Maintenance cabin intelligent system based on multi-modal fusion and dynamic cooperative control

Abstract

The invention relates to the technical field of military equipment maintenance support, in particular to a maintenance cabin intelligent system based on multi-modal data fusion and dynamic cooperative control, which comprises a multi-modal sensor module, a dynamic decision module and an actuator module, the multi-modal sensor module, the dynamic decision-making module and the actuator module are connected through a closed-loop control link, and original data of the multi-modal sensor module are input into the dynamic decision-making module; a control instruction output by the dynamic decision module drives the actuator module to act; and the execution state of the actuator module is fed back to the dynamic decision module to dynamically adjust the task priority. According to the method, a three-dimensional map is constructed through multi-modal sensor fusion, and task allocation is dynamically optimized through reinforcement learning and Nash equilibrium; the GAN generates adaptive camouflage, improves the YOLOv7, improves the damage detection precision, carries out real-time path planning and dynamic power management, and achieves the high environmental adaptability, efficient execution, and reliable and stable system performance.

Description

Technical Field

[0001] The present invention relates to the technical field of military equipment maintenance support, and particularly relates to an intelligent system for a maintenance cabin based on multi-modal data fusion and dynamic cooperative control. Background Art

[0002] Current military maintenance support technologies mainly focus on the optimization and upgrading of single-functional modules, using optical chromic materials to achieve static camouflage adaptation, or relying on a preset terrain database to adjust camouflage patterns; UAV delivery systems are mostly based on GPS navigation and fixed-rule path planning, such as the shortest path algorithm; environmental perception generally uses a single sensor modality, such as optical or infrared devices for target recognition. Although such technologies have basic functions in specific scenarios, they are difficult to meet the requirements of complex and changeable battlefield environments, especially in multi-task cooperation, dynamic decision-making, and real-time response, where there are obvious shortcomings.

[0003] Currently, the isolated operation of functional modules leads to resource allocation conflicts, such as the delivery task and the deployment of camouflage competing for power resources, and the task priority depends on manually preset rules, and it cannot dynamically respond to sudden threats or terrain changes. Single-sensor data is easily interfered by the environment, such as dust occlusion or sunlight reflection causing false detections, and the measured false judgment rate exceeds 15%. The path planning algorithm lacks the ability of real-time obstacle avoidance, and the re-planning takes more than 500 ms, which is difficult to meet the requirements of emergency tasks. In addition, battle damage assessment relies on manual experience judgment, and the power management adopts a fixed-priority allocation strategy, with insufficient intelligence level, seriously restricting the actual combat effectiveness and adaptability of the maintenance support system. Therefore, there is an urgent need to design an intelligent system for a maintenance cabin based on multi-modal data fusion and dynamic cooperative control. Summary of the Invention

[0004] The purpose of the present invention is to overcome the above problems and provide an intelligent system for a maintenance cabin based on multi-modal fusion and dynamic cooperative control. To achieve the above purpose, the present invention adopts the following technical solutions:

[0005] An intelligent system for a maintenance cabin based on multi-modal fusion and dynamic cooperative control includes a multi-modal sensor module, a dynamic decision module, and an actuator module; the multi-modal sensor module, the dynamic decision module, and the actuator module are connected through a closed-loop control link, and the original data of the multi-modal sensor module is input into the dynamic decision module; the control instructions output by the dynamic decision module drive the actuator module to act; the execution status of the actuator module is fed back to the dynamic decision module to dynamically adjust the task priority; the multi-modal sensor module includes an optical camera, an infrared thermal imager, a terrain radar, and a sonic sensor.

[0006] Further, the dynamic decision module includes environmental perception data fusion, dynamic task priority allocation, and multi-task conflict resolution;

[0007] The environmental perception data is fused through the following steps:

[0008] S11. Perform HSV color space conversion on the optical camera data, and extract the environmental main hue H and texture saturation S, where H ∈ [0°, 360°] and S ∈ [0, 1];

[0009] S12. Perform threshold segmentation on the infrared thermal imager data, and set the temperature threshold Extract the pixel coordinate set of the high-temperature area {(x i , y i )};

[0010] S13. Use the RANSAC algorithm to fit the terrain surface for the terrain radar point cloud data, and output the terrain curvature C and slope θ, where R is the curvature radius of the fitted surface;

[0011] S14. Input H, S, T high , C, θ into a convolutional neural network (CNN) to output the environmental three-dimensional semantic map M semantic .

