An air-ground cooperative rescue system and method based on intention recognition

Abstract

The present application relates to a kind of air-ground collaborative rescue system and method based on intention recognition, including first heterogeneous agent, second heterogeneous agent and wireless communication network;First heterogeneous agent and second heterogeneous agent are all run heterogeneous behavior cooperation strategy model;Heterogeneous behavior cooperation strategy model obtains observation data, calculates information entropy of characterization intention prediction certainty, and feature mutation measure of characterization local environment sudden change risk;If both do not reach trigger condition, maintain communication silence, each heterogeneous agent is independently generated action distribution based on local dimension reduction observation feature, and control instruction is output to physical execution mechanism;When detecting that time sequence input fault is caused when losing packet, each heterogeneous agent is generated virtual observation feature vector based on history hidden layer state and physical kinematics equation extrapolation;If any index reaches trigger condition, break communication silence, carry out global state broadcast by network to restore distributed collaboration.

Description

Technical Field

[0001] This invention relates to the field of air-ground coordinated rescue technology, and in particular to an air-ground coordinated rescue system and method based on intent recognition. Background Technology

[0002] In recent years, with the deep integration of artificial intelligence and distributed cooperative control technology, multi-agent reinforcement learning (MARL) technology has been widely applied in heterogeneous intelligent agent systems (such as air-ground cooperative rescue systems composed of aerial drones and ground unmanned vehicles). In these systems, heterogeneous intelligent agents acquire local observation data through onboard multimodal environmental perception components, rely on wireless communication networks for state interaction, and utilize deep neural networks (such as temporal networks and attention mechanisms) for cross-domain intent prediction and joint decision-making, thereby achieving cooperative operations in complex unstructured environments.

[0003] However, most existing multi-agent cooperative strategies are built in ideal computer simulation environments, and their theories heavily rely on assumptions of environmental stability, instantaneous communication, and strong synchronous clocks. When these theoretical models are translated into actual engineering physical systems, they face stringent communication and hardware constraints in the real physical world, resulting in performance defects and security risks, specifically in the following three aspects: First, existing algorithms lack fault tolerance under weak communication network conditions, making them highly susceptible to cascading error amplification. In actual disaster relief and other engineering environments, wireless communication networks are often affected by multipath effects, signal fading, or complex terrain obstruction, exhibiting weak communication characteristics such as high latency and random packet loss. Existing collaborative strategies generally assume that data interaction is instantaneous. Once instantaneous packet loss occurs, it can easily lead to discontinuities or numerical abrupt changes in the temporal input of the temporal feature encoder (such as GRU networks). If network transmission delays occur, the system, lacking a confidence assessment mechanism, will assign equal weight to severely lagging historical data. This temporal discontinuity and feature distortion perceived at the front end will directly lead to incorrect weight allocation by the attention mechanism, resulting in significant deviations in intent prediction. In a cascaded decision architecture, this unconstrained high-dimensional latent feature error will be directly substituted into the lower-level action network, amplified into unexpected yaw, sudden braking, and other abnormal divergent commands, leading to serious physical collision risks.

[0004] Second, existing fixed-step decision-making mechanisms suffer from frequency mismatches with the control frequencies of heterogeneous physical actuators. Existing collaborative decision-making algorithms typically employ a globally unified absolute time step (fixed clock synchronization mechanism) for network updates and action distribution. However, in real-world heterogeneous systems, the flight control system of airborne equipment is a high-frequency actuator (e.g., 100Hz), while the control chassis of ground equipment is a low-frequency actuator (e.g., 20Hz), leading to synchronization conflicts due to non-integer multiples of their control cycles. If a fixed clock synchronization is forcibly adopted, the high-frequency equipment must wait for lagging data from the low-frequency equipment within a very short period, directly causing flight attitude lag or mechanical oscillations; while the low-frequency equipment faces the accumulation and blockage of high-frequency data, reducing the smoothness and robustness of physical control.

[0005] Third, periodic network-wide broadcasting and high-frequency model inference lead to communication congestion and overload of onboard resources. Existing multi-agent systems typically employ periodic state broadcasting mechanisms to maintain decision consistency. In local area networks with limited bandwidth, such high-frequency concurrency of massive amounts of indiscriminate data can easily cause network paralysis. On the other hand, heterogeneous agents need to wake up the upper-layer large-scale neural network for complex matrix inference calculations within each extremely short fixed period. This not only consumes the computing power of the processor 302 but also causes a surge in heat generation and rapid power consumption of the main control chip. For airborne mobile nodes with strictly limited battery capacity, this will significantly shorten their effective loitering time.

[0006] In summary, how to break the dependence of traditional algorithms on fixed synchronous clocks and perfect communication networks, solve the decision distortion caused by timing input gaps and data delay accumulation under weak communication conditions, eliminate physical stuttering caused by fixed step-size synchronization of heterogeneous hardware, and effectively reduce network bandwidth usage and airborne unit power consumption are key technical problems that urgently need to be solved in the field of air-ground collaborative decision-making and engineering implementation.

