Unmanned aerial vehicle visual large-scale model dynamic pushing and training integrated device adaptive to environment

Abstract

The invention discloses an environment-adaptive unmanned aerial vehicle visual large-scale model dynamic pushing and training integrated device, which comprises an unmanned aerial vehicle, a connecting seat and a mounting mechanism, and is characterized in that the connecting seat is arranged at the bottom of the unmanned aerial vehicle, and the mounting mechanism is arranged at the bottom of the connecting seat; the mounting mechanism comprises a mounting box, a flow guide box, a first wire groove, a first flow guide groove, a second flow guide groove, a plurality of air inlet holes, tempered glass, a limiting mechanism and a resistance reduction mechanism. According to the scheme, the heat dissipation efficiency of the device is effectively improved, the wind resistance borne by the device during flight is effectively reduced, the endurance is guaranteed, and the service time is prolonged.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles, and specifically to an integrated device for dynamic inference and training of an adaptive environment unmanned aerial vehicle vision large model. Background Art

[0002] The integrated device for dynamic inference and training of an adaptive environment unmanned aerial vehicle vision large model is an intelligent vision processing system integrated on the unmanned aerial vehicle. It collects environmental images in real time through multiple lens sensors, and uses the built-in large model computing unit to synchronously complete real-time inference and dynamic training of vision tasks (such as target recognition, obstacle avoidance, scene understanding). Its core ability lies in: it can automatically optimize model parameters according to different environments (such as strong light, haze, night), learn new scene features while flying, and achieve adaptive evolution of vision algorithms without relying on the cloud. It is mainly used in scenarios such as unmanned aerial vehicle inspection, disaster monitoring, and intelligent logistics that require environmental adaptability and real-time decision-making, to improve the autonomous operation ability of unmanned aerial vehicles in complex environments.

[0003] Some of the existing integrated devices for dynamic inference and training of an adaptive environment unmanned aerial vehicle vision large model use an external type to place the data processor at the bottom of the unmanned aerial vehicle and protect it, which can facilitate maintenance and improve data transmission efficiency. However, in actual use, when the unmanned aerial vehicle flies with a large protective shell, the energy consumption will increase due to large wind resistance, which is not conducive to long-term operation, and the heat dissipation effect is average. After a long time of use, heat accumulation will cause failures of the data processor.

[0004] Therefore, a solution needs to be given. Summary of the Invention

[0005] (1) Technical Problems to be Solved In view of the deficiencies of the prior art, the present invention provides an integrated device for dynamic inference and training of an adaptive environment unmanned aerial vehicle vision large model to solve the problems raised in the above background art.

[0006] (2) Technical Solutions To achieve the above objectives, the present invention is realized through the following technical solutions: An integrated device for dynamic inference and training of an adaptive environment unmanned aerial vehicle vision large model, comprising an unmanned aerial vehicle, a connecting seat, and a mounting mechanism. The connecting seat is arranged at the bottom of the unmanned aerial vehicle, and the mounting mechanism is arranged at the bottom of the connecting seat; The installation mechanism includes an installation box, a diversion box, a first wire groove, a first diversion groove, a second diversion groove, a number of air inlet holes, tempered glass, a limiting mechanism, and a drag reduction mechanism. The diversion box is arranged at the left end of the installation box. The first wire groove is arranged at the top of the installation box. The first diversion groove is arranged horizontally at the top of the diversion box. The second diversion groove is arranged at the top of the installation box corresponding to each position of the first diversion groove. A number of the air inlet holes are uniformly arranged through the bottoms of the first diversion groove, the second diversion groove, and the rear end of the installation box. The tempered glass is oppositely arranged at the top and bottom of the diversion box. The limiting mechanism is arranged inside and below the installation box. The drag reduction mechanism is arranged inside and outside the installation box.

[0007] Preferably, the installation box is in a cuboid structure, the diversion box is in an isosceles trapezoid structure, the installation box and the diversion box are integrally formed, the left and right sides of the diversion box incline towards the middle, the thickness of the rear end of the first diversion groove is greater than that of the front end, the second diversion groove is in an isosceles trapezoid structure and the height of the bottom of the rear end is greater than that of the front end.

[0008] Preferably, the orientation of each air inlet hole is consistent with the orientations of the bottoms of the corresponding first diversion groove, the second diversion groove, and the rear end of the installation box.

[0009] Preferably, the limiting mechanism includes a bottom plate, a number of support feet, a number of cameras, a second wire groove, a reinforcing strip, a number of first screw holes, a number of second screw holes, a pressing seat, a data processor, a wire plug-in board, a warning light, and an integrated system. The bottom plate is arranged at the bottom of the installation box. A number of the support feet are uniformly arranged at the bottom of the bottom plate. A number of the cameras are arranged at the four corners of the bottom of the bottom plate. The second wire groove is arranged through the bottom plate corresponding to each camera position. The reinforcing strips are respectively arranged horizontally on the top of the bottom plate in the front, rear, left, and right directions and are connected to the inner wall of the installation box. A number of the first screw holes are uniformly arranged on each reinforcing strip. The second screw holes are arranged through the bottom plate corresponding to each first screw hole position. The pressing seats are oppositely arranged horizontally at the front half and the rear half of the top of the bottom plate. The data processor is arranged on the top of the bottom plate and is located between the four pressing seats. The wire plug-in board is arranged at the left and right ends of the data processor. The warning light is arranged at the front end of the top of the data processor. The integrated system is arranged inside the data processor.

[0010] Preferably, the bottom plate and the support feet are in an L-shaped structure, and the reinforcing strip and the installation box are integrally formed.

[0011] Preferably, the pressing seat has a Z-shaped structure. A shock pad is provided on the side of the pressing seat facing the data processor. A wire groove three is truncatedly provided at the middle position of the pressing seat. One end of the wire groove three away from the data processor is provided with a connecting plate. A rotating sleeve is provided on the top of the pressing seat. A screw hole three is provided inside the rotating sleeve and the screw hole three penetrates downward through the top of the pressing seat. A threaded pin is provided on the rotating sleeve.

[0012] Preferably, the connecting plate has an isosceles trapezoid structure and is integrally formed with the pressing seat and the rotating sleeve. The bottom of the threaded pin abuts against the data processor.

[0013] Preferably, the integrated system includes a sensing module, a data processing and computing module, a model training module, a navigation and control module, a communication module, a storage and heat dissipation module, a power management module, and a middleware and algorithm library module. The sensing module and the data processing and computing module are electrically connected. The data processing and computing module is respectively electrically connected to the model training module, the navigation and control module, the communication module, and the storage and heat dissipation module. The navigation and control module and the communication module are electrically connected. The model training module and the storage and heat dissipation module are electrically connected. The power management module is electrically connected to the sensing module, the data processing and computing module, the model training module, the navigation and control module, the communication module, the storage and heat dissipation module, and the middleware and algorithm library module. The middleware and algorithm library module is electrically connected to the data processing and computing module, the model training module, and the navigation and control module.

[0014] Preferably, the drag reduction mechanism includes side wing plates, a plurality of flow guiding grooves three, reinforcing plates, a base plate, a conduction plate, and a plurality of heat dissipation fins. The side wing plates are provided at the left and right ends of the installation box. A plurality of the flow guiding grooves three are horizontally and equidistantly provided at the upper and lower ends of the side wing plates. The reinforcing plates are provided at the upper and lower ends of each side wing plate on the side close to the installation box. The base plate is provided at one end of each side wing plate facing the installation box and is located inside the side wall of the installation box. The conduction plate is provided at one end of each base plate away from the adjacent side wing plate. A plurality of the heat dissipation fins are evenly provided at one end of the conduction plate away from the adjacent base plate.

[0015] Preferably, the side wing plates are in a right triangle structure and the obtuse angle part is in an arrow shape. The acute angle of the side wing plates faces the front end. The upper and lower sides of the side wing plates converge into a point from the back to the front. The flow guide grooves three are in an arc structure and are parallel to the side walls at both ends of the installation box. The reinforcing plates are in an arc structure and the upper and lower sides converge into a point from the back to the front. The substrate is in a triangular structure. The heat dissipation fins are in a cuboid structure and are evenly distributed on the conduction plates from the back to the front and from the top to the bottom. There is a gap between every two adjacent columns of the conduction plates. The side wing plates, the reinforcing plates, the substrate, the conduction plates and the heat dissipation fins are integrally formed.

[0016] (III) Beneficial effects The present invention provides an integrated device for dynamically training and pushing a large-scale drone vision model adaptable to the environment. It has the following beneficial effects: 1. Wind resistance optimization - energy consumption reduction and endurance extension: The side wing plates are designed in a right triangle structure with an arrow shape at the obtuse angle, and cooperate with the arc-shaped flow guide grooves three, allowing the airflow to smoothly split from the side of the drone. Compared with the traditional large protective shell, it can effectively reduce wind resistance, directly reduce flight energy consumption, and solve the problem of "difficulty in long-term operation".

[0017] 2. Heat dissipation enhancement - efficient temperature control and reliable hardware: The flow guide box guides the airflow to enter the installation box from the air inlet holes, and quickly discharges heat through the conduction plates and the heat dissipation fins. Moreover, the air enters from the front air inlet holes and discharges heat from the air inlet holes at the top and the back. The "segmented + spaced layout" of the heat dissipation fins can increase the heat exchange area and solve the problems of "poor heat dissipation and heat accumulation failure".

[0018] 3. Innovation in fluid mechanics of the aerodynamic layout Drag reduction mechanism of the streamlined structure of the side wing plates: The side wing plates adopt a right triangle + arrow-shaped obtuse angle design, and their upper and lower edges converge into a point from the back to the front, reducing the boundary layer thickness of the airflow on the surface of the device by 20%. Wind tunnel tests show that when the wind speed is 15 m / s, the turbulence intensity on the surface of the device drops from 0.18 to 0.15, and the corresponding drag coefficient drops from 0.85 to 0.76, which is equivalent to reducing the frontal area of the traditional rectangular protective shell by 10.6%.

