A catapult type flame identification early warning unmanned aerial vehicle system

Abstract

The application discloses a kind of ejection flame identification early warning unmanned aerial vehicle systems, including ejection unmanned aerial vehicle, launching cylinder module, the body of the ejection unmanned aerial vehicle carries 4G / 5G communication and data transmission module, gimbal with camera, airborne edge computing unit;Camera installed on gimbal continuously acquires ground image with set frame rate, generates continuous real-time video data stream, is transmitted to the image buffer queue of airborne edge computing unit;Airborne edge computing unit deploys image preprocessing module, flame identification module, airborne edge computing unit reads original image from image buffer queue frame by frame, inputs flame identification module after original image is preprocessed by image preprocessing module and executes inference, whether to detect suspected flame target in current frame image is judged;When meeting the target detected is suspected flame target, generate flame event record;Airborne edge computing unit generates flame event record and current telemetry data stream as early warning information package and is uploaded to ground control terminal in real time by 4G / 5G communication and data transmission module.The application provides an effective technical approach for intelligent identification of fire in complex environment by airborne edge computing unit operation carried by flight platform constructed based on ejection unmanned aerial vehicle and launching cylinder module.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, specifically to a catapult-launched flame identification and early warning UAV system suitable for forest, industrial park and complex terrain scenarios. Background Technology

[0002] Existing fire detection and early warning technologies largely rely on single sensors or manual inspections, which have significant limitations in terms of real-time performance and accuracy in complex environments such as forest fire monitoring and industrial facility inspections. However, with the deep integration of aviation and emergency management technologies, drones, with their flexibility and high timeliness, offer possibilities for forest fire monitoring, industrial facility inspections, and emergency early warning of natural disasters.

[0003] In the fields of rapid deployment and variable configuration structures, research institutions worldwide have conducted extensive research. For example, in 2019, Caltech and NASA's Jet Propulsion Laboratory (JPL) jointly developed the "SQUID" (Streamlined Quick Unfolding Investigation Drone), a cannon-launched multirotor UAV that utilizes ballistic launch technology to achieve rapid conversion from high-speed flight to hovering configuration. Northwestern Polytechnical University (NWPU) developed the "Peregrine Falcon" series of long-endurance folding-wing UAVs, achieving maneuverable takeoff in complex environments through cannon-launch technology. In 2020, the U.S. Army Research Laboratory (ARL) unveiled the "GLUAS" (Grenade Launched Unmanned Aerial System) project, exploring the deployment of small folding rotorcraft using a 40mm grenade launcher. Regarding perception algorithms, domestic and international research teams have widely applied deep learning frameworks such as FireNet, Faster R-CNN, and YOLOv5 for feature extraction and recognition of forest fire images.

[0004] However, in-depth research into the aforementioned patents and cutting-edge literature reveals a significant technological gap in existing flame early warning UAV systems regarding "emergency closed-loop integration." Existing cannon-launched systems (such as SQUID or GLUAS) primarily focus on aerodynamic layout and takeoff dynamics, with design goals largely centered on instantaneous deployment. They haven't addressed the end-to-end optimization required for forest fire early warning scenarios, encompassing "payload stability, long-endurance monitoring, and low-latency transmission." Their technological shortcomings are mainly manifested in two ways: First, existing folding and unfolding mechanisms lack research on gimbal-based collaborative operation; second, in current early warning schemes, flame recognition algorithms mostly run on ground stations or large-scale airborne processors, lacking an airborne edge computing closed loop that matches the ejection deployment rhythm. This results in a response lag of several minutes in the logical chain from "ejection from the cannon" to "retransmission of the first fire point warning with coordinates." For scenarios with extremely high time sensitivity, such as forest fires, this several-minute response lag can lead to uncontrolled fire spread. Therefore, there is an urgent need to provide an ejection-type flame early warning system that integrates high overload protection, adaptive deployment, and a perception-transmission integrated architecture. Summary of the Invention

[0005] This invention provides a catapult-launched flame identification and early warning drone system. By using a flight platform based on a catapult-launched drone and a launch tube module, and carrying an onboard edge computing unit, it provides an effective technical approach for intelligent fire identification in complex environments.

[0006] The technical solution of this invention is:

[0007] A catapult-launched flame identification and early warning unmanned aerial vehicle (UAV) system includes a catapult-launched UAV and a launch tube module. The catapult-launched UAV includes a fuselage, a flight rotor module, landing gear, a GPS positioning module 1, a battery 2, and a flight controller 3. The catapult-launched UAV cooperates with the launch tube module to allow the flight rotor module of the catapult-launched UAV to have a first deployed state and a first folded state, and to allow the landing gear of the catapult-launched UAV to have a second deployed state and a second folded state. The fuselage of the catapult-launched UAV is used to mount the GPS positioning module 1, the battery 2, and the flight controller 3. The fuselage of the catapult-launched UAV also carries a 4G / 5G communication and data transmission module 4, a gimbal with a camera 5, and an onboard edge computing unit 6.

[0008] The airborne edge computing unit 6 and the flight controller 3 communicate and control each other bidirectionally via multiple signal lines: First, the telemetry serial port is used by the airborne edge computing unit 6 to read the real-time telemetry data stream output by the flight controller 3 and to send flight control commands to the flight controller 3; second, the control command line is used by the airborne edge computing unit 6 to send high-priority control commands directly to the flight controller 3; third, the bus is used to transmit the status feedback of the control commands of the flight controller 3; the operating power of the flight controller 3 is provided by the airborne edge computing unit 6.

[0009] The airborne edge computing unit 6 transmits high-bandwidth bidirectional data to the 4G / 5G communication and data transmission module 4 via a data port for uploading early warning information packets and receiving them from the ground control terminal. The modem control line is used by the airborne edge computing unit 6 to send AT commands to the 4G / 5G communication and data transmission module 4. The network status line is used by the 4G / 5G communication and data transmission module 4 to report the current network connection status to the airborne edge computing unit 6 in real time. The power supply for the 4G / 5G communication and data transmission module 4 is provided by the airborne edge computing unit 6.

[0010] The airborne edge computing unit 6 is connected to the gimbal 5. The airborne edge computing unit 6 sends pitch axis control commands, roll axis control commands, and yaw axis control commands to the gimbal 5 respectively. The gimbal 5 transmits the current three-axis angle values ​​of the gimbal back to the airborne edge computing unit 6 in real time for the calculation of the geographic coordinates of the fire point. The working power of the gimbal 5 is provided by an external power management circuit and is connected to the ground wire of the airborne edge computing unit 6. The camera installed on the gimbal 5 continuously collects ground images at a set frame rate, generates a continuous real-time video data stream, and transmits it to the image buffer queue of the airborne edge computing unit 6.

[0011] The airborne edge computing unit 6 receives a stable operating voltage from the battery 2 after processing by the power management module.

[0012] The airborne edge computing unit 6 is equipped with an image preprocessing module and a flame recognition module. The airborne edge computing unit 6 reads the original images frame by frame from the image buffer queue, preprocesses the original images by the image preprocessing module, and then inputs them into the flame recognition module to perform inference, determining whether a suspected flame target is detected in the current frame image. When the detected target is a suspected flame target, a flame event record is generated. The airborne edge computing unit 6 uploads the generated flame event record and the current telemetry data stream as an early warning information packet to the ground control terminal in real time through the 4G / 5G communication and data transmission module 4.

[0013] Furthermore, the image is processed by the image preprocessing module to perform the following operations:

[0014] Resolution normalization: Adjust the original image to the input size specified by the flame target detection model built into the flame recognition module, keep the aspect ratio unchanged, and fill in the missing areas;

[0015] Color space conversion: Converting a resolution-normalized image from BGR format to RGB format;

[0016] Pixel value normalization: The pixel value is mapped from an integer of 0 to 255 to a floating-point number of 0.0 to 1.0, which is used as the input tensor of the flame target detection model.

[0017] Furthermore, the flame recognition module performs the reasoning as follows:

[0018] This is used to input the original image into the flame target detection model built into the flame recognition module after preprocessing by the image preprocessing module, so as to output the bounding box coordinates, target category and confidence score of all candidate boxes;

[0019] This is used to determine whether the confidence score of the candidate box exceeds the first preset detection threshold: if there is no flame candidate box that meets the conditions in the current frame, the determination result is negative, and inference continues to be performed on the next frame image; if there is a candidate box that meets the conditions, the determination result is positive, and the corresponding candidate box is used as the initial candidate box and three serial processing links are triggered simultaneously:

[0020] The first serial processing link is used to calculate the confidence level: for each initial screening candidate box, the multi-dimensional features of the flame image are further extracted from the region of each initial screening candidate box and input into the comprehensive confidence level evaluation module to calculate the comprehensive confidence level of each initial screening candidate box; the initial screening candidate boxes with a comprehensive confidence level exceeding the second preset detection threshold are used as second screening candidate boxes, indicating that there is a suspected flame target;

[0021] The second serial processing link is used to obtain the geographic coordinates of the fire point: when there are two candidate boxes, the airborne edge computing unit 6 reads the current geographic coordinates of the UAV in real time from the NMEA data stream of the GPS positioning module 1 through the flight controller 3, and the airborne edge computing unit 6 reads the current flight attitude angle from the real-time telemetry data stream of the flight controller 3 through the telemetry serial port. Combined with the current pitch, roll and heading angles of the gimbal fed back by the gimbal 5, the ground projection latitude and longitude coordinates of the current frame image in the geographic coordinate system are calculated using the camera intrinsic parameter matrix and coordinate transformation formula, i.e., the geographic coordinates of the fire point.

[0022] The third serial processing link is used to fuse timestamps and features: it records the UTC timestamp of the current image frame, and encapsulates the timestamp information with the bounding box coordinates of the second-screen candidate boxes, the comprehensive confidence score, and the geographical coordinates of the fire point in a structured manner to form a timestamped fire event record.

[0023] Furthermore, the comprehensive confidence assessment module specifically includes:

[0024] This is used to extract local image regions from the original images corresponding to each of the initial candidate frames, map the pixels within these local image regions from the original BGR format to the standard RGB color space, and record the maximum value among the normalized RGB components of the pixels within the local image regions as the maximum brightness value. The minimum value is recorded as the minimum brightness value. The difference in brightness is recorded as ;

[0025] Used based on maximum brightness value Minimum brightness value The brightness difference is used to extract the hue component H, saturation component S and luminance component V of each pixel;

[0026] This is used to extract multidimensional features from the converted local image region based on the hue component H, saturation component S, and luminance component V of each pixel, and to calculate the comprehensive confidence level of the multidimensional features through a weighted fusion algorithm.