[0007] Furthermore, on the one hand, there are differences in understanding among those skilled in the art; on the other hand, the applicant studied a large number of documents and patents when making this invention, but due to space limitations, not all details and contents were listed in detail. However, this does not mean that the present invention does not possess the features of these prior art. On the contrary, the present invention already possesses all the features of the prior art, and the applicant reserves the right to add relevant prior art to the background art. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides an air-ground collaborative rescue system based on intent recognition, comprising at least one first heterogeneous intelligent agent, at least one second heterogeneous intelligent agent, and a wireless communication network. The first and second heterogeneous intelligent agents are connected via the wireless communication network and both operate a heterogeneous behavior collaboration strategy model. The heterogeneous behavior collaboration strategy model is configured to: acquire observation data, calculate the information entropy representing the determinism of intent prediction, and the feature mutation metric representing the risk of drastic changes in the local environment; when neither the information entropy nor the feature mutation metric reaches the first and second triggering conditions respectively, communication silence is maintained, and each heterogeneous intelligent agent generates a virtual observation feature vector based on the historical hidden layer state and physical kinematic equations, independently generates an action distribution based on the virtual observation feature vector, and outputs control commands to the physical execution mechanism; when any indicator reaches the corresponding triggering condition, the communication silence is broken, global state broadcast is performed through the network to restore distributed collaboration, and physical control commands are output to the physical execution mechanism to complete the collaborative action.

[0009] This scheme introduces information entropy and feature mutation metrics as communication trigger indicators to construct a dynamic communication silencing and wake-up mechanism. When the environment is stable and the intent is clear, each heterogeneous agent operates independently based on local data, effectively reducing the bandwidth occupancy of the wireless communication network and decreasing the computational load on the onboard computing unit. When the environment changes drastically or the prediction confidence is insufficient, the system can break the communication silence and restore global state synchronization, ensuring the accuracy of collaborative decision-making by heterogeneous agents in complex environments. This mechanism balances the conservation of communication resources with the security of collaborative operations.

[0010] According to a preferred embodiment, the method of acquiring observation data includes: performing sliding window statistics based on round-trip delay to calculate an adaptive decision threshold; when the arrival delay of a data packet is detected to be greater than or equal to the decision threshold, intercepting and suspending the original input path using a finite state machine; substituting the historical hidden layer states extracted from the heterogeneous agent into the physical kinematic equations to perform state extrapolation calculations to generate a virtual observation feature vector; and inputting the virtual observation feature vector into the timing coding module.

[0011] This invention uses a sliding window to calculate an adaptive threshold based on round-trip time delay, accurately identifying instantaneous packet loss in weak communication network environments. It utilizes a finite state machine to intercept the original input path and combines historical hidden layer states with physical kinematic equations for state extrapolation, generating continuous virtual observation feature vectors during packet loss. This approach avoids feature distortion caused by temporal input discontinuities, prevents the action network from generating unexpected divergent instructions, and improves the system's perception fault tolerance under harsh communication conditions.

[0012] According to a preferred embodiment, the intent recognition method includes: extracting hardware timestamps to calculate communication delay time difference; determining an exponential penalty factor based on the communication delay time difference; calculating the original attention score matrix, multiplying the exponential penalty factor into the original attention score matrix, and then performing renormalization calculation through Softmax to output the attention weight matrix.

[0013] This invention extracts hardware timestamps to calculate the actual communication latency difference, transforms it into an exponential penalty factor, and constructs a dynamic confidence evaluation mechanism. This penalty factor is multiplied into the original attention score matrix for renormalization, and the system adaptively reduces the weight of lagging historical data in intent recognition based on network latency. This operational logic effectively weakens the interference of high-latency data on joint decision-making, avoids attention allocation errors, and ensures the data alignment accuracy and decision reliability of heterogeneous agents in the intent prediction stage.

[0014] According to a preferred embodiment, the information entropy is specifically Shannon information entropy; the method for determining whether the triggering condition is met and breaking the communication silence includes: logically comparing the Shannon information entropy with a preset entropy threshold to determine whether the first triggering condition is met; performing a moving average filter on the cosine similarity of the observation coding vectors of adjacent time steps; inputting the filtered cosine similarity into a hysteresis comparator for determination; when the cosine similarity after moving average filtering meets the low threshold determination condition, determining that the second triggering condition is met; when the first triggering condition or the second triggering condition is met, generating a hardware interrupt signal to break the communication silence and performing a global broadcast.

[0015] This invention employs Shannon information entropy to quantify the absolute confidence of intent prediction and utilizes the cosine similarity of the observed encoded vectors to characterize abrupt changes in local environmental features. By combining moving average filtering and a hysteresis comparator for logical decision-making, the system effectively filters out transient noise interference and prevents frequent communication state fluctuations. When a set trigger threshold is reached, a hardware interrupt signal is generated to forcibly break the communication silence, enabling the system to switch states and synchronize global data. This decision-making logic improves the stability and response sensitivity of communication state switching.

[0016] According to a preferred embodiment, during the communication silence period of the wireless communication network, each heterogeneous agent independently runs an action strategy module, specifically including: extracting local dimensionality-reduced observation features; acquiring preset heterogeneous agent role codes as hard constraint bias terms; and independently driving the action strategy module to generate action distributions based on the local dimensionality-reduced observation features and the hard constraint bias terms. This invention uses heterogeneous agent role codes as hard constraint bias terms during the communication silence period to guide agents to independently generate action distributions based on local dimensionality-reduced observation features. This configuration enables airborne and ground equipment to still perform local obstacle avoidance or path following tasks according to preset cooperative role positioning even in the absence of global data interaction. Heterogeneous nodes maintain operation by relying on independently driven action strategy modules, eliminating fixed clock synchronization conflicts caused by control frequency mismatches and improving the control smoothness of heterogeneous actuators.

