Rescue emergency system and method for unmanned aerial vehicle and robot dog

Abstract

The invention relates to the technical field of artificial intelligence rescue, and discloses an unmanned aerial vehicle and robot dog rescue emergency system and method, and the system comprises an unmanned aerial vehicle cluster module which is used for air data collection, communication guarantee and emergency material distribution; the robot dog cluster module is used for ground environment detection, sample collection and special task execution; the communication cooperation module is used for ensuring information intercommunication and task cooperation between air-ground equipment; the central control module provides task instructions for the unmanned aerial vehicle cluster and the robot dog cluster; the human-computer interaction interface module is used for realizing effective communication between rescue workers and the system; the modules work cooperatively, multi-source data acquisition, three-dimensional scene reconstruction, semantic segmentation, risk assessment, fire spreading prediction and decision support information generation of a disaster site are realized, and a reliable basis is provided for rescue command decision.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence rescue, and more specifically, it relates to a rescue emergency system and method for drones and robotic dogs. Background Art

[0002] In the fire rescue scenario, the rescue command center needs to have a comprehensive and intuitive perception of the fire scene environment to make accurate decisions. However, the indoor environment at the fire scene is often difficult to perceive due to factors such as thick smoke and insufficient light. Traditional two-dimensional displays and static models cannot express complex dynamic environments.

[0003] The current main technical problems include: serious degradation of visual sensor data caused by the smoke environment, and traditional reconstruction algorithms cannot effectively reason in data missing areas; there are spatio-temporal matching errors and representation consistency problems in the fusion of heterogeneous data from different sources (thermal imaging, lidar, RGB, etc.); existing technologies lack the ability to simulate the physical process of fire and are difficult to express and predict the dynamic characteristics of fire; the collaborative perception and information sharing mechanism among traditional rescue equipment is insufficient, and the optimal allocation of resources cannot be achieved. Summary of the Invention

[0004] The present invention provides a rescue emergency system and method for drones and robotic dogs to solve the technical problems of data perception degradation, heterogeneous data fusion, dynamic fire prediction, and collaborative scheduling of rescue resources in related technologies.

[0005] The present invention provides a rescue emergency system for drones and robotic dogs, including: An air-ground collaborative equipment module, including reconnaissance drones, communication relay drones, material delivery drones, light reconnaissance robotic dogs, and heavy-duty operation robotic dogs, for aerial data collection, communication guarantee, emergency material distribution, ground environment detection, sample collection, and execution of special tasks; A communication collaborative module, for establishing a collaborative communication network between drones and robotic dogs, realizing real-time data transmission and instruction issuance, and ensuring information interconnection and task coordination between air-ground equipment; An intelligent analysis and decision-making module, for receiving multi-source heterogeneous data transmitted by the communication collaborative module, performing scene reconstruction, semantic segmentation, risk assessment, and fire trend prediction, and generating rescue decision-making information to provide task instructions for the air-ground collaborative equipment; A human-computer interaction interface module, for receiving the decision-making information generated by the intelligent analysis and decision-making module, real-time displaying the three-dimensional rescue scene, risk distribution, and assisting in command and decision-making, supporting multi-terminal access and interactive operations, and realizing effective communication between rescue personnel and the system.

[0006] In a preferred embodiment, the air-ground collaborative equipment module includes: Reconnaissance UAV, equipped with a high-definition visible light camera, an infrared thermal imager, and a lidar, for aerial data collection and scene reconnaissance; Communication relay UAV, equipped with a directional antenna and multi-band communication equipment, for establishing an aerial communication network to ensure the communication stability of the system; Material delivery UAV, equipped with a controllable release mechanism and a precise positioning system, for delivering emergency supplies and rescue equipment to the trapped area.

[0007] In a preferred embodiment, the air-ground cooperation equipment module includes: Lightweight reconnaissance robot dog, equipped with a small sensor kit and a flexible mobility system, for entering narrow spaces for environmental detection and sample collection; Heavy-duty operation robot dog, equipped with a robotic arm and special tools, for performing special rescue tasks.

[0008] In a preferred embodiment, the communication cooperation module includes: Self-organizing mesh communication network, for constructing a dynamic communication network covering the rescue field, supporting direct communication and multi-hop routing between devices; Time synchronization mechanism, for ensuring the time consistency of multi-source heterogeneous data and providing a basis for subsequent data fusion; Bandwidth adaptive allocation algorithm, which dynamically adjusts the communication resource allocation according to the task priority and data importance to ensure the priority transmission of critical information.

[0009] In a preferred embodiment, the intelligent analysis and decision-making module includes: Neural implicit representation system, for fusing and reconstructing multi-source heterogeneous data into a unified three-dimensional scene representation to support high-precision scene reconstruction; Conditional random field semantic segmentation model, for semantic understanding of the reconstructed scene to identify fire sources, safety channels, dangerous goods storage areas, and other key environmental elements; Fire source propagation simulation method based on physical models and machine learning, for predicting the development trend of fires and evaluating potential risks; Risk assessment system based on the Bayesian network framework, for calculating the risk levels of different regions in the scene and generating a risk heat map; Decision support information generation module, for generating rescue decision suggestions and action plans based on the scene understanding and risk assessment results.

[0010] In a preferred embodiment, the neural implicit representation system includes: Position encoding module, for mapping three-dimensional space coordinates to a high-dimensional feature space to enhance the network's ability to express spatial details; A multi-layer perceptron network for learning the mapping relationship from spatial points to scene attributes, including color, geometry, semantic labels, and physical properties; A cross-modal feature alignment algorithm for processing heterogeneous data from different sensors, learning the correlations between different modalities, and achieving adaptive fusion; A dynamic neural radiance field model for representing and predicting the time-varying state of a scene and supporting the simulation of dynamic processes.

[0011] In a preferred embodiment, the conditional random field semantic segmentation model is used for semantic understanding of the reconstructed scene, identifying fire sources, safe passages, dangerous goods storage areas, and other key environmental elements, and achieving high-precision scene segmentation through multi-modal feature fusion and hierarchical optimization, providing semantic information support for rescue decision-making.

[0012] In a preferred embodiment, the risk assessment system of the Bayesian network framework includes: A key area identification module for identifying key areas in the rescue scene based on semantic segmentation results and expert rules; A risk factor assessment module for calculating the structural stability, fire situation, concentration of toxic gases, and other risk factors in each area; A risk level calculation module that uses the fuzzy comprehensive evaluation method to map the monitoring data of each factor to the membership function and obtains the comprehensive risk index through weighted averaging; A risk map generation module for generating an intuitive risk heat map based on the risk assessment results and overlaying it with the three-dimensional scene model for display; A dynamic update mechanism for continuously updating the risk assessment results according to changes in environmental conditions, with a higher update frequency for high-risk areas than for low-risk areas.

