduling and attitude control parameters for firefighting drones.

[0039] Specifically, after receiving the fire source location output from the above steps, the visualization platform initiates the scheduling of firefighting drones, including: based on the pre-built digital twin model of the photovoltaic power station, combined with the fire source location, the distribution of firefighting drone equipment and obstacle information in the photovoltaic power station, using route planning algorithms to plan the optimal firefighting drone firefighting route, and simultaneously determining the firefighting drone takeoff point, flight altitude and bombing preparation point.

[0040] After the dispatch command is issued, the firefighting drone flies along the flight path and automatically loads the fire extinguishing bombs through the modular hangar mounting mechanism (or other mechanisms with automatic loading functions). The modular hangar mounting mechanism refers to a modular mounting mechanism adopted in the hangar design, which allows different equipment or components to be quickly installed and disassembled in a modular manner, thereby automatically completing the loading of the fire extinguishing bombs.

[0041] To ensure the stability of the firefighting drone during flight along its designated route, a quaternion attitude calculation architecture is employed, fusing multi-sensor data collected by the drone's onboard attitude sensors. This data is then dynamically adjusted using PID control parameters to maintain the drone's attitude stability. Specifically, the quaternion attitude calculation architecture calculates the current attitude angle using real-time attitude data collected by gyroscopes, accelerometers, and magnetometers, and employs an extended Kalman filter algorithm. The calculated result is compared with the target attitude to generate an attitude deviation signal, which is then input to an adaptive PID controller. The adaptive PID controller dynamically adjusts the control parameters according to the flight phase (cruise, hovering, and bomb preparation phase), controlling the motor speed via PWM signals to maintain the drone's attitude stability.

[0042] In addition, during flight, the active phased array radar and binocular visual obstacle avoidance data mounted on the firefighting drone are combined to correct the flight trajectory in real time, ensuring that the drone arrives at the bombing preparation point stably. During this time, the flight status data is transmitted back to the data center for dynamic monitoring.

[0043] S4. Implement firefighting operations and conduct effectiveness evaluations and network optimization to form a closed-loop control.

[0044] Specifically, based on the location of the fire source and the flight path of the fire-fighting drone, the bombing operation is initiated when the fire-fighting drone reaches the fire source location. The location where the drone reaches the fire source location can be selected based on the fire source's combustion range and temperature distribution characteristics (such as the fire source location center or a random location) as the bombing preparation point. Once the fire-fighting drone reaches the bombing preparation point, the bombing operation is initiated.

[0045] To ensure the fire extinguishing bombs accurately target the fire source, before initiating the bombing operation, a laser rangefinder mounted on the fire extinguishing drone measures the distance between the drone and the fire source in real time. The optimal bombing angle and timing are then calculated using a parabolic trajectory model of the fire extinguishing bomb. The drone's hovering attitude is adjusted according to the optimal bombing angle, and the bombing command is triggered at the optimal moment. The specific steps for calculating the optimal bombing angle and timing include: Initialization: After the drone reaches the fire source location, it hovers. Loop Measurement: The laser rangefinder measures the slant distance (R), obtaining the drone's altitude (H) and horizontal distance (L). Based on the parabolic trajectory model of the fire extinguishing bomb, parameters are calculated. B=-L, C=AH; Calculate the discriminant: Check the feasibility, if There is no solution if the drone's position needs to be adjusted. Then continue calculating the value of u. Obtain the corresponding high and low trajectories, and calculate the optimal bombing angle using the low trajectory. In calculation And when the situation stabilizes (e.g., the deviation of several consecutive calculated values ​​is less than the threshold), immediately adjust the launching device to the desired angle. The system can then trigger bomb release. Alternatively, the bomb release trajectory can be pre-simulated in a digital twin model (e.g., the release point needs to be 1.5 meters in advance when the wind speed is 3 m / s), outputting the optimal release angle and timing. Furthermore, when the mechanism carrying the fire extinguishing bomb includes ropes, the swing state of the fire extinguishing drone's payload can be monitored in real time during the release process. The swing can be eliminated by adjusting the tension or length of the slings, ensuring the fire extinguishing bomb accurately targets the fire source area.