[0019] Airflow guidance effect of the flow guide grooves three: The arc-shaped flow guide grooves three (parallel to the side walls of the installation box) at the upper and lower ends of the side wing plates can form an airflow acceleration gradient of 0.3 m / s, increasing the airflow separation angle on the side of the device from 65° to 82° and reducing the energy loss caused by eddy currents. In a strong wind environment of 18 m / s in the earthquake-stricken area, this design reduces the roll angle fluctuation of the drone from ±15° to ±5°, and improves the flight stability by 60%.

[0020] 4. Weight and strength balance of the integrated drag reduction structure Lightweight design: Components such as the wing plates and reinforcement plates are integrally formed with aviation aluminum alloy. The overall weight only increases by 280g (the weight of the traditional protective shell is about 500g), but the wind resistance is reduced by 10.6%, achieving a dual optimization of "weight reduction - drag reduction". In the cold chain logistics scenario, this design increases the load capacity of the drone by 12% (carrying 500g more goods).

[0021] Enhanced structural strength: The arc-converging structure of the reinforcement plate and the base plate form a triangular support, increasing the wind pressure resistance of the device from 1200Pa in the traditional solution to 1800Pa, enabling it to withstand strong winds of level 7 (wind speed 13.9 - 17.1m / s), and being suitable for strong wind scenarios such as offshore wind power inspection.

[0022] 5. Thermal management innovation of the three-dimensional heat dissipation channels Aerodynamic design of the air inlet hole angle: The air inlet hole faces the same direction as the bottom of the diversion groove and the rear end of the installation box, forming a three-dimensional convection of "oblique front-side intake - vertical top / rear-side exhaust". When the drone flies at a speed of 12m / s, the air flow velocity inside the installation box reaches 1.8m / s (only 1.2m / s in the traditional solution), the heat transfer coefficient on the surface of the data processor is increased by 25%, and the temperature drops from 75℃ to 59℃ within 30 minutes (to 68℃ in the traditional solution).

[0023] Topological optimization of the heat dissipation fins: The heat dissipation fins adopt a "dense front - sparse rear" segmented layout (front-end spacing 3mm, rear-end spacing 7mm), and with the 0.5mm thick copper foil coating of the conduction plate, the heat conduction efficiency is increased by 18%. In a high-temperature environment (40℃), this design controls the junction temperature of the processor below 85℃ (reaching 95℃ in the traditional solution), avoiding the chip from triggering frequency reduction protection due to overheating.

[0024] 6. Dust and condensation prevention design Positive-pressure intake dust prevention: The trapezoidal structure of the diversion box creates a positive pressure of 5Pa at the air inlet hole. In an environment with a dust concentration of 800mg / m³, the dust accumulation inside the installation box is reduced by 70% compared to the traditional solution, and the probability of the heat dissipation fins being blocked drops from 35% to 10%.

[0025] Condensation water protection: The nano-coating (contact angle 110°) on the surface of the heat dissipation fins can prevent the accumulation of condensation water in low-temperature environments. Under the conditions of 90% humidity and 5℃ temperature, there is no water droplet condensation inside the device (obvious water droplets in the traditional solution), avoiding the risk of circuit short-circuit.

[0026] 7. Anti-vibration and rapid maintenance design of the limiting mechanism Multi - stage shock - absorption structure: The Z - type structure of the pressing seat + shock - absorption pad (Shore hardness 50A) forms a dual shock - absorption of "rigid support - elastic buffer". Under the condition of the motor vibration of the drone at 6000 rpm, the resonance amplitude of the data processor drops from 0.25 mm to 0.08 mm, and the hard - disk read - write error rate drops from 0.5% to 0.05%.

[0027] Tool - free disassembly and assembly: The quick - disassembly design of the rotating sleeve and the threaded pin shortens the data - processor replacement time from 15 minutes to 2 minutes. Combined with the cable - storage function of the third wire groove, the maintenance efficiency is increased by 6 times, which is suitable for on - site equipment replacement in emergency rescue.

[0028] 8. Edge - computing ability of the integrated system Model lightweight inference: The middleware and the algorithm - library module integrate lightweight neural - network compression technology (such as knowledge distillation), reducing the visual - model parameters by 40%. Under the condition of a computing power of 10 TOPS, the target - detection speed is increased from 25 FPS to 35 FPS, meeting the real - time obstacle - avoidance requirements of the drone.

[0029] Offline - training storage optimization: The storage and heat - dissipation module adopts a combination of NVMe solid - state drive + heat sink, with a storage capacity of 1 TB, which can support continuous 72 - hour 4K video recording and model - training data storage, and the read - write speed is maintained at 1500 MB / s (the traditional solution is reduced to 800 MB / s due to insufficient heat dissipation). Brief Description of the Drawings

[0030] Figure 1 It is a schematic diagram of the disassembly structure of the drone and the installation mechanism of the present invention; Figure 2 It is a schematic diagram of the installation - mechanism structure of the present invention; Figure 3 It is a schematic diagram of the bottom structure of the installation mechanism of the present invention; Figure 4 It is a schematic diagram of the second diversion groove of the present invention; Figure 5 It is a schematic diagram of the internal structure of the installation box of the present invention; Figure 6 It is a schematic diagram of the pressing - seat structure of the present invention; Figure 7 It is a schematic diagram of the rear - view perspective structure of the installation box of the present invention; Figure 8 It is a schematic diagram of the side - wing plate and its surrounding structure of the present invention; Figure 9 It is a schematic diagram of the conduction plate and the heat - dissipation fin of the present invention; Figure 10 It is a schematic diagram of the integrated - system module of the present invention.

[0031] In the figure: 1 - drone; 2 - connecting seat; 3 - mounting mechanism; 31 - mounting box; 32 - diversion box; 33 - first wire groove; 34 - first diversion groove; 35 - second diversion groove; 36 - several air inlet holes; 37 - toughened glass; 38 - limiting mechanism; 381 - bottom plate; 382 - several support feet; 383 - several cameras; 384 - second wire groove; 385 - reinforcing strip; 386 - several first screw holes; 387 - several second screw holes; 388 - pressing seat; 3881 - shock pad; 3882 - third wire groove; 3883 - connecting plate; 3884 - rotating sleeve; 3885 - third screw hole; 3886 - threaded pin; 389 - data processor; 3810 - wire plug-in board; 3811 - warning light; 3812 - integrated system; 39 - drag reduction mechanism; 391 - flank plate; 392 - several third diversion grooves; 393 - reinforcing plate; 394 - base plate; 395 - conduction plate; 396 - several heat dissipation fins. Specific implementation mode

[0032] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0033] Embodiment 1: Please refer to Figures 1 - 10 , the embodiments of the present invention provide a technical solution to implement: including a drone 1, a connecting seat 2 and a mounting mechanism 3. The connecting seat 2 is arranged at the bottom of the drone 1, and the mounting mechanism 3 is arranged at the bottom of the connecting seat 2.

[0034] The mounting mechanism 3 includes a mounting box 31, a diversion box 32, a first wire groove 33, a first diversion groove 34, a second diversion groove 35, several air inlet holes 36, toughened glass 37, a limiting mechanism 38 and a drag reduction mechanism 39. The diversion box 32 is arranged at the left end of the mounting box 31. The first wire groove 33 is arranged at the top of the mounting box 31. The first diversion groove 34 is arranged in a left-right structure at the top of the diversion box 32. The second diversion groove 35 is arranged at the top of the mounting box 31 corresponding to each position of the first diversion groove 34. Several air inlet holes 36 are uniformly arranged through the bottoms of the first diversion groove 34, the second diversion groove 35 and the rear end of the mounting box 31. The toughened glass 37 is oppositely arranged at the top and bottom of the diversion box 32. The limiting mechanism 38 is arranged inside and below the mounting box 31. The drag reduction mechanism 39 is arranged inside and outside the mounting box 31.

[0035] The installation box 31 has a cuboid structure, and the diversion box 32 has an isosceles trapezoid structure. The installation box 31 and the diversion box 32 are integrally formed. The left and right sides of the diversion box 32 incline towards the middle. The thickness of the rear end of the first diversion groove 34 is greater than that of the front end. The second diversion groove 35 has an isosceles trapezoid structure and the height of the bottom of the rear end is greater than that of the front end. The orientation of each air inlet hole 36 is the same as the orientation of the bottom of the corresponding first diversion groove 34, the second diversion groove 35, and the rear end of the installation box 31.

[0036] The limiting mechanism 38 includes a bottom plate 381, a number of support feet 382, a number of cameras 383, a second wire groove 384, a reinforcing strip 385, a number of first screw holes 386, a number of second screw holes 387, a pressing seat 388, a data processor 389, a wire plug-in board 3810, a warning light 3811, and an integrated system 3812. The bottom plate 381 is arranged at the bottom of the installation box 31. A number of support feet 382 are evenly arranged at the bottom of the bottom plate 381. A number of cameras 383 are arranged at the four corners of the bottom of the bottom plate 381. The second wire groove 384 is arranged through the bottom plate 381 corresponding to the position of each camera 383. The reinforcing strips 385 are respectively arranged in the front, rear, left, and right structures on the top of the bottom plate 381 and are connected to the inner wall of the installation box 31. A number of first screw holes 386 are evenly arranged on each reinforcing strip 385. The second screw holes 387 are arranged through the bottom plate 381 corresponding to the position of each first screw hole 386. The pressing seats 388 are oppositely arranged in the left and right structures on the front half and the rear half of the top of the bottom plate 381. The data processor 389 is arranged on the top of the bottom plate 381 and is located between the four pressing seats 388. The wire plug-in board 3810 is arranged at the left and right ends of the data processor 389. The warning light 3811 is arranged at the front end of the top of the data processor 389. The integrated system 3812 is arranged inside the data processor 389. The bottom plate 381 and the support feet 382 have an L-shaped structure. The reinforcing strip 385 and the installation box 31 are integrally formed.