[0017] According to a preferred embodiment, the heterogeneous behavior cooperative strategy model is constructed based on multi-agent reinforcement learning. The value function calculation and training method of the heterogeneous behavior cooperative strategy model includes: constructing a continuous-time discounted value function of a semi-Markov decision process model using a variable step-size discount coefficient; calculating the target Q value by performing a time-decayed weighted integral on the single-step execution reward within the communication silence interval based on the variable step-size discount coefficient; expanding the data structure of the experience replay pool to a six-tuple containing the dimension of the real communication time interval during the model training phase; and calculating and setting the sampling weights that decay over time based on the real communication time interval during sampling.

[0018] This invention constructs a continuous-time discount value function based on a variable step-size discount coefficient in the reinforcement learning model, reconstructing the fixed discount factor into a decay function based on the actual communication time interval. The target Q-value is calculated by weighted integral of the single-step execution reward, and a time interval dimension is added to the experience replay pool to adjust the sampling weights. The system matches sample values ​​according to the actual time sequence. This reduces the impact of samples with long delays or high noise on model updates, solves the credit allocation problem under multiple time scales, and ensures the convergence of the algorithm.

[0019] According to a preferred embodiment, when the heterogeneous behavior cooperation strategy model calculates the communication delay time difference, the calculation method includes: achieving synchronization of the global clock reference through the synchronization pulse signal output by the global satellite navigation system; extracting the hardware timestamp in the MAC layer register of the communication interface, wherein the hardware timestamp includes a receive timestamp and a send timestamp; and linearly correcting the hardware timestamp based on the clock frequency offset to calculate the actual communication delay time difference.

[0020] This invention establishes a global clock reference using the synchronization pulse signal of a global navigation satellite system and calculates it by extracting the receive and transmit timestamps from the MAC layer registers of the communication interface. Combined with linear correction of clock frequency offset, the system eliminates time jitter interference caused by application layer software processing and operating system scheduling. This calculation mechanism obtains high-precision real communication delay time difference at the microsecond level, providing accurate timescale input for subsequent dynamic confidence assessment and attention weight penalty, thus improving the system's data alignment accuracy.

[0021] According to a preferred embodiment, the heterogeneous behavior collaborative strategy model further includes a feature mapping bridging module to generate virtual observation feature vectors during state extrapolation calculations. The feature mapping bridging module includes a physical state extraction unit, a cross-domain mapping unit, a joint inference unit, and an observation generation unit: the physical state extraction unit is used to obtain the real kinematic state; the cross-domain mapping unit is used to align high-dimensional, implicit mathematical features with mechanism-driven low-dimensional, explicit physical models using linear projection matrices and biases; the joint inference unit is used to fuse the mapped features and combine them with kinematic differential equations to predict the physical state; and the observation generation unit is used to back-map the predicted physical state back to the observation feature space to generate virtual observation feature vectors.

[0022] The feature mapping bridging module in this invention aligns the dimensions of high-dimensional implicit features with the mechanism-driven low-dimensional explicit physical model through a linear projection matrix and bias. It combines kinematic differential equations to jointly deduce the physical state and maps the predicted state back to the observation feature space, thus achieving a deep fusion of the data-driven model and the physical mechanism model. This system architecture ensures that the generated virtual observation feature vectors have a kinematic basis during communication packet loss, improving the confidence of the compensation data and the rationality of the physical actions.

[0023] According to a preferred embodiment, the heterogeneous behavior collaboration strategy model is further configured with a boundary protection and recovery module. The boundary protection and recovery module is configured to: determine the validity of the virtual observation feature vector based on a preset time window limit and feature deviation threshold; when the validity range is exceeded, freeze the current heterogeneous agent role and only perform local obstacle avoidance and attitude maintenance actions, or terminate the collaborative task and control the heterogeneous agent to park or hover in place; when the wireless communication network is restored, use linear weighted fusion to gradually replace the virtual observation feature vector, and ensure the continuity of instructions through a transition smoothing function.

[0024] The boundary protection and recovery module of this invention performs real-time verification of the validity of virtual observation vectors. When the extrapolation time or feature deviation exceeds the set safety range, the system actively freezes the role and triggers takeover actions such as parking or hovering, avoiding the risk of error accumulation and physical collision caused by long-term open-loop extrapolation. After the communication network is restored, the virtual observation features are gradually replaced using linear weighted fusion and transition smoothing functions. This design prevents numerical step jumps in the observation feature space and ensures a smooth transition of physical commands.

[0025] This invention provides, from a second aspect, an air-ground collaborative rescue method based on intent recognition, applied to a system including at least one first heterogeneous intelligent agent, at least one second heterogeneous intelligent agent, and a wireless communication network. The first and second heterogeneous intelligent agents are connected via the wireless communication network and both run heterogeneous behavior cooperation strategy models. The method is executed by the processors of the first and second heterogeneous intelligent agents and includes: acquiring observation data, calculating information entropy representing the determinism of intent prediction, and feature mutation metric representing the risk of drastic changes in the local environment; when the information entropy does not reach a first triggering condition and the feature mutation metric does not reach a second triggering condition, communication silence is maintained, and each heterogeneous intelligent agent independently generates an action distribution based on local dimensionality-reduced observation features and outputs control commands to the physical execution mechanism; when a transient packet loss is detected causing a temporal input discontinuity, each heterogeneous intelligent agent generates a virtual observation feature vector based on the historical hidden layer state and physical kinematic equations extrapolation; when the information entropy reaches the first triggering condition or the feature mutation metric reaches the second triggering condition, the communication silence is broken, global state broadcast is performed through the network to restore distributed collaboration, and physical control commands are output to the physical execution mechanism to complete the collaborative action.