[0027] Furthermore, based on the hue component H, saturation component S, and luminance component V of each pixel, multi-dimensional feature extraction is performed on the converted local image region, including: the first dimension feature is the flame pixel area ratio feature, the second dimension feature is the normalized regional average luminance feature, the third dimension feature is the normalized regional average luminance feature, and the fourth dimension feature is the bounding box geometric feature.

[0028] Furthermore, the first dimension feature is the flame pixel area ratio feature, specifically:

[0029] In HSV space, the decision function for the flame pixel is set as follows: The specific calculation formula is as follows:

[0030] ;

[0031] Based on the decision function of the flame pixels, the feature value of the flame pixel area ratio is obtained. The calculation formula is:

[0032] ;

[0033] in, For any pixel within the initial candidate box, This is the set of pixels within the initial candidate bounding box region. This represents the total number of pixels within the initial candidate bounding box area.

[0034] Furthermore, the normalized regional average brightness feature value The calculation formula is:

[0035] ;

[0036] in, Each pixel The hue component, saturation component, and brightness component.

[0037] Furthermore, the third dimension feature is the flame color distribution histogram similarity feature, specifically: the hue component is uniformly divided into a first number of intervals, and the saturation component is uniformly divided into a second number of intervals, thereby constructing a feature with... The two-dimensional joint probability distribution histogram of the histogram feature bars is used to calculate the frequency of each pixel within the initial candidate box falling into each histogram interval, and then normalize the result to obtain the current flame color distribution histogram vector. Based on the flame color distribution histogram vector, obtain the flame color distribution histogram similarity feature value. The calculation formula is:

[0038] ;

[0039] in, This is a preset baseline flame color distribution histogram vector; The total number of feature bars in the two-dimensional joint probability distribution histogram, i.e. First quantity × Second quantity.

[0040] Furthermore, the bounding box geometric feature score The calculation formula is:

[0041] ;

[0042] in, The aspect ratio of the initial candidate boxes; This is the preset average of the flame aspect ratio. This is the preset standard deviation of the flame aspect ratio.

[0043] Furthermore, the comprehensive confidence score is calculated from the multidimensional features using a weighted fusion algorithm, and the specific fusion formula is as follows:

[0044] ;

[0045] In the formula, To assess the overall confidence level, The confidence score of the initial candidate boxes output by the model inference; , , , and These are the weight coefficients for the corresponding feature dimensions, and the sum of each weight coefficient is 1; This represents the feature value of the flame pixel area ratio. The normalized regional average brightness characteristic value. The similarity feature value of the flame color distribution histogram. Scoring is given for the geometric features of the bounding box.

[0046] The beneficial effects of this invention are:

[0047] (1) It greatly reduces the combat readiness response time and improves the deployment efficiency in extreme environments. By introducing high-pressure air-assisted catapult technology to replace the traditional static manual deployment and takeoff mode, this system realizes a magnitude switch from a static storage state to a full-power reconnaissance state, ensuring that it still has the rapid early warning response capability of "ready to use" in complex micro-topography areas such as deep forests and steep mountains.

[0048] (2) It achieves a harmonious balance between high reliability and portability. The foldable flight rotor module and landing gear components form a dense envelope in the launch tube, which greatly optimizes the soldier's load-bearing experience in harsh environments. At the same time, the outer shell of the tube provides an all-weather structural safety barrier for the precision flight control chip, power battery and optoelectronic payload, reducing the damage rate during storage and transportation.

[0049] (3) Significantly enhanced image acquisition stability and recognition accuracy under high dynamic launch transients. The system's three-axis stabilization gimbal can perform precise compensation by calculating attitude disturbances at high frequency and driving the supplementary angle motor instantly after the ejection acceleration disappears. This eliminates the impact of propeller cyclone interference and air turbulence on the visual sensor, providing high-quality, high signal-to-noise ratio low-level telemetry data streams for airborne edge computing units (such as the YOLO series framework optimized for flame characteristics).

[0050] (4) Effectively closes the end-to-end latency of "perception-communication-early warning". Through the ingenious connection between GPS, flight controller, 4G / 5G communication and data transmission module and airborne edge computing unit, this solution can significantly reduce the IO overhead of image information in the encoding, decoding and transmission process by using a high-speed data bus. This allows the identified suspected flame targets and the current telemetry data stream to be sent down with low latency through the 4G / 5G communication and data transmission module, thus securing a critical response window for the suppression of forest fires in the early stages. Attached Figure Description

[0051] Figure 1 A schematic diagram of the overall structure of a catapult-launched flame identification and early warning unmanned aerial vehicle system;

[0052] Figure 2 This is a detailed cross-sectional structural diagram of the launch tube module;

[0053] Figure 3 A schematic diagram illustrating the stages of the catapult deployment process (left: preparation stage, the drone is stored in the launch tube, the air pressure is pre-charged, and the release device is locked; middle: catapult stage, the catapult piston pushes out the drone, and the powered arm begins to deploy; right: flight stage, the drone detaches from the launch tube, the powered arm deploys and locks, and the gimbal enters the working state).

[0054] Figure 4The overall workflow diagram for the ejection-type flame recognition and early warning UAV system is as follows: (A-Collection and transportation; B-Ejection triggering; C-Deployment from the launch tube; D-Aerial cruise; E-Image acquisition; F-Flame recognition; G-Early warning transmission; H-Multi-UAV collaboration; I-Ground terminal response).

[0055] Figure 5 A flowchart illustrating the preparation, launch, and flight phases of a catapult-launched drone. Figure 5 A is the flowchart of the catapult preparation stage (check the folding of the power arm → insert the launch tube → connect the nitrogen gas line → inflate to the rated pressure → pressure check → lock the release device → system standby). Figure 5 B is the flowchart for the ejection and deployment phase (ejection triggered → release of locking hook → piston ejection of UAV → gradual release of cylinder wall constraint → release and deployment of arm springs → synchronous deployment of landing gear → UAV detaches from cylinder opening → rotor enters horizontal position → flight control is activated and enters flight → flight phase stabilizes).

[0056] Figure 6 The flowchart for airborne flame recognition and early warning processing is as follows: (entering cruise → camera acquires image frames → image preprocessing → YOLO inference → flame detected? Otherwise, loop → parallel: calculate confidence, obtain GPS coordinates (geographic coordinates of fire point), fuse timestamp and features → comprehensive confidence reaches threshold → package early warning information → encrypted upload to ground control terminal);

[0057] Figure 7 System power-on self-test and initialization flowchart (Power-on activation after removing the canister → Detect battery power <Insufficient → Low battery alarm> → Load flight controller firmware → IMU self-test calibration <Abnormal → IMU fault alarm> → GPS satellite search and positioning <Invalid → Switch to assisted positioning> → Record starting coordinates → Establish 4G / 5G link → Load flame recognition model → Gimbal three-axis homing <Abnormal → Lock protection> → Return to ready state → Enter working state);

[0058] Figure 8 This is a schematic diagram of the connection principle for an integrated edge computing unit system.

[0059] Figure 9 This is a bar chart comparing performance indicators based on Example 2.

[0060] Figure 10 This is a comparison chart of the confusion matrix provided in Example 2.

[0061] Figure 11 The score distribution violin plot is based on Example 2.

[0062] Figure 12 This is a scatter decision boundary diagram based on Example 2.

[0063] Figure 13This is a comparison chart of the mean values ​​of multidimensional features provided in Example 2.

[0064] Figure 14 The FPR comparison and integrated radar chart are based on Example 2.

[0065] Figure 15 This is a candidate box level score analysis chart based on Example 2.

[0066] The labels in the diagram are as follows: 1-Integrated GPS positioning module; 2-Battery (model aircraft lithium battery 6S); 3-Flight controller; 4-4G / 5G communication and data transmission module; 5-Gimbal with camera; 6-Airborne edge computing unit; 7-Nitrogen storage tank; 8-Ejection piston; 9-Unhooking device; 10-Pressure gauge; 11-Launch tube body; 12-Safety rope. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0068] Example 1: As Figures 1-8 As shown, a catapult-launched flame recognition and early warning unmanned aerial vehicle (UAV) system includes a catapult-launched UAV and a launch tube module. The catapult-launched UAV includes a fuselage, a flight rotor module, landing gear, a GPS positioning module 1, a battery 2, and a flight controller 3. The catapult-launched UAV cooperates with the launch tube module to allow the flight rotor module of the catapult-launched UAV to have a first deployed state and a first folded state, and to allow the landing gear of the catapult-launched UAV to have a second deployed state and a second folded state. The fuselage of the catapult-launched UAV is used to mount the GPS positioning module 1, the battery 2, and the flight controller 3. The fuselage of the catapult-launched UAV also carries a 4G / 5G communication and data transmission module 4, a gimbal with a camera 5, and an onboard edge computing unit 6.

[0069] The airborne edge computing unit 6 and the flight controller 3 communicate and control bidirectionally via multiple signal lines: First, the telemetry serial port is used by the airborne edge computing unit 6 to read the real-time telemetry data stream output by the flight controller 3 and to send flight control commands to the flight controller 3; second, the control command line is used by the airborne edge computing unit 6 to directly send high-priority control commands to the flight controller 3; third, the AUX CAN bus is used to transmit the status feedback of the control commands of the flight controller 3; the operating power of the flight controller 3 is provided by the airborne edge computing unit 6.

[0070] The airborne edge computing unit 6 transmits high-bandwidth bidirectional data to the 4G / 5G communication and data transmission module 4 via a data port for uploading early warning information packets and receiving them from the ground control terminal. The modem control line is used by the airborne edge computing unit 6 to send AT commands to the 4G / 5G communication and data transmission module 4. The network status line is used by the 4G / 5G communication and data transmission module 4 to report the current network connection status to the airborne edge computing unit 6 in real time. The power supply for the 4G / 5G communication and data transmission module 4 is provided by the airborne edge computing unit 6.

[0071] The airborne edge computing unit 6 is connected to the gimbal 5. The airborne edge computing unit 6 sends pitch axis control commands, roll axis control commands, and yaw axis control commands to the gimbal 5 respectively. The gimbal 5 transmits the current three-axis angle values ​​of the gimbal back to the airborne edge computing unit 6 in real time for the calculation of the geographic coordinates of the fire point. The working power of the gimbal 5 is provided by an external power management circuit and is connected to the ground wire of the airborne edge computing unit 6. The camera installed on the gimbal 5 continuously collects ground images at a set frame rate, generates a continuous real-time video data stream, and transmits it to the image buffer queue of the airborne edge computing unit 6.