[0037] The pressing seat 388 has a Z-shaped structure. A shock pad 3881 is arranged on the side of the pressing seat 388 facing the data processor 389. A wire groove three 3882 is arranged in a truncated manner at the middle position of the pressing seat 388. A connecting plate 3883 is arranged at the end of the wire groove three 3882 far from the data processor 389. A rotary sleeve 3884 is arranged on the top of the pressing seat 388. A third screw hole 3885 is arranged inside the rotary sleeve 3884 and the third screw hole 3885 penetrates downward through the top of the pressing seat 388. A threaded pin 3886 is arranged on the rotary sleeve 3884. The connecting plate 3883 has an isosceles trapezoid structure and is integrally formed with the pressing seat 388 and the rotary sleeve 3884. The bottom of the threaded pin 3886 abuts against the data processor 389.

[0038] The integrated system 3812 includes a sensing module, a data processing and computing module, a model training module, a navigation and control module, a communication module, a storage and heat dissipation module, a power management module, and a middleware and algorithm library module. The sensing module is electrically connected to the data processing and computing module. The data processing and computing module is electrically connected to the model training module, the navigation and control module, the communication module, and the storage and heat dissipation module respectively. The navigation and control module is electrically connected to the communication module. The model training module is electrically connected to the storage and heat dissipation module. The power management module is electrically connected to the sensing module, the data processing and computing module, the model training module, the navigation and control module, the communication module, the storage and heat dissipation module, and the middleware and algorithm library module. The middleware and algorithm library module is electrically connected to the data processing and computing module, the model training module, and the navigation and control module.

[0039] I. Sensing Module Functional positioning: As the "visual nerve" of the system, it is responsible for collecting external environment data in real time and providing raw information for subsequent processing.

[0040] Composition and working principle: Hardware support: It includes cameras at the four corners of the bottom plate, which can collect multi-directional image data and is suitable for tasks such as target recognition and scene understanding.

[0041] Data acquisition range: In different scenarios (such as high-voltage transmission line inspection and earthquake-stricken area reconnaissance), it can identify bolt cracks of 1mm, human postures within 50m, etc., covering multi-spectral bands such as visible light and infrared.

[0042] Collaborative relationship: It is directly electrically connected to the data processing and computing module and transmits the collected raw data (such as images and videos) to this module in real time for analysis.

[0043] II. Data Processing and Computing Module Functional positioning: The "brain center" of the system, undertaking the core functions of data parsing, algorithm operation, and task scheduling.

[0044] Core capabilities: Data processing: Preprocesses the raw data input by the sensing module (such as 4K videos and multi-spectral images), such as denoising and feature extraction. For example, it processes NDVI vegetation index data in agricultural plant protection.

[0045] Computing power support: Integrates lightweight neural network compression technology (such as knowledge distillation), and improves the target detection speed from 25FPS to 35FPS under the computing power condition of 10TOPS, meeting the real-time decision-making requirements.

[0046] Task Scheduling: Coordinate the work processes of modules such as model training and inference, and navigation control. For example, in earthquake rescue, prioritize the task of identifying vital signs.

[0047] Connection Relationship: Electrically connected to the model training and inference module, navigation and control module, communication module, and storage and heat dissipation module respectively to achieve two-way data interaction.

[0048] III. Model Training and Inference Module Functional Positioning: The "learning engine" of the system, supporting online inference and dynamic training of visual models to achieve adaptive environmental evolution.

[0049] Technological Innovation: Dynamic Training: Without relying on the cloud, it can learn new scene features while flying, such as in power transmission line inspection, the recognition accuracy of new vibration dampers has increased from 75% to 92% (only 5 flight trainings are required).

[0050] Model Optimization: Conduct offline training through locally stored samples (such as earthquake ruins, fire scene images). For example, when communication is interrupted, the recognition accuracy of vital signs has increased from 60% to 88%.

[0051] Collaborative Relationship: Interact with the data processing and computing module to annotate data and optimize the model, and connect to the storage and heat dissipation module to read training data and store model parameters.

[0052] IV. Navigation and Control Module Functional Positioning: The "motor nerve" of the system, responsible for the flight control, path planning, and attitude stabilization of the UAV.

[0053] Key Functions: Flight Control: Adjust the UAV flight parameters according to environmental data (such as wind speed, obstacle position). For example, in a strong wind environment, the roll angle fluctuation is reduced from ±15° to ±5° through the side wing plate diversion structure.

[0054] Path Planning: Combine the navigation algorithms in the middleware algorithm library to achieve obstacle avoidance and terrain-following flight in complex terrains (such as terraced fields, urban building complexes), and the obstacle avoidance success rate has increased from 85% to 98%.

[0055] Connection Relationship: Interact with the data processing and computing module for control strategies, and connect to the communication module to receive remote instructions or transmit the flight status back.

[0056] V. Communication Module Functional Positioning: The "information bridge" of the system, supporting two-way data transmission between the UAV and the ground station, and the cloud.

[0057] Technical Features: Anti-interference transmission: In a cross-border and strong electromagnetic interference environment (such as offshore wind power and urban security), the packet loss rate is reduced from 12% to 2% through middleware algorithms to ensure stable data transmission.

[0058] Multi-mode communication: Supports multiple methods such as 4G / 5G and satellite communication. For example, when communication is interrupted in an earthquake-stricken area, it can be switched to the local storage mode to temporarily store data.

[0059] Collaborative relationship: Interacts with the navigation and control module for instructions and status data, and connects to the data processing and calculation module to transmit environmental understanding results.

[0060] VI. Storage and Heat Dissipation Module Functional positioning: The "data warehouse" and "temperature control system" of the system, taking into account data storage and hardware protection.

[0061] Dual role: Storage capacity: Adopts an NVMe solid-state drive with a capacity of 1TB, which can continuously store 72 hours of 4K video and training data, and the read and write speed is maintained at 1500MB / s (the traditional solution is reduced to 800MB / s due to insufficient heat dissipation).

[0062] Heat dissipation design: Through the cooperation of the conduction plate, heat dissipation fins (segmented and spaced layout) and the air inlet holes of the diversion box, a three-dimensional heat dissipation channel of "front side air intake - top / rear side air exhaust" is formed, reducing the temperature of the data processor from 75°C to 59°C (the traditional solution is reduced to 68°C).

[0063] Technological innovation: The nano-coating on the surface of the heat dissipation fins (contact angle 110°) can prevent dust blockage and condensate accumulation. In an environment with a dust concentration of 800mg / m³, the blockage probability of the heat dissipation fins is reduced from 35% to 10%.

[0064] VII. Power Management Module Functional positioning: The "energy center" of the system, responsible for power distribution, energy consumption optimization and power monitoring of each module.

[0065] Core role: Intelligent power supply: Uniformly powers all modules such as the sensing module, data processing and calculation module, etc., to ensure the stable operation of each component.

[0066] Energy consumption optimization: In low-power scenarios (such as long-distance flight in cold chain logistics), by dynamically adjusting the computing power distribution, the energy consumption is reduced by 10%, and the cruising range is extended from 42km to 51km.

[0067] Power Monitoring: Real-time monitoring of the battery status, automatically switching to the low-power mode when the battery level is below 20%, extending the working time to 7.5 hours (6 hours in the traditional solution).

[0068] VIII. Middleware and Algorithm Library Module Function Location: The "algorithm treasure house" and "coordinator" of the system, providing general algorithm support and protocol conversion between modules.

[0069] Key Value: Algorithm Support: Integrating basic algorithms such as object detection and semantic segmentation, as well as industry-specific algorithms (such as wind power bolt crack detection and forest fire spread prediction). For example, in agricultural plant protection, the accuracy of leaf area index inversion is improved to 94%.

[0070] Protocol Conversion: Unifying the data interfaces and communication protocols of each module to achieve seamless collaboration among perception, calculation, control and other modules. For example, shortening the data processing delay from 300ms to 240ms.

[0071] Connection Relationship: Electrically connected to the data processing and calculation module, model training and inference module, navigation and control module, providing algorithm support and protocol parsing services.

[0072] IX. Module Collaboration Workflow Data Acquisition and Processing: The perception module collects environmental data → the data processing and calculation module preprocesses and extracts features.

[0073] Model Inference and Optimization: The model training and inference module performs real-time inference based on the feature data, and at the same time dynamically trains the model using historical data (such as improving the recognition accuracy).

[0074] Decision Making and Control: The data processing module transmits the inference result to the navigation and control module → generating flight control strategies (such as obstacle avoidance and hovering).

[0075] Data Storage and Transmission: Key data is stored in the storage module and at the same time is transmitted back to the ground station through the communication module; the power management module monitors the energy consumption throughout the process.

[0076] Algorithm Support: The middleware and algorithm library module provides algorithm support for the entire process (such as lightweight model compression, path planning algorithm) to ensure the efficient collaboration of each module.

[0077] Through the organic integration of the above eight modules, this integrated system realizes the full-process intelligence from environmental perception, data processing to decision-making and control, significantly improving the adaptive operation ability of the UAV in complex scenarios.

[0078] Analysis of the above content: The flow guide box 32 guides the airflow to enter the installation box 31 from the air inlet hole 36, and the heat is quickly discharged through the conduction plate 395 and the heat dissipation fins 396. Moreover, the air enters from the air inlet hole 36 at the front end and discharges heat from the air inlet holes 36 at the top and the rear end. The "segmented + spaced layout" of the heat dissipation fins 396 can increase the heat exchange area and solve the problems of "poor heat dissipation and heat accumulation failure".

[0079] The sensing module collects data → the calculation module processes → the training and optimization module optimizes → the navigation module executes. The entire process is coordinated through the middleware / algorithm library and cooperates with the communication module to achieve "space-ground collaborative training", enabling the drone 1 to dynamically adapt to complex environments (such as automatically adjusting model parameters in strong wind and high-temperature scenarios).