[0026] This method establishes a closed-loop processing logic at the process level, encompassing state monitoring, silent maintenance, packet loss compensation, and coordinated wake-up. This method enables heterogeneous agents to autonomously switch operating modes based on network latency and environmental abrupt changes. Under normal conditions, the system relies on independent control to reduce resource consumption caused by high-frequency synchronization. When communication is interrupted, the system maintains short-term operation through mechanism extrapolation. In complex situations, it proactively restores distributed coordination. This process breaks away from the dependence of traditional algorithms on synchronous clocks and ideal communication networks, enhancing the feasibility of system deployment. Attached Figure Description

[0027] Figure 1 This is a simplified schematic diagram of the module connection relationship of the air-ground collaborative rescue system based on intent recognition provided by the present invention; Figure 2 This is a schematic diagram of the internal module structure of the heterogeneous behavior collaboration strategy model provided by the present invention; Figure 3This is a flowchart of the air-ground collaborative rescue method based on intent recognition provided by the present invention.

[0028] List of reference numerals 100: First heterogeneous intelligent agent; 101: First airborne computing unit; 102: Visual perception device; 103: Active environmental detection module; 104: Control chassis; 200: Second heterogeneous intelligent agent; 201: Second airborne computing unit; 202: Wide-angle visual perception device; 203: Flight control system; 300: Wireless communication network; 301: Communication interface; 302: Processor; 303: Data buffer; 401: Action strategy module; 402: Timing coding module; 403: Finite state machine; 404: Hysteresis comparator; 405: Feature mapping bridging module; 406: Boundary protection and recovery module; 407: Intent recognition module; 408: Information entropy calculation module. Detailed Implementation

[0029] The following is a detailed explanation with reference to the accompanying drawings.

[0030] Example In applications such as air-to-ground collaborative rescue, heterogeneous intelligent agents possess completely different physical motion characteristics. To address the single point of failure issue caused by centralized control, an air-to-ground collaborative rescue system based on intent recognition adopts a distributed autonomous collaborative architecture. This embodiment provides an air-to-ground collaborative rescue system based on intent recognition. The overall architecture of the intent recognition-based air-to-ground collaborative rescue system includes at least one first heterogeneous intelligent agent 100, at least one second heterogeneous intelligent agent 200, and a wireless communication network 300.

[0031] Specifically, the first heterogeneous intelligent agent 100 can be a ground-based unmanned vehicle. The first heterogeneous intelligent agent 100 is equipped with a first onboard computing unit 101, and features a visual perception device 102 (forward-facing camera) and an active environmental detection module 103 (radar sensor). The physical movement of the first heterogeneous intelligent agent 100 is executed by a control chassis 104, which typically operates at a low frequency. The visual perception device 102, the active environmental detection module 103, and the control chassis 104 are all physically or electrically connected to the first onboard computing unit 101. The visual perception device 102 and the active environmental detection module 103 provide environmental detection data to the first onboard computing unit 101, and the control chassis 104 receives physical control commands issued by the first onboard computing unit 101.

[0032] The second heterogeneous intelligent agent 200 can be an aerial unmanned aerial vehicle (UAV). The second heterogeneous intelligent agent 200 is equipped with a second airborne computing unit 201 and a wide-angle visual perception device 202 (downward-facing wide-view camera). The physical motion of the second heterogeneous intelligent agent 200 is executed by a flight control system 203, which typically has a high control frequency. Both the wide-angle visual perception device 202 and the flight control system 203 are physically or electrically connected to the second airborne computing unit 201. Because the control frequencies of the control chassis 104 and the flight control system 203 differ significantly and are non-common divisors, the control chassis 104 and the flight control system 203 constitute heterogeneous physical motion execution mechanisms with different control frequencies. Both the first airborne computing unit 101 and the second airborne computing unit 201 internally include a communication interface 301, a processor 302, and a data buffer 303. The first airborne computing unit 101 and the second airborne computing unit 201 independently run heterogeneous behavior cooperation strategy models. The first heterogeneous intelligent agent 100 and the second heterogeneous intelligent agent 200 are connected to the wireless communication network 300 through the communication interfaces 301 inside the first airborne computing unit 101 and the second airborne computing unit 201, respectively, thereby realizing physical communication connections and information interaction between heterogeneous nodes. In addition, the first airborne computing unit 101 and the second airborne computing unit 201 are also equipped with BeiDou / GPS receiver modules, which receive PPS second pulse signals to achieve synchronization of the global clock reference. The communication interface 301 can adopt various physical or logical implementation methods, such as an Ethernet interface, a wireless local area network (WLAN) adapter, a cellular mobile communication module, a near-field communication module, or other interface units with data transmission and reception functions. This embodiment preferably uses an Ethernet interface or a wireless network card to achieve stable transmission of high-frequency status data.

[0033] Based on the aforementioned physical hardware architecture, the heterogeneous behavior cooperation strategy model is constructed using multi-agent reinforcement learning. This model is deployed as a computer program in the data buffers 303 of the first onboard computing unit 101 and the second onboard computing unit 201. The heterogeneous behavior cooperation strategy model is executed by the processors 302 of the first onboard computing unit 101 and the second onboard computing unit 201.