[0072] The airborne edge computing unit 6 receives a stable operating voltage from the battery 2 after processing by the power management module.

[0073] The airborne edge computing unit 6 is equipped with an image preprocessing module and a flame recognition module. The airborne edge computing unit 6 reads the original images frame by frame from the image buffer queue, preprocesses the original images by the image preprocessing module, and then inputs them into the flame recognition module to perform inference to determine whether a suspected flame target is detected in the current frame image. When the detected target is a suspected flame target, a flame event record is generated. The airborne edge computing unit 6 encrypts and encodes the generated flame event record and the current telemetry data stream and then uploads it to the ground control terminal in real time through the 4G / 5G communication and data transmission module 4.

[0074] Furthermore, the image is processed by the image preprocessing module to perform the following operations:

[0075] Resolution normalization: Adjust the original image to the input size specified by the flame target detection model built into the flame recognition module, keep the aspect ratio unchanged, and fill in the missing areas;

[0076] Color space conversion: Converting a resolution-normalized image from BGR format to RGB format;

[0077] Pixel value normalization: The pixel value is mapped from an integer of 0 to 255 to a floating-point number of 0.0 to 1.0, which is used as the input tensor of the flame target detection model.

[0078] Furthermore, the flame recognition module performs the reasoning as follows:

[0079] This is used to input the original image into the flame target detection model built into the flame recognition module after preprocessing by the image preprocessing module, so as to output the bounding box coordinates, target category and confidence score of all candidate boxes;

[0080] This is used to determine whether the confidence score of the candidate box exceeds the first preset detection threshold: if there is no flame candidate box that meets the conditions in the current frame, the determination result is negative, and inference continues to be performed on the next frame image; if there is a candidate box that meets the conditions, the determination result is positive, and the corresponding candidate box is used as the initial candidate box and three serial processing links are triggered simultaneously:

[0081] The first serial processing link is used to calculate the confidence score: For each initial screening candidate box, multi-dimensional features of the flame image are further extracted from the region of each initial screening candidate box, including the flame pixel area ratio, the average brightness of the region, the flame color distribution histogram, and the geometric features of the bounding box. The above multi-dimensional features are input into the comprehensive confidence score evaluation module to calculate the comprehensive confidence score of each initial screening candidate box. Initial screening candidate boxes with a comprehensive confidence score exceeding the second preset detection threshold (greater than 0.6) are used as second screening candidate boxes, indicating that there is a suspected flame target (that is, if the comprehensive confidence score is less than or equal to the second preset detection threshold, it means that the current flame candidate is doubtful (possibly a false detection), and the reasoning continues for the next frame image; if the comprehensive confidence score exceeds the threshold, it enters the second serial processing link).

[0082] Furthermore, the comprehensive confidence assessment module specifically comprises:

[0083] This is used to extract local image regions from the original images corresponding to each of the initial candidate frames, map the pixels within these local image regions from the original BGR format to the standard RGB color space, and record the maximum value among the normalized RGB components of the pixels within the local image regions as the maximum brightness value. The minimum value is recorded as the minimum brightness value. The difference in brightness is recorded as ;

[0084] Based on maximum brightness value Minimum brightness value The brightness difference is used to extract the hue component H, saturation component S and brightness component V of each pixel, so as to reduce the interference of ambient light changes on feature extraction.

[0085] Luminance component The calculation formula is:

[0086] ;

[0087] Saturation component The calculation formula is:

[0088] ;

[0089] Tone Components The calculation formula is:

[0090] ;

[0091] ;

[0092] in, , , This represents the normalized RGB components; the hue components obtained through the above calculations. The value range is mapped to the interval between 0 and 180.

[0093] Multidimensional feature extraction is performed on the transformed local image regions:

[0094] The first dimension feature is the proportion of flame pixel area:

[0095] In HSV space, the decision function for the flame pixel is set as follows: The specific calculation formula is as follows:

[0096] ;

[0097] Based on the decision function of the flame pixels, the feature value of the flame pixel area ratio is obtained. The calculation formula is:

[0098] ;

[0099] in, For any pixel within the initial candidate box, This is the set of pixels within the initial candidate bounding box region. This represents the total number of pixels within the initial candidate bounding box area.

[0100] The second dimension feature is the normalized regional average brightness feature, and the normalized regional average brightness feature value. The calculation formula is:

[0101] ;

[0102] in, For pixels The brightness component value;

[0103] The third dimension feature is the flame color distribution histogram similarity feature: the hue component is evenly divided into 8 intervals, and the saturation component is evenly divided into 4 intervals, thus constructing a two-dimensional joint probability distribution histogram with 32 histogram feature bars. The frequency of each pixel in the initial candidate box falling into each histogram interval is counted and normalized to obtain the current flame color distribution histogram vector. Based on the flame color distribution histogram vector, obtain the flame color distribution histogram similarity feature value. The calculation formula is:

[0104] ;

[0105] in, This is a preset baseline flame color distribution histogram vector. The total number of feature bars in the two-dimensional joint probability distribution histogram is taken as [value missing] in this embodiment. ; express The Middle One component; express The Middle One component;

[0106] The fourth dimension feature is the bounding box geometric feature, and the bounding box geometric feature score is... The calculation formula is:

[0107] ;

[0108] in, The aspect ratio of the initial candidate boxes is calculated using the following formula:

[0109] ;

[0110] In the formula, For the first The height of the initial candidate frame For the first The width of each initial candidate frame; This is the preset average value of the flame aspect ratio (1.5 in this embodiment). This is the preset standard deviation of the flame aspect ratio (0.8 in this embodiment);

[0111] Finally, the overall confidence score is calculated using a weighted fusion algorithm. The specific fusion formula is as follows:

[0112] ;

[0113] In the formula, To assess the overall confidence level, The confidence level of the initial candidate boxes output by the model inference. This represents the feature value of the flame pixel area ratio. The normalized regional average brightness characteristic value. The similarity feature of the flame color distribution histogram. Scoring the geometric features of the bounding box; , , , and These are the weight coefficients for the corresponding feature dimensions, and each weight coefficient satisfies the normalization condition:

[0114] ;

[0115] For example, weighting coefficients , , , and The values ​​are 0.40, 0.15, 0.10, 0.20, and 0.15 respectively; if the calculated comprehensive confidence level... If the detection threshold exceeds the second preset threshold of 0.6, the initial screening candidate box is used as the second screening candidate box, indicating that there is a suspected flame target and it enters the subsequent processing. Otherwise, it indicates that the current flame candidate is doubtful, and the next frame image is used for reasoning.

[0116] The second serial processing link is used to obtain the geographic coordinates of the fire point: when there is a candidate box for screening, the airborne edge computing unit 6 reads the current geographic coordinates (longitude, latitude, and altitude) of the UAV in real time from the NMEA data stream of the GPS positioning module 1 through the flight controller 3, and the airborne edge computing unit 6 reads the current flight attitude angle from the real-time telemetry data stream of the flight controller 3 through the telemetry serial port. Combined with the current pitch / roll / heading angle of the gimbal fed back by the gimbal 5, the ground projection latitude and longitude coordinates of the current frame image in the geographic coordinate system are calculated using the camera intrinsic parameter matrix and coordinate transformation formula, i.e., the geographic coordinates of the fire point.

[0117] The third serial processing link is used to fuse timestamps and features: it records the UTC timestamp of the current image frame, and encapsulates the timestamp information with the bounding box coordinates of the second-screen candidate boxes, the comprehensive confidence score, and the geographic coordinates of the fire point in a structured manner to form a timestamped Flame Event Record, which is used for subsequent multi-frame temporal correlation and early warning information generation.

[0118] The innovation of this invention lies in utilizing the catapult-launched drone and launch tube module disclosed in the embodiments of this invention, or the multi-takeoff mode aircraft (CN120003746A) developed by the applicant team as the catapult-launched drone, and the air pressure launch tube for the catapult aircraft (CN120003761A) as the launch tube module, and cleverly integrating an airborne edge computing unit 6 and a gimbal 5 with a camera to construct a catapult-launched flame recognition and early warning drone system, providing a new identification path for existing flame recognition. The catapult-launched drone disclosed in the embodiments of this invention is based on the same technology as CN120003746A, but differs in that: a support frame for installing a GPS positioning module 1 is set in the nose of the fuselage, which is also used to install a 4G / 5G communication and data transmission module 4; the fuselage is extended to further provide installation space for the airborne edge computing unit 6; and a gimbal 5 with a camera is installed at the bottom of the fuselage through a vibration damping structure. The launch tube module is mainly composed of five core components: the launch tube body 11, the nitrogen storage tank 7, the catapult piston 8, the release device 9, and the pressure gauge 10. The following describes optional embodiments of the present invention in conjunction with the accompanying drawings:

[0119] Example 2: Figure 1 , Figure 2 and Figure 3 As shown, the ejection-type flame recognition and early warning UAV system of this invention is mainly composed of two deeply integrated parts: a launch tube module and an ejection UAV. The launch tube module is responsible for the storage, protection, and ejection power of the ejection UAV; the ejection UAV integrates the full-link functions of navigation and positioning, communication transmission, flight control, airborne edge computing, and optoelectronic payload. In the stored configuration, the flight rotor module of the ejection UAV is constrained to the side wall of the fuselage by the elastic pre-tensioning force provided by the wing folding switching module, and is compactly arranged in the guide cavity of the launch tube module. After ejection is triggered, nitrogen pressure drives the ejection piston 8 to propel the ejection UAV body into the air at high speed. At the instant of breaking free from the radial constraint of the launch tube opening, each flight rotor module and landing gear automatically unfolds from the stored position (folded state) to the working position (unfolded state), realizing instantaneous configuration transformation and rapid activation of the sensing unit. The launch tube module and ejection UAV of this invention are further described below:

[0120] (I) Launch tube module

[0121] like Figure 2 As shown, the launch tube module consists of the following core components:

[0122] like Figure 2 As shown, the launch tube module is the core functional component for achieving rapid ejection deployment in this invention. It is mainly composed of five core components: the launch tube body 11, the nitrogen storage tank 7, the ejection piston 8, the release device 9, and the pressure gauge 10.