[0013] In a preferred embodiment, the decision support information generation module includes a multi-UAV and robot dog collaborative decision-making algorithm based on reinforcement learning for autonomously planning the optimal search and rescue path and task allocation.

[0014] In a preferred embodiment, a rescue and emergency method for UAVs and robot dogs, which is used to implement a rescue and emergency system for UAVs and robot dogs, includes the following steps: Data acquisition and transmission, collecting multi-source heterogeneous data through the air-ground collaborative device module and using the communication collaborative module to achieve real-time data transmission; Intelligent analysis and processing, processing data in the intelligent analysis and decision-making module, performing scene reconstruction, semantic segmentation, risk assessment, and fire prediction, and generating rescue decision support information; Information visualization, real-time displaying the three-dimensional rescue scene, risk distribution, and auxiliary command and decision-making information through the human-computer interaction interface module; Collaborative decision-making execution, based on a collaborative decision-making algorithm, dynamically adjusts the task allocation and execution strategies of drones and robotic dogs.

[0015] The beneficial effects of the present invention are as follows: Through the collaborative perception of drones and robotic dogs, in a thick smoke environment with low visibility, it is still possible to construct a three-dimensional scene model with high accuracy, improve the semantic classification accuracy and perception coverage rate, and solve the problem of sensor data degradation in a smoke environment.

[0016] The dynamic scene reconstruction technology with physical constraints can predict the spread direction and intensity of the fire in advance, improve the prediction accuracy, advance the arrival time of rescue resources, reduce the fire control time, and solve the problem of dynamic expression and prediction in a fire environment.

[0017] The decision-making assistance system combining the Bayesian network framework and multiple expert knowledge bases effectively solves the problem of multi-source heterogeneous data fusion, shortens the decision-making response time, and improves the resource utilization efficiency.

[0018] The collaborative decision-making algorithm based on reinforcement learning solves the problem of collaborative planning of multiple rescue devices, improves the task completion rate in a communication-constrained environment, and increases the resource utilization rate.

[0019] The technical solution of the present invention does not rely on fixed infrastructure, is applicable to various complex fire environments, and has the characteristics of rapid deployment, strong adaptability, and good scalability, bringing a revolutionary improvement to urban fire rescue. Brief Description of the Drawings

[0020] Figure 1 It is a module diagram of a rescue emergency system of drones and robotic dogs according to the present invention. Detailed Embodiments

[0021] Now, the subject matter described herein will be discussed with reference to exemplary embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein. Without departing from the protection scope of the content of this specification, changes can be made to the functions and arrangements of the elements discussed. Each example can omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples can also be combined in other examples.

[0022] In at least one embodiment of the present invention, a rescue emergency system of drones and robotic dogs is disclosed, as Figure 1 shown, including: An air-ground collaborative device module, including reconnaissance drones, communication relay drones, material delivery drones, light reconnaissance robotic dogs, and heavy-duty operation robotic dogs, for aerial data collection, communication guarantee, emergency material distribution, ground environment detection, sample collection, and special task execution; Specifically, it includes the following: Reconnaissance UAV: Equipped with a high-definition visible light camera and thermal imaging equipment for quickly obtaining the overall situation of the disaster site.

[0023] Carry a zoom optical camera with a resolution of not less than 4K, supporting 30x optical zoom; Integrate an infrared thermal imager with a resolution of 640×512 and a wavelength of 8 - 14μm, with a temperature resolution better than 0.05°C; Have omnidirectional obstacle avoidance ability, with a detection distance of not less than 20 meters; The endurance time is not less than 45 minutes, with the ability to resist level 8 winds; Communication relay UAV: Equipped with multi-band communication equipment to establish an aerial relay network in communication-restricted areas.

[0024] Carry 5G, WiFi, ZigBee, and low-frequency emergency communication modules; The communication coverage radius is not less than 3 kilometers; The hovering accuracy is better than ±0.5 meters, and the maximum hovering time is not less than 90 minutes; Support the algorithm for automatically finding the best relay position; Material delivery UAV: Have the ability of precise positioning and delivery, and can transport light rescue materials such as first aid kits and communication equipment.

[0025] The maximum load is not less than 3 kilograms; The delivery accuracy is better than ±1 meter; Support fully automatic takeoff and landing and target tracking; Equipped with a release mechanism to ensure the safe delivery of materials; The UAV cluster adopts a distributed cooperative control architecture internally, and realizes information sharing and task coordination through a self-organizing network. The cluster scale can be dynamically adjusted according to rescue needs, and the standard configuration is 4 reconnaissance UAVs, 2 communication relay UAVs, and 2 material delivery UAVs.

[0026] Communication cooperation module, used to establish a cooperative communication network between UAVs and robotic dogs, realize real-time data transmission and instruction issuance, and ensure information interconnection and task cooperation between air and ground equipment; Specifically, it includes the following: Lightweight reconnaissance robotic dog: Weighing less than 15 kilograms, mainly used for exploration in narrow spaces and preliminary hazard assessment; Equipped with a multi-spectral camera system, integrating visible light, infrared, and fluorescence imaging capabilities; Carry a gas sensor array, which can detect 、 、 、 and other toxic and harmful gases; It has a 12-degree-of-freedom motion system and a maximum speed of 3 meters per second; The climbing ability is not less than 35 degrees, and it can climb over obstacles up to 30 cm in height.

[0027] Heavy-duty robot dogs: weighing 30 to 45 kg, used for rescue equipment transportation and environmental intervention operations; The maximum load is not less than 20 kg; Equipped with a 1-DOF robotic arm, capable of basic operations such as grasping, pushing, pulling, and rotating; Integrated high-precision environmental parameter acquisition system, including temperature, humidity, air pressure, radiation and other sensors; With protection grade up to IP67, it can work continuously in shallow water, muddy and dusty environment; The robot dog cluster adopts a master-slave control structure, consisting of a master robot dog and multiple slave robot dogs. The master robot dog is responsible for local decision-making and task allocation, and the slave robot dogs perform specific reconnaissance and operation tasks. The standard configuration is 4 light reconnaissance robot dogs and 2 heavy operation robot dogs.

[0028] The communication coordination module adopts heterogeneous network fusion technology and integrates multiple communication methods to ensure reliable data transmission in complex environments. The system architecture includes: Physical layer: Integrates multiple wireless communication technologies such as 5G, WiFi, ZigBee and satellite communications to form a redundant backup mechanism.

[0029] 5G networks are used for high-bandwidth data transmission, such as high-definition video streaming and three-dimensional scene models; The WiFi network builds a local high-speed communication network to cover the core operation area; ZigBee network is used for low-power, high-reliability sensor data collection and control command transmission; Satellite communications serve as the last resort, ensuring that basic command capabilities are maintained when all ground communications fail; Link layer: Uses dynamic routing protocols to automatically select the optimal communication path based on communication quality and bandwidth requirements.