[0012] Furthermore, the dynamic task priority allocation includes the following steps:

[0013] S21. Define the reinforcement learning state space S = {L threat , A damage , R resource},

[0014] where L threat ∈ {1, 2, 3, 4, 5}, representing the threat level; A damage ∈ [0%, 100%], representing the battle damage assessment result; R resource = (E battery , Q stock ), representing the remaining power E battery and the equipment inventory Q stock ;

[0015] S22. Define the action space A = {I camouflage , P path , Q delivery},

[0016] where I camouflage ∈ [0%, 100%], representing the camouflage intensity; P path = {p1, p2,..., p n}, representing the set of path points generated by the A * strategy; Q delivery ∈ N + , representing the number of delivery equipment;

[0017] S23. The reward function is defined as:

[0018] R = 0.6·I camouflage + 0.3·Accuracy - 0.1·T delivery ;

[0019] where Accuracy ∈ [0%, 100%] represents the detection accuracy; Tdelivery ∈ R+ represents the delivery time.

[0020] Furthermore, the multi-task conflict resolution includes the following steps:

[0021] S31. Calculate the task benefit function U i = w i ·(L threat ·A damage ), where the weight coefficient w camouflage = 0.4, w scout = 0.5, w delivery = 0.3;

[0022] S32. Solve the optimal resource allocation scheme based on the Nash equilibrium model; U i s.t. ∑R resource ≤ R max , where R max is the maximum available resource of the system.

[0023] Furthermore, the actuator module includes a camouflage material driving device, and its control logic is camouflage generation and driving, including the following steps:

[0024] S41. The generative adversarial network (GAN) generates a camouflage pattern,

[0025] where the generator G is a U-Net architecture, taking the input environmental semantic map M semantic and outputting a 1024×1024 pixel camouflage image I camouflage ; the discriminator D is a PatchGAN structure, outputting a local consistency score D score ∈ [0, 1];

[0026] S42. Adjust the deployment angle of the driving mechanism based on the PID control algorithm:

[0027]

[0028] where e(t) = θ target - θ current , θ target is the target deployment angle, and θ current is the current angle.

[0029] Furthermore, the actuator module includes a drone delivery platform, and the battle damage assessment method for the drone delivery platform is as follows:

[0030] S51. Improve the YOLOv7 model to detect the damaged parts: Add the CBAM attention mechanism to the backbone network and output the attention weight w attn ∈[0,1]; The loss function uses CloULoss:

[0031]

[0032] where ρ is the Euclidean distance between the center points of the predicted bounding box b pred and the ground truth bounding box b gt ; c is the length of the diagonal of the minimum enclosing box: represents the aspect ratio consistency;

[0033] S52. Process the lidar point cloud data; Perform voxel filtering for noise reduction, with the voxel size V size = 0.1m 3 ,

[0034] The calculation formula for the damaged volume is: where N is the number of point clouds in the damaged area.

[0035] S53. If V damage ≥ 200cm 3 , mark it as a high-priority maintenance task.

[0036] Furthermore, the path planning of the drone delivery platform includes global path planning, local real-time obstacle avoidance, and path replanning trigger conditions;

[0037] The method for the global path planning is as follows:

[0038] The cost function of the A* algorithm is defined as:

[0039] f(n) = g(n) + h(n) + 0.5·C threat

[0040] where g(n) is the actual movement cost from the starting point to node n; h(n) is the Euclidean distance from node n to the end point; C threat ∈[0,1] is the threat coefficient, which is calculated in real time by the millimeter-wave radar;

[0041] The local real-time obstacle avoidance includes the parameter settings of the Dynamic Window Approach (DWA); The maximum linear velocity v max = 2m / s; The maximum angular velocity ω max = 1.5rad / s; The safety distance threshold d safe ≥ 1.5m.