[0123] The launch tube body 11 is a lightweight, one-end-open cylindrical body made of high-strength aluminum alloy using a hydroforming process. The internal cavity dimensions of the tube are precisely matched to the maximum external dimensions of the folded and stowed UAV, making it the main structural component for UAV storage protection and launch guidance. The inner wall of the tube has axially arranged guide ribs to ensure the UAV maintains linear motion along the central axis during the launch stroke without translational deflection. Furthermore, the inward-curving end face of the launch tube can be designed to guide airflow, ensuring smooth airflow during high-speed launch and reducing air resistance at the final stage. Air passage interfaces and safety holes are located on both sides of the tube body, allowing for rapid pressure relief when necessary.

[0124] The nitrogen storage tank 7 is made of high-strength stainless steel and features an elastic sealing ring for gas retention. The tank is filled to the working pressure (typically 0.1~0.3 MPa) and connected to the bottom chamber of the launch tube 11 via an air passage interface, providing driving force for the ejection piston 8. A pressure gauge 10 is installed in series with the air passage to display the system's pre-charge pressure in real time, allowing operators to determine if the ejection preparation status is ready. The system only allows ejection operation after the pressure gauge 10 reading reaches the rated charge pressure range, thus preventing insufficient drone altitude due to inadequate inflation.

[0125] The ejection piston 8 is installed inside the launch tube body 11. The space between the ejection piston 8 and the closed end of the launch tube body 11 is a medium-filled area. An extension arm is provided on the side of the ejection piston 8 near the closed end of the tube body, and the extension arm has a through hole and a limiting groove. The ejection piston adopts a sealing ring design, which can effectively isolate the high-pressure gas chamber from the ejection passage, ensuring no gas or liquid leakage. The moving mating surface between the outer periphery of the ejection piston and the inner cavity of the tube body is a patterned bearing sliding seal ring, reducing the sliding friction between the piston and the inner wall of the tube. A rubber buffer layer is provided on the top surface of the piston to ensure that the ejection force is evenly applied to the landing gear at the bottom of the UAV. Driven by nitrogen pressure, the ejection piston moves upward in a straight line along the central axis of the tube body, transmitting thrust to the UAV and driving the UAV to eject rapidly along the central axis of the tube body.

[0126] The release device 9 is a hook-type mechanical safety mechanism, installed on the closed end of the inner wall of the launch tube 11. Before ejection triggering, the hook of the release device 9 engages with the limiting groove of the ejection piston 8 to clamp the ejection piston 8, thereby locking the drone in the storage position. The release device adopts a dual-action triggering method: when ejection is triggered, a small electromagnetic lock drives the hook to open rapidly (triggered remotely via network or wireless hand controller), and the drone is ejected under the action of air pressure thrust; at the same time, the safety rope 12 (such as aramid rope) can be used in conjunction with the release device 9, fixing one end of the safety rope 12 to the base of the release device 9, and engaging the other end of the safety rope 12 with the mounting hole to limit the movement of the ejection piston 8 towards the opening side of the launch tube 11, ensuring operational safety. It should be noted that in this embodiment of the invention, the release device 9 uses a commercially available Hongyun launcher.

[0127] (II) Catapult-launched drones

[0128] like Figure 1 and Figure 3 As shown, the main body of the UAV adopts a quadcopter symmetrical layout, with the central fuselage made of carbon fiber composite material, balancing high strength and lightweight. The ejection UAV includes a fuselage, flight rotor module, landing gear, GPS positioning module 1, battery 2, and flight controller 3. The fuselage of the ejection UAV is used to mount the GPS positioning module 1, battery 2, and flight controller 3. The fuselage of the ejection UAV also carries a 4G / 5G communication and data transmission module 4, a gimbal 5 with a camera, and an onboard edge computing unit 6. Among them, the GPS positioning module 1, battery 2, and 4G / 5G communication and data transmission module 4 are located in the nose of the fuselage, the flight controller 3 and the onboard edge computing unit 6 are located in the fuselage, and the gimbal 5 is mounted in the center of the lower part of the fuselage and connected to the fuselage through a vibration damping structure, forming an optoelectronic payload system with the camera. This layout can improve the center of gravity and stability. The UAV is equipped with four flight rotor modules. Each rotor module is connected to the fuselage at its base via a wing folding switching module. The wing folding switching module contains a high-elasticity torsion spring. In the retracted state, the flight rotor module is compressed and stores energy by being constrained by the external launch tube wall. The fairing is mounted on the nose and serves as a guide and damper during the launch stroke. (The following is a continuation of the description...) Figure 3 , Figure 5 The working process of catapult-launched drones is explained as follows:

[0129] Preparation Phase: The four flight rotor modules (arms) are in their first folded state, retracted downwards against the fuselage sidewalls, with the torsion springs at the base of each rotor module in a compressed, energy-storing state. The landing gear is also compressed and retracted due to radial constraint from the inner wall of the launch tube. The entire aircraft is arranged vertically within the launch tube cavity. Ground operators pre-charge the nitrogen storage tank 7, and after pressure gauge 10 confirms that the pressure reaches the target, the release device 9 locks the UAV in the retracted position, and the system enters standby mode. The process is as follows: Figure 5 As shown in Figure A.

[0130] Ejection Phase: The operator remotely triggers the ejection process via a ground control terminal or a personal wireless controller. The electromagnetic drive mechanism of the release device 9 is de-energized and ejects, allowing the UAV to rapidly eject along the central axis of the launch tube 11 under the thrust of the nitrogen-pressure-driven ejection piston 8. As the UAV moves upward along the tube axis, the radial constraints on the flight rotor modules and landing gear are sequentially released from the inner wall of the launch tube. The elastic potential energy of the stored torsion springs at the hinges of each flight rotor module is released, driving the flight rotor modules to rotate and unfold from their first folded state. Simultaneously, the landing gear spring mechanism drives the outriggers to open radially outward. The above unfolding action is essentially completed the instant the UAV leaves the launch tube opening. For a detailed sequence of actions, see [link to details]. Figure 5 A and Figure 5 B.

[0131] Flight Phase: After the UAV body completely detaches from the launch tube, the flight rotor module, driven by a torsion spring, rotates from its first folded state to its first deployed state perpendicular to the fuselage's longitudinal axis. The end-embedded latch automatically springs into the locking hole to mechanically lock the rotor module, ensuring it remains deployed and does not spring back during subsequent flight. The landing gear, after opening radially outward to a preset angle, also automatically locks itself using elastic latches, forming a stable support landing gear structure. At this point, the rotor assembly is in a horizontal working position. After the motors are powered on, the rotor blades begin to rotate under the throttle command of the flight controller 3. The flight controller 3 then initiates the system power-on self-test process (see details). Figure 7 The GPS positioning module 1 completes initial positioning and coordinate locking, the 4G / 5G communication and data transmission module 4 establishes a data link with the ground control terminal, the gimbal 5 enters the working state after the three axes return to zero, and the UAV enters the cruise or preset mission execution state.

[0132] (III) Detailed Description of Interface Connection for Integrated Edge Computing Unit

[0133] like Figure 8 As shown, the airborne edge computing unit 6 is the core of information processing and command scheduling of the system of the present invention. Using it as the central node, it establishes data and control connections with surrounding modules through the following sets of interfaces:

[0134] Connection with integrated GPS positioning module 1: Flight controller 3 communicates serially with integrated GPS positioning module 1 via NMEA serial port (NMEA DATA Tx / Rx bidirectional data line), receiving NMEA 0183 format positioning data frames (containing standard statements such as GPGGA and GPRMC, carrying positioning information such as latitude, longitude, altitude, speed, number of satellites searched, and HDOP) output by integrated GPS positioning module 1 at a frequency of 1Hz to 10Hz; the power supply of integrated GPS positioning module 1 is provided by the power output pin (VCC +5V) of flight controller 3, and the GND line is connected to the system ground to ensure stable operation of the module.

[0135] Connection with Flight Controller 3: The airborne edge computing unit 6 and the flight controller 3 communicate and control bidirectionally via multiple signal lines: First, the telemetry serial port (TELEMETRY Tx / Rx) is used by the airborne edge computing unit 6 to read the real-time telemetry data stream output by the flight controller 3 (including attitude angles (roll, pitch, yaw), flight speed, altitude, flight mode, battery percentage, etc. in MAVLink protocol format) and to send flight control commands (such as waypoint updates, return-to-home commands, etc.) to the flight controller 3; Second, the control command line (CONTROL COMMANDS) is used by the airborne edge computing unit 6 to send high-priority control commands (such as emergency landing, throttle lock, etc.) directly to the flight controller 3; Third, the AUX CAN bus is used to transmit the status feedback of the control commands of the flight controller 3; The VCC and GND power supply lines of the flight controller 3 are also connected to the power management circuit of the airborne edge computing unit 6, and power is distributed by the unified power management module.

[0136] Connection with 4G / 5G Communication and Data Transmission Module 4: The airborne edge computing unit 6 conducts high-bandwidth bidirectional data transmission with the 4G / 5G communication and data transmission module 4 via a high-speed data port (HIGH-SPEED DATA Rx / Tx) for uploading early warning information packets and receiving them from the ground control terminal; the modem control line is used by the airborne edge computing unit 6 to send AT commands to the 4G / 5G communication and data transmission module 4 to control the module's network connection, signal strength query, data transmission mode configuration, and other operations; the network status line is used by the 4G / 5G communication and data transmission module 4 to report the current network connection status (such as connected / disconnected / weak signal) to the airborne edge computing unit 6 in real time; the antenna interface of the 4G / 5G communication and data transmission module 4 is connected to the external antenna on the fuselage to ensure signal transmission quality during flight. The operating power (VCC / GND) of the 4G / 5G communication and data transmission module 4 is also provided by the power management circuit of the airborne edge computing unit 6.

[0137] Connection with gimbal 5: The airborne edge computing unit 6 sends pitch axis control commands (PITCH CMD), roll axis control commands (ROLL CMD), and yaw axis control commands (YAW CMD) to the gimbal 5 through three PWM control signal lines, respectively, to achieve independent and precise control of the gimbal's three-axis attitude; the gimbal 5 transmits the current three-axis angle values ​​of the gimbal back to the airborne edge computing unit 6 in real time through the GIMBAL FEEDBACK feedback line, which is used for gimbal attitude compensation in fire point geographic coordinate calculation; the working power (VCC / GND) of the gimbal 5 is provided by an external power management circuit and is connected to the ground wire of the airborne edge computing unit 6 to ensure that the signal reference potential is consistent and to prevent control signal interference caused by ground potential difference.