[0030] Real-time monitoring of the signal strength, delay and reliability of each communication link; Perform link switching and load balancing based on preset strategies and real-time status; Support link aggregation technology to combine multiple low-bandwidth links into a high-bandwidth link; Network layer: implements a self-organizing mesh network to support direct communication and multi-hop forwarding between drones and robot dogs.

[0031] The communication relay drone automatically locates to the best position to expand the network coverage; The robotic dog swarm forms a ground mobile relay network to solve the problems of signal occlusion and attenuation; Supports dynamic network topology adjustment to adapt to changing environmental conditions; Application layer: Implements data priority management and bandwidth allocation to ensure real-time transmission of critical information.

[0032] Control instructions and alarm information obtain the highest transmission priority; Video streams and sensor data dynamically adjust the resolution and sampling rate according to importance; Supports data compression and incremental transmission to reduce bandwidth occupancy.

[0033] The intelligent analysis and decision-making module is used to receive multi-source heterogeneous data transmitted by the communication cooperation module, perform scene reconstruction, semantic segmentation, risk assessment, and fire prediction, and generate rescue decision-making information to provide task instructions for air-ground cooperation equipment; Specifically, it includes the following: Scene reconstruction and semantic understanding: First, obtain multi-angle and multi-scale image data of the disaster site through the drone swarm, construct a three-dimensional scene model, and perform semantic segmentation and target recognition. Specifically, it includes: Multi-source data collection: Dispatch reconnaissance drones to obtain RGB images, infrared thermal images, and laser point cloud data of the site according to the preset flight path. Data collection follows the strategy from macro to micro and from periphery to core to ensure comprehensive coverage of the target area.

[0034] Three-dimensional scene reconstruction: Use visual SLAM (Simultaneous Localization and Mapping) technology to construct a high-precision three-dimensional scene model based on multi-source heterogeneous data. This application uses an improved ORB-SLAM3 algorithm to enhance the robustness of feature point extraction and matching, making the reconstruction accuracy reach the centimeter level. The reconstruction process is expressed as: ; Among them, is the three-dimensional scene model, representing the complete three-dimensional spatial structure after reconstruction; is the visible light image sequence, containing the conventional color image data collected by the drone; is the infrared image sequence, used to capture thermal radiation information, especially suitable for night or smoky environments; is the laser point cloud data, providing high-precision distance and shape information; is the time series, recording the timestamps of data collection, used for multi-source data synchronization and dynamic scene reconstruction; is the improved SLAM algorithm, responsible for fusing and processing multi-source heterogeneous data and generating a consistent three-dimensional model.

[0035] Semantic understanding: Perform semantic segmentation and object recognition on the 3D scene model to identify key elements such as channels, doors and windows, obstacles, dangerous areas, and potential trapped persons.

[0036] This application uses a multi-modal fusion deep learning network that combines an improved ResNeXt backbone network and an FPN (Feature Pyramid Network) structure, which can effectively process RGB-D-IR multi-modal input data.

[0037] Dynamic update: As the rescue operation progresses, the system continuously obtains new sensing data to update the 3D scene model and semantic understanding results in real time. The update adopts an incremental processing strategy, only reconstructing and analyzing the changed areas, thereby reducing the computational burden and improving the system response speed.

[0038] Through this step, the system obtains a 3D scene understanding result with rich semantic information, providing a basis for subsequent path planning and decision-making.

[0039] Path planning and task allocation: Based on the scene understanding result, the system generates an optimized collaborative search and rescue path and allocates tasks according to the respective characteristics of the drones and robot dogs. The specific steps include: Region division: According to the scene complexity and rescue priority, the target area is divided into several sub-regions. The division strategy considers factors such as physical boundaries, functional continuity, and risk distribution, and uses an improved superpixel segmentation algorithm: ; Among them, is the set of sub-regions, representing the multiple sub-regions into which the entire rescue area is divided, , , represent the 1st, 2nd, and nth sub-regions respectively; n represents the number of sub-regions; is the 3D scene model, which contains the scene geometric structure and spatial information obtained by the drones; is the risk assessment result, which contains the danger level and type assessment data of each region; is the region division algorithm, a function used to divide the rescue area into multiple sub-regions according to the 3D scene model and risk assessment result.

[0040] Search and rescue path planning: For each sub-region, the system generates an optimally covered search and rescue path. This application proposes a multi-objective optimization method, considering four objectives: coverage rate, time efficiency, energy consumption, and risk avoidance. The path planning problem is defined as: ; Among them, is the set of optimal paths, representing the best search and rescue path combination calculated by the system; is the coverage function, which measures the coverage of a path to the target area. The smaller the value, the higher the coverage rate; is the time efficiency function, which measures the time required to complete the search and rescue mission. The smaller the value, the higher the time efficiency; is the energy consumption function, which measures the energy consumption during the execution of the path. The smaller the value, the higher the energy utilization efficiency; is the risk function, which measures the safety risks faced during the execution of the path. The smaller the value, the lower the risk; , , , respectively represent the weight coefficients of the probability function, time efficiency function, energy consumption function, and risk function, which are used to balance the importance of different objectives in decision-making and can be dynamically adjusted according to the rescue scenario; represents finding the set of paths that minimize the objective function .

[0041] The solution uses an improved ant colony optimization algorithm, which introduces heuristic information and local search strategies to accelerate the convergence process.

[0042] Task allocation: Based on the capabilities and characteristics of the drones and robotic dogs, as well as their current status and location, the system dynamically allocates search and rescue tasks to each execution unit. The allocation model is based on the market auction mechanism, where each search and rescue unit bids on tasks according to its capabilities and status, and the system selects the allocation plan with the maximum overall utility: ; Constraints: ; ; Among them, is the optimal allocation matrix, representing the best task allocation plan calculated by the system; is the utility value of unit executing task , which measures the suitability and expected effect of a search and rescue unit executing a specific task; is a binary decision variable, which takes the value of 1 when task is allocated to unit , and 0 otherwise; is the number of search and rescue units, including the total number of all available drones and robotic dogs; is the number of tasks, representing the total number of search and rescue subtasks to be allocated; represents the double summation over all search and rescue units and all tasks; represents finding the allocation matrix that maximizes the objective function ; The first constraint indicates that each search and rescue unit can perform at most one task; the second constraint indicates that each task must be assigned to a search and rescue unit.

[0043] Dynamic adjustment: During the execution process, the system dynamically adjusts the path planning and task allocation according to the real-time obtained scene change information and task execution situation. The adjustment strategy is based on the rolling horizon optimization method, which ensures a rapid response to environmental changes while maintaining global optimality.

[0044] Through this step, the system realizes the efficient cooperation between the UAV and the robotic dog, improving the search and rescue coverage rate and time efficiency.