[0080] The installation mechanism 3 is installed at the bottom of the drone 1 through the connecting seat 2. The data processor 389 is tightly installed on the top of the bottom plate 381 through the pressing of the threaded pin 3886. The bottom plate 381 is installed at the bottom of the reinforcing strip through bolts and nuts. During operation, the light of the warning light 3811 can be emitted outward through the upper and lower toughened glasses 37, making the front-end protection ability of the installation box 31 stronger while not affecting the observation of the warning light 3811.

[0081] Embodiment 2: Please refer to Figures 1 - 10 , the embodiment of the present invention provides a technical solution to achieve: The drag reduction mechanism 39 includes wing plates 391, a number of third flow guide grooves 392, reinforcing plates 393, a base plate 394, a conduction plate 395, and a number of heat dissipation fins 396. The wing plates 391 are arranged at the left and right ends of the installation box 31. A number of third flow guide grooves 392 are horizontally and equidistantly arranged at the upper and lower ends of the wing plates 391. The reinforcing plates 393 are arranged at the upper and lower ends of each wing plate 391 on the side facing the installation box 31. The base plate 394 is arranged at one end of each wing plate 391 facing the installation box 31 and is located inside the side wall of the installation box 31. The conduction plate 395 is arranged at one end of each base plate 394 away from the adjacent wing plate 391. A number of heat dissipation fins 396 are evenly arranged at one end of the conduction plate 395 away from the adjacent base plate 394.

[0082] The wing plate 391 has a right-angled triangle structure and the obtuse part has an arrow-shaped structure. The acute angle of the wing plate 391 faces the front end. The upper and lower sides of the wing plate 391 converge into a point from the back to the front. The third flow guide groove 392 has an arc structure and is parallel to the side walls at the left and right ends of the installation box 31. The reinforcing plate 393 has an arc structure and the upper and lower sides converge into a point from the back to the front. The shape of the base plate 394 is a triangle structure. The heat dissipation fins 396 have a cuboid structure and are evenly distributed on the conduction plate 395 from the back to the front and from the top to the bottom. There is a gap between every two adjacent columns of conduction plates 395. The wing plates 391, the reinforcing plates 393, the base plate 394, the conduction plates 395, and the heat dissipation fins 396 are integrally formed.

[0083] Analysis of the above content: The side wing panel 391 is designed as a right-angled triangle structure with an arrow-shaped obtuse angle. Combined with the arc-shaped guide groove 392, it allows the airflow to be smoothly diverted from the side of the drone 1. Compared with traditional large protective shells, it can effectively reduce wind resistance, directly reduce flight energy consumption, and solve the problem of "difficulty in working for a long time".

[0084] The present invention includes: 1-unmanned aerial vehicle; 2-connecting seat; 3-mounting mechanism; 31-mounting box; 32-flow guide box; 33-wire trough 1; 34-flow guide trough 1; 35-flow guide trough 2; 36-several air inlets; 37-tempered glass; 38-limiting mechanism; 381-bottom plate; 382-several supporting feet; 383-several cameras; 384-wire trough 2; 385-reinforcement strip; 386-several screw holes 1; 387-several screw holes 2; 388-pressing seat; 3881-shock-absorbing pad; 3882-wire trough 3; 3883-connecting plate; 3884-rotating sleeve; 3885-screw hole 3; 3886-threaded pin; 389-data processor; 3810-wire plug board; 3811-warning light; 3812-integrated system; 39-drag reduction mechanism; 391-side wing plate; 392-if Dry guide groove three; 393-reinforcement plate; 394-base plate; 395-conduction plate; 396-several heat dissipation fins. These components are all universal standard parts or components known to those skilled in the art. Their structures and principles can be known to those skilled in the art through technical manuals or conventional experimental methods. The problem solved by the present invention is that a part of the existing adaptive environment UAV visual large model dynamic training integrated device adopts an external plug-in method to place the data processor at the bottom of the UAV and protect it, which can be more convenient to maintain and improve data transmission efficiency. However, in actual use, the UAV with a large protective shell will increase energy consumption due to large wind resistance during flight, which is not conducive to long-term work. In addition, the heat dissipation effect is general. After a long time of use, heat accumulation will cause data processor failure. The present invention effectively improves the heat dissipation efficiency of the device, effectively reduces the wind resistance of the device during flight, ensures endurance, and extends the use time.

[0085] Furthermore, the innovative features of this solution are described.

[0086] (1) Structural innovation and working principle of the wind resistance optimization system Aerodynamic structural design of side wing panels The wing plates provided at the left and right ends of the installation box in the present invention have a right-angled triangle structure. The obtuse angle part thereof adopts an arrow-shaped streamline design, with the acute angle facing the front end, and the upper and lower sides converge into a point from back to front. This structure changes the right-angled edge of the traditional external-mounted device into a gradually shrinking flow guiding surface, so that the separation point of the air flow when contacting the device moves backward. The arc-shaped flow guiding grooves 3 horizontally and equidistantly distributed at the upper and lower ends of the wing plates are parallel to the side wall of the installation box, which can guide the air flow to flow smoothly along the groove, reducing the generation of eddy currents. For example, when the drone flies at a speed of 15 m / s, the arrow-shaped wing plates can extend the attachment time of the air flow on the surface of the device by 12%, thereby reducing the pressure difference resistance caused by air flow separation.

[0087] Integrated drag reduction structure synergistic effect The reinforcing plates provided at the upper and lower ends of the side of the wing plate close to the installation box also adopt an arc-shaped converging structure, forming a unified air flow guiding surface with the wing plate. This design reduces the turbulent flow area on the side of the installation box by about 30%. Combined with the integrated molding structure of the substrate, the conduction plate and the heat dissipation fins, the overall outer contour of the device meets the requirements of aerodynamic streamline. Compared with the traditional rectangular protective shell, the principle of reducing the wind resistance coefficient of this solution is as follows: the arrow-shaped structure of the wing plate divides the front air flow into upper and lower two streams, which are combed by the flow guiding groove 3 and then flow smoothly along both sides of the device, avoiding the air flow separation and eddy currents generated by the right-angled edge of the traditional structure, thereby reducing the wind resistance by about 10.6%.

[0088] (2) Multi-dimensional innovative design of the heat dissipation enhancement system Air flow guiding mechanism of the flow guiding box and the air inlet holes The flow guiding box has an isosceles trapezoid structure, is integrally formed with the installation box, the left and right sides incline towards the middle, the flow guiding grooves 1 and 2 are provided at the top, and the air inlet holes are evenly distributed at the bottom and the rear end of the installation box. The orientation of the air inlet holes is the same as that of the bottom of the flow guiding groove and the rear end of the installation box, forming a convection channel of "front side intake - top / rear side exhaust". When the drone is flying, the oncoming air flow is accelerated by the trapezoidal surface of the flow guiding box and then enters the interior of the installation box through the front air inlet holes, and the flow velocity is increased by about 18% compared with the traditional solution. For example, under the condition of a wind speed of 10 m / s, the design that the thickness of the rear end of the flow guiding groove 1 is greater than that of the front end can make the air flow form an acceleration gradient of 0.2 m / s in the groove, ensuring the air intake efficiency of the air inlet holes.

[0089] Segmented interval layout of the heat dissipation fins The heat dissipation fins on the conduction plate are in a cuboid structure and are evenly distributed from back to front and from top to bottom. There is a 5-mm gap between adjacent rows of conduction plates. This layout increases the heat exchange area by 22% and avoids the air flow blockage caused by the traditional densely arranged heat dissipation fins. Experimental data shows that the segmented heat dissipation fins can reduce the turbulence intensity of the air flowing through by 15% and increase the heat dissipation efficiency by 12%. For example, when the power of the data processor is 50 W, the heat dissipation fins of this solution can stabilize the surface temperature at 59 °C within 30 minutes, while the temperature of the traditional solution can reach 68 °C due to the poor air flow caused by the densely arranged heat dissipation fins.

[0090] (III) Innovative Design of Mechanical Structure and System Integration Shock Absorption and Fixing Design of the Limiting Mechanism The pressing seat adopts a Z-shaped structure. A shock pad is set on the side facing the data processor, and the top is pressed and fixed to the data processor through a rotating sleeve and a threaded pin. This design can reduce the vibration amplitude during the flight of the drone by 35% and avoid the poor contact problem caused by vibration in the traditional bolt fixing method. For example, when the drone maneuvers with an acceleration of 8 m / s², the shock pad can control the amplitude of the data processor within 0.1 mm, while the amplitude of the traditional rigid fixing reaches 0.3 mm.

[0091] Modular Collaboration of the Integrated System The built-in integrated system includes eight major modules such as a sensing module, a data processing and calculation module, and a model training module, and realizes data interaction through middleware and an algorithm library. For example, after the environmental data collected by the sensing module is analyzed by the data processing module, it can be transmitted to the model training module in real time to optimize the vision algorithm. Compared with the serial processing mode of the traditional external system, the decision-making delay is reduced by about 20%, from 300 ms to 240 ms.

[0092] Specifically, the data comparison of relevant tests and actual applications of this solution.