[0046] After the bombs are dropped, the camera devices (including infrared thermal imaging devices) on the fire-fighting drone continuously monitor the temperature and smoke data of the fire-fighting area, analyze and evaluate the fire-fighting effect, determine the fire source location error and provide feedback; the fire-fighting area can be determined based on the range of the fire-fighting bombs captured by the camera; the fire-fighting effect can be evaluated according to preset fire-fighting effect evaluation rules or a fire-fighting effect evaluation model can be built using a deep learning network, and the fire-fighting effect evaluation model can be trained by using the flame and smoke data of the fire-fighting area labeled with the fire-fighting effect, and the fire-fighting effect evaluation result can be output. The fire extinguishing effect assessment results include: completely extinguished, partially extinguished, and with a risk of reignition. If the fire extinguishing effect assessment result is completely extinguished, it indicates successful fire extinguishing, and the fire extinguishing drone is dispatched to return. If the fire extinguishing effect assessment result is partially extinguished or with a risk of reignition, a second bombing is triggered. When a second bombing is triggered, steps S1-S4 are repeated, namely, multi-source data acquisition, multi-source data fusion processing and precise fire source location, fire extinguishing drone dispatch and attitude control parameter configuration, fire extinguishing operation, and effect assessment, until the fire extinguishing effect assessment result indicates no risk of reignition. The fire source location error is the difference between the fire source location data and the actual fire source location (which can be determined by manual upload or fire extinguishing drone photography). Finally, the layout of ground sensing equipment and the drone patrol routes are adjusted and optimized based on the feedback on fire extinguishing effectiveness and fire source location errors. Specifically, if the fire extinguishing effectiveness assessment result is complete extinguishment or the fire source location error is small compared to the preset error, it indicates that the current layout of ground sensing equipment and drone patrol routes are reasonable. If the fire extinguishing effectiveness assessment result is partial extinguishment or there is a risk of reignition, or the fire source location error is not less than the preset error, it indicates a risk of inaccurate fire source location, requiring further optimization of the ground sensing equipment layout and drone patrol routes. This can be achieved by acquiring fire source location error data, calculating the location error coverage of the current layout, and adjusting the ground sensing equipment layout accordingly to cover all location errors. Alternatively, multi-objective optimization can be used, based on historical fire distribution and historical error location data, to set a multi-objective optimization framework that maximizes the fire source detection probability and minimizes the location error, thus solving for the optimal route.

[0047] In a specific embodiment, to address fire risks in different areas in a targeted manner, further improve the fire suppression response capability and prevention effect of photovoltaic power plants, and reduce the possibility of fire and losses, the method further includes: Real-time collection of environmental data, electricity consumption data, and regional functional data of photovoltaic power plants; among which, environmental data includes weather factors such as temperature, humidity, and wind speed, electricity consumption data includes peak and off-peak electricity consumption periods, and regional functional data includes functional data of various areas such as battery array area, inverter area, power distribution room area, and cable trench area.

[0048] The real-time collected environmental data, electricity consumption data, and regional functional data of the photovoltaic power station are input into a pre-constructed fire risk prediction model to predict the fire risk level of each area of ​​the photovoltaic power station in the future, such as low risk, medium risk, and high risk. The photovoltaic power station can be divided into multiple levels of areas according to function and location. The fire risk prediction model adopts a neural network model, which is generated by collecting historical environmental data, electricity consumption data, and regional functional data of each area of ​​the photovoltaic power station marked with fire risk level.

[0049] To better align the sensing network with actual fire risk conditions and improve the accuracy of fire source location and the efficiency of multi-device collaborative emergency response, ground-based multimodal fire source sensing devices and aerial deployment devices are configured differently based on the predicted fire risk levels of various areas of the photovoltaic power station in the future, and the constructed integrated ground-air multi-source sensing network is dynamically adjusted. The differentiated configuration of ground-based multimodal fire source sensing devices and aerial deployment devices includes: adapting the type, density, and operating mode of ground-based multimodal fire source sensing devices to the risk level; adapting the payload and inspection strategy of aerial deployment drones to the risk level; and adapting the linkage mechanism between ground-based and aerial deployment devices to the risk level to execute preset early warning and response linkage measures under corresponding risk levels.