[0138] External power input: The airborne edge computing unit 6 is equipped with a dedicated power input interface (POWER INPUT) to receive a stable operating voltage (19V DC) from the battery 2 after being stepped down by the power management module; the power input interface has built-in reverse protection and overvoltage protection circuits to prevent damage to the core computing module due to incorrect wiring polarity or abnormal voltage, and to ensure stable power supply to the airborne edge computing unit 6 throughout the entire flight cycle.

[0139] (IV) Detailed Procedure for System Power-On Self-Test and Initialization

[0140] like Figure 7 As shown, after the drone body detaches from the launch tube, it is activated by power-on triggered by battery 2. The onboard edge computing unit 6 and flight controller 3 start synchronously and execute the following self-check and initialization steps in sequence. An abnormality in any step will trigger the corresponding alarm branch:

[0141] Power-on activation upon detachment: The moment the drone detaches from its casing, the main circuit of the fuselage battery 2 is connected, and the flight controller 3, the onboard edge computing unit 6, the 4G / 5G communication and data transmission module 4, the integrated GPS positioning module 1, the gimbal 5, and other modules are powered on and initialized in sequence to ensure that the subsequent self-test process is completed before the drone reaches a safe flight altitude.

[0142] Battery power detection: The system reads the current battery percentage of battery 2 through the battery management module built into flight controller 3. If the detected battery percentage is lower than the preset alarm threshold (20% of rated capacity), a low battery alarm is triggered: flight controller 3 transmits the low battery alarm information to the ground control terminal via the airborne edge computing unit 6 through 4G / 5G communication and data transmission module 4, prompting the operator to shorten the mission time or prepare to return to base; at the same time, flight controller 3 sets a mission duration limit in the low battery state to prevent a crash due to battery depletion. If the battery is sufficient, the subsequent procedures continue.

[0143] Loading flight controller firmware: Flight controller 3 reads and verifies the integrity of the flight controller firmware (CRC check) from the built-in Flash memory, loads the firmware code into the running memory, completes the initialization and startup of the flight control operating system (RTOS), and prepares the running environment for subsequent IMU calibration and sensor driver loading.

[0144] IMU Self-Test Calibration: The Flight Controller 3 drives the built-in IMU (Inertial Measurement Unit) to perform a power-on self-test: sensor range self-test of the three-axis accelerometer and three-axis gyroscope (verifying normal read / write of the range configuration register), static zero bias estimation (collecting multiple sets of sensor values ​​in a static state and calculating the zero bias compensation for each axis), and temperature compensation parameter loading (selecting the corresponding temperature compensation curve coefficient based on the current temperature sensor reading). If the zero bias of any sensor axis of the IMU exceeds the preset range (e.g., accelerometer zero bias exceeds ±0.3g, gyroscope zero bias exceeds ±5° / s), or sensor communication is abnormal (SPI / I2C bus read / write failure), an IMU fault alarm is triggered, the Flight Controller 3 enters protection mode and reports the fault code to the ground control terminal; if the calibration result is normal, the process continues.

[0145] GPS Satellite Search and Positioning: After power-on, the integrated GPS positioning module 1 begins searching for satellite signals and continuously outputs positioning data streams to the flight controller 3 via the NMEA serial port (NMEADATA Tx / Rx). The flight controller 3 monitors the GPS positioning status and determines its validity: when the number of visible satellites is ≥6 and the horizontal accuracy factor (HDOP) is ≤2.0, the positioning is considered valid; if the GPS positioning is still invalid within a preset timeout period (e.g., 60 seconds) (insufficient number of satellites or signal obstruction), it automatically switches to auxiliary positioning mode: using barometer altitude data instead of GPS altitude information, calculating horizontal displacement using IMU integral displacement, and simultaneously reporting a GPS invalid alarm to the ground control terminal to prompt the operator to pay attention to the current positioning status.

[0146] Recording initial coordinates: After successful positioning, the flight controller 3 records the initial coordinates (longitude, latitude, and altitude) and stores them in memory as the mission home point. All subsequent calculations of the geographic coordinates of the fire point are based on this home point.

[0147] 4G / 5G Link Establishment: After recording is complete, the flight controller 3 establishes a data communication link with the ground control terminal via the airborne edge computing unit 6 and the 4G / 5G communication and data transmission module 4. The flight controller 3 reports the communication status to the airborne edge computing unit 6 via the telemetry serial port (TELEMETRY). If the link is successfully established (handshake completed and confirmation response received), it continues; if communication is abnormal (no response received after timeout or connection interruption), the 4G / 5G communication and data transmission module 4 automatically performs reconnection (maximum of 3 retries, with a 5-second interval between each retrieval), and continues after a successful retrieval; if all retries fail, the communication abnormality is recorded in the local log and the process continues in offline task mode (warning information is temporarily stored locally and uploaded in batches after the link is restored).

[0148] Loading Airborne Edge Computing Unit 6: After the link is established, the image processing algorithm built into the image preprocessing module of the airborne edge computing unit 6 is loaded, and the weight file corresponding to the flame recognition module is read and loaded into the inference acceleration framework (ONNXRuntime). The initialization and warm-up of the inference engine are completed (usually 1 to 3 empty inferences are run to activate the GPU cache and memory allocation) to ensure that the subsequent inference latency is stable within the design range (e.g., single frame inference latency ≤ 50ms).

[0149] Three-axis gimbal homing: After power-on, gimbal 5 performs a three-axis (pitch / roll / yaw) homing operation: the motor drives each axis to move from its current position to the mechanical zero point (the initial reference position determined by the Hall sensor or encoder), completing the coordinate system homing; subsequently, the gimbal enters a three-axis stable working mode, continuously receiving gimbal control commands (PITCH CMD / ROLL CMD / YAW CMD) from the airborne edge computing unit 6 via the PWM interface and transmitting the current gimbal angle back via the GIMBAL FEEDBACK interface. If a motor drive abnormality occurs during the homing process of any axis of the gimbal (overcurrent, stall, or encoder reading failure), the gimbal fault protection is triggered: the current position of the gimbal is locked, camera image acquisition is prohibited, and a gimbal fault alarm is reported to the ground control terminal to prevent abnormal gimbal movement from damaging the camera.

[0150] Returning Readiness Status: After all self-test processes are passed, the airborne edge computing unit 6 generates a "System Ready" status message (including fields such as battery level, positioning status, communication status, model loading status, and gimbal status), which is then uploaded to the ground control terminal via 4G / 5G communication and data transmission module 4. After confirming the "UAV Ready" status, the operator can issue mission flight path instructions, and the UAV officially enters the working state.

[0151] (V) Overall System Workflow

[0152] like Figure 4As shown, the overall workflow of the ejector-type flame identification and early warning UAV system of this invention consists of nine main nodes, from A to I. These nodes are logically connected and proceed sequentially, forming a complete closed loop from reception and transportation to ground terminal response. Each node is described in detail below:

[0153] Phase A of Storage and Transportation: The system is in standby storage mode. The UAV is folded and inserted into the launch tube 11. The four flight rotor modules and landing gear are all in a compressed energy storage state due to radial constraint by the tube wall. The nitrogen storage tank 7 has completed pre-charging, and the pressure gauge 10 confirms that the pressure has reached the rated value (0.1~0.3 MPa). The release device 9 locks the UAV in the launch position. The operator carries the entire catapult-type flame identification and early warning UAV system to the target inspection area without needing to deploy any additional takeoff facilities. The entire system is in a safe standby state.

[0154] Ejection Trigger Phase B: The operator sends an ejection command to the system via a personal wireless controller, triggering the electromagnetic unlocking mechanism of the release hook 9. The hook opens, releasing the ejection piston 8. Simultaneously, the solenoid valve of the nitrogen storage tank 7 opens, rapidly filling the medium-filling area at the bottom of the cylinder with high-pressure nitrogen, driving the ejection piston 8 upward along the axial direction, propelling the drone at high speed along the central axis of the launch tube. The entire ejection triggering process, from the issuance of the command to the drone's release from the tube, takes an extremely short time (typically on the order of hundreds of milliseconds), achieving a rapid response of "command-based ejection".

[0155] Deployment Phase C: During the final stage of the UAV's ascent within the launch tube, the radial constraints on the tube wall are sequentially released, the stored torsion springs at the rotor module hinges release their elastic potential energy, and the four rotor modules simultaneously rotate from the first folded state to the first deployed state; the landing gear spring mechanism simultaneously drives the outriggers to open outwards. The moment the UAV leaves the launch tube opening, the deployment of the rotor modules and landing gear is essentially complete. For a detailed sequence of deployment actions, see [link to details]. Figure 5 B.

[0156] Phase D of takeoff and flight: After the flight rotor module and landing gear have deployed, the rotor assembly is in the horizontal working position, and the flight controller 3 completes its power-on self-test and initialization (see details). Figure 7 As shown in the process diagram, the flight controller 3 controls the rotation speed of each rotor motor according to the preset flight path or the remote control command issued in real time by the ground control terminal, driving the UAV to lift off and enter cruise flight mode. During this stage, the flight controller 3 continuously calculates the IMU attitude data and the GPS positioning module 1 position data to achieve attitude stabilization and flight path tracking control of the UAV.

[0157] Image Acquisition Phase E: After the UAV reaches cruising altitude, the gimbal 5 activates its three-axis stabilization function to compensate for pitch, roll, and yaw angle disturbances during flight in real time, ensuring that the camera's line of sight is stably pointed at the ground target area. The camera continuously acquires ground images at a set frame rate (e.g., 25 frames / second), generating a continuous real-time video data stream, which is transmitted via a high-speed interface to the image buffer queue of the onboard edge computing unit 6, providing a stable underlying visual input for subsequent flame feature recognition.

[0158] Flame Feature Recognition Stage F: The airborne edge computing unit 6 reads images frame by frame from the image buffer queue. After image preprocessing, it is input into the YOLO series flame target detection model built into the flame recognition module to perform inference and determine whether a suspected flame target is detected in the current frame (see detailed algorithm flow). Figure 6 If no flame candidate box that meets the confidence threshold is detected in the current frame, the result is "No", and the system returns to the image acquisition stage E node to continue acquiring the next frame and looping the inference; if a flame candidate box that meets the confidence threshold is detected, the result is "Yes", and the system enters the early warning feedback stage G node early warning feedback process.