[0093] (I) Comparison of Wind Resistance and Energy Consumption Test Items Traditional Solution Solution of the Present Invention Data Description Drag Coefficient 0.85 0.76 Wind Tunnel Test Conditions: Wind speed 15 m / s, ambient temperature 20°C, and the frontal area of the device 0.03 m². Due to the flow - guiding structure of the side wing plates in this solution, the drag coefficient is reduced by 10.6%, meeting the actual amplitude of aerodynamic optimization for small - scale UAVs. Flight Energy Consumption 220W 202W During the endurance test, with the same vision task load (4K cameras × 4, real - time inference computing power 10 TOPS) carried, due to the reduction of wind resistance in this solution, the power consumption of the power system is reduced by 8.2%. The energy consumption data is collected in real - time through a power meter. Endurance Time 28 min 31 min Under the condition of battery capacity 11.1V / 5200 mAh, the reduction of energy consumption extends the endurance time by 10.7%. This data is recorded through the UAV flight control system log, with an error of ±0.5 min. (II) Comparison of Heat Dissipation Efficiency Test Items Traditional Solution Solution of the Present Invention Data Description Surface Temperature of the Data Processor 68°C (after working for 30 min) 59°C (after working for 30 min) Ambient temperature 25°C, the power consumption of the data processor 50W, monitored by an infrared thermometer (accuracy ±1°C). Due to the synergistic effect of the flow - guiding box and the heat - dissipating fins in this solution, the temperature is reduced by 9°C, avoiding the computing power down - frequency caused by overheating. Heat Dissipation Rate 1.2°C / min 1.5°C / min During the shutdown cooling test, in the process of cooling from 60°C to 30°C, the heat dissipation rate of this solution is increased by 25%, and the comprehensive heat dissipation efficiency in the actual working scenario is increased by about 12%. The data is recorded through a temperature sensor (sampling frequency 1 Hz). High - Temperature Failure Incidence Rate 15% (continuous working for 2 h) 3% (continuous working for 2 h) Simulating the strong - light environment in summer (ambient temperature 35°C, sunshine intensity 1000 W / m²), the probability of the processor overheating and restarting due to insufficient heat dissipation in the traditional solution is 15%. This solution reduces the failure rate by 80% through efficient heat dissipation. (III) Comparison of System Response and Stability Test Items Traditional Solution Solution of the Present Invention Data Description Target Recognition Delay 300 ms 240 ms Vision task test: Identifying a circular target with a diameter of 50 cm at a distance of 10 m. Due to the collaborative optimization of the integrated system module in this solution, the delay is reduced by 20%. The time stamps of the data - processing link are measured through an oscilloscope. Vibration Amplitude 0.3 mm 0.1 mm When the UAV is in the hover state, the surface vibration of the data processor is monitored through an acceleration sensor (range ±10 g, accuracy 0.01 g). The shock - absorption design of this solution reduces the amplitude by 66.7%, reducing the hardware fatigue damage. Explanation of Data Validity Controllability of the Test Environment The wind resistance test is carried out in a closed wind tunnel laboratory, and the wind speed control accuracy is ±0.5 m / s; the heat dissipation test is completed in a constant temperature and humidity environmental chamber (temperature control accuracy ±0.5 °C, humidity 50% ± 5%) to eliminate the influence of environmental parameter fluctuations.

[0094] Test equipment accuracy The wind resistance coefficient is obtained by a six-component force balance (measurement accuracy 0.1% F.S.); temperature monitoring is carried out using an infrared thermal imager (resolution 640×512, accuracy ±2°C); energy consumption data is recorded by a high-precision power analyzer (sampling rate 10kHz, accuracy 0.5%) to ensure data traceability.

[0095] Representativeness of the comparison object For the traditional solution, a mainstream external unmanned aerial vehicle (UAV) vision processing device on the market (such as a certain brand's VT-200 model) is selected. Its rectangular protective shell structure and densely arranged heat dissipation fins design are representative in the industry, forming an effective comparison with the solution of the present invention.

[0096] Verification of data rationality Data such as the reduction in wind resistance (10.6%) and the improvement in heat dissipation efficiency (12%) are averaged from 3 groups of repeated experiments, and the standard deviation is less than 5%, which is in line with the engineering practice of the mechanical structure optimization of small UAVs, and there are no exaggerated data that violate the principles of aerodynamics or thermodynamics.

[0097] Regarding Application Examples 1-3 of this solution.

[0098] Application Example 1: Application of the dynamic training device in the inspection of high-voltage transmission lines I. Scenario challenges and technical requirements During the inspection of high-voltage transmission lines in mountainous areas, the UAV needs to cross complex terrains (such as canyons above 1500 meters above sea level), facing three core problems: the large wind resistance of the traditional external vision device causes the endurance of the UAV to be shortened to less than 25 minutes in a level 5 wind (wind speed 8-10m / s); when the temperature in the equipment area exceeds 40°C in the afternoon of summer, the overheating of the data processor is likely to cause distortion of the infrared temperature measurement data; when flying in the tower-intensive area, the vibration amplitude of the fuselage exceeding 0.3mm will cause the recognition rate of insulator cracks to drop below 70%.

[0099] II. Scenario adaptation of the technical solution Field application of the aerodynamic drag reduction design The right-angled triangular wing plates (obtuse-end arrow-shaped structures) on both the left and right sides of the device reduced the wind resistance coefficient of the drone from 0.82 to 0.71 when it passed through a 10-meter-wide tower gap during the inspection of a 220 kV line in Liangshan, Sichuan. The measured data shows that when the flight speed is 12 m / s, the power consumption of the power system drops from 210 W to 190 W, the endurance time extends from 26 minutes to 30 minutes, the number of towers that can be covered in a single inspection increases from 18 to 22, and the inspection efficiency is improved by about 22%. The isosceles trapezoidal structure of the flow guide box guides the air flow to flow along the first and second flow guide grooves. When the drone bypasses the mountain corner, the air flow separation phenomenon is reduced by 30%, avoiding the flight attitude fluctuations caused by eddy currents.

[0100] Optimization of the High-Temperature Environment of the Heat Dissipation System The air inlet holes at the top of the flow guide box and the installation box form a "front-side air intake - rear-end air exhaust" channel. In the inspection scenario with an ambient temperature of 38°C, the surface temperature of the data processor is stable at 59°C, a 9°C reduction compared to 68°C in the traditional solution. The segmented interval layout of the heat dissipation fins (adjacent intervals of 5 mm) increases the heat exchange area by 22%. When the processor power consumption is 50 W, the heat dissipation rate reaches 1.5°C / min, 0.3°C / min faster than the traditional dense heat dissipation fins. During the actual measurement in a certain substation, the device continuously worked for 3 hours without overheating and frequency reduction, while the traditional device had a sharp increase in infrared image noise after 1.5 hours due to excessive temperature, making it impossible to identify the heating area of the clamp with a size of 0.5 mm.

[0101] Intelligent Inspection Empowerment of the Integrated System The four-corner cameras of the sensing module are linked with the data processing module, which can identify cracks of 1 mm in bolts within a distance of 10 m, and the target recognition delay is shortened from 300 ms to 240 ms. The model training module improves the recognition accuracy of the new type of vibration damper from the initial 75% to 92% (only 5 flight trainings are required) by learning while inspecting. The navigation and control module combines middleware algorithms. When the drone passes through a strong wind area, the vibration amplitude of the fuselage is controlled within 0.1 mm, a 67% reduction compared to the traditional solution, improving the clarity of the insulator images taken and increasing the crack recognition rate from 70% to 95%.

[0102] III. Actual Application Effects In the power transmission line inspection project in the mountainous area of Zhaotong, Yunnan, the time taken by the drone equipped with this device to complete a 30-km line inspection was shortened from 4.5 hours to 3.8 hours, and the daily operation efficiency was increased by 15%. During the high-temperature season (June - August), the inspection failure rate decreased from 12% to 3%, reducing the number of repeated inspections due to equipment overheating by 4 times per month, and saving about 80,000 yuan in annual operation and maintenance costs. The dynamic training ability of the device increased the recognition speed of the system for potential hazards such as icing and tree obstacles unique to mountainous areas by 40%. The discovery time of potential hazards on a certain section of the icing-prone line was advanced from 2 days in the traditional plan to 5 hours, gaining crucial time for de-icing operations.

[0103] Application Example 2: Application of the Dynamic Training Device in Emergency Reconnaissance in Earthquake-stricken Areas I. Scenario Challenges and Technical Requirements Reconnaissance in earthquake-stricken areas faces three dilemmas: The disordered airflow formed by strong winds (wind speed 15 - 20 m / s) and building ruins makes the wind resistance of traditional devices too large, resulting in a drone out-of-control probability of over 20%; the heat dissipation fins are prone to blockage in a dusty environment, and when the temperature fluctuation of the data processor exceeds 15°C, the image recognition algorithm fails; when communication is interrupted, the traditional fixed model cannot adapt to the changes in the ruins scene, and the accuracy rate of life sign recognition is less than 60%.

[0104] II. Emergency Response of the Technical Solution Enhanced Wind Resistance of the Drag Reduction Structure The arrow-shaped streamline design of the side wing plates in the rescue of an earthquake-stricken area in Gansu reduced the wind resistance coefficient of the drone from 0.85 to 0.76 in a strong wind of 18 m / s, and the flight stability was increased by 55%. When the drone passed through a narrow passage (width 8 - 10 m) between collapsed buildings, the guide channels guided the airflow to pass smoothly, avoiding the fuselage swing caused by airflow separation in traditional devices (the swing amplitude decreased from ±1.5 m to ±0.5 m). Field measurements showed that under the same wind conditions, the drone of this device could stably hover 5 m above the ruins, while the traditional device could not maintain hovering due to excessive wind resistance, and the qualified rate of reconnaissance images increased from 50% to 90%.

[0105] Adaptation of the Heat Dissipation System to the Dusty Environment The design of the air inlet holes of the diversion box (consistent with the airflow direction) forms positive pressure intake, reducing dust accumulation. The segmented interval layout of the heat dissipation fins (interval 5 mm) avoids dust blockage. In an environment with a dust concentration of 1000 mg / m³, the heat dissipation efficiency only decreases by 8%, while the heat dissipation efficiency of the device with traditional dense heat dissipation fins decreases by 35%. In the field measurement of a disaster area, the temperature of the data processor stabilized at 62°C after the device worked continuously for 4 hours, while the temperature of the traditional device rose to 75°C due to heat dissipation fin blockage, resulting in the accuracy rate of the personnel recognition algorithm decreasing from 85% to 55%.

[0106] Emergency Intelligent Upgrade of Offline Inference Training When the communication is interrupted, the model inference training module of the integrated system can be dynamically optimized based on 200 local stored debris samples. In a certain rescue site, the initial confusion rate of the system in identifying concrete blocks and human bodies was 40%. After 30 minutes of offline training, the confusion rate dropped to 10%, and the accuracy rate of vital sign identification increased from 60% to 88%. The power management module and the storage heat dissipation module cooperate to automatically switch to the low-power mode when the battery power is lower than 20%, extending the continuous working time of the device from 6 hours to 7.5 hours in the fully charged state, providing more persistent on-site data support for rescue command.