[0050] Specifically, the types, density, and operating modes of ground-deployed multimodal fire source sensing devices are adapted according to risk level, including: For low-risk levels, the types of ground-deployed multimodal fire source sensing devices include basic types such as temperature sensors and smoke sensors; the density is set at 30m / device, and 50m / device in non-critical areas (such as edge arrays); the operating mode is to use low-frequency data acquisition (30s / time), with some sensors in sleep mode during non-working periods (such as nighttime); For medium-risk levels, the types of ground-deployed multimodal fire source sensing devices include enhanced types such as temperature and humidity composite sensors and infrared thermal imaging devices. The density is set at an interval of 15-20m / unit, and 10m / unit for non-critical areas (such as edge arrays); the operation mode uses low-frequency acquisition (10s / time), and short-term high-frequency acquisition (1s / time) is triggered by abnormal parameters (such as a sudden temperature rise of 5℃); for high-risk level corresponding adaptation, the types of multimodal fire source sensing equipment deployed on the ground include basic type sensors, infrared thermal imaging, thermal imaging observation type photoelectric turntable and binocular intelligent monitoring gimbal, etc.; the density is set at an interval of 5-10m / unit, and 3m / unit for non-critical areas (such as edge arrays); the operation mode uses low-frequency acquisition (1-5s / time), and is active at all times.

[0051] Specifically, the aerial deployment of UAV payloads and UAV inspection strategies are adapted according to risk levels as follows: For low-risk levels, the aerial deployment of inspection UAVs uses lightweight payloads, equipped only with dual-light cameras, and employs a fixed route for one full-coverage inspection per day; For medium-risk levels, the aerial deployment of inspection UAVs uses standard payloads, equipped with dual-light cameras and depth cameras, and employs a dynamic route (covering medium-risk areas) for one inspection per day, with three inspections per day; For high-risk levels, the aerial deployment of inspection UAVs uses professional payloads, equipped only with dual-light cameras, depth cameras, and multispectral cameras, and employs a regional cyclical inspection strategy.

[0052] Specifically, the linkage mechanism between ground-deployed and air-deployed equipment is adapted according to the risk level to implement preset early warning response linkage measures under the corresponding risk level. These measures include: For low-risk levels, if the ground sensor detects multiple abnormal warnings consecutively, the inspection drone is allowed to delay its first-period response flight for inspection; for medium-risk levels, if the ground sensor detects multiple abnormal warnings consecutively, the inspection drone is allowed to delay its second-period response flight for inspection (less than the first period); for high-risk levels, when the ground sensor alarms, coordinates are immediately sent to the inspection drone, which immediately responds, interrupting its current task to prioritize the inspection.

[0053] To improve the accuracy and adaptability of preliminary fire source location, thereby enhancing the accuracy of fire source location and the ability to respond to different scenarios, feature fusion of multi-source data and fire source location are performed in a differentiated manner based on the predicted fire risk level of each area of ​​the photovoltaic power station in the future time period.

[0054] Specifically, feature fusion of multi-source data is differentiated, including: dynamically adjusting feature fusion weights based on features extracted using an improved residual network, according to risk level and the scenario type of each area of ​​the photovoltaic power station; each combination of risk level and scenario type of each area of ​​the photovoltaic power station is pre-adapted with preset feature fusion weights; the scenario type of each area is determined based on environmental data, electricity consumption data, and regional functional data, including: scenario types for each area, basic / harsh environment scenario types, and peak and off-peak electricity consumption scenarios. For example, the scenario type of the battery array area and the high-risk combination are pre-adapted with a preset feature fusion weight of 0.3:0.7 for the fusion of other features and temperature features collected by ground sensing equipment; the scenario type of peak electricity consumption and the high-risk combination are pre-adapted with a preset feature fusion weight of 0.3:0.7 for the fusion of feature data collected by ground sensing equipment and image features collected by inspection drones.

[0055] Specifically, differentiated initial fire source localization is implemented. This includes: in addition to the existing initial fire source localization method using an improved anchorless detection algorithm based on an optimized loss function, a new method is introduced using a pre-built rule-based judgment model (with pre-defined rules such as: locations with temperature > 200℃ + CO > 50ppm are fire source locations) and a YOLOv8 model. The combination and weights of these initial fire source localization methods are dynamically selected according to the risk level to achieve differentiated initial fire source localization. Specifically, for high-risk levels, a combination of the improved anchorless detection algorithm based on an optimized loss function, the pre-built rule-based judgment model, and the YOLOv8 model is used, with a weight ratio of 0.5:0.2:0.3; for medium-risk levels, a combination of the improved anchorless detection algorithm based on an optimized loss function and the pre-built rule-based judgment model is used, with a weight ratio of 0.6:0.4; and for low-risk levels, only the improved anchorless detection algorithm based on an optimized loss function is used.