[0045] Key area identification and risk assessment: The system can automatically identify the key areas at the rescue site and evaluate the risk levels. The specific implementation methods include: Key area identification: Based on deep learning object detection and scene parsing techniques, automatically identify the following types of key areas: Vital sign area: Locations where trapped people may be present; Structural hazard area: Building structures at risk of collapse; Hazard source area: Fire sources, toxic gas leakage points, flammable and explosive substances, etc.; Key passage: Passageways that can be used for rescue personnel and equipment to enter; Safe area: Areas that can be used for temporary placement and treatment; An improved Faster R-CNN model is used for key area identification. This model integrates multi-scale features and attention mechanisms, and can adapt to rescue scene targets with large size differences and complex appearance changes. The model output is a set of region bounding boxes, their class labels, and confidence scores: ; Among them, is the region bounding box, representing the position and range of the key area in the image or three-dimensional space, defined by a set of coordinate points; is the class label, representing the type of the region, such as vital sign area, structural hazard area, hazard source area, etc.; is the confidence score, representing the certainty degree of the model's classification result for this region. The value range is [0, 1], and the higher the value, the more reliable the recognition result; represents the index number of the key area; is the number of identified key areas, representing the total number of all key areas detected by the system in the current rescue scene.

[0046] Risk level assessment: Based on multi-factor comprehensive analysis, conduct risk level assessment for each identified key area. The following factors are considered in the assessment: Structural stability: Evaluated based on structural cracks, deformation, and inclination degree; Fire situation: Analyzed based on thermal imaging data to determine the location, temperature, and spread trend of the fire source; Toxic gas concentration: Assessed based on gas sensor data to evaluate the safety of air quality; Slipperiness: Evaluated based on on-site conditions to assess the safety of ground passage; Visibility: Evaluated based on smoke density and lighting conditions to assess visual reliability; The risk level is calculated using the fuzzy comprehensive evaluation method. The monitoring data of each factor is mapped to a membership function in the interval [0, 1], and then the comprehensive risk index is obtained through weighted averaging: ; Among them, is the comprehensive risk index of the th area, indicating the overall danger level of this area; is the weight coefficient of the th risk factor, used to represent the importance of different risk factors in the evaluation; is the membership function of the th factor, mapping the original monitoring data to the risk level in the interval [0, 1]; is the monitoring value of the th factor in the th area, that is, the original data obtained from the sensor; is the total number of evaluation factors, including the number of all considered risk factors such as structural stability, fire situation, and toxic gas concentration.

[0047] Risk map generation: Based on the identification of key areas and the results of risk assessment, the system generates an intuitive risk heat map, using different colors to represent the spatial distribution of risk levels. The risk map is superimposed on the 3D scene model to provide an intuitive risk perception for rescue personnel.

[0048] Dynamic update: As the rescue operation progresses and environmental conditions change, the system continuously updates the identification of key areas and the results of risk assessment. The update frequency is dynamically adjusted according to the importance of the area and the risk level. The update frequency of high-risk areas is significantly higher than that of low-risk areas.

[0049] Through this step, the system can provide accurate risk perception capabilities, helping rescue personnel avoid danger and optimize rescue routes and strategies.

[0050] Fire source propagation simulation method: The system integrates a fire source propagation simulation method that combines physical models and machine learning to predict the development trend of fires. This method includes four main modules: physical parameter configuration, boundary condition setting, propagation calculation, and visualization.

[0051] Based on the material recognition results, the physical parameter configuration module automatically assigns corresponding combustion characteristic parameters to different objects in the scene.

[0052] The system maintains a physical parameter database containing common building materials and indoor items, including: heat capacity, thermal conductivity, ignition temperature, combustion heat value, and smoke generation coefficient, etc.

[0053] For objects whose materials cannot be determined, the system uses a material recognition algorithm based on image features to map the objects to the most similar known material categories.

[0054] Material recognition uses a deep convolutional neural network, and the model architecture is a variant of DenseNet-201, which is fine-tuned on the material recognition dataset through transfer learning methods.

[0055] The boundary condition setting module determines the initial conditions and boundary conditions for simulation calculations based on real-time environmental monitoring data. Key parameters include: initial fire source location, initial fire source temperature, environmental temperature, air humidity, ventilation condition, and wind speed and direction, etc. For areas where direct measurement is not possible, the system uses a spatial interpolation method based on physical laws for parameter estimation. The boundary conditions are expressed as: ; Among them, represents the boundary condition, is the initial fire source temperature, used to represent the thermal energy level at the fire starting point; is the fire source location coordinates, representing the exact location of the fire starting point in three-dimensional space; is the environmental temperature, representing the environmental temperature around the fire area; is the air humidity, representing the water vapor content in the environment; is the ventilation rate, representing the speed of air circulation; is the wind direction, representing the flow direction of the wind; is the oxygen concentration, representing the volume percentage of oxygen in the environment; and are the carbon monoxide and carbon dioxide concentrations respectively, representing the content of combustion products in the air; is the atmospheric pressure, which affects the combustion process; is the combustible mass, representing the quantity of combustible materials; is the air density, which affects the heat transfer efficiency.

[0056] The propagation calculation module uses a hybrid model method, combining Computational Fluid Dynamics (CFD) and the heat conduction equation, to simulate the processes of flame spread, heat transfer, and smoke diffusion.

[0057] Considering the real-time requirements, the system adopts a simplified regional grid model, discretizing the space into an interconnected network of cells. The state update of each cell is based on the following control equations: ; where, is the temperature of the cell , representing the thermal energy level at a specific location in space; is the partial derivative of temperature with respect to time, representing the rate of change of temperature over time; is the thermal diffusivity, a physical quantity representing the heat conduction ability of the material; is the Laplacian operator of temperature, representing the second-order derivative of temperature in space and describing the diffusion of heat in space; is the heat source term, representing the heat generated per unit volume; is the density, representing the ratio of the mass to the volume of the material; is the specific heat capacity, representing the amount of heat required to raise the temperature of a unit mass of a substance by a unit temperature; is the fluid velocity vector, representing the direction and rate of fluid motion; is the temperature gradient, representing the rate of change and direction of temperature in space; is the convection term, representing the heat transfer due to fluid motion.

[0058] To improve the computational efficiency, the system adopts an adaptive grid refinement strategy, using a finer grid division in the flame front region and a coarser grid in the region far from the fire source.

[0059] The visualization module converts the simulation results into an intuitive three-dimensional dynamic scene, showing the flame spread path, temperature distribution, and smoke density. The system supports two visualization modes: prediction mode and comparison mode. The prediction mode shows the fire development status at different future time points; the comparison mode compares the simulation results with actual observed data to verify the model accuracy and perform adaptive adjustment.

[0060] Decision-aided information generation: Provide a decision-aided system based on multi-source information fusion, which organically integrates semantic understanding, risk assessment results, and fire spread prediction to generate structured decision-aided information and provide data support for rescue command decisions.

[0061] The decision-aided information generation method includes four main modules: information integration, priority ranking, resource allocation, and visualization presentation.