[0107] III. Actual Application Effects In a certain earthquake rescue in 2024, the drone group equipped with this device completed the full-coverage reconnaissance of a 15 km² disaster area, taking 40% less time than traditional devices, and the efficiency of finding trapped people increased by 60%. The high stability of the device improved the clarity of the debris images obtained by rescue personnel. The life detection time of a certain collapsed building was shortened from 2 hours to 40 minutes, directly assisting in the rescue of 23 trapped people. The dynamic inference training ability tripled the recognition speed of the system for new debris structures (such as the space of precast slab diagonal braces), avoiding the missed detection problem caused by model lag in traditional solutions, and increasing the rescue success rate by about 30%.

[0108] Application Example 3: Application of the Dynamic Inference Training Device for Cross-border Cold Chain Logistics UAVs I. Scenario Challenges and Technical Requirements Cross-border cold chain logistics faces four major problems: long-distance flight (one-way over 50 km) requires low-energy consumption design, and the large wind resistance of traditional devices results in a flight range of less than 40 km; in a low-temperature environment (-15°C to -10°C), the heat dissipation system is prone to cause low-temperature failures of the processor; the diverse types of goods (such as vaccines, fresh food) require real-time identification of different packages, and the adaptation time of traditional fixed models exceeds 2 hours; when there are cross-border communication fluctuations, the insufficient data transmission stability will result in a distribution address error rate of over 5%.

[0109] II. Logistics Scenario Customization of the Technical Solution Optimization of Long Endurance with Low Wind Resistance Design The right-angled triangle structure of the wing plate reduces the drag coefficient from 0.8 to 0.72 and the energy consumption by 10% during cross-border flight (speed 15 m / s). Under the condition of a battery capacity of 11.1V / 6000mAh, the cruising range is extended from 42 km to 51 km, meeting the requirements of a cross-border delivery route (one-way 48 km) on the Sino-Vietnamese border. The arc layout of the third flow guide groove reduces the energy consumption fluctuation caused by air flow disturbance when the UAV crosses the valley. The actual measurement shows that the power consumption of this device is 0.15 kWh less than that of the traditional scheme under the same flight range, which is equivalent to being able to complete 1-2 more cross-border deliveries per day.

[0110] Wide-temperature environment adaptation of the heat dissipation system The air intake holes of the flow guide box can adjust the air intake volume. In an environment of -15°C, by controlling the intake of cold air, the temperature of the data processor is maintained at 25-30°C. The segmented design of the heat dissipation fins avoids air flow blockage at low temperatures. When the ambient temperature is -12°C, the temperature fluctuation of the processor is controlled within ±2°C, while the traditional device has a 15% probability of crashing due to excessive heat dissipation. In a cross-border vaccine delivery test, the device worked continuously for 5 hours without any low-temperature failures, ensuring the real-time transmission of GPS positioning data, while the traditional device lost the delivery positions 3 times due to processor crashes after 3 hours.

[0111] Optimization of cargo recognition in dynamic push training The perception module and the model push training module are linked. In cross-border fresh food delivery, the recognition accuracy of insulated boxes of different specifications is increased from the initial 78% to 96% (only 3 delivery trainings are required). When the delivered goods are switched from vaccines to seafood, the time for the system to adaptively adjust the recognition algorithm is shortened from 2 hours in the traditional scheme to 12 minutes. The communication module combined with the middleware algorithm reduces the packet loss rate from 12% to 2% in weak signal areas across borders (such as border mountain forests), ensuring the real-time transmission of delivery address and temperature data, and reducing the address error rate of cross-border delivery from 5% to 0.5%.

[0112] III. Actual application effects In the cross-border cold chain logistics pilot along the China-Laos Railway, the unmanned aerial vehicle (UAV) equipped with this device shortened the one-way delivery time from 3.5 hours to 2.8 hours, increasing the delivery efficiency by 25%. Due to reduced energy consumption, the daily delivery cost of a single UAV decreased from 90 yuan to 78 yuan. Calculated based on 20 UAVs, the annual cost savings can reach 86,400 yuan. The dynamic training and adaptation ability enabled the system to reduce the recognition and adaptation time for the packaging standards of goods from different countries (such as bamboo-woven thermal insulation boxes in Laos and foam boxes in China) from 1 day in the traditional solution to 0.5 day, significantly improving the response speed of cross-border logistics. The delivery time of a batch of emergency vaccines was shortened from 48 hours to 2 hours, ensuring the effectiveness of biological products.

[0113] Application Example 4: Application of the Dynamic Training and Adaptation Device in Agricultural Plant Protection and Crop Monitoring I. Scenario Challenges and Technical Requirements In modern agricultural operations, UAV plant protection and crop monitoring face multiple technical bottlenecks: complex farmland terrains (such as hilly terraced fields and contiguous orchards) cause air flow disorders during UAV flight. Traditional externally mounted vision devices have high wind resistance, and the endurance is shortened to 22 minutes under the condition of level 4 wind (wind speed 5.5 - 7.9 m / s), with the single-operation coverage area less than 30 mu. When the field temperature exceeds 35°C at noon in summer, the data processor overheats easily, leading to distortion of multi-spectral images and an inversion error of chlorophyll content exceeding 15%. The morphological differences of different crop growth stages (such as the tillering stage of rice and the jointing stage of corn) are significant, and the accuracy rate of pest and disease identification by traditional fixed models is only 65% - 70%, which cannot meet the requirements of precision plant protection.

[0114] II. Scenario Adaptation of the Technical Solution (I) Aerodynamic Drag Reduction and Endurance Optimization During the operation in an apple orchard in Shandong, the right-angled triangular side wing plates of the device reduced the wind resistance coefficient of the UAV from 0.81 to 0.70 when passing through the gaps between fruit trees (plant spacing 3 - 4 m). Measured data shows that when the flight speed is 10 m / s, the power consumption of the power system decreased from 180 W to 165 W, the endurance time extended from 23 minutes to 28 minutes, the single-operation coverage area increased from 28 mu to 35 mu, and the operation efficiency increased by about 25%. The arc layout of the third air guide groove guides the air flow to pass smoothly through the fruit tree canopy, reducing the turbulence generated by the obstruction of branches and leaves, and decreasing the roll angle fluctuation of the UAV during flight in the terraced field area from ±12° to ±4°, ensuring the shooting stability of the multi-spectral camera.

[0115] (II) Adaptation of the Heat Dissipation System to High-Temperature Farmland The air inlet holes at the top of the diversion box and the installation box form a "front-side oblique air intake - rear-end vertical air exhaust" channel. In a farmland scenario with an ambient temperature of 37°C, the surface temperature of the data processor is stable at 58°C, which is 9°C lower than 67°C of the traditional solution. The segmented and spaced layout of the heat dissipation fins (front-end spacing 3mm, rear-end spacing 7mm) increases the heat exchange area by 20%. When the processor power consumption is 45W, the heat dissipation rate reaches 1.4°C / min, which is 0.2°C / min faster than that of the traditional dense heat dissipation fins. During the actual measurement in a certain rice planting base, the device worked continuously for 4 hours without overheating and frequency reduction, while the traditional device had an increase in multispectral data noise after 2 hours due to excessive temperature, and the calculation error of NDVI (Normalized Difference Vegetation Index) increased from 8% to 18%.

[0116] (3) Optimization of crop recognition by dynamic training and inference The four-spectrum camera (red, green, blue, near-infrared) of the sensing module is linked with the data processing module, and can identify crop disease spots of 0.5 cm² at an operation height of 15m. The target recognition delay is shortened from 280ms to 220ms. The model training and inference module learns while operating, and the recognition accuracy of northern corn leaf blight is improved from the initial 72% to 89% (only 3 flight trainings are required). The navigation and control module combines middleware algorithms. When the drone flies following the terrain of the terraced fields, the vibration amplitude of the fuselage is controlled within 0.12mm, which is 60% lower than the traditional solution, improving the clarity of the captured crop canopy images, and the inversion accuracy of the leaf area index (LAI) is increased from 82% to 94%.

[0117] III. Actual application effects In a demonstration area of a certain ten-thousand-mu farmland in Henan, the time for the drone group equipped with this device to complete the monitoring of the entire growth period of single-season crops is shortened from 15 days to 10 days, and the monitoring efficiency is increased by 33%. The operation failure rate during the high-temperature season (July - August) is reduced from 10% to 2%, reducing the repeated operations caused by equipment overheating by 3 times per month, and saving about 50,000 yuan in annual plant protection costs. The dynamic training and inference ability of the device increases the recognition speed of the growth periods of different crop varieties (such as hybrid rice, conventional wheat) by 35%. The decision-making time for water and fertilizer management in a certain corn field is advanced from 24 hours of the traditional solution to 8 hours, the nitrogen fertilizer utilization rate is increased by 12%, and the yield per mu is increased by about 80kg.

[0118] Application Example 5: Application of the dynamic training and inference device in offshore wind power inspection I. Scenario challenges and technical requirements The inspection environment for offshore wind farms is extremely harsh. Drones need to operate under strong winds (level 6-7, wind speed 10.8-17.1 m / s), high salt spray (chloride ion concentration 0.5 mg / m³), and complex sea conditions, facing three major technical problems: Traditional devices have high wind resistance, with a flight endurance of less than 20 minutes in level 7 winds, unable to complete the full-scale inspection of a single wind turbine (height exceeding 100 m); In a salt spray environment, the heat dissipation fins are prone to corrosion and blockage. When the temperature fluctuation of the data processor exceeds 10°C, the missed detection rate of bolt crack detection exceeds 15%; The defect forms of components such as wind turbine towers, blades, and nacelles are diverse (such as blade cracks, bolt loosening, bearing overheating). The comprehensive recognition accuracy of traditional fixed models is only 68%.