[0056] In addition, based on the predicted fire risk levels of various areas of the photovoltaic power station in the future, different methods are used for fire source location verification, including: for high-risk areas, the unscented Kalman filter algorithm is used for fire source location verification; for medium-risk areas, the extended Kalman filter algorithm is retained for fire source location verification; and for low-risk areas, the simplified Kalman filter algorithm is used for fire source location verification.

[0057] In a specific embodiment, to further improve the rationality of firefighting resource allocation and the safety and efficiency of firefighting operations, the method further includes: Considering that many photovoltaic power plants have ground-based fire suppression auxiliary devices, such as fire sprinklers, determine the location and operation data of the ground-based fire suppression auxiliary devices in the photovoltaic power plant (e.g., whether they are in operation, their operating coverage area, etc.).

[0058] Based on the initial fire source location data and environmental data, the future fire spread range can be predicted. An LSTM neural network model can be used, with historical fire source location data and fire data (including the initial fire range and the fire spread range) and environmental data as training data. The trained LSTM model can then predict the future fire spread range based on the input initial fire source location data and environmental data.

[0059] According to preset fire resource matching rules, the required number of firefighting drones and ground firefighting auxiliary devices are matched based on the predicted future fire spread range. The preset fire resource matching rules include: for fires predicted to spread within the first spread range, which are considered small-scale fires, the required number of firefighting drones and ground firefighting auxiliary devices is 1 main firefighting drone, 1 auxiliary firefighting drone, and 0 ground firefighting auxiliary devices; for fires predicted to spread within the second spread range, which are considered medium-scale fires, 2-3 firefighting drones and 2-3 ground firefighting auxiliary devices are required; for fires predicted to spread within the third spread range, which are considered large-scale fires, 5-10 firefighting drones and 10 ground firefighting auxiliary devices are required.

[0060] Based on the required number of firefighting drones and ground firefighting auxiliary devices, the current operating data of the firefighting drones and the operating data of the ground firefighting auxiliary devices, a scheduling task for firefighting drones and ground firefighting auxiliary devices that meets the requirements is generated, and the firefighting drones and ground firefighting auxiliary devices are scheduled according to the generated scheduling task.

[0061] Based on the scheduling tasks of firefighting drones and ground-based firefighting auxiliary devices, during the planning of drone flight paths using a digital twin model of a photovoltaic power station, spatial and temporal conflicts are detected among the generated drone flight paths. If spatial and temporal conflicts are detected, the flight altitude and flight time of the conflicting drones are adjusted to avoid the conflict. For example, the system automatically detects trajectory intersections (such as two drones potentially colliding above the inverter area) and adjusts the flight altitude difference (e.g., a 5-meter interval) or flight sequence (e.g., drone A flies to the bombing point first, and drone B takes off 10 seconds later). Furthermore, to avoid conflicts arising from the combined operation of bombing and ground-based firefighting auxiliary devices, the system monitors whether there are overlapping operational areas between these two operations. If overlapping areas exist, the operation time of the ground-based firefighting auxiliary devices is adjusted to stagger the bombing operation time, prioritizing bombing and using ground-based firefighting auxiliary devices as a supplement.

[0062] In a specific embodiment, considering that traditional PID controllers cannot cope with sudden environmental disturbances such as sudden gusts of wind, model predictive control (MPC) needs to be added. This MPC is adjusted in advance using multi-sensor data to ensure stable UAV attitude and accurate projectile trajectory during bomb release. The method also includes: An attitude prediction model is established based on the dynamic equations; taking the roll channel as an example, the pitch / yaw channels are similarly constructed, and the dynamic equation formula is as follows:

[0063] In the formula, , which represents the change in angular velocity between the future moment and the present moment; The motor control torque for the rolling channel; This is the air drag torque; The wind disturbance moment is calculated from future wind speed and direction.

[0064] Using the next N steps (t=1 to t=N) as the prediction window, the system inputs environmental data for the future time period obtained from a pre-built environmental prediction model, the current operational status data of the inspection (firefighting) drone, and the target attitude trajectory of the inspection (firefighting) drone. These inputs are then substituted into the attitude prediction model to predict the attitude for the future time period. The pre-built environmental prediction model employs a neural network model, and the output predicted environmental data for the future time period includes wind speed, wind direction, and air density for the next N steps. The current operational status data of the inspection drone includes its current attitude angle, angular velocity, motor torque, and remaining battery power. The target attitude trajectory of the inspection (firefighting) drone includes its target attitude angle and angular velocity for the next N steps.