[0042] The path replanning trigger condition is: when the threat level L threat ≥2 and the threat distance d threat <50m; the path replanning response time t response ≤200ms.

[0043] Furthermore, the power management method of the system includes dynamic priority allocation and an overload protection mechanism;

[0044] In the dynamic priority allocation, the device priority weight is defined as: w = {reconnaissance drone: 5, main control unit: 3, camouflage device: 1, other devices: 0.5};

[0045] The output power distribution formula of the 30kW generator:

[0046]

[0047] In the overload protection mechanism, if the current I > 1.2I for three consecutive cycles rated , trigger the fuse protection; after the non-critical equipment is powered off, every t recover = 60s attempts to restore power supply automatically.

[0048] The advantages of the present invention are as follows:

[0049] 1. The present invention realizes multi-modal data fusion by integrating optical, infrared, terrain radar and acoustic sensors. The optical camera extracts the environmental hue and saturation, the infrared thermal imager identifies the high-temperature area, the terrain radar fits the terrain curvature and slope, and combines with the convolutional neural network to generate a three-dimensional semantic map. It significantly improves the perception accuracy and comprehensiveness of complex environments and provides a reliable basis for dynamic decision-making.

[0050] 2. Based on the reinforcement learning framework and the Nash equilibrium model, the system evaluates the threat level, battle damage rate and resource status in real time, and dynamically adjusts the camouflage intensity, path planning and material delivery priority. Under resource constraints, it preferentially executes high-threat and high-battle-damage tasks, avoiding efficiency losses caused by multi-task conflicts. Compared with the static task allocation scheme, the resource utilization rate is improved and the response speed of key tasks is accelerated.

[0051] 3. The present invention uses a generative adversarial network to generate high-resolution camouflage patterns, combines the PID algorithm to accurately adjust the driving mechanism; improves the YOLOv7 model by integrating the CBAM attention mechanism to enhance the damage detection accuracy; uses the dynamic window method to achieve real-time obstacle avoidance, and shortens the path replanning response time to 200ms. Power management ensures continuous power supply for critical devices through dynamic priority allocation and overload protection mechanisms. The system has strong adaptability and stability in complex environments, and its execution efficiency and reliability are significantly better than traditional fixed-strategy systems. Description of the Drawings

[0052] The accompanying drawings, which form a part of this application, are used to provide a further understanding of this application, making other features, objectives, and advantages of this application more apparent. The schematic diagrams of the exemplary embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation of this application.

[0053] In the accompanying drawings:

[0054] Figure 1 It is a system architecture diagram of the intelligent maintenance cabin system based on multi-modal data fusion and dynamic cooperative control in Embodiment 1.

[0055] Figure 2 It is a structural diagram of the improved YOLOv7 model of the intelligent maintenance cabin system based on multi-modal data fusion and dynamic cooperative control in Embodiment 1.

[0056] Figure 3 It is a path planning flowchart of the intelligent maintenance cabin system based on multi-modal data fusion and dynamic cooperative control in Embodiment 1. Detailed Embodiments

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Generally, the components of the embodiments of the present invention described and illustrated in the accompanying drawings here can be arranged and designed in various different configurations.

[0058] The present invention will be introduced in detail and specifically below through specific embodiments to better understand the present invention. However, the following embodiments do not limit the protection scope of the present invention.

[0059] Embodiment 1

[0060] As Figures 1-3 shown, the intelligent maintenance cabin system based on multi-modal fusion and dynamic cooperative control includes a multi-modal sensor module, a dynamic decision-making module, and an actuator module; the multi-modal sensor module, the dynamic decision-making module, and the actuator module are connected through a closed-loop control link, and the original data of the multi-modal sensor module is input into the dynamic decision-making module; the control instructions output by the dynamic decision-making module drive the actuator module to act; the execution status of the actuator module is fed back to the dynamic decision-making module to dynamically adjust the task priority; the multi-modal sensor module includes an optical camera, an infrared thermal imager, a terrain radar, and an acoustic wave sensor.