[0062] The information integration module uses a Bayesian network framework to probabilistically fuse scene information from different sources to solve data uncertainty and conflict problems; The key information includes the results of scene semantic segmentation, identification of key areas, risk level assessment, prediction of fire spread, and status of available rescue resources, etc.

[0063] The priority ranking module ranks the rescue targets based on the Comprehensive Risk Index (CRI): ; Among them, is the comprehensive risk index, representing the overall priority score of the rescue target; represents the danger degree factor, quantifying the danger degree within the area, including fire intensity, concentration of toxic gases, etc.; represents the time urgency factor, evaluating the time sensitivity of the rescue operation, and the higher the value, the faster the response is required; represents the accessibility factor, evaluating the difficulty for rescue equipment to reach the target area, and the lower the value, the more difficult it is to reach; represents the survival probability factor, evaluating the survival possibility of the trapped people under the current conditions; 、 、 、 represent the weight coefficients of the danger degree factor, time urgency factor, accessibility factor, and survival probability factor respectively, satisfying .

[0064] The resource allocation module uses an integer linear programming model to solve the optimal resource allocation plan, with the objective function of maximizing the rescue benefit, and the constraint conditions including total resource limit, time window requirement, and technical feasibility, etc.

[0065] The visualization presentation module generates a three-dimensional semantic map with spatial annotations, intuitively showing information such as the location of the fire source, spread direction, safe passage, dangerous area, and key infrastructure, etc.; at the same time, it generates a hierarchical risk level assessment report to quantify the risk types, levels, and urgencies of different areas.

[0066] Collaborative decision-making algorithm: Provides a collaborative decision-making algorithm for multiple unmanned aerial vehicles and robotic dogs based on reinforcement learning, which is used for autonomous planning of the optimal search and rescue path and task allocation. This algorithm models the entire rescue process as a Partially Observable Markov Decision Process (POMDP), and uses the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) method for solution.

[0067] This algorithm includes the following key components: Status Representation: The state vector of each agent (drone or robot dog) contains the following information: Self-state: including position coordinates , movement speed , remaining energy , current load , etc.

[0068] Observation Information: including the local feature map of the surrounding environment , detected target positions and types , where , , respectively represent the spatial coordinate sets of the 1st, 2nd, th targets detected by the agent; , , respectively represent the type identification sets of the 1st, 2nd, th targets; represents the number of targets detected at the current moment.

[0069] Task Information: including the currently assigned task type , priority and completion degree .

[0070] Team Information: including the positions of other agents , task assignments and communication status .

[0071] Action Space: The action space of the agent includes: Movement Actions: continuous action vectors for changing position and speed: ; where represents the movement action vector, respectively represent the position change amounts of the agent in the x-axis, y-axis, and z-axis directions in three-dimensional space; respectively represent the speed change amounts of the agent in the x-axis, y-axis, and z-axis directions in three-dimensional space.

[0072] Task Actions: discrete action sets for performing specific tasks: ; where represents the discrete action set for performing specific tasks, represents the search task, and the agent actively explores unknown areas; Indicates a detection task, where the agent conducts detailed perception and recognition of a specific area or target; Indicates a pickup task, where the agent picks up or carries a specific item; Indicates a transportation task, where the agent transports an item to a designated location; Indicates a communication task, where the agent exchanges information with other units.

[0073] Collaborative actions: A set of discrete actions coordinated with other agents: ; Among them, Indicates a set of discrete actions coordinated with other agents, Indicates information sharing, where the agent actively shares the environmental information it perceives with other agents; Indicates a request for help, where the agent sends a request for assistance to other agents; Indicates providing help, where the agent actively provides assistance to other agents; Indicates task handover, where the agent transfers the currently executed task to other agents to continue completion.

[0074] The overall action vector is represented as: ; Among them, Indicates the overall action vector of the agent which contains all executable action combinations; Indicates the movement action vector, which controls the position and speed changes of the agent; Indicates the task action, which refers to the specific task type executed by the agent, such as search, detection, pickup, etc.; Indicates the collaborative action, which refers to the coordination behavior between the agent and other agents, such as information sharing, request for help, etc.

[0075] Reward function: The reward function consists of multiple sub-goals: Exploration reward: Encourages the agent to explore unknown areas: ; Among them, Indicates the exploration reward value, Indicates the weight coefficient of the exploration reward, which is used to adjust the importance of the exploration behavior in the overall reward, Indicates the area of the newly explored area by the agent at the current time step, that is, the size of the environmental area newly discovered by the agent within the current decision-making cycle and not explored before.

[0076] Target discovery reward: A positive reward when a target is successfully discovered: ; Among them, Represents the reward value obtained when the agent discovers a target. Represents the weight coefficient of the target discovery reward, used to adjust the importance of target discovery in the overall reward. Represents the summation over all discovered target types. Is an indicator function, indicating whether a target of type is discovered. Represents the importance weight of a target of type .

[0077] Task completion reward: Positive reward for successfully completing the assigned task: ; Among them, Represents the reward value obtained when the agent completes the task. Is the weight coefficient of the task completion reward. Represents the priority score of the task. The higher the value, the more important the task. Represents the degree of task completion, with a value range of 0 - 1, where 1 means complete completion. This reward function is designed such that higher rewards are obtained for completing high-priority tasks, and the reward is proportional to the degree of completion.

[0078] Energy penalty: Negative reward for consuming energy: ; Among them, Represents the negative reward value received by the agent for consuming energy. Represents the weight coefficient of the energy penalty, used to adjust the importance of energy consumption in the overall reward. Represents the amount of energy consumed by the agent in the current decision cycle, and the unit can be joules or percentage. Indicates that this is a penalty term. When the agent consumes energy, the overall reward will be reduced, thus encouraging the agent to adopt energy-saving strategies.

[0079] Collaboration reward: Reward for successfully collaborating with other agents: ; Among them, Represents the reward value obtained when the agent successfully collaborates with other agents. Represents the weight coefficient of the collaboration reward, used to adjust the importance of collaborative behavior in the overall reward. Represents the number of times the agent has successfully collaborated with other agents in the current decision cycle, including the cumulative number of collaborative behaviors such as information sharing, task handover, and joint execution. This reward mechanism aims to encourage collaborative work among agents and improve the overall rescue efficiency.

[0080] Risk Penalty: Negative Reward for Entering High-Risk Areas: ; Among them, represents the negative reward value received by the agent for entering the high-risk area; represents the weight coefficient of the risk penalty, used to adjust the importance of the risk penalty in the overall reward; represents the risk level evaluation value at the position coordinates in three-dimensional space. The higher the value, the greater the risk at that position; indicates that this is a penalty term, and the agent entering the risk area will reduce the overall reward.