[0065] To prevent the predicted attitude adjustment from being too large and exceeding constraints such as the maximum motor torque and the maximum range of attitude angles, an optimization objective is set. Within the next N prediction steps, the objective is to minimize the deviation between the actual attitude (future time period attitude) and the target attitude, and to minimize the change in the control quantity (torque). Under the constraints, the optimal attitude adjustment for the future time period is obtained by solving the problem.

[0066] The system utilizes PID control to calculate motor control quantities and maintain UAV attitude stability based on the optimal attitude adjustment value for future time periods and the error of the actual attitude angle calculated by fusing multi-sensor data using a quaternion attitude calculation architecture. The dynamic adjustment of PID parameters for attitude control is divided into three stages: cruise, hovering, and bomb preparation. Each stage is configured with an independent set of proportional-integral-derivative parameters, and different stages correspond to future time periods of preset duration.

[0067] like Figure 2 As shown, this application provides a UAV fire source location and extinguishing control system based on multi-source sensing, including: Data acquisition module 101 is used to construct an integrated ground-air multi-source sensing network and collect multi-source data based on the terrain geographic information data of photovoltaic power stations; the construction of the integrated ground-air multi-source sensing network includes: deploying multimodal fire source sensing equipment on the ground, and deploying inspection drones and fire-fighting drones in the air; The fire source localization module 102 is used to perform fusion processing of multi-source data and precise fire source localization. The data fusion processing and precise fire source localization include: extracting fire source features using an improved residual network based on embedded adaptive attention, fusing multi-scale feature maps through a feature pyramid network; identifying fire sources and outputting preliminary fire source localization using an improved anchor-free detection algorithm based on an optimized loss function; and correcting fire source position deviations by using extended Kalman filtering combined with data from ground sensing devices and data collected by inspection drones to obtain the fire source localization. The fire suppression scheduling module 103 is used to execute the scheduling and attitude control parameter configuration of fire suppression drones. The fire suppression drone scheduling and attitude control include: planning the flight path of the fire suppression drone based on the pre-built digital twin model of the photovoltaic power station, and automatically loading fire extinguishing bombs using the modular mounting mechanism of the hangar on the fire suppression drone; during the flight according to the flight path, a quaternion attitude calculation architecture is used to integrate multi-sensor data collected by the attitude sensors on the fire suppression drone, and the attitude stability of the fire suppression drone is maintained by dynamically adjusting the PID control parameters. The fire extinguishing execution module 104 is used to implement fire extinguishing operations and perform effect evaluation and network optimization to form a closed-loop control. The real-time fire extinguishing operation and effect evaluation include: based on the fire source location and the flight path location of the fire extinguishing drone, initiating bomb dropping operations when the drone reaches the fire source location; after bomb dropping, using the camera device on the fire extinguishing drone to continuously monitor the temperature and smoke data of the fire extinguishing area, analyze and evaluate the fire extinguishing effect, determine the fire source location error and provide feedback; the network optimization in the closed-loop control includes: adjusting and optimizing the layout of ground sensing equipment and the drone inspection route based on the feedback fire extinguishing effect and fire source location error.

[0068] In one specific embodiment, the data acquisition module 101 in the system is also used to collect environmental data, electricity consumption data, and regional function data of the photovoltaic power station in real time; input the real-time collected environmental data, electricity consumption data, and regional function data of the photovoltaic power station into a pre-constructed fire risk prediction model to predict the fire risk level of each area of ​​the photovoltaic power station in the future; and based on the predicted fire risk level of each area of ​​the photovoltaic power station in the future, differentiate the configuration of ground multimodal fire source sensing equipment and air-deployed equipment to dynamically adjust the constructed ground-air integrated multi-source sensing network.

[0069] In one specific embodiment, the fire source location module 102 in the system is also used to perform feature fusion of multi-source data and fire source location in a differentiated manner based on the predicted fire risk level of each area of ​​the photovoltaic power station in the future period; and to perform feature fusion of multi-source data in a differentiated manner.