[0034] Specifically, such as Figure 2As shown, the heterogeneous behavior collaboration strategy model internally includes an action strategy module 401, a timing encoding module 402, a finite state machine 403, a hysteresis comparator 404, a feature mapping bridging module 405, a boundary protection and recovery module 406, an intent recognition module 407, and an information entropy calculation module 408. The aforementioned action strategy module 401, timing encoding module 402, finite state machine 403, hysteresis comparator 404, feature mapping bridging module 405, and boundary protection and recovery module 406 are all located in the first onboard computing unit 101 and the second onboard computing unit 201. The aforementioned action strategy module 401, timing encoding module 402, finite state machine 403, hysteresis comparator 404, feature mapping bridging module 405, and boundary protection and recovery module 406 perform data flow and processing through the first onboard computing unit 101 and the second onboard computing unit 201.

[0035] like Figure 2 As shown, the intent recognition module 407 is connected to the information entropy calculation module 408. The intent recognition module 407 receives raw observation data from the heterogeneous intelligent agent and performs intent recognition. The intent recognition module 407 generates an intent recognition probability distribution and sends it to the information entropy calculation module 408. The information entropy calculation module 408 is connected to the hysteresis comparator 404. The information entropy calculation module 408 receives the intent recognition probability distribution information from the intent recognition module 407 and calculates the information entropy representing the determinism of intent prediction. The information entropy calculation module 408 generates information entropy and sends it to the hysteresis comparator 404.

[0036] Finite state machine 403 is connected to feature mapping bridging module 405. Finite state machine 403 receives data packet information from the heterogeneous agent input, intercepts it, and suspends the original input path. Finite state machine 403 generates an extrapolation trigger command and sends it to feature mapping bridging module 405. To provide a temporal feature basis, temporal encoding module 402 is connected to feature mapping bridging module 405, action policy module 401, and hysteresis comparator 404. Temporal encoding module 402 receives raw observation data, virtual observation feature vectors, and lower-level role feature information from heterogeneous agent input, feature mapping bridging module 405, and action policy module 401, and performs moving average filtering on the cosine similarity of observation encoding vectors at adjacent time steps. Temporal encoding module 402 generates historical hidden layer states and feature mutation metrics, sends the historical hidden layer states to feature mapping bridging module 405, and sends the feature mutation metrics to hysteresis comparator 404.

[0037] The feature mapping bridging module 405 is connected to the timing encoding module 402 and the boundary protection and recovery module 406, respectively. The feature mapping bridging module 405 receives extrapolation trigger commands and historical hidden layer state information from the finite state machine 403 and the timing encoding module 402, and performs physical state prediction by combining kinematic differential equations. The feature mapping bridging module 405 generates a virtual observation feature vector, virtual observation features, and extrapolation time, sends the virtual observation feature vector to the timing encoding module 402, and sends the virtual observation features and extrapolation time to the boundary protection and recovery module 406. Subsequently, the boundary protection and recovery module 406 receives the virtual observation features and extrapolation time information from the feature mapping bridging module 405 and determines the validity of the virtual observation feature vector. The boundary protection and recovery module 406 generates trigger freeze / park / hover commands and outputs them externally to control the heterogeneous intelligent agent to perform local obstacle avoidance and attitude maintenance actions.

[0038] Hysteresis comparator 404 is connected to information entropy calculation module 408 and timing coding module 402 respectively. Hysteresis comparator 404 receives the aforementioned information entropy and feature mutation measurement information from information entropy calculation module 408 and timing coding module 402, and performs low-threshold judgment condition determination. Hysteresis comparator 404 generates a hardware interrupt signal / breaks communication silence, outputs it externally to break communication silence, and performs global broadcast. Simultaneously, action strategy module 401 runs independently, independently generating action distributions. Action strategy module 401 generates lower-level role characteristics and outputs physical control commands, sends the lower-level role characteristics feedback to timing coding module 402, and outputs the physical control commands externally to the physical actuator to complete the coordinated action.

[0039] During routine inspection, the heterogeneous behavior collaboration strategy model maintains communication silence on the wireless communication network 300. During this communication silence period, the first heterogeneous agent 100 and the second heterogeneous agent 200 independently run the action strategy module 401. For example... Figure 3 As shown, the heterogeneous behavior cooperative strategy model extracts local dimensionality-reduced observation features and obtains pre-set heterogeneous agent role codes as hard constraint bias terms. Subsequently, based on the local dimensionality-reduced observation features and hard constraint bias terms, the heterogeneous behavior cooperative strategy model independently drives the action strategy module 401 to generate action distributions. The action strategy module 401 outputs two-dimensional driving commands or three-dimensional flight commands to the control chassis 104 or the flight control system 203.

[0040] To overcome strong synchronous clock dependence and achieve adaptive cooperative task triggering, the heterogeneous behavior cooperation strategy model acquires observation data in real time for intent recognition. The model calculates the information entropy of the intent recognition probability distribution and simultaneously calculates the feature mutation metric of local observation features. When determining whether the triggering condition is met or communication silence is broken, the heterogeneous behavior cooperation strategy model compares the Shannon information entropy with a preset entropy threshold to determine whether the first triggering condition has been met. Figure 3 As shown. Simultaneously, the heterogeneous behavior cooperation strategy model applies a moving average filter to the cosine similarity of the observed encoded vectors at adjacent time steps. The model then inputs the filtered cosine similarity into a hysteresis comparator for judgment. When the cosine similarity after the moving average filter meets the low-threshold judgment condition, the heterogeneous behavior cooperation strategy model determines that the second trigger condition has been met. When the information entropy does not reach the first trigger condition and the feature mutation measurement does not reach the second trigger condition, the heterogeneous behavior cooperation strategy model maintains communication silence in the wireless communication network 300, as shown. Figure 3 As shown. When the information entropy reaches the first trigger condition or the feature mutation metric reaches the second trigger condition, the hysteresis comparator 404 in the heterogeneous behavior cooperation strategy model generates a hardware interrupt signal to break the communication silence, such as... Figure 3 As shown. After breaking the communication silence, the heterogeneous behavior cooperation strategy model is broadcast globally through the wireless communication network 300, thereby enabling the first heterogeneous agent 100 and the second heterogeneous agent 200 to enter a distributed autonomous cooperation state, and output physical control commands to the physical actuator to complete the cooperative action.