[0061] The system consists of a multi-modal sensor module, a dynamic decision-making module, and an actuator module to form a closed-loop control link. The multi-modal sensor module integrates an optical camera, an infrared thermal imager, a terrain radar, and a sonic sensor, and is responsible for collecting optical, thermal radiation, terrain structure, and sonic data of the environment. The dynamic decision-making module receives the original sensor data, generates control instructions to drive the actuator module to act, and at the same time dynamically adjusts the task priority according to the real-time feedback of the execution status, forming a closed-loop optimization mechanism. The actuator module includes a camouflage device, a drone delivery platform, etc., and directly executes the maintenance task.

[0062] Furthermore, the dynamic decision-making module includes environmental perception data fusion, dynamic task priority allocation, and multi-task conflict resolution;

[0063] The environmental perception data fusion includes the following steps:

[0064] S11. Perform HSV color space conversion on the optical camera data, and extract the main environmental hue H and texture saturation S, where H ∈ [0°, 360°], S ∈ [0, 1];

[0065] S12. Perform threshold segmentation on the infrared thermal imager data, and set the temperature threshold Extract the pixel coordinate set of the high-temperature area {(x i , y i )};

[0066] S13. Use the RANSAC algorithm to fit the terrain surface for the terrain radar point cloud data, and output the terrain curvature C and slope θ, where R is the curvature radius of the fitted surface;

[0067] S14. Input H, S, T high , C, θ into a convolutional neural network (CNN), and output the environmental three-dimensional semantic map M semantic .

[0068] The environmental perception data fusion of the dynamic decision-making module is achieved through multi-step processing: the optical camera data is converted by HSV to extract the environmental hue (H) and saturation (S); the infrared thermal imager identifies the high-temperature area coordinates through 80°C threshold segmentation; the terrain radar point cloud data fits the surface through the RANSAC algorithm, and outputs the curvature (C) and slope (θ). The above parameters are input into a convolutional neural network (CNN) to generate a three-dimensional semantic map (Msemantic) containing terrain, temperature, and optical features, providing a unified environmental model for subsequent task decisions.

[0069] Furthermore, the dynamic task priority allocation includes the following steps:

[0070] S21. Define the reinforcement learning state space S = {L threat , A damage,R resource},

[0071] Among them, L threat ∈ {1, 2, 3, 4, 5}, representing the threat level; A damage ∈ [0%, 100%], representing the battle damage assessment result; R resource = (E battery , Q stock ), representing the remaining power E battery and the equipment inventory Q stock ;

[0072] S22. Define the action space A = {I camouflage , P path , Q delivery},

[0073] Among them, I camouflage ∈ [0%, 100%], representing the camouflage intensity; P path = {p1, p2,..., p n}, representing the set of path points generated by the A * tactic method; Q delivery ∈ N + , representing the number of delivery equipment;

[0074] S23. The reward function is defined as:

[0075] R = 0.6·I camouflage + 0.3·Accuracy - 0.1·T delivery ;

[0076] Among them, Accuracy ∈ [0%, 100%] represents the detection accuracy rate; T delivery ∈ R + represents the delivery time consumption.

[0077] The dynamic task priority allocation is based on the reinforcement learning framework. The state space (S) includes the threat level (L threat ), the battle damage rate (A damage ) and the resource status (R resource ); the action space (A) covers the camouflage intensity (I camouflage ), the set of path points (P path ) and the delivery quantity (Q delivery ). The reward function (R) synthesizes the camouflage intensity, the detection accuracy rate and the delivery time consumption, driving the system to balance concealment, efficiency and resource consumption during task execution.

[0078] Furthermore, the multi-task conflict resolution includes the following steps:

[0079] S31. Calculate the benefit function U i = wi ·(L threat ·A damage ), where the weight coefficients w camouflage = 0.4, w scout = 0.5, w delivery = 0.3;

[0080] S32. Solve the optimal resource allocation scheme based on the Nash equilibrium model; U i s.t. ∑R resource ≤R max , where R max is the maximum available resource of the system.