[0081] The overall reward function is: ; Among them, represents the overall reward value of the agent , which is the sum of all sub-reward terms; represents the exploration reward, encouraging the agent to explore unknown areas; represents the target discovery reward, the positive reward obtained when the agent successfully discovers the target; represents the task completion reward, the positive reward obtained when the agent successfully completes the assigned task; represents the energy penalty, the negative reward obtained when the agent consumes energy; represents the cooperation reward, the positive reward obtained when the agent successfully cooperates with other agents; represents the risk penalty, the negative reward obtained when the agent enters the high-risk area.

[0082] Policy Network: Adopt an Actor-Critic network architecture based on the attention mechanism. The Actor network outputs the action probability distribution, and the Critic network evaluates the value of the state-action pair. Both networks use the multi-head attention mechanism to process information from different sources (self-state, environmental observation, task information, and team information). The policy network is trained using the MADDPG algorithm based on experience replay and target network to solve the training problem in a non-stationary environment.

[0083] Task Coordination: Adopt a distributed task allocation mechanism based on auctions. When a new task is discovered or the status of an existing task changes, the system starts the task auction process. Each agent calculates the bid value based on its own capabilities, current status, and task characteristics and submits it to the coordinator. The coordinator (which can be a central control module or a temporarily selected agent) synthesizes the bid values to generate a globally optimal task allocation plan.

[0084] In some embodiments, the collaborative decision-making algorithm may integrate a Graph Neural Network (GNN) to enhance communication and collaboration among agents. Each agent is regarded as a node in the graph, and the communication links among agents form the edges of the graph. Through the graph convolutional network, each agent can aggregate information from all agents within its communication range: ; where represents the feature representation of agent at the -th layer, that is, the updated agent state information; represents the feature representation of agent at the -th layer, that is, the current agent state information; represents the feature representation of agent at the -th layer, that is, the state information of neighboring agents; represents the set of communication neighbors of agent , that is, all agents that can communicate directly with agent ; represents the set of neighbors of agent plus the agent itself; is an aggregation function for integrating neighbor node information, which can be an average function (taking the average of all neighbor features), a max function (selecting the maximum of all neighbor features), or an attention-weighted sum (weighted summing neighbor features according to importance); is the weight matrix of the -th layer, used for linear transformation of the aggregated features; is a non-linear activation function for introducing non-linear transformation to enhance the model's expressive power.

[0085] Through multi-layer graph convolution, agents can obtain global information in a wider range, improving the effect of collaborative decision-making. For example, two-layer graph convolution enables agents to obtain information about second-order neighbors (neighbors of neighbors), and three-layer graph convolution can obtain information about third-order neighbors, thus forming a more comprehensive environmental perception.

[0086] To address the possible communication interruption problem in the rescue environment, the algorithm can integrate predictive planning capabilities. When making decisions, each agent not only considers its current state but also predicts the possible behaviors of other agents in case of communication interruption: ; where represents the predicted action of agent , that is, the estimated value of the possible actions that other agents may take by the system; Represents a prediction model for predicting the behavior of an agent, which is a function that maps states to actions; Represents the current state of the agent, including information such as its position, energy level, task status, etc.; Represents the agent 's current state, containing information such as its position, energy level, task status, etc.; Represents the set of parameters of the prediction model, including the weights and biases of the neural network, etc. These parameters are learned by observing the historical behavior data of the agent 's historical behavior data.

[0087] When making decisions, the agent will take these predictions of the behavior of other agents into account, thereby generating a more robust collaboration strategy, especially maintaining effective team collaboration even in the case of communication interruption.

[0088] In some highly dynamic rescue scenarios, the collaborative decision-making algorithm can integrate the Meta-Reinforcement Learning (Meta-RL) framework to improve the system's ability to quickly adapt to new scenarios.

[0089] Meta-Reinforcement Learning learns a general adaptation strategy through pre-training in multiple different environments, enabling the agent to quickly adjust its behavior when facing a new environment: ; Among them, Represents the optimal model parameters of meta-learning, which is a set of parameters obtained through the meta-learning process and can quickly adapt to new tasks; Represents the model parameters, including trainable parameters such as the weights and biases of the neural network; Represents finding the parameters that minimize the objective function ; Represents the mathematical expectation, which is used to calculate the average value of multiple task losses; Represents different rescue task scenarios, such as different types of tasks like fire rescue, earthquake rescue, flood rescue, etc.; Represents the task distribution, that is, the probability distribution of different rescue tasks occurring; Represents sampling a task from the task distribution; Represents the task 's loss function, which is used to measure the performance of the model on a specific task; Represents the parameterized policy function, that is, the decision-making model determined by the parameter and maps states to actions; Represents the policy function on the task 's loss value.

[0090] By training in various simulation scenarios, the meta-learning model can extract general knowledge independent of tasks and quickly adapt with only a small number of samples when facing new scenarios.

[0091] In addition, in the case of extremely limited resources, the collaborative decision-making algorithm can adopt a task replanning mechanism based on multi-objective optimization. When detecting shortages of critical resources (such as energy and communication bandwidth), this mechanism automatically triggers the process of re-evaluating task priorities and reallocating resources.

[0092] The replanning uses a multi-objective evolutionary algorithm to solve the Pareto optimal front, balancing three objectives: task completion rate, time efficiency, and resource consumption: ; where, represents the objective function vector of the multi-objective optimization problem, containing three objective functions that need to be minimized simultaneously; represents the decision variable vector, representing the resource allocation scheme and task planning strategy; represents the negative value of the task completion rate. Taking the negative value is to convert the maximization of the task completion rate into a minimization problem; represents the execution time, that is, the total time required to complete all tasks; represents the resource consumption, including the total consumption of energy, computing, and communication resources; represents the th inequality constraint, such as energy limit, time limit, etc.; represents the index of the inequality constraint, represents the total number of inequality constraints; represents the th equality constraint, such as the hard requirement that tasks must be completed; represents the index of the equality constraint, represents the total number of equality constraints; represents the vector transpose symbol, organizing the objective function into a column vector form; represents the minimization operation, that is, finding the decision variable that makes the objective function reach the minimum value; represents "subject to" or "the constraint condition is", indicating the constraint conditions that the optimization problem must satisfy; represents "and", connecting multiple constraint conditions that must be satisfied simultaneously.

[0093] The system selects the most suitable solution from the Pareto front according to the current situation to achieve dynamic optimal allocation of resources. The Pareto front refers to the situation in a multi-objective optimization problem where it is impossible to improve at least one objective without sacrificing another.

[0094] The human-machine interaction interface module is used to receive the decision-making information generated by the intelligent analysis and decision-making module, display the 3D rescue scene, risk distribution, and assist in command and decision-making in real time, support multi-terminal access and interactive operations, and achieve effective communication between rescue personnel and the system; Specifically, it includes the following: Situation awareness view: Integrating three perspectives of bird's-eye view, first-person view, and 3D scene model to provide all-round on-site situation awareness.

[0095] The bird's-eye view shows the plane layout and resource distribution of the entire rescue area; The first-person view shows the front view of the drone or robot dog in real time; The 3D scene model supports zooming and rotating at any angle for easy in-depth observation of specific areas; Task planning and monitoring: Displays the current task planning and execution status, and supports manual intervention and adjustment.