[0041] In a distributed autonomous collaborative state, to address the problem of intent recognition distortion caused by network transmission latency, the heterogeneous behavior collaboration strategy model calculates the communication delay time difference when performing intent recognition.

[0042] The data processing steps for calculating the communication delay time difference in the intent recognition module 407 of the heterogeneous behavior cooperation strategy model include: the intent recognition module 407 synchronizes the global clock reference using the synchronization pulse signal output by the Global Navigation Satellite System; the heterogeneous behavior cooperation strategy model extracts the hardware timestamp from the MAC layer register of the communication interface 301, the hardware timestamp including the receive timestamp and the transmit timestamp; then, the heterogeneous behavior cooperation strategy model performs linear correction on the hardware timestamp based on the clock frequency offset to calculate the actual communication delay time difference; during intent recognition, the heterogeneous behavior cooperation strategy model determines the exponential penalty factor based on the actual communication delay time difference; subsequently, the heterogeneous behavior cooperation strategy model calculates the original attention score matrix, multiplies the exponential penalty factor into the original attention score matrix, performs renormalization calculation through a Softmax step, and outputs the attention weight matrix.

[0043] like Figure 3 As shown, the heterogeneous behavior collaboration strategy model determines whether a temporal input discontinuity occurs due to transient packet loss. If not, it proceeds with the intent recognition step.

[0044] If so, i.e., the wireless communication network 300 experiences momentary packet loss, the heterogeneous behavior cooperation strategy model acquires observation data through adaptive extrapolation fault tolerance and secure recovery steps, such as... Figure 3 As shown.

[0045] Specifically, the heterogeneous behavior cooperative strategy model performs sliding window statistics based on round-trip time delay to calculate an adaptive decision threshold. When the heterogeneous behavior cooperative strategy model detects that the arrival delay of a data packet is greater than or equal to the decision threshold, it uses a finite state machine 403 to intercept and suspend the original input path. Subsequently, the heterogeneous behavior cooperative strategy model initiates the feature mapping bridging module 405 to generate a virtual observation feature vector. Preferably, the physical state extraction unit of the feature mapping bridging module 405 acquires the real kinematic state. The cross-domain mapping unit of the feature mapping bridging module 405 uses a linear projection matrix and bias to align the dimensions of the historical hidden layer state with the real kinematic state. The joint inference unit of the feature mapping bridging module 405 fuses the mapped features, and the joint inference unit combines the kinematic differential equations to predict the physical state; the observation generation unit of the feature mapping bridging module 405 maps the predicted physical state back to the observation feature space to generate a virtual observation feature vector. The heterogeneous behavior cooperative strategy model inputs the virtual observation feature vector into the temporal coding module 402. During the extrapolation calculation of the state, the boundary protection and recovery module 406 determines the validity of the virtual observation feature vector based on a preset time window limit and feature deviation threshold. Specifically, it determines whether the extrapolated state exceeds the valid range. If it does not exceed the valid range, the intent recognition step continues based on the generated virtual observation feature vector. If it is determined that the valid range is exceeded, the boundary protection and recovery module 406 freezes the current heterogeneous agent role and only performs local obstacle avoidance and attitude maintenance actions, or the boundary protection and recovery module 406 terminates the collaborative task and controls the first heterogeneous agent 100 or the second heterogeneous agent 200 to park or hover, thereby ending the current collaborative process. When the wireless communication network 300 recovers, the boundary protection and recovery module 406 uses linear weighted fusion to gradually replace the virtual observation feature vector, and the boundary protection and recovery module 406 ensures instruction continuity through a transition smoothing function.

[0046] Furthermore, the value function calculation and training method of the heterogeneous behavior collaborative strategy model includes specific data processing steps. Specifically, the heterogeneous behavior collaborative strategy model uses a variable step-size discount coefficient to construct the continuous-time discounted value function of the semi-Markov decision process model; based on the variable step-size discount coefficient, the heterogeneous behavior collaborative strategy model performs a time-decayed weighted integral on the single-step execution reward within the communication quiescent interval to calculate the target Q value. During the model training phase, the heterogeneous behavior collaborative strategy model expands the data structure of the experience replay pool to a six-tuple containing the dimension of the actual communication time interval. During sampling, the heterogeneous behavior collaborative strategy model calculates and sets sampling weights that decay over time based on the actual communication time interval.

[0047] Traditional algorithms rely on the assumption of a strongly synchronized clock with a globally uniform absolute time step. In the aforementioned environment, forced synchronization can lead to control lag or mechanical oscillations; it can also cause data congestion due to low-frequency devices, and periodic network-wide broadcasts can easily cause network congestion. To address this issue, a heterogeneous behavior cooperation strategy model implements an air-to-ground collaborative rescue method based on intent recognition. This method is applied to a distributed system containing a first heterogeneous agent and a second heterogeneous agent, and is executed by the heterogeneous behavior cooperation strategy model. The specific implementation steps are as follows.