[0081] Multi-task conflict resolution is achieved through the benefit function (U i ) and the Nash equilibrium model. The task benefit is weighted by the product of the threat level and the battle damage rate (camouflage weight 0.4, reconnaissance 0.5, delivery 0.3), and the total benefit is maximized under the total resource constraint. This model ensures that when resources are limited, high-threat and high-battle-damage tasks are preferentially allocated, while also taking into account the collaborative requirements of camouflage and delivery.

[0082] Furthermore, the actuator module includes a camouflage material driving device, and its control logic is camouflage generation and driving, including the following steps:

[0083] S41. The generative adversarial network (GAN) generates a camouflage pattern,

[0084] where the generator G is a U-Net architecture, taking the input environmental semantic map M semantic and outputting a 1024×1024 pixel camouflage image I camouflage ; the discriminator D is a PatchGAN structure, outputting a local consistency score D score ∈[0,1];

[0085] S42. Adjust the deployment angle of the driving mechanism based on the PID control algorithm:

[0086]

[0087] where e(t)=θ target -θ current , θ target is the target deployment angle, and θ current is the current angle.

[0088] The camouflage material driving device adopts a generative adversarial network (GAN) and PID control. The generator generates a 1024×1024 pixel camouflage image according to the environmental semantic map; the discriminator (PatchGAN) evaluates local consistency. The PID algorithm adjusts the angle error of the driving mechanism through proportional (0.8), integral (0.2), and differential (0.05) coefficients to achieve the precise deployment of the camouflage pattern.

[0089] Furthermore, the actuator module includes a drone delivery platform, and the battle damage assessment method of the drone delivery platform is as follows:

[0090] S51. Improve the YOLOv7 model to detect the damaged part: Add the CBAM attention mechanism to the backbone network and output the attention weight w attn ∈[0,1]; The loss function adopts CloU Loss:

[0091]

[0092] where ρ is the Euclidean distance between the center points of the predicted bounding box b pred and the ground truth bounding box b gt ; c is the length of the diagonal of the smallest enclosing box: represents the aspect ratio consistency;

[0093] S52. Process the lidar point cloud data; perform voxel filtering for noise reduction, and the voxel size V size =0.1m 3 ,

[0094] The calculation formula for the damaged volume is: where N is the number of point clouds in the damaged area.

[0095] S53. If V damage ≥200cm 3 , it is marked as a high-priority maintenance task.

[0096] The battle damage assessment of the drone delivery platform combines the improved YOLOv7 model and lidar data processing. CBAM attention mechanism is added to YOLOv7, and the damage detection accuracy is improved through the CloU loss function; after the lidar point cloud is denoised by 0.1m 3 voxel filtering, the damaged volume is calculated. If V damage ≥200cm 3 , it is marked as a high-priority maintenance task to ensure a quick response to major damages.

[0097] Furthermore, the path planning of the drone delivery platform includes global path planning, local real-time obstacle avoidance, and the triggering conditions for path replanning;

[0098] The method of the global path planning is as follows:

[0099] The cost function of the A* algorithm is defined as:

[0100] f(n) = g(n) + h(n) + 0.5·C threat

[0101] Wherein, g(n) is the actual movement cost from the starting point to node n; h(n) is the Euclidean distance from node n to the end point; C threat ∈[0,1] is the threat coefficient, which is calculated in real time by the millimeter-wave radar;

[0102] The local real-time obstacle avoidance includes the parameter setting of the Dynamic Window Approach (DWA); the maximum linear velocity v max = 2m / s; the maximum angular velocity ω max = 1.5rad / s; the safety distance threshold d safe ≥1.5m.

[0103] The triggering condition of the path replanning is: when the threat level L threat ≥2 and the threat distance d threat <50m; the path replanning response time t response ≤200ms.

[0104] The path planning is divided into three parts: global, local and replanning. The global path adopts an improved A* algorithm, and the cost function (f(n)) integrates the actual cost (g(n)), the Euclidean distance (h(n)) and the threat coefficient (Cthreat). The local obstacle avoidance sets the maximum linear velocity of 2m / s, the angular velocity of 1.5rad / s and the safety distance of 1.5m through the Dynamic Window Approach (DWA). When the threat level ≥2 and the distance <50m, the path replanning is triggered within 200ms to ensure the safety and timeliness of the UAV in the dynamic environment.