[0070] In a specific embodiment, the fire suppression dispatch module 103 is further used to determine the location and operation data of the ground fire suppression auxiliary devices within the photovoltaic power station in the system; predict the future fire spread range based on the acquired preliminary fire source location data and environmental data; match the required number of fire suppression drones and ground fire suppression auxiliary devices according to preset fire resource matching rules and the predicted future fire spread range; generate a dispatch task for fire suppression drones and ground fire suppression auxiliary devices that meets the requirements based on the required number of fire suppression drones and ground fire suppression auxiliary devices, the current operation data of the fire suppression drones, and the operation data of the ground fire suppression auxiliary devices; and dispatch the fire suppression drones and ground fire suppression auxiliary devices according to the generated dispatch task.

[0071] In a specific embodiment, the fire suppression dispatch module 103 is also used to establish an attitude prediction model based on dynamic equations; it inputs environmental data for future periods and current operational status data of the inspection drone into the attitude prediction model based on a pre-built environmental prediction model to obtain the attitude for future periods as the actual attitude; combined with the target attitude trajectory of the inspection drone, it sets the goal of minimizing the deviation between the actual attitude and the target attitude and minimizing the change in motor control quantity within the future prediction period, and solves for the optimal attitude adjustment value for future periods under the constraints of attitude angle, angular velocity, and torque; using PID to calculate the motor control quantity and control the drone to maintain attitude stability based on the optimal attitude adjustment value for future periods and the error of the actual attitude angle calculated by fusing multi-sensor data using a quaternion attitude calculation architecture.

[0072] In one specific embodiment, the fire extinguishing execution module 104 in the system is also used to measure the distance between the fire extinguishing drone and the fire source in real time using a laser rangefinder mounted on the fire extinguishing drone before the fire extinguishing drone reaches the fire source location and starts the bomb dropping operation, calculate the optimal bomb dropping angle and time in combination with the parabolic trajectory model of the fire extinguishing bomb, adjust the hovering attitude of the fire extinguishing drone according to the optimal bomb dropping angle, and trigger the bomb dropping command at the optimal time.

[0073] In a specific embodiment, the fire extinguishing execution module 104 in the system is also used to trigger a second bombing when the fire extinguishing effect assessment result is that the fire is partially extinguished or there is a risk of reignition. When the second bombing is triggered, multi-source data acquisition, multi-source data fusion processing and precise fire source location are carried out again, fire extinguishing drone scheduling and attitude control parameter configuration are executed, fire extinguishing operations are carried out and effect assessment is conducted until the fire extinguishing effect assessment result is that there is no risk of reignition.

[0074] This application also discloses a computer-readable storage medium.

[0075] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed, such as the above-described method for locating and extinguishing fire sources from unmanned aerial vehicles based on multi-source perception. The computer-readable storage medium includes, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0076] This application also discloses a computer device.

[0077] Specifically, the computer device includes a memory and a processor. The memory stores a computer program that can be loaded by the processor and executed to implement the aforementioned method for locating and extinguishing fire sources using unmanned aerial vehicles (UAVs) based on multi-source perception.