[0096] The task queue panel shows all tasks to be executed, in execution, and completed; The resource status panel shows the positions, energy, load, and task assignments of all drones and robot dogs; The intervention control panel supports task replanning, priority adjustment, and manual control switching; Data analysis and decision support: Provides data visualization and decision-making suggestion functions.

[0097] The risk heat map intuitively shows the risk levels in different areas; The path planning map shows the optimal search and rescue path and alternative plans; The decision-making suggestion panel provides action suggestions and risk warnings based on the current situation; Interaction methods: Support multiple interaction modes to adapt to different usage scenarios.

[0098] Touch operation: Complete operations on tablets or touchscreens by clicking, dragging, and gestures; Voice commands: Support natural language commands, such as "View the perspective of robot dog 2 in area A", "Dispatch drone 3 to area B for support", etc.; Augmented reality: Superimpose key information on the field of vision of rescue personnel through AR glasses to achieve an intuitive interaction of "seeing is believing"; The interface design follows the principles of "simple, intuitive, and efficient", and is optimized for the rapid operation requirements in emergency situations. All control elements adopt a hierarchical layout, and important information and frequently used functions maintain high visibility and accessibility. The system also supports hierarchical permission management, and different roles (such as the general commander, area leader, operator, etc.) can access different levels of information and control functions.

[0099] Specific application examples: Example 1: Chemical fire rescue application in industrial parks; An explosion and fire occurred in a chemical warehouse in a chemical park, and the fire quickly spread to the surrounding areas. There were many dangerous chemicals at the scene of the accident, which produced a lot of thick smoke, extremely low visibility, and the risk of secondary explosion. The structure in the factory was complex, with crisscrossing pipelines, and some passages were blocked by fire and collapsed objects. The fire commander needed to quickly understand the situation at the fire scene, determine the development trend of the fire, determine the location of the trapped people, and plan the best rescue route.

[0100] System deployment: After the fire rescue team arrives at the scene, they will immediately deploy the drone and robot dog rescue emergency system of this application: Deploy 4 reconnaissance drones, 2 communication relay drones, and 2 material delivery drones; Deploy 4 light reconnaissance robot dogs and 2 heavy duty robot dogs; Establish a command and control center and activate edge computing units and cloud computing servers; Deploy a self-organizing mesh communication network to ensure full-field communication coverage.

[0101] System operation process: Multi-source data collection and scene reconstruction: The reconnaissance drone quickly obtains the overall picture of the fire scene according to the preset flight path, and collects high-altitude bird's-eye view images and thermal imaging data. The light reconnaissance robot dog enters the safe and accessible area to obtain high-definition images, infrared thermal imaging and laser point cloud data from the ground perspective. The system uses a time synchronization mechanism to unify multi-source heterogeneous data to the same time reference.

[0102] The neural implicit representation system receives multi-source data and maps spatial coordinates to a high-dimensional feature space. In the specific implementation, a 32-dimensional position encoding is used to convert the three-dimensional coordinates into Mapped to a 512-dimensional feature vector, the intermediate processing layer consists of 8 fully connected layers, each with 256 neurons, using the GELU activation function. The decoding layer outputs RGB color, geometric distance field, semantic label and temperature field respectively.

[0103] The correlation weight matrix learned by the cross-modal feature alignment algorithm shows that in areas of thick smoke, the weight of thermal imaging data is automatically increased to 0.65, while the weight of visible light images is reduced to 0.15, and the weight of laser point cloud data remains at 0.20, reflecting the system's ability to adaptively fuse data from different sensors.

[0104] Fire spread prediction and semantic understanding: The system uses a dynamic neural radiation field model to predict fire spread based on the reconstructed three-dimensional scene. The model integrates Navier-Stokes equation constraints and takes into account physical parameters such as material properties, ventilation conditions, and ambient wind direction of the chemical warehouse.

[0105] The prediction results show that if no measures are taken, the fire will spread to the adjacent high-pressure gas storage area within 20 minutes, creating a risk of large-scale chain explosions. The system calculates the three most likely paths for the fire to spread and predicts the corresponding time windows.

[0106] The conditional random field model performs semantic segmentation on the three-dimensional scene and identifies 12 key elements, including: main fire source area, secondary fire source, safe passage, blocked passage, structural support, hazardous materials storage area, control equipment, water source, possible location of trapped persons, safe area, evacuation passage and potential explosion risk area.

[0107] The segmentation model enhanced by contrastive learning still maintains high accuracy in areas with dense smoke. In areas with visibility less than 1 meter, the system successfully identified an unmarked flammable liquid storage cabinet through thermal imaging data and geometric structure features and marked it as a high-risk point.

[0108] Collaborative search and rescue and decision support: Based on the scene understanding and risk assessment results, the system divides the entire fire scene into 8 sub-areas and assigns risk levels and rescue priorities to each area. The decision support information generation module calculates the priority rescue order of the most likely location of the trapped people and plans the main rescue path and alternative paths for each target.

[0109] The system detected that four workers were trapped in the control room in the northwest corner of the plant, but the direct access to the area had been cut off by the fire. After comprehensive analysis by the Bayesian network framework, two light reconnaissance robot dogs and one heavy operation robot dog were assigned to this high-priority rescue target, and an alternative route was planned through the exhaust system of the adjacent building.

[0110] The collaborative decision-making algorithm automatically assigned the drone a continuous monitoring task to track the development of the fire and provide communication relay. At the same time, the robot dog was assigned precise search and rescue and guided evacuation tasks. A communication relay drone was adjusted to the best position to ensure that the robot dog in the trapped area could maintain stable communication.

[0111] Rescue execution and real-time adjustment: During the rescue process, the system detected that the fire suddenly accelerated and threatened the robot dog's scheduled return path. The collaborative decision-making algorithm immediately replanned the route and delivered cooling equipment to the robot dog through a material delivery drone, buying time for the trapped people and the robot dog to evacuate.

[0112] The 3D scene visualization interface shows the positions, statuses, and surrounding environments of each rescue unit in real time. The commander can obtain detailed information about each area through the interactive query interface. The hierarchical risk assessment report generated by the system shows that the cooling system on the east side of the factory area is about to fail due to high temperature, which may lead to a chain reaction. Based on this warning, the commander dispatched additional fire forces to reinforce the defense line.