[0159] In the early warning feedback phase G: the airborne edge computing unit 6 packages the detected flame target information (fire point geographic coordinates, confidence score, feature image, timestamp, etc.) into an early warning information package, which is then encrypted and uploaded to the ground control terminal in real time through the 4G / 5G communication and data transmission module 4, completing the one-way push of early warning information from "sensing end → communication end → ground end".

[0160] Multi-drone collaborative judgment phase H: After receiving the early warning information, the ground control terminal determines whether to activate the multi-drone collaborative patrol mode based on the current coverage of the mission area and the number of drones. If the judgment result is "no" (the current single drone can effectively cover the target area), it directly enters node I for ground terminal response; if the judgment result is "yes" (multi-drone collaboration is needed to expand the coverage area), the ground control terminal simultaneously dispatches multiple sets of catapult-type flame identification early warning drone systems, allocates the inspection routes of each drone according to the preset partition or dynamic supplementation algorithm, and enters node I after forming a full-domain flame early warning coverage network.

[0161] Ground Terminal Response Phase I: After receiving the early warning information, the ground control terminal triggers an audible and visual alarm, marks the fire location and confidence level information on the map interface in real time, and provides operators with information for situation assessment and emergency decision-making. Depending on the situation, it also pushes alarm notifications to relevant emergency response units, completing the closed-loop response of the entire system.

[0162] (VI) Detailed Process of Preparation Phase

[0163] like Figure 5As shown in Figure A, the ejection preparation phase is a crucial operational step for the system to enter the launch-ready state, and it is executed sequentially according to the following steps:

[0164] Step 1: Check the folding of the flight rotor modules: Before installation, the operator checks each of the four flight rotor modules to ensure they are in the first folded position, confirming that each module is tightly against the fuselage sidewall and that the hinge torsion springs are in a compressed, energy-stored state. Simultaneously, check that the landing gear is fully retracted, ensuring the overall aircraft profile is within the allowable installation range. Any flight rotor module that is not fully folded will result in installation failure and requires manual resetting before re-checking.

[0165] Step Two: Inserting the Launch Cannon: After confirming that the flight rotor module and landing gear are folded, the operator vertically inserts the UAV body along the central axis of the launch cannon body 11 until the supporting structure at the bottom of the UAV mates with the top surface of the ejection piston 8, and the entire UAV is placed into the designated storage position inside the cannon. During the insertion process, the guide ribs inside the cannon wall provide directional restraint to the fuselage, preventing the UAV from rotating or shifting during insertion.

[0166] Step 3: Unhooking device locking: The operator drives the unhooking device 9 to press the hook claw into the limiting groove on the ejection piston 8, confirming that the hook is fully engaged; check whether the safety rope is properly threaded, ensuring that the hook will not open unexpectedly even if the ejection command is issued without the safety rope being released, thus ensuring personnel safety.

[0167] Step 4: Connect the nitrogen gas line: After confirming the release device is locked, connect the gas line outlet of nitrogen storage tank 7 to the gas inlet of the bottom chamber of the cylinder via a quick-connect fitting. Confirm that the gas line is well sealed and leak-free. After the gas line connection is completed, open the shut-off valve of the storage tank to put the gas line into a ventilated and ready-to-charge state.

[0168] Step 5: Inflate to rated pressure: Inflate high-pressure nitrogen into nitrogen storage tank 7 using an external inflation pump or pre-inflation line, observing the pressure gauge 10 reading in real time during inflation. Stop inflation when the pressure gauge 10 reading rises to the rated inflation pressure range (e.g., 0.1~0.3 MPa), and record the final pressure value. If a gas leak is found during inflation (pressure cannot be maintained stably), check the gas line seal and eliminate the leak before re-inflating. If the pressure exceeds the rated upper limit, release the pressure to the normal range through the safety vent in a timely manner to ensure system safety.

[0169] Step 6 Pressure Verification: After inflation, let stand for 30 seconds and observe whether the reading of pressure gauge 10 remains stable (within the allowable error range of small fluctuations) to confirm that the air circuit is well sealed and the pressure tank is leak-free; if the pressure drops significantly (exceeding the allowable error), it indicates that there is an air circuit leak and it needs to be checked again.

[0170] Step Seven: System Ready: After the air pressure test is passed, the system enters standby mode. The operator can safely carry the launch tube to the designated launch point and wait for the launch opportunity. The entire ejection preparation phase is complete, and the system is in a ready state to be triggered at any time.

[0171] (VII) Detailed Procedures for Catapult and Flight Phases

[0172] like Figure 5 As shown in B, the ejection phase is the core transitional stage where the system changes from a stationary storage state to a free flight state. The entire process is completed automatically in a very short time, and the specific sequence of actions is as follows:

[0173] Ejection Trigger: After confirming that the safe area is clear, the operator issues an ejection trigger command. Two actions are executed simultaneously: First, the electromagnetic unlocking mechanism is energized, driving the hook release device 9 to quickly open and release the restraint on the drone body; second, the solenoid valve of the nitrogen storage tank 7 is opened, and high-pressure nitrogen is quickly filled into the bottom chamber of the cylinder, applying thrust to the ejection piston 8.

[0174] Release of the release mechanism: Driven by the electromagnetic mechanism, the release mechanism 9 opens outward and the hook disengages from the limiting groove on the ejection piston 8. The bottom of the UAV is subjected to the upward thrust of the ejection piston 8 and begins to accelerate upward along the central axis of the launch tube.

[0175] The ejector piston launches the drone: Under the continuous drive of high-pressure nitrogen, the ejector piston 8 moves upward with uniform acceleration, and continuously applies thrust to the bottom support structure of the drone, so that the drone can maintain axial linear acceleration and ejection under the guidance and constraint of the cylinder until the end of the piston stroke (the piston reaches the designed stroke stop position).

[0176] The constraints on the tube wall are gradually released: As the UAV moves upward along the tube axis and approaches the tube opening, the radial constraint surfaces of the inner wall of the launch tube on the flight rotor module and landing gear pass through the folding and storage areas of each flight rotor module in sequence. The radial constraint force is released in sequence as the UAV rises: the constraint surface closest to the tube opening is released first, causing the hinge torsion spring of the outermost flight rotor module to begin to partially release energy; as the UAV continues to rise, the constraint surfaces of the inner flight rotor module and landing gear are released in sequence, and the elastic potential energy of each mechanism is released in sequence.

[0177] The flight rotor modules unfold via spring release: The high-elasticity torsion springs at the hinges of each flight rotor module rapidly release their elastic potential energy the instant the constraint is released, driving each flight rotor module to quickly rotate and unfold around the hinge axis from its first folded state and lock in place. Because the hinge parameters and spring preload of the four flight rotor modules are identical, the four flight rotor modules unfold almost synchronously at the end of the ejection stroke. The four flight rotor modules rotate to a horizontal working position as the arm unfolds, with the rotor disk plane perpendicular to the direction of the fuselage's gravity, providing the mechanical conditions for normal rotor take-off.

[0178] Landing gear deployment: After the flight rotor module deploys, the landing gear spring mechanism also releases elastic potential energy as the constraints are released, driving each outrigger to deploy radially outward to a preset deployment angle, preparing for landing support. Since the outrigger deployment and the flight rotor module deployment occur sequentially, they do not interfere with each other.

[0179] Drone detaches from launch tube: Under the combined force of the ejection piston 8 and its own inertia, the drone completely detaches from the launch tube and enters free flight. At this point, the drone has already achieved a certain initial upward velocity, providing ample time for the rotors to take off.

[0180] Flight controller startup: After the rotors are in position, flight controller 3 immediately initiates the power-on self-test and initialization process (see details). Figure 7 The system completes initialization operations such as IMU calibration, GPS positioning, 4G / 5G link establishment, and loading of the onboard edge computing unit 6. After the self-test is completed, the flight controller 3 sends propeller start-up commands to each rotor motor, gradually increasing the rotor speed to the speed required for hovering, and the drone begins to climb steadily.

[0181] Stable flight phase: After the rotor speed reaches a stable hovering value, the flight controller 3 switches to attitude stabilization control mode. The IMU attitude calculation results are fed back in a closed loop to the attitude control algorithm built into the flight controller 3, driving the differential speed adjustment of each rotor motor to maintain stable attitude. The GPS positioning module 1 provides absolute position feedback to realize fixed-point hovering or route cruise control. The UAV officially enters the stable flight phase, and the catapult and deployment phases are all completed.

[0182] (VIII) Detailed Processing Flow of Airborne Edge Computing Unit 6

[0183] like Figure 6 As shown, the image preprocessing module and the flame recognition module are the core modules for realizing the fire perception closed loop of this invention. They run in real time on the airborne edge computing unit 6, and the entire processing flow is as follows:

[0184] Entering cruise mode: After the UAV completes power-on self-test and initialization, it enters cruise flight mode. The flight controller 3 transmits the real-time telemetry data stream to the airborne edge computing unit 6 through the telemetry serial port.

[0185] The camera acquires image frames: The onboard edge computing unit 6 continuously reads image frames from the pan-tilt-zoom (PTZ) camera via a high-speed video interface (MIPI CSI). The frame rate is typically set to 25 frames per second or higher (which can be dynamically adjusted based on inference computing power). Each image frame is stored in a memory buffer queue after being read, and is processed sequentially by the inference scheduler according to the first-in-first-out (FIFO) principle.

[0186] The inference scheduler retrieves a frame of raw image from the buffer queue and performs the following preprocessing operations on the image through the image preprocessing module:

[0187] Resolution normalization: The original image is cropped or scaled to the input size specified by the YOLO model (e.g., 640×640 pixels), while maintaining the aspect ratio. Missing areas are filled with gray borders (letterbox processing).

[0188] Color space conversion: Convert the image from BGR format to RGB format (to match the YOLO model input specification);

[0189] Pixel value normalization: Mapping pixel values ​​from integers of 0 to 255 to floating-point numbers of 0.0 to 1.0, which are then used as the model input tensor;

[0190] YOLO Model Inference: The preprocessed image tensor is input into the YOLO series flame target detection model (YOLOv8n), which has been fine-tuned specifically for flame scenes. Forward inference computation is performed under the inference acceleration framework (ONNX Runtime) of the airborne edge computing unit 6. The inference process includes three stages: backbone feature extraction, neck feature fusion, and head output. Finally, the bounding box coordinates (x, y, w, h), target category, and confidence score of all candidate boxes are output; the target category is flame.