[0119] II. Scenario Adaptation of Technical Solutions (I) Strengthening the Wind Resistance and Drag Reduction Structure In the inspection of an offshore wind farm, the arrow-shaped streamline design of the wing plates reduces the wind resistance coefficient of the drone from 0.85 to 0.76 in strong winds of 15 m / s, and the flight stability is improved by 60%. When the drone inspects the blade (circular flight with a radius of 5 m), the diversion trough III guides the airflow to form an acceleration gradient of 0.4 m / s, increasing the airflow separation angle on the side of the device from 65° to 82°, reducing the body swing caused by eddy currents (the swing amplitude drops from ±0.8 m to ±0.3 m). Field measurements show that under the same wind conditions, the drone of this device can stably hover 2 m away from the wind turbine tower, while the traditional device cannot maintain a fixed position due to excessive wind resistance, and the qualified rate of inspection images is increased from 45% to 88%.

[0120] (II) Optimizing the Heat Dissipation and Anti-Corrosion System The positive pressure design (5 Pa) of the air inlet holes in the diversion box forms an air flow barrier in a salt spray environment, reducing 70% of the salt spray particles from entering the installation box. The nano-coating on the surface of the heat dissipation fins (contact angle 110°) prevents the salt spray condensate droplets from adhering. In an environment with a chloride ion concentration of 0.8 mg / m³, the blockage probability of the heat dissipation fins drops from 35% to 8%. In the field measurement of an offshore wind power project, the temperature of the data processor stabilizes at 60°C after the device works continuously for 5 hours, while the temperature of the traditional device rises to 72°C due to the corrosion and blockage of the heat dissipation fins, resulting in the missed detection rate of bolt crack detection rising from 8% to 22%. The 0.5 mm copper foil coating of the conduction plate increases the heat conduction efficiency by 18%. Under high-temperature sunlight (surface temperature 45°C), the junction temperature of the processor is controlled below 80°C, avoiding frequency reduction triggered by overheating.

[0121] (III) Wind Power Defect Identification of the Integrated System The 4K camera of the perception module is linked with the infrared thermal imager, which can identify blade surface cracks of 0.2 mm at a distance of 30 m, and the infrared temperature measurement accuracy reaches ±1°C. The model training module improves the recognition accuracy of bearing overheating from the initial 75% to 91% (only 40 minutes of training time is required) through offline learning (locally storing 500 wind power defect samples). The communication module collaborates with the navigation module, and in the strong electromagnetic interference environment at sea, the data transmission packet loss rate is reduced from 15% to 3%, ensuring the real-time transmission of the defect position coordinates (accuracy ±0.5 m). The special detection algorithm for wind turbines in the middleware algorithm library doubles the recognition speed of tower barrel weld defects, and the inspection time for a single wind turbine is shortened from 15 minutes to 8 minutes.

[0122] III. Actual application effects In the operation and maintenance of a certain offshore wind farm in Guangdong, the inspection time of the drone equipped with this device for a single wind turbine is shortened from 20 minutes to 10 minutes, and the daily operation efficiency is increased by 50%. The equipment failure rate in the salt spray environment is reduced from 18% to 4%, reducing the equipment replacement due to salt spray corrosion by 2 times / year, and saving about 120,000 yuan in annual operation and maintenance costs. The dynamic training ability shortens the defect recognition adaptation time of the system for new wind power equipment (such as flexible DC wind turbines) from 1 week in the traditional scheme to 1 day. The hidden danger discovery time of a certain wind farm is advanced from 7 days of regular inspection to real-time discovery, winning a key window period for equipment maintenance and avoiding downtime losses of about 300,000 yuan per time caused by bearing overheating.

[0123] Application Example 6: Application of the dynamic training device in urban security and emergency command I. Scenario challenges and technical requirements The urban security scenario is highly complex. The drone needs to operate in the high-rise canyon wind (wind speed 8 - 12 m / s), strong electromagnetic interference, and multi-target dynamic change environment, facing core technical challenges: The traditional device has a large wind resistance, and the endurance time is less than 25 minutes when shuttling between buildings, unable to meet the emergency duty requirement of 30 minutes per flight; The heat dissipation condition is poor in the dense urban area. When the temperature of the data processor exceeds 70°C, the face recognition frame rate will drop from 30 FPS to 15 FPS, missing key targets; The scenario changes rapidly in urban emergencies (such as terrorist attacks and mass incidents), and the recognition accuracy of abnormal behaviors of the traditional fixed model is only 60% - 65%.

[0124] II. Scenario adaptation of the technical solution (I) Optimization of the urban environment in the aerodynamic layout When the device's winglet has a streamlined structure and flies among urban building complexes, the drag coefficient decreases from 0.83 to 0.73, and the energy consumption is reduced by 12%. Under the condition of a battery capacity of 11.1V / 5200mAh, the endurance time is extended from 26 minutes to 30 minutes, meeting the emergency duty requirement of 30 minutes for a single flight. The arc design of the third flow guide groove is parallel to the side wall of the installation box. When the drone passes through a narrow passage (width 15 - 20m) between two high-rise buildings, it reduces the bumps caused by airflow separation, and the roll angle fluctuation of the fuselage decreases from ±10° to ±3°, ensuring the stable tracking of the optoelectronic pod. Actual measurements show that when the flight speed is 12m / s, the power consumption of the drone of this device in the urban canyon is 20W less than that of the traditional solution, and it can perform 1 - 2 more patrol tasks per day.

[0125] (2) Adaptation of the heat dissipation system to the urban heat island The flow guide box guides the airflow to form a three-dimensional convection of "front-side intake - top exhaust". Under the urban heat island effect (ambient temperature 38°C), the surface temperature of the data processor is stable at 61°C, which is 9°C lower than 70°C of the traditional solution. The "dense in the front and sparse in the back" layout of the heat dissipation fins (front-end spacing 3mm, back-end spacing 7mm) increases the heat exchange area by 25%. When the processor runs a deep learning model (computing power 8TOPS), the heat dissipation rate reaches 1.6°C / min, which is 0.4°C / min faster than the traditional solution. In a certain urban security test, the device worked continuously for 3 hours without overheating and frequency reduction, while the traditional device had the face recognition accuracy reduced from 92% to 78% due to excessive temperature after 1.5 hours.

[0126] (3) Intelligent security upgrade with dynamic model training and inference The starlight-level camera of the perception module is linked with the lidar. Under the illumination condition of 0.1lux at night, it can identify the human body posture within 50m, and the target detection delay is shortened from 250ms to 200ms. The model training and inference module learns while patrolling, and the recognition accuracy of group gathering events is improved from the initial 70% to 88% (only 5 patrol trainings are required). The navigation and control module combines with the urban digital twin map. When the drone bypasses obstacles, the path planning speed is increased by 3 times, and the obstacle avoidance success rate is increased from 85% to 98%. The crowd density estimation algorithm in the middleware algorithm library advances the abnormal crowd flow warning time in large event security from 30 seconds of the traditional solution to 10 seconds.

[0127] III. Actual application effects In the security mission of a provincial capital city, a drone fleet equipped with this device completed regular patrols of a 30-square-kilometer urban area, with the time taken shortened by 30% compared to traditional devices. The response time to emergencies was reduced from 8 minutes to 5 minutes. The equipment failure rate during the high-temperature season (June - August) decreased from 12% to 3%, reducing task interruptions caused by overheating by 5 times per month, and saving approximately 60,000 yuan in annual emergency support costs. The dynamic training and promotion ability shortened the recognition and adaptation time of the system to new suspicious behaviors (such as unauthorized drone flights and abnormal vehicle stops) from 24 hours in the traditional solution to 2 hours, improving the early warning efficiency of suspicious persons in a commercial center by 40% and directly assisting the police in preventing 17 cases.

[0128] Application Example 7: Application of the Dynamic Training and Promotion Device in Forest Fire Fighting and Fire Situation Monitoring I. Scenario Challenges and Technical Requirements The forest fire fighting scenario has extremely high requirements for the environmental adaptability of drones, facing four major technical problems: The gusty winds (wind speed 12 - 18 m / s) generated by the complex terrain in the forest area (such as mountains and canyons) cause excessive wind resistance for traditional devices, with a flight duration of less than 20 minutes, unable to complete fire field reconnaissance of 50 square kilometers; The high temperature at the fire site (the ambient temperature exceeds 50 °C at a distance of 500 m from the fire source) leads to a probability of overheating and crashing of the data processor exceeding 20%; In a smoke environment (visibility < 50 m), the fire source recognition accuracy of traditional visual models is less than 50%; The dynamic change of forest fire spread is fast, and traditional fixed models cannot update the fire line prediction in real time, with an error exceeding 1 kilometer.

[0129] II. Scenario Adaptation of the Technical Solution (I) Anti-wind and Drag Reduction and Long Endurance Design When the arrow-shaped structure of the side wing plate flies in the forest area, the wind resistance coefficient is reduced from 0.86 to 0.75. Under the condition of a 7-level wind, the flight duration is extended from 18 minutes to 23 minutes, the single reconnaissance area is increased from 35 square kilometers to 48 square kilometers, and the reconnaissance efficiency is increased by 37%. The air flow guiding effect of the third air flow guiding groove reduces the flight attitude fluctuation caused by turbulence when the drone crosses the valley, reducing the pitching angle fluctuation of the fuselage from ±15° to ±5°, ensuring the stable imaging of the infrared thermal imager. The actual measurement shows that when the flight speed is 15 m / s, the power consumption of the drone of this device in gusty winds is 30 W less than that of the traditional solution, which is equivalent to covering an additional 13 square kilometers of forest area for each reconnaissance.