[0048] S100: The intention recognition module 407 in the heterogeneous behavior collaboration strategy model acquires observation data for intention recognition, and the information entropy of the intention recognition probability distribution is calculated by the information entropy calculation module 408 in the heterogeneous behavior collaboration strategy model; at the same time, the feature mutation measure of the local observation features is calculated by the temporal coding module 402 in the heterogeneous behavior collaboration strategy model.

[0049] The heterogeneous behavior collaboration strategy model acquires observation data and performs extrapolation compensation for instantaneous packet loss.

[0050] During system operation, the heterogeneous behavior cooperation strategy model, based on an internally pre-built finite state machine 403, continuously monitors the arrival status of data packets from the heterogeneous agents at the other end. Finite state machine 403 employs a round-trip delay sliding statistics method, dynamically calculating an adaptive decision threshold under a sliding window constraint of 20 to 50 frames. The formula for calculating the adaptive decision threshold is: .

[0051] In the above formula, when the time delay exceeds the adaptive decision threshold, the hierarchical finite state machine (FSM) intercepts the original input path and triggers the temporal dynamic extrapolation compensation mechanism. This represents the measured round-trip time (in milliseconds) for frame t. This represents the average round-trip time within the sliding window. This represents the standard deviation of the round-trip delay within the sliding window. This is the safety margin factor, with a value ranging from 1.5 to 2.0.

[0052] Finite state machine 403 is configured with global state flags. When the data packet arrival delay meets... At this time, the finite state machine 403 determines it as normal network jitter, and the heterogeneous behavior cooperation strategy model executes forward encoding calculation normally under the drive of real observation data. When the arrival delay is detected to meet the condition... When this occurs, it is determined to be an instantaneous packet loss and disconnection. Under this condition, the heterogeneous behavior collaboration strategy model uses a finite state machine 403 to intercept and suspend the original input path. Under the constraints of stopping reading from the data buffer 303 and freezing the pointer, the feature mapping bridging module 405 is started to generate virtual observation feature vectors, and read-only locking is performed on the original observation buffer during the disconnection.

[0053] The specific execution process is as follows: The CPU pin of the onboard computing unit receives the second pulse signal output by the global positioning system. The CPU executes the hardware interrupt service routine and reads the media access control layer timestamp register of the network interface controller. The CPU calculates the transmission delay of the wireless communication network 300 based on the hardware timestamp. When the delay value exceeds the adaptive threshold set based on the standard deviation, the interrupt controller generates a hardware interrupt. The operating system kernel takes over the interrupt request and suspends the daemon process of the network driver receive queue. The memory management unit modifies the page table attributes of the data buffer 303, sets the access permission of the corresponding memory page to read-only, and stops updating the read and write pointers. This process prevents network packets from overwriting the valid data in the data buffer 303 and maintains the time consistency of the data read by the timing feature extraction module.

[0054] Preferably, the feature mapping bridging module 405 includes a physical state extraction unit, a cross-domain mapping unit, a joint inference unit, and an observation generation unit. These units are collectively configured within the onboard computing unit; they are interconnected via the onboard computing unit's system data bus and shared memory address. When performing data loss compensation, the physical state extraction unit radially transmits the extracted time-series state matrix to the cross-domain mapping unit along the data stream.

[0055] The physical state extraction unit acquires the true kinematic state. The cross-domain mapping unit extracts the high-dimensional hidden state of the temporal coding module 402 at time t-1. Using linear projection matrix With bias The mapping formula is executed to align the dimensions of the historical hidden layer states with the actual kinematic states. This represents an m-row, d-column real matrix. d represents the high-dimensional hidden state of the input. The dimension of is m, where m is the output dimension after projection.

[0056] The mapping formula is: .

[0057] The feature fusion formula for the features after fusion mapping of the joint simulation units is as follows: , .

[0058] Where β represents the fusion coefficient, and its value range is [0, 1]; Represents the true kinematic state; This represents the pseudo-physical state obtained by mapping the historical hidden layer state.

[0059] As the fundamental mechanism for state deduction, the general discrete kinematics iterative equation is: .

[0060] In the above formula, Indicates the minimum sampling step size of the system; Represents the kinematic functions of a rigid body; This indicates the control command from the previous moment.

[0061] The fused features Substituting these equations into the general discrete kinematics iterative equations for joint derivation, the physical state at time t is predicted using the following formula: .

[0062] In the above formula, Let be the predicted physical state value at time t.

[0063] The observation generation unit sets the inverse mapping network matrix. With bias The predicted physical state is mapped back to the observation feature space to generate a virtual observation feature vector. The formula for this reverse mapping is: .

[0064] In the above formula, This indicates that the system has a unified observation coding dimension, which is completely consistent with the actual observation dimension; the generated virtual observation feature vector Dimensions, range and actual observations Completely equivalent.

[0065] The heterogeneous behavior collaboration strategy model directly inputs the virtual observation feature vector into the temporal coding module 402 for forward computation and update. Its temporal feature update formula is as follows: .

[0066] The high-dimensional hidden state at the current moment is obtained through the above calculations. This ensures that the output of downstream physical actions remains continuous even when there is a lack of input.

[0067] During the execution of the state extrapolation calculation, the boundary protection and recovery module 406 determines the validity of the virtual observation feature vector based on a preset time window limit and a feature deviation threshold. The Euclidean distance calculation formula for the feature deviation is: .

[0068] In the above formula, This represents the Euclidean distance deviation between virtual observations and historical real observations; This represents the virtual observation feature vector generated by extrapolation at the current time. Represents the actual observation of a historical time step.