[0078] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for fire source positioning and fire extinguishing control of unmanned aerial vehicle based on multi-source perception, characterized in that, include: Based on the topographical geographic information data of photovoltaic power stations, an integrated ground-air multi-source sensing network is constructed and multi-source data is collected. The construction of the integrated ground-air multi-source sensing network includes: deploying multimodal fire source sensing equipment on the ground, and deploying inspection drones and firefighting drones in the air; Multi-source data is fused and processed for precise fire source localization. The data fusion and precise fire source localization include: extracting fire source features using an improved residual network based on embedded adaptive attention, fusing multi-scale feature maps through a feature pyramid network; identifying fire sources and outputting preliminary fire source localization using an improved anchorless detection algorithm based on an optimized loss function; and correcting fire source location deviations by combining extended Kalman filtering with data from ground sensing devices and data collected by inspection drones to obtain the fire source localization. The system executes the scheduling and attitude control parameter configuration for firefighting drones. The scheduling and attitude control of firefighting drones include: planning the flight path of the firefighting drone based on a pre-built digital twin model of a photovoltaic power station; automatically loading fire extinguishing bombs using the modular mounting mechanism of the hangar on the firefighting drone; and maintaining the attitude stability of the firefighting drone by dynamically adjusting the PID control parameters during flight according to the flight path using a quaternion attitude calculation architecture that integrates multi-sensor data collected by the attitude sensors on the firefighting drone. The system implements firefighting operations, conducts effectiveness evaluations, and optimizes the network to form a closed-loop control system. Real-time firefighting operations and effectiveness evaluation include: based on the fire source location and the flight path of the firefighting drone, initiating bomb dropping operations when the drone reaches the fire source location; after bomb dropping, continuously monitoring temperature and smoke data in the firefighting area using cameras onboard the drone, analyzing and evaluating the firefighting effect, determining the fire source location error, and providing feedback; network optimization in the closed-loop control includes: adjusting and optimizing the layout of ground sensing equipment and the drone's inspection routes based on the feedback on the firefighting effect and the fire source location error. 2.The multi-source perception based unmanned aerial vehicle fire source positioning and extinguishing control method according to claim 1, characterized in that, Also includes: Real-time collection of environmental data, electricity consumption data, and regional functional data of photovoltaic power plants; The environmental data, electricity consumption data and regional functional data of the photovoltaic power station collected in real time are input into the pre-built fire risk prediction model to predict the fire risk level of each area of ​​the photovoltaic power station in the future. Based on the predicted fire risk levels of various areas of the photovoltaic power station in the future, ground-based multimodal fire source sensing devices and aerial deployment devices are configured differently, and the constructed integrated ground-air multi-source sensing network is dynamically adjusted. The differentiated configuration of ground-based multimodal fire source sensing devices and aerial deployment devices includes: adapting the type, density, and operating mode of ground-based multimodal fire source sensing devices to the risk level; adapting the aerial deployment drone payload and drone inspection strategy to the risk level; and adapting the linkage mechanism between ground-based and aerial deployment devices to the risk level to execute preset early warning response linkage measures under corresponding risk levels. 3.The multi-source perception based UAV fire source positioning and extinguishing control method according to claim 2, characterized in that, Also includes: Based on the predicted fire risk levels of various areas of the photovoltaic power station in the future, feature fusion of multi-source data and fire source location are carried out in a differentiated manner. Differentiated feature fusion of multi-source data includes: based on features extracted by the improved residual network, the feature fusion weights are dynamically adjusted according to the risk level and the scenario type of each region of the photovoltaic power station; each combination of risk level and scenario type of each region of the photovoltaic power station is pre-adapted with preset feature fusion weights; the scenario type of each region is determined based on the environmental data, electricity consumption data and regional functional data of each region. Differentiated initial fire source location includes: based on the existing fire source location method using an improved anchorless detection algorithm based on an optimized loss function, a new fire source location method based on a pre-built rule-based judgment model and a YOLOv8 model is introduced, and the combination and weight of the fire source location methods are dynamically selected according to the risk level to achieve differentiated initial fire source location. 4.The multi-source perception based unmanned aerial vehicle fire source positioning and extinguishing control method according to claim 1, characterized in that, Also includes: Determine the location and operational data of the ground-based fire suppression auxiliary devices within the photovoltaic power station; Based on the preliminary location data of the fire source and environmental data obtained, the range of fire spread in the future is predicted; According to preset fire resource matching rules, the required number of firefighting drones and ground firefighting auxiliary devices are matched based on the predicted future fire spread range; based on the required number of firefighting drones and ground firefighting auxiliary devices... Based on the current operational data of firefighting drones and ground firefighting auxiliary devices, dispatch tasks for firefighting drones and ground firefighting auxiliary devices that meet the requirements are generated, and firefighting drones and ground firefighting auxiliary devices are dispatched according to the generated dispatch tasks. Based on the scheduling tasks of firefighting drones and ground firefighting auxiliary devices, during the planning of firefighting drone routes using a digital twin model of a photovoltaic power station, spatial and temporal conflicts are detected between the generated multiple firefighting drone routes. If spatial and temporal conflicts are detected among multiple firefighting drone routes, the flight altitude and flight time of the conflicting firefighting drones are adjusted to avoid conflicts. The system also monitors whether there are overlapping operation areas between bombing operations and ground firefighting auxiliary device operations. If there are overlapping operation areas, the operation time of the ground firefighting auxiliary device is adjusted to stagger the bombing operation time.