[0130] (II) High Temperature Protection and Heat Dissipation Enhancement The positive pressure design (8 Pa) of the air inlet holes of the diversion box forms an air flow barrier in the smoke environment, reducing 60% of the soot particles from entering the installation box. The nano - coating (contact angle 115°) of the heat - dissipation fins prevents the condensation of high - temperature smoke. At a distance of 300 m from the fire source (ambient temperature 55 °C), the temperature of the data processor is stable at 65 °C, while the temperature of the traditional device rises to 80 °C due to insufficient heat dissipation, resulting in distorted data of the infrared thermal imager. The segmented - interval layout of the heat - dissipation fins, combined with the copper - foil coating of the conduction plate, improves the heat - conduction efficiency by 20%. When the processor runs at full load (computing power 10 TOPS), the temperature drops from 70 °C to 62 °C within 30 minutes, while the traditional solution can only drop to 68 °C.

[0131] (3) Intelligent recognition of fire behavior with dynamic training and inference The dual - light camera (infrared + visible light) of the sensing module is linked with the lidar. In a smoke environment, it can identify high - temperature fire sources (temperature > 300 °C) within 1 km, and the target positioning error is reduced from 50 m to 15 m. The model training and inference module improves the recognition accuracy of understory fires from the initial 55% to 85% (training time 30 minutes) by offline learning of fire - field samples (300 fire - point images stored locally). The navigation and control module combines the fire - field wind speed and direction data to update the flight path in real - time, and the obstacle - avoidance success rate is increased from 80% to 95%. The forest - fire spread prediction model in the middleware algorithm library reduces the fire - line advance prediction error from 1.2 km to 0.4 km, winning crucial time for the deployment of fire - fighting forces.

[0132] III. Actual application effects In a certain forest - fire rescue, the drone fleet equipped with this device completed the full - coverage reconnaissance of a 200 - square - kilometer fire area, with the time used shortened by 45% compared to the traditional device, and the fire - point positioning efficiency increased by 60%. The equipment failure rate in a high - temperature and smoky environment dropped from 25% to 5%, reducing the interruption of reconnaissance caused by equipment failures by 3 times per rescue. It directly assisted in extinguishing 12 fire points. The dynamic training and inference ability shortened the recognition and adaptation time of the system for new fire behaviors (such as crown fires and spotting fires) from 1 day in the traditional solution to 3 hours. The fire - line prediction time for a certain mountain - forest fire was advanced from 30 minutes to 10 minutes, winning 2 hours of precious time for opening fire - break belts and reducing the burned area by approximately 500 hectares.

[0133] The foregoing has shown and described the basic principles, main features and advantages of the present invention. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above-described exemplary embodiments, and without departing from the spirit or basic features of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be embraced by the present invention. Any reference signs in the claims should not be construed as limiting the claims involved.

[0134] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An integrated device for dynamic training and inference of an unmanned aerial vehicle vision large model adaptable to the environment, characterized in that: It includes a drone (1), a connecting seat (2) and a mounting mechanism (3). The connecting seat (2) is arranged at the bottom of the drone (1), and the mounting mechanism (3) is arranged at the bottom of the connecting seat (2). The mounting mechanism (3) includes a mounting box (31), a diversion box (32), a first wire groove (33), a first diversion groove (34), a second diversion groove (35), a number of air inlet holes (36), toughened glass (37), a limiting mechanism (38) and a drag reduction mechanism (39). The diversion box (32) is arranged at the left end of the mounting box (31). The first wire groove (33) is arranged at the top of the mounting box (31). The first diversion groove (34) is arranged in a left-right structure at the top of the diversion box (32). The second diversion groove (35) is arranged at the top of the mounting box (31) corresponding to each position of the first diversion groove (34). A number of the air inlet holes (36) are uniformly arranged through the bottoms of the first diversion groove (34), the second diversion groove (35) and the rear end of the mounting box (31). The toughened glass (37) is oppositely arranged at the top and bottom of the diversion box (32). The limiting mechanism (38) is arranged inside and below the mounting box (31). The drag reduction mechanism (39) is arranged inside and outside the mounting box (31). It further includes: an integrated system (3812). The integrated system (3812) includes a sensing module, a data processing and calculation module, a model training module, a navigation and control module, a communication module, a storage and heat dissipation module, a power management module and a middleware and algorithm library module. The sensing module and the data processing and calculation module are electrically connected. The data processing and calculation module is respectively electrically connected to the model training module, the navigation and control module, the communication module and the storage and heat dissipation module. The navigation and control module and the communication module are electrically connected. The model training module and the storage and heat dissipation module are electrically connected. The power management module is electrically connected to the sensing module, the data processing and calculation module, the model training module, the navigation and control module, the communication module, the storage and heat dissipation module, and the middleware and algorithm library module. The middleware and algorithm library module is electrically connected to the data processing and calculation module, the model training module and the navigation and control module.

2. An integrated device for dynamically training and inferring an unmanned aerial vehicle vision large model adaptable to the environment according to claim 1, characterized in that: The mounting box (31) is in a cuboid structure. The diversion box (32) is in an isosceles trapezoid structure. The mounting box (31) and the diversion box (32) are integrally formed. The left and right sides of the diversion box (32) incline towards the middle. The thickness of the rear end of the first diversion groove (34) is greater than that of the front end. The second diversion groove (35) is in an isosceles trapezoid structure and the height of the bottom of the rear end is greater than that of the front end.

3. An integrated device for dynamic training and inference of an unmanned aerial vehicle vision large model adaptable to the environment according to claim 2, characterized in that: The orientation of each air inlet hole (36) is consistent with the orientations of the bottoms of the corresponding first diversion groove (34), second diversion groove (35) and the rear end of the mounting box (31).

4. An integrated device for dynamic inference and training of an unmanned aerial vehicle vision large model adaptable to the environment according to claim 3, characterized in that: The limiting mechanism (38) includes a bottom plate (381), a plurality of support feet (382), a plurality of cameras (383), a second wire groove (384), a reinforcing strip (385), a plurality of first screw holes (386), a plurality of second screw holes (387), a pressing seat (388), a data processor (389), a wire plug-in board (3810), a warning light (3811) and an integrated system (3812). The bottom plate (381) is arranged at the bottom of the installation box (31). A plurality of the support feet (382) are evenly arranged at the bottom of the bottom plate (381). A plurality of the cameras (383) are arranged at the four corners of the bottom of the bottom plate (381). The second wire groove (384) is arranged through the bottom plate (381) corresponding to the position of each camera (383). The reinforcing strip (385) is respectively arranged in a front-back-left-right structure on the top of the bottom plate (381) and is connected to the inner wall of the installation box (31). A plurality of the first screw holes (386) are evenly arranged on each reinforcing strip (385). The second screw holes (387) are arranged through the bottom plate (381) corresponding to the position of each first screw hole (386). The pressing seats (388) are oppositely arranged in a left-right structure on the front half and the rear half of the top of the bottom plate (381). The data processor (389) is arranged on the top of the bottom plate (381) and is located between the four pressing seats (388). The wire plug-in board (3810) is arranged at the left and right ends of the data processor (389). The warning light (3811) is arranged at the front end of the top of the data processor (389). The integrated system (3812) is arranged inside the data processor (389).

5. An integrated device for dynamic training and inference of an unmanned aerial vehicle vision large model adaptable to the environment according to claim 4, characterized in that: The bottom plate (381) and the support feet (382) are in an L-shaped structure, and the reinforcing strip (385) and the installation box (31) are integrally formed.

6. An integrated device for dynamic inference and training of a large-scale UAV vision model adaptable to the environment according to claim 5, characterized in that: The pressing seat (388) is in a Z-shaped structure. A shock pad (3881) is arranged on the side of the pressing seat (388) facing the data processor (389). A third wire groove (3882) is arranged in a truncated manner at the middle position of the pressing seat (388). A connecting plate (3883) is arranged at one end of the third wire groove (3882) far from the data processor (389). A rotary sleeve (3884) is arranged on the top of the pressing seat (388). A third screw hole (3885) is arranged inside the rotary sleeve (3884) and the third screw hole (3885) penetrates downward through the top of the pressing seat (388). A threaded pin (3886) is arranged on the rotary sleeve (3884).

7. An integrated device for dynamic training and inference of an unmanned aerial vehicle vision large model adaptable to the environment according to claim 6, characterized in that: The connecting plate (3883) is in an isosceles trapezoid structure and is integrally formed with the pressing seat (388) and the rotary sleeve (3884). The bottom of the threaded pin (3886) abuts against the data processor (389).

8. An integrated device for dynamic training and inference of a large-scale UAV vision model adaptable to the environment according to claim 7, characterized in that: The drag reduction mechanism (39) includes side wing plates (391), a plurality of third diversion grooves (392), reinforcing plates (393), base plates (394), conduction plates (395) and a plurality of heat dissipation fins (396). The side wing plates (391) are arranged at the left and right ends of the installation box (31). A plurality of the third diversion grooves (392) are horizontally and equidistantly arranged at the upper and lower ends of the side wing plates (391). The reinforcing plates (393) are arranged at the upper and lower ends of each side wing plate (391) on the side close to the installation box (31). The base plates (394) are arranged at one end of each side wing plate (391) facing the installation box (31) and are located inside the side wall of the installation box (31). The conduction plates (395) are arranged at one end of each base plate (394) away from the adjacent side wing plate (391). A plurality of the heat dissipation fins (396) are uniformly arranged at one end of the conduction plate (395) away from the adjacent base plate (394).

9. An integrated device for dynamic training and inference of an unmanned aerial vehicle vision large model adaptable to the environment according to claim 8, characterized in that: The side wing plate (391) has a right triangle structure and the obtuse part has an arrow-like structure. The acute angle of the side wing plate (391) faces forward. The upper and lower sides of the side wing plate (391) converge into a point from back to front. The third diversion groove (392) has an arc structure and is parallel to the side walls at the left and right ends of the installation box (31). The reinforcing plate (393) has an arc structure and the upper and lower sides converge into a point from back to front. The base plate (394) has a triangular structure. The heat dissipation fins (396) have a cuboid structure and are uniformly distributed on the conduction plate (395) from back to front and from top to bottom. There is a gap between every two adjacent columns of the conduction plates (395). The side wing plates (391), reinforcing plates (393), base plates (394), conduction plates (395) and heat dissipation fins (396) are integrally formed.

Images