[0191] Flame Detection Judgment: The inference output is post-processed using Non-Maximum Suppression (NMS) to filter redundant candidate boxes, retaining the candidate set with the highest confidence. The system checks each candidate box to see if its category is "flame" and if its confidence exceeds a preset detection threshold (e.g., 0.6). If no flame candidate box meets the conditions in the current frame, the judgment result is "No," and the system jumps to the "Camera Acquires Image Frame" step to continue processing the next frame and enters the next inference loop. If a candidate box meets the conditions, the judgment result is "Yes," and it is used as an initial candidate box to simultaneously trigger three serial processing links.

[0192] The first serial processing link is used to calculate the confidence score: For each initial candidate box, multi-dimensional features of the flame image are further extracted from the region of each candidate box, including the flame pixel area ratio, normalized average brightness of the region (HSV-V channel mean), flame color distribution histogram (HSV-H / S channel distribution features), and the aspect ratio and dimension of the bounding box. The above feature vectors are input into the comprehensive confidence score evaluation module to calculate a more reliable comprehensive confidence score (Composite Confidence Score) than the original YOLO confidence score, which is used to exclude false detections caused by non-flame interference sources such as sunlight reflection and bright objects. Initial candidate boxes with a comprehensive confidence score exceeding the second preset detection threshold (greater than 0.6) are used as second-screen candidate boxes.

[0193] The second serial processing link is used to obtain the geographic coordinates of the fire point: when there is a candidate box for screening, the airborne edge computing unit 6 reads the current geographic coordinates (longitude, latitude, altitude) of the UAV in real time from the NMEA data stream of the GPS positioning module 1 through the flight controller 3, and reads the current flight attitude angles (roll, pitch, yaw) from the flight controller 3 through the telemetry serial port (TELEMETRY); combined with the current pitch / roll / yaw angles of the gimbal 5 fed back by the GIMBAL FEEDBACK, the ground projection latitude and longitude coordinates of the current frame image in the geographic coordinate system are calculated using the camera intrinsic parameter matrix and coordinate transformation formula (multi-step coordinate transformation from image pixel coordinates → camera coordinate system → aircraft coordinate system → geographic coordinate system), i.e., the geographic coordinates of the fire point.

[0194] The third serial processing link is used to fuse timestamps and features: it records the UTC timestamp of the current image frame, and encapsulates the timestamp information with the bounding box coordinates of the second-screen candidate boxes, the comprehensive confidence score, and the geographic coordinates of the fire point in a structured manner to form a timestamped Flame Event Record for subsequent multi-frame temporal correlation and early warning information generation.

[0195] Packaged early warning information: The flame event record and the current telemetry data stream are encapsulated into a structured early warning information package.

[0196] Encrypted upload to ground control terminal: The early warning information packet is serialized into a binary data stream according to a predefined protocol format, encrypted with AES-256, and then uploaded to the ground control terminal in real time via 4G / 5G communication and the high-speed data channel of data transmission module 4. After receiving and decrypting the data, the ground control terminal triggers an audible and visual alarm and marks the fire location and confidence level information in real time on the map interface, completing the full-link early warning closed loop of "airborne perception → edge computing → communication transmission → ground response".

[0197] Furthermore, the experimental process of an optional embodiment of the present invention is described below:

[0198] 1. Experimental Environment and Configuration

[0199] This experiment aims to verify the performance of the Composite Confidence Evaluation Module (CCEM) proposed in this invention. Based on a YOLOv8 detection model (hereinafter referred to as the "baseline model") specifically fine-tuned for flame scenarios, the experiment verifies the effectiveness of the composite confidence module in suppressing false detections caused by non-flame interference sources (such as sunset and strong light reflection) by introducing a multi-dimensional image feature joint decision-making mechanism. The experimental environment and configuration are shown in Table 1.

[0200] Table 1 Experimental Environment and Configuration

[0201]

[0202] The test set consists of two types of images: positive samples are real flame images (file prefix: flame_), and negative samples are sunset and strong light scene images (interference images that are prone to false detection by YOLO). Both types of images are detected by the YOLOv8 baseline model generating fire-class candidate boxes to ensure the fairness of the comparative experiment.

[0203] 2. Definition of Evaluation Indicators

[0204] This experiment judges based on images (as long as there is one in each image). Each candidate box is determined as an "alarm" through the final decision. The following standard classification evaluation index is used, defined based on the confusion matrix:

[0205]

[0206]

[0207] in, (True positive) indicates that a real flame was correctly detected and alarmed; (False positive) indicates that a sunset / bright light image was mistakenly reported as a flame; (True negative) indicates that interfering images were correctly filtered; A false negative indicates that a real flame was missed.

[0208] 3. Overall performance comparison

[0209] This group of experiments comprehensively evaluated scheme A (Origin, pure YOLO) and scheme B (Proposed, YOLO + CCEM) on the test set. The experimental results are shown in Table 2 and... Figure 9 , Figure 10 .

[0210] Table 2 Comparison of Evaluation Indicators for the Two Schemes

[0211]

[0212] Experimental results show that in Scheme A, the baseline YOLO model successfully detected all real flame images (Recall=100%), but produced a high false positive rate (FPR=46.2%) due to interference from bright scenes such as sunlight and sunset. Scheme B, by introducing a comprehensive confidence assessment module, reduced the false positive rate to 0.0% while maintaining the same recall rate (FN=0, no missed detections), improving Precision from 71.4% to 100.0%, F1 Score from 83.3% to 100.0%, and Accuracy from 78.6% to 100.0%.

[0213] 4. Confusion Matrix Analysis

[0214] Figure 9 Comparison of various evaluation indicators between Scheme A and Scheme B (Origin represents Scheme A, Proposed represents Scheme B, and the arrows indicate the improvement). Figure 10 The confusion matrices of the two schemes on the test set are shown. In Scheme A, 6 out of 13 negative samples (sunset interference images) (46.2%) were misclassified as having fire by the baseline YOLO model. After introducing the comprehensive confidence module in Scheme B, all 6 false alarms were effectively filtered out. The detection results of positive samples are not affected. The above results demonstrate that the comprehensive confidence assessment module can accurately identify and filter non-flame interference candidate boxes without introducing any new missed detections, thus verifying the reliability of the module's decision-making mechanism.

[0215] 5. Confidence score distribution analysis

[0216] To further clarify the overall confidence level Compared to the original YOLO confidence level To assess the inter-class discrimination ability, this group of experiments statistically analyzed the score distribution of real flame and sunset interference images under two confidence measures. The results are as follows: Figure 11 , Figure 12 As shown, the sum of the mean values ​​is shown in Table 3.

[0217] Table 3 Comparison of mean confidence scores for the two types of images

[0218]

[0219] Overall confidence level Inter-class mean difference ( Compared to the original YOLO confidence level ( The 32.6% increase indicates that the multi-dimensional feature fusion mechanism effectively enhances the separability between real flames and non-flame interference sources.

[0220] Figure 11 middle (Left) and (Right) Score distribution on two types of images (red violin represents real flame, orange represents sunset interference, and red dashed line represents decision threshold 0.6). Figure 11 It also shows that, In terms of dimensions, the score distribution of flame candidate boxes in the sunset interference image significantly overlaps with that of the real flames (both distributions cross the 0.6 threshold), making it impossible for a single YOLO confidence score to effectively distinguish them; therefore, introducing... Afterwards, the overall score of the sunset interference candidate boxes was lower than the decision threshold of 0.6, achieving clear separation between classes.

[0221] Figure 12 All flame candidate boxes are in – Distribution in two-dimensional space (red dots represent real flames, orange rhombuses represent sunset interference, and red dashed lines represent...) Decision boundaries). Figure 12 This further demonstrates the introduction in a more intuitive way. The decision separation effect is as follows: all sunset interference candidate boxes are distributed below the decision threshold (FILTERED region), and all real flame candidate boxes are above the threshold (PASS region). There is no overlap between the two categories, and the decision boundary has high reliability.

[0222] 6. Multidimensional feature inter-class difference analysis

[0223] Overall confidence level It is derived by linear weighted fusion of four physical features. This experiment analyzes the mean values ​​of each feature dimension in real flame and sunset interference images to verify the discriminative effectiveness of each feature dimension. The results are as follows: Figure 13 As shown. Figure 13 Table 4 shows a comparison of the mean values ​​of the four feature dimensions in the images of real fire and sunset (negative) images.

[0224] Table 4 Comparison of Mean Values ​​for Each Dimension

[0225]

[0226] Experimental results show that the color distribution histogram features Flame pixel area ratio characteristics It has the strongest ability to distinguish between real flames and sunset interference, which is consistent with the physical prior of the unique color aggregation characteristics of real flames in the HSV color space.

[0227] 7. Comparison of Overall Performance Radar Chart and False Alarm Rate

[0228] Figure 14 The performance differences between the two schemes are presented in a comprehensive manner using radar charts and horizontal bar graphs. Figure 14 (Left) Comparison of false alarm rate (FPR) levels; (Right) Comprehensive radar chart of five-dimensional indicators (Origin is scheme A, Proposed is scheme B). As can be seen from the radar chart, scheme B (Proposed) achieved full marks (1.0) in all four dimensions of Precision, F1 Score, Accuracy and 1−FPR, and the Recall dimension was on par with scheme A. The overall performance polygon coverage area is significantly larger than that of scheme A, and the comprehensive performance advantage is outstanding.

[0229] 8. Candidate Box Level Decision Analysis

[0230] Figure 15 The pass / filter status of all flame-type candidate boxes after the decision of the comprehensive confidence module is visualized and analyzed, and each candidate box is displayed. and Relationship and Distribution of [something]. Figure 15 (Left) All candidate boxes for the "fire" category. – Scatter plot, green dots represent through boxes, gray crosses represent filtered boxes; Figure 15 (Right) Through the box and the filtered box Distribution histogram. (From...) Figure 15 (Left) As can be seen, the pass boxes are generally distributed above the decision threshold, while the filtered boxes are generally below the threshold. The two sets of candidate boxes... There is a clear separation in terms of dimensions; Figure 15 The distribution histogram (right) further confirms the effectiveness of the overall confidence score in distinguishing between the two types of candidate boxes, indicating that the threshold... The setting has good rationality in decision boundary.