5. The multi-source perception based unmanned aerial vehicle fire source positioning and extinguishing control method according to claim 4, characterized in that, Also includes: An attitude prediction model is established based on the dynamic equations; The environmental data for future periods and the current operating status data of the inspection drone are obtained based on the pre-built environmental prediction model and input into the attitude prediction model to obtain the attitude for future periods as the actual attitude. Based on the target attitude trajectory of the inspection drone, the goal is to minimize the deviation between the actual attitude and the target attitude and minimize the change in motor control quantity within the future prediction period. Under the constraints of attitude angle, angular velocity and torque, the optimal attitude adjustment value for the future period is obtained by solving. The system utilizes PID control to calculate motor control quantities and maintain UAV attitude stability based on the optimal attitude adjustment value for future time periods and the error of the actual attitude angle calculated by fusing multi-sensor data using a quaternion attitude calculation architecture. The dynamic adjustment of PID parameters for attitude control is divided into three stages: cruise, hovering, and bomb preparation. Each stage is configured with an independent set of proportional-integral-derivative parameters, and different stages correspond to future time periods of preset duration. 6.The multi-source perception based UAV fire source positioning and extinguishing control method according to claim 4, wherein, Also includes: Before the fire-fighting drone reaches the fire source location and starts the bomb-dropping operation, the distance between the fire-fighting drone and the fire source is measured in real time using a laser rangefinder mounted on the fire-fighting drone. The optimal bomb-dropping angle and time are calculated in combination with the parabolic trajectory model of the fire-fighting bomb. The hovering attitude of the fire-fighting drone is adjusted according to the optimal bomb-dropping angle, and the bomb-dropping command is triggered at the optimal time.

7. The multi-source perception based unmanned aerial vehicle fire source positioning and extinguishing control method according to claim 1, characterized in that, Also includes: If the assessment of the fire extinguishing effect determines that the fire is partially extinguished or that there is a risk of reignition, a second bombing will be triggered. When a second bombing is triggered, multi-source data acquisition, multi-source data fusion processing and precise fire source location are carried out again, fire-fighting drone scheduling and attitude control parameter configuration are executed, fire-fighting operations are carried out and the effect is evaluated, until the fire-fighting effect evaluation result is that there is no risk of reignition.

8. A multi-source perception based unmanned aerial vehicle fire source positioning and fire extinguishing control system, characterized in that, include: The data acquisition module is used to construct an integrated ground-air multi-source sensing network and collect multi-source data based on the terrain and geographic information data of photovoltaic power stations; The construction of the integrated ground-air multi-source sensing network includes: deploying multimodal fire source sensing equipment on the ground, and deploying inspection drones and firefighting drones in the air; The fire source localization module is used to fuse multi-source data and accurately locate fire sources. The data fusion processing and accurate fire source localization include: extracting fire source features using an improved residual network based on embedded adaptive attention, fusing multi-scale feature maps through a feature pyramid network; identifying fire sources and outputting preliminary fire source localization using an improved anchor-free detection algorithm based on an optimized loss function; and correcting fire source position deviations by using extended Kalman filtering combined with data from ground sensing devices and data collected by inspection drones to obtain the fire source localization. The fire suppression scheduling module is used to execute the scheduling and attitude control parameter configuration of fire suppression drones. The fire suppression drone scheduling and attitude control include: planning the flight path of the fire suppression drone based on a pre-built digital twin model of a photovoltaic power station, and automatically loading fire extinguishing bombs using the modular mounting mechanism of the hangar on the fire suppression drone; during flight according to the flight path, a quaternion attitude calculation architecture is used to integrate multi-sensor data collected by the attitude sensors on the fire suppression drone, and the attitude stability of the fire suppression drone is maintained by dynamically adjusting the PID control parameters. The fire suppression execution module is used to carry out fire suppression operations and perform effect evaluation and network optimization to form a closed-loop control. The real-time fire suppression operation and effect evaluation include: based on the fire source location and the flight path of the fire suppression drone, initiating bomb dropping operations when the drone reaches the fire source location; after bomb dropping, continuously monitoring the temperature and smoke data of the fire suppression area using the camera device on the fire suppression drone, analyzing and evaluating the fire suppression effect, determining the fire source location error, and providing feedback; the network optimization in the closed-loop control includes: adjusting and optimizing the layout of ground sensing equipment and the drone inspection route based on the feedback of the fire suppression effect and the fire source location error.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the method as described in any one of claims 1 to 7.

10. A computer device, comprising: The computer device includes a memory, a processor, and a program stored in and executable on the memory, the program being executed by the processor to implement the steps of the method as described in any one of claims 1 to 7.

Images