[0105] Furthermore, the power management method of the system includes dynamic priority allocation and an overload protection mechanism;

[0106] In the dynamic priority allocation, the device priority weight is defined as: w = {reconnaissance UAV: 5, main control unit: 3, camouflage device: 1, other devices: 0.5};

[0107] The output power distribution formula of the 30kW generator:

[0108]

[0109] In the overload protection mechanism, if the current I > 1.2I for 3 consecutive cycles rated , the fuse protection is triggered; after the non-critical equipment is powered off, it attempts to self-recover power supply every t recover = 60s.

[0110] Power management is achieved through dynamic priority allocation and overload protection. The device priority weights are: reconnaissance drone 5, main control unit 3, camouflage device 1, and other devices 0.5. The 30kW generator power is allocated proportionally. The overload protection melts when the current exceeds the rated value by 20% for three consecutive cycles. After the non-critical devices are powered off, they attempt self-recovery every 60 seconds to ensure continuous power supply for critical tasks and system stability.

[0111] The specific embodiments of the present invention have been described in detail above, but they are only examples. The present invention is not equivalent to the specific embodiments described above. For those skilled in the art, any equivalent modifications and substitutions to the present invention are also within the scope of the present invention. Therefore, all equivalent transformations and modifications made without departing from the spirit and scope of the present invention should be covered within the scope of the present invention.

Claims

1. A maintenance cabin intelligent system based on multimodal fusion and dynamic collaborative control, characterized in that, It includes a multi-modal sensor module, a dynamic decision-making module, and an actuator module; the multi-modal sensor module, the dynamic decision-making module, and the actuator module are connected through a closed-loop control link, and the original data of the multi-modal sensor module is input into the dynamic decision-making module; the control instructions output by the dynamic decision-making module drive the actuator module to act; the execution state of the actuator module is fed back to the dynamic decision-making module to dynamically adjust the task priority; the multi-modal sensor module includes an optical camera, an infrared thermal imager, a terrain radar, and a sonic sensor.

2. The intelligent system for maintenance cabin based on multi-modal fusion and dynamic cooperative control according to claim 1, wherein The dynamic decision-making module includes environmental perception data fusion, dynamic task priority allocation, and multi-task conflict resolution; The environmental perception data fusion includes the following steps: S11. Perform HSV color space conversion on the optical camera data, and extract the environmental main hue H and texture saturation S, where H ∈ [0°, 360°], S ∈ [0, 1]; S12. Perform threshold segmentation on the data of the infrared thermal imager and set the temperature threshold Extract the pixel coordinate set of the high-temperature area {(x i , y i )}; S13. Use the RANSAC algorithm to fit the terrain surface for the terrain radar point cloud data, and output the terrain curvature C and slope θ, where R is the radius of curvature of the fitted surface; S14. Input H, S, T high , C, θ into a Convolutional Neural Network (CNN) and output the 3D environmental semantic map M semantic .

3. The intelligent maintenance cabin system based on multimodal fusion and dynamic cooperative control according to claim 1, wherein, The dynamic task priority allocation includes the following steps: S21. Define the reinforcement learning state space S = {L threat , A damage , R resource}, Among them, L threat ∈ {1, 2, 3, 4, 5}, representing the threat level; A damage ∈ [0%, 100%], representing the battle damage assessment result; P resource = (E [[ID=?]] battery , Q stock ), representing the remaining power E battery and the equipment inventory Q stock ; It seems there is a formatting issue with the "∈" symbol in the original text where it's not properly separated in the translation. The "∈" should be formatted as "∈" in the translation as well. Also, there is an unclear "?" in the translation where the original symbol should be properly represented. Please check and correct the original text for a more accurate translation. The corrected translation should be: Among them, L threat ∈ {1, 2, 3, 4, 5}, representing the threat level; A damage ∈ [0%, 100%], representing the battle damage assessment result; P resource = (E battery , Q stock ), representing the remaining power E battery and the equipment inventory Q stock ; S22. Define the action space A = {I camouflage , P path , Q delivery}, Among them, I camouflage ∈ [0%, 100%], representing the camouflage intensity; P path = {p1, p2,..., p n}, representing the set of path points generated by the A * tactic method; Q delivery ∈ N + , representing the number of delivery equipment; S23. The reward function is defined as: R = 0.6·I camouflage + 0.3·Accuracy - 0.1·T delivery ; Among them, Accuracy∈[0%,100%] represents the detection accuracy; T delivery ∈R + Indicates the delivery time.