[0113] Rescue effect: The system assisted the rescue team in evacuating all trapped personnel within 45 minutes after the accident, successfully preventing the fire from spreading to the high-pressure gas storage area and avoiding a larger-scale disaster. Compared with traditional rescue methods, the system achieved the following benefits: The situation awareness time was shortened from the traditional 30 to 40 minutes to 8 minutes, improving the rescue speed; The prediction accuracy of the fire development reached 91.5%, providing a reliable basis for command and decision-making; Three unknown hazard sources were successfully identified in the low visibility environment with thick smoke, avoiding potential casualties; The robotic dog successfully rescued 4 trapped personnel through the alternative path planned by the system, which was difficult to complete by conventional rescue methods within the given time window; The overall rescue efficiency was increased by 68%, and the resource utilization rate was increased by 45%; This embodiment fully demonstrates the application value of the drone and robotic dog rescue emergency system of the present application in complex disaster environments, especially having significant advantages in situation awareness, risk assessment, collaborative decision-making, and precise execution.

[0114] The above describes the embodiments of the present invention. However, these embodiments are not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative and not restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make more equivalent embodiments in various forms, all of which fall within the protection scope of this embodiment.

Claims

1. A rescue and emergency system for a drone and a robotic dog, characterized in that, Including: The air-ground cooperation equipment module, including reconnaissance UAVs, communication relay UAVs, material delivery UAVs, light reconnaissance robot dogs and heavy-duty operation robot dogs, is used for aerial data collection, communication guarantee, emergency material distribution, ground environment detection, sample collection and special task execution; The communication cooperation module is used to establish a cooperative communication network between UAVs and robot dogs, realize real-time data transmission and instruction issuance, and ensure information interconnection and task cooperation between air-ground equipment; The intelligent analysis and decision-making module is used to receive multi-source heterogeneous data transmitted by the communication cooperation module, perform scene reconstruction, semantic segmentation, risk assessment and fire prediction, and generate rescue decision-making information to provide task instructions for the air-ground cooperation equipment; The human-computer interaction interface module is used to receive the decision-making information generated by the intelligent analysis and decision-making module, display the three-dimensional rescue scene, risk distribution and auxiliary command decision-making in real time, support multi-terminal access and interactive operations, and realize effective communication between rescue personnel and the system.

2. The rescue and emergency system for an unmanned aerial vehicle and a robotic dog according to claim 1, wherein, The air-ground cooperation equipment module includes: Reconnaissance UAVs, equipped with high-definition visible light cameras, infrared thermal imagers and lidar, are used for aerial data collection and scene reconnaissance; Communication relay UAVs, equipped with directional antennas and multi-band communication equipment, are used to establish an aerial communication network to ensure the communication stability of the system; Material delivery UAVs, equipped with a controllable release mechanism and a precise positioning system, are used to deliver emergency materials and rescue equipment to the trapped area.

3. The rescue and emergency system of a drone and a robotic dog according to claim 1, characterized in that, The air-ground cooperation equipment module includes: Light reconnaissance robot dogs, equipped with a small sensor kit and a flexible mobility system, are used to enter narrow spaces for environment detection and sample collection; Heavy-duty operation robot dogs, equipped with robotic arms and special tools, are used to perform special rescue tasks.

4. The rescue and emergency system of a drone and a robotic dog according to claim 1, wherein The communication cooperation module includes: The self-organizing mesh communication network is used to build a dynamic communication network covering the rescue field, supporting direct communication and multi-hop routing between devices; The time synchronization mechanism is used to ensure the time consistency of multi-source heterogeneous data and provide a basis for subsequent data fusion; The bandwidth adaptive allocation algorithm dynamically adjusts the communication resource allocation according to the task priority and data importance to ensure the priority transmission of key information.

5. The rescue and emergency system for an unmanned aerial vehicle and a robotic dog according to claim 1, wherein, The intelligent analysis and decision-making module includes: The neural implicit representation system is used to fuse and reconstruct multi-source heterogeneous data into a unified three-dimensional scene representation to support high-precision scene reconstruction; The conditional random field semantic segmentation model is used to perform semantic understanding on the reconstructed scene, identify fire sources, safe passages, dangerous goods storage areas and other key environmental elements; The fire source propagation simulation method based on physical models and machine learning is used to predict the development trend of fires and evaluate potential risks; The risk assessment system based on the Bayesian network framework is used to calculate the risk levels of different regions in the scene and generate a risk heat map; The decision support information generation module is used to generate rescue decision-making suggestions and action plans based on the scene understanding and risk assessment results.

6. The rescue and emergency system of a drone and a robotic dog according to claim 5, characterized in that, The neural implicit representation system includes: The position encoding module is used to map three-dimensional space coordinates to a high-dimensional feature space to enhance the network's ability to express spatial details; A multi-layer perceptron network for learning the mapping relationship from spatial points to scene attributes, including color, geometric shape, semantic label, and physical attributes; A cross-modal feature alignment algorithm for processing heterogeneous data from different sensors, learning the correlation between different modalities, and achieving adaptive fusion; A dynamic neural radiance field model for representing and predicting the state of the scene changing over time, and supporting the simulation of dynamic processes.

7. The rescue and emergency system of a drone and a robotic dog according to claim 5, characterized in that, The conditional random field semantic segmentation model is used to perform semantic understanding on the reconstructed scene, identify fire sources, safety channels, dangerous goods storage areas, and other key environmental elements, and achieve high-precision scene segmentation through multi-modal feature fusion and hierarchical optimization, providing semantic information support for rescue decision-making.

8. The rescue and emergency system of a drone and a robotic dog according to claim 5, characterized in that, The risk assessment system of the Bayesian network framework includes: A key area identification module for identifying key areas in the rescue scene based on semantic segmentation results and expert rules; A risk factor assessment module for calculating the structural stability, fire situation, concentration of toxic gases, and other risk factors in each area; A risk level calculation module that uses the fuzzy comprehensive evaluation method to map the monitoring data of each factor to the membership function and obtains the comprehensive risk index through weighted average; A risk map generation module for generating an intuitive risk heat map based on the risk assessment results and overlaying it with the 3D scene model for display; A dynamic update mechanism for continuously updating the risk assessment results according to changes in environmental conditions, with a higher update frequency for high-risk areas than for low-risk areas.

9. The rescue and emergency system for a drone and a robotic dog according to claim 5, wherein, The decision support information generation module includes a multi-UAV and robot dog collaborative decision-making algorithm based on reinforcement learning for autonomously planning the optimal search and rescue path and task allocation.

10. A rescue and emergency response method for an unmanned aerial vehicle and a robotic dog, which is used to implement a rescue and emergency response system for an unmanned aerial vehicle and a robotic dog according to any one of claims 1-9, characterized in that, Including the following steps: Data collection and transmission, collecting multi-source heterogeneous data through the air-ground collaborative equipment module and using the communication collaborative module to achieve real-time data transmission; Intelligent analysis and processing, processing data in the intelligent analysis and decision-making module, performing scene reconstruction, semantic segmentation, risk assessment, and fire prediction, and generating rescue decision support information; Information visualization, real-time displaying the 3D rescue scene, risk distribution, and auxiliary command decision-making information through the human-computer interaction interface module; Collaborative decision execution, dynamically adjusting the task allocation and execution strategies of UAVs and robot dogs based on the collaborative decision-making algorithm.

Images