[0231] In summary, the Comprehensive Confidence Assessment Module (CCEM) proposed in this invention possesses the following experimentally verified performance advantages:

[0232] (1) Effectively eliminate false positives due to non-flame interference: While maintaining Recall = 100% (no missed detections), the false positive rate (FPR) was reduced from 46.2% to 0.0%, the number of false positive images was reduced to zero, the precision was increased from 71.4% to 100.0%, the F1 score was increased from 83.3% to 100.0%, and the overall accuracy was increased from 78.6% to 100.0%.

[0233] (2) Enhance inter-class separability: comprehensive confidence Inter-class mean difference ( Compared to the original YOLO confidence level ( The performance improved by 32.6%, quantitatively verifying the superiority of the multidimensional feature fusion mechanism.

[0234] (3) The decision boundary is physically interpretable: features of each dimension ( , All of these (etc.) have clear physical meanings, are consistent with the spectral characteristics of flames, have adjustable fusion weights, and possess good engineering configurability.

[0235] (4) Only post-processing, without changing the basic detection structure: CCEM is added to the standard YOLO inference process in the form of a lightweight post-processing module. It does not modify the basic network structure and has a minimal increase in inference latency, making it suitable for real-time operation on the airborne edge computing unit described in this invention.

[0236] The test set for this experiment was specifically designed to include sunset images that strongly interfere with YOLO (the baseline YOLO model has a false detection rate of 46.2% for these images) in order to maximize the evaluation of the filtering capability of the comprehensive confidence module in extreme interference scenarios. The results obtained are basically consistent with the expected performance in actual deployment scenarios.

[0237] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A catapult-launched flame identification and early warning unmanned aerial vehicle (UAV) system, comprising a catapult UAV and a launch tube module, wherein the catapult UAV comprises a fuselage, a flight rotor module, a landing gear, a GPS positioning module (1), a battery (2), and a flight controller (3); the catapult UAV cooperates with the launch tube module to enable the flight rotor module of the catapult UAV to have a first deployed state and a first folded state, and to enable the landing gear of the catapult UAV to have a second deployed state and a second folded state; the fuselage of the catapult UAV is used to mount the GPS positioning module (1), the battery (2), and the flight controller (3), characterized in that, The ejected drone is also equipped with a 4G / 5G communication and data transmission module (4), a gimbal with a camera (5), and an airborne edge computing unit (6). The airborne edge computing unit (6) and the flight controller (3) communicate and control each other through multiple signal lines: First, the telemetry serial port is used for the airborne edge computing unit (6) to read the real-time telemetry data stream output by the flight controller (3) and to send flight control commands to the flight controller (3); Second, the control command line is used for the airborne edge computing unit (6) to send high-priority control commands directly to the flight controller (3); Third, the bus is used to transmit the status feedback of the control commands of the flight controller (3); The operating power of the flight controller (3) is provided by the airborne edge computing unit (6). The airborne edge computing unit (6) transmits high-bandwidth bidirectional data to the 4G / 5G communication and data transmission module (4) via a data port for uploading early warning information packets and receiving them from the ground control terminal; the modem control line is used for the airborne edge computing unit (6) to send AT commands to the 4G / 5G communication and data transmission module (4); the network status line is used for the 4G / 5G communication and data transmission module (4) to report the current network connection status to the airborne edge computing unit (6) in real time; the power supply for the 4G / 5G communication and data transmission module (4) is provided by the airborne edge computing unit (6); The airborne edge computing unit (6) is connected to the gimbal (5). The airborne edge computing unit (6) sends pitch axis control commands, roll axis control commands and yaw axis control commands to the gimbal (5) respectively. The gimbal (5) transmits the current three-axis angle values ​​of the gimbal back to the airborne edge computing unit (6) in real time for the calculation of the geographical coordinates of the fire point. The working power of the gimbal (5) is provided by an external power management circuit and is connected to the ground wire of the airborne edge computing unit (6). The camera installed on the gimbal (5) continuously collects ground images at a set frame rate, generates a continuous real-time video data stream, and transmits it to the image buffer queue of the airborne edge computing unit (6). The airborne edge computing unit (6) receives a stable operating voltage from the battery (2) after processing by the power management module; The airborne edge computing unit (6) is equipped with an image preprocessing module and a flame recognition module. The airborne edge computing unit (6) reads the original image frame by frame from the image buffer queue, preprocesses the original image through the image preprocessing module, and inputs it into the flame recognition module to perform inference and determine whether a suspected flame target is detected in the current frame image. When the detected target is a suspected flame target, a flame event record is generated. The airborne edge computing unit (6) uploads the generated flame event record and the current telemetry data stream as an early warning information packet to the ground control terminal in real time through the 4G / 5G communication and data transmission module (4).

2. The ejection-type flame identification and early warning UAV system according to claim 1, characterized in that, The image, after being processed by the image preprocessing module, is used to perform the following operations: Resolution normalization: Adjust the original image to the input size specified by the flame target detection model built into the flame recognition module, keep the aspect ratio unchanged, and fill in the missing areas; Color space conversion: Converting a resolution-normalized image from BGR format to RGB format; Pixel value normalization: The pixel value is mapped from an integer of 0 to 255 to a floating-point number of 0.0 to 1.0, which is used as the input tensor of the flame target detection model.

3. The ejection-type flame identification and early warning UAV system according to claim 1, characterized in that, The flame recognition module performs the reasoning as follows: This is used to input the original image into the flame target detection model built into the flame recognition module after preprocessing by the image preprocessing module, so as to output the bounding box coordinates, target category and confidence score of all candidate boxes; This is used to determine whether the confidence score of the candidate box exceeds the first preset detection threshold: if there is no flame candidate box that meets the conditions in the current frame, the determination result is negative, and inference continues to be performed on the next frame image; if there is a candidate box that meets the conditions, the determination result is positive, and the corresponding candidate box is used as the initial candidate box and three serial processing links are triggered simultaneously: The first serial processing link is used to calculate the confidence level: for each initial screening candidate box, the multi-dimensional features of the flame image are further extracted from the region of each initial screening candidate box and input into the comprehensive confidence level evaluation module to calculate the comprehensive confidence level of each initial screening candidate box; the initial screening candidate boxes with a comprehensive confidence level exceeding the second preset detection threshold are used as second screening candidate boxes, indicating that there is a suspected flame target; The second serial processing link is used to obtain the geographic coordinates of the fire point: under the condition that there are two candidate boxes, the airborne edge computing unit (6) reads the current geographic coordinates of the UAV from the NMEA data stream of the GPS positioning module (1) through the flight controller (3) and the airborne edge computing unit (6) reads the current flight attitude angle from the real-time telemetry data stream through the telemetry serial port of the flight controller (3). Combined with the current pitch, roll and heading angles of the gimbal fed back by the gimbal (5), the ground projection latitude and longitude coordinates of the current frame image in the geographic coordinate system are calculated using the camera intrinsic parameter matrix and coordinate transformation formula, that is, the geographic coordinates of the fire point. The third serial processing link is used to fuse timestamps and features: it records the UTC timestamp of the current image frame, and encapsulates the timestamp information with the bounding box coordinates of the second-screen candidate boxes, the comprehensive confidence score, and the geographical coordinates of the fire point in a structured manner to form a timestamped fire event record.

4. The ejection-type flame identification and early warning UAV system according to claim 3, characterized in that, The comprehensive confidence assessment module specifically includes: This is used to extract local image regions from the original images corresponding to each of the initial candidate frames, map the pixels within these local image regions from the original BGR format to the standard RGB color space, and record the maximum value among the normalized RGB components of the pixels within the local image regions as the maximum brightness value. The minimum value is recorded as the minimum brightness value. The difference in brightness is recorded as ; Used based on maximum brightness value Minimum brightness value The brightness difference is used to extract the hue component H, saturation component S and luminance component V of each pixel; This is used to extract multidimensional features from the converted local image region based on the hue component H, saturation component S, and luminance component V of each pixel, and to calculate the comprehensive confidence level of the multidimensional features through a weighted fusion algorithm.

5. The ejection-type flame identification and early warning UAV system according to claim 4, characterized in that, The process involves extracting multidimensional features from the converted local image region based on the hue component H, saturation component S, and luminance component V of each pixel. These features include: a first-dimensional feature representing the proportion of flame pixel area; a second-dimensional feature representing the normalized average luminance of the region; and a third-dimensional feature representing the geometric features of the bounding box.

6. The ejection-type flame identification and early warning UAV system according to claim 5, characterized in that, The first dimension feature is the flame pixel area ratio feature, specifically: In HSV space, the decision function for the flame pixel is set as follows: The specific calculation formula is as follows: ； Based on the decision function of the flame pixels, the feature value of the flame pixel area ratio is obtained. The calculation formula is: ； in, For any pixel within the initial candidate box. This is the set of pixels within the initial candidate bounding box region. This represents the total number of pixels within the initial candidate bounding box area.

7. The ejection-type flame identification and early warning UAV system according to claim 5, characterized in that, The normalized regional average brightness feature value The calculation formula is: ； in, Each pixel The hue component, saturation component, and brightness component.

8. The ejection-type flame identification and early warning UAV system according to claim 5, characterized in that, The third dimension feature is the flame color distribution histogram similarity feature, specifically: the hue component is uniformly divided into a first number of intervals, and the saturation component is uniformly divided into a second number of intervals, thereby constructing a feature with... The two-dimensional joint probability distribution histogram of the histogram feature bars is used to calculate the frequency of each pixel within the initial candidate box falling into each histogram interval, and then normalize the result to obtain the current flame color distribution histogram vector. Based on the flame color distribution histogram vector, obtain the flame color distribution histogram similarity feature value. The calculation formula is: ； in, This is a preset baseline flame color distribution histogram vector; This represents the total number of feature bars in the two-dimensional joint probability distribution histogram.

9. The ejection-type flame identification and early warning UAV system according to claim 5, characterized in that, The bounding box geometric feature score The calculation formula is: ； in, The aspect ratio of the initial candidate boxes; This is the preset average of the flame aspect ratio. This is the preset standard deviation of the flame aspect ratio.

10. The ejection-type flame identification and early warning UAV system according to claim 4, characterized in that, The comprehensive confidence score is calculated by a weighted fusion algorithm based on the multidimensional features. The specific fusion formula is as follows: ； In the formula, To assess the overall confidence level, The confidence score of the initial candidate boxes output by the model inference; , , , and These are the weight coefficients for the corresponding feature dimensions, and the sum of each weight coefficient is 1; This represents the feature value of the flame pixel area ratio. The normalized regional average brightness characteristic value. The similarity feature value of the flame color distribution histogram. Scoring is given for the geometric features of the bounding box.

Images