4. The maintenance cabin intelligent system based on multimodal fusion and dynamic collaborative control according to claim 1 is characterized in that: The multi-task conflict resolution includes the following steps: S31. Calculate the task revenue function U i = w i ·(L threat ·A damage ), where the weight coefficient w camouflage = 0.4, w scout = 0.5, w delivery = 0.3; S32. Solve the optimal resource allocation scheme based on the Nash equilibrium model; s.t. ∑R resource ≤R max , where R max is the maximum available resource of the system.

5. The intelligent maintenance cabin system based on multi-modal fusion and dynamic collaborative control according to claim 1, characterized in that, The actuator module includes a camouflage material driving device, and its control logic is camouflage generation and driving, including the following steps: S41. The generative adversarial network (GAN) generates a camouflage pattern, Among them, the generator G is of the U-Net architecture, and the input is the environmental semantic map M semantic , and the output is a camouflage image I of 1024×1024 pixels camouflage ; the discriminator D is of the PatchGAN structure, and the output is the local consistency score D sc ore∈[0,1]; S42. Adjust the deployment angle of the driving mechanism based on the PID control algorithm: where e(t) = θ target -θ current , θ target is the target deployment angle, and θ current is the current angle.

6. The intelligent maintenance cabin system based on multi-modal fusion and dynamic collaborative control according to claim 1, wherein The actuator module includes a drone delivery platform, and the battle damage assessment method of the drone delivery platform is: S51. Improve the YOLOv7 model to detect the damaged part: Add the CBAM attention mechanism to the backbone network and output the attention weight w attn ∈[0, 1]; The loss function uses CloULoss: where ρ is the Euclidean distance between the center points of the predicted bounding box b pred and the ground truth bounding box b gt ; c is the length of the diagonal of the minimum enclosing box: represents the aspect ratio consistency; S52, LiDAR point cloud data processing; voxel filtering for noise reduction, voxel size V size = 0.1 m 3 , The calculation formula for the damaged volume is as follows: where N is the number of point clouds in the damaged area. S53. If V damage ≥ 200 cm 3 , it is marked as a high-priority maintenance task.

7. The intelligent maintenance cabin system based on multi-modal fusion and dynamic collaborative control according to claim 1, wherein, The path planning of the drone delivery platform includes global path planning, local real-time obstacle avoidance, and path replanning trigger conditions; The method of the global path planning is as follows: The cost function of the A* algorithm is defined as: f(n) = g(n) + h(n) + 0.5·C threat Among them, g(n) is the actual movement cost from the starting point to node n; h(n) is the Euclidean distance from node n to the end point; C threat ∈[0, 1] is the threat coefficient, which is calculated in real time by the millimeter-wave radar; The local real-time obstacle avoidance includes the parameter setting of the Dynamic Window Approach (DWA); the maximum linear velocity v max = 2 m / s; the maximum angular velocity ω max = 1.5 rad / s; the safety distance threshold d safe ≥ 1.5 m. The path replanning trigger condition is: when the threat level L threat ≥ 2 and the threat distance d threat < 50 m; the path replanning response time t response ≤ 200 ms.

8. The intelligent maintenance cabin system based on multimodal fusion and dynamic cooperative control according to claim 1, wherein The power management method of the system includes dynamic priority allocation and an overload protection mechanism; In the dynamic priority allocation, the device priority weight is defined as: w = {reconnaissance drone: 5, main control unit: 3, camouflage device: 1, other devices: 0.5}; The output power distribution formula of the 30kW generator: In the overload protection mechanism, if the current I > 1.2I for three consecutive cycles rated , the fuse protection is triggered; after the non-critical equipment is powered off, it attempts to resume power supply automatically every t recover = 60 s.

Images