[0069] The conditional expression for the boundary protection and recovery module 406 to determine whether the extrapolated state exceeds the limit is: or .

[0070] in, Indicates the cumulative extrapolation duration; This indicates the maximum permissible extrapolation time window, with an engineering value range of 200 to 500 ms; This indicates the preset deviation threshold.

[0071] When the cumulative extrapolation duration Greater than or equal to the maximum allowed extrapolation time window Right now or characteristic deviation Greater than or equal to the deviation threshold If the effective range is exceeded, the boundary protection and recovery module 406 triggers a security degradation mechanism. Specifically, the boundary protection and recovery module 406 freezes the current heterogeneous agent role and only controls the current heterogeneous agent to perform local obstacle avoidance and attitude maintenance actions; or, the boundary protection and recovery module 406 terminates the cooperative task and controls the first heterogeneous agent 100 or the second heterogeneous agent 200 to park or hover in place.

[0072] After the wireless communication network 300 recovers, the boundary protection and recovery module 406 gradually replaces the virtual observation feature vector using linear weighted fusion. Within the set transition frame range, the formula for calculating the fusion coefficient used for dynamic adjustment is as follows: .

[0073] In the above formula, This represents the fusion coefficient of the k-th frame; This indicates the preset number of transition frames, with an engineering value range of 10 to 30 frames; k represents the current transition frame number, and satisfies... .

[0074] The boundary protection and restoration module 406 initiates a transition smoothing function based on the fusion coefficient to gradually transition virtual observation features to real observation features. The transition smoothing formula is as follows: .

[0075] In the above formula, This represents the observed features that are finally input into the network after fusion. This represents the actual observations that were received again after the network was restored; This represents the virtual observation feature vector generated by extrapolation.

[0076] Throughout the smooth transition, the heterogeneous behavior collaboration strategy model does not modify the timing coding module 402 and the upper-level role decision logic, thereby ensuring the continuity of physical control commands and effectively avoiding mechanical mechanism oscillations caused by data space step jumps.

[0077] The heterogeneous behavior cooperation strategy model performs intent recognition and calculates the dynamic confidence decay of communication delay. The model calculates the communication delay time difference by receiving the PPS second pulse signal (synchronization pulse signal) through the BeiDou / GPS receiver module (acting as a global satellite navigation system) to achieve global clock reference synchronization, and extracts the hardware timestamp from the MAC layer hardware register of communication interface 301. The formula for calculating network transmission delay is: .

[0078] In the above formula, Indicates network transmission latency (representing only pure network transmission latency, excluding application layer jitter); Indicates the received timestamp Indicates the sending timestamp.

[0079] The actual communication delay time difference is calculated by linearly correcting the hardware timestamp based on clock frequency skew. The correction formula for the actual communication delay time difference is: .

[0080] In the above formula, This represents the actual communication delay time difference; This indicates the clock frequency offset.

[0081] When performing intent recognition, the heterogeneous behavior collaboration strategy model determines an exponential penalty factor that monotonically decreases over time based on the actual communication latency difference. The formula for calculating the exponential penalty factor is: .

[0082] In the above formula, Let λ represent the exponential penalty factor corresponding to the i-th agent, and let λ represent the exponential decay coefficient, satisfying λ>0; This represents the actual communication delay time difference corresponding to the i-th agent.

[0083] The heterogeneous behavior collaboration strategy model calculates the original attention score matrix, multiplies the exponential penalty factor into the original attention score matrix, and the updated attention score calculation formula is as follows: .

[0084] in, This indicates the attention score after the punishment; This represents the original attention score of the i-th agent.

[0085] Subsequently, the heterogeneous behavior collaboration strategy model undergoes renormalization calculation through a Softmax step, outputting the attention weight matrix. The renormalization formula is as follows: .

[0086] In the above formula, represents the attention weight of the i-th agent after renormalization; N represents the total number of agents.

[0087] During the model training phase, the data structure of the experience replay pool is expanded to include a six-tuple containing the dimension of the actual communication time interval. A robust penalty term is added to the loss function. The formula for calculating the robust penalty term is: .

[0088] In the above formula, Indicates the delay robustness penalty; Let represent the penalty weight, and satisfy η > 0; This represents the variance of the attention weights; the greater the latency, the higher the variance of the attention weights. This guides the agent to distinguish between policy bias and network fading interference.

[0089] The heterogeneous behavior collaboration strategy model calculates the information entropy and feature mutation measure for intent recognition. It logically compares the Shannon information entropy with a preset entropy threshold to determine whether the first triggering condition has been met. The model uses Shannon information entropy to transform relative weights into absolute confidence scalars; the formula for Shannon information entropy is: ; .

[0090] In the above formula, This represents Shannon information entropy, used to assess absolute confidence. As an absolute confidence assessment metric, higher entropy indicates a more dispersed attention distribution, higher uncertainty, and lower confidence; conversely, lower entropy indicates higher confidence. The preset entropy threshold is... ,when When the interrupt signal is set to 1, it indicates insufficient confidence, triggering communication and role reselection; when Setting the interrupt signal to 0 indicates that the confidence level is normal and communication remains silent.

[0091] Simultaneously, the heterogeneous behavior cooperative strategy model calculates the cosine similarity of the observation encoding vectors at adjacent time steps. The formula for calculating the cosine similarity is: .

[0092] In the above formula, Indicates cosine similarity; Represents the observation encoding vector at time t; This represents the observation encoding vector at time t-1. The lower the similarity between the observation encodings of two adjacent frames, the more drastic the feature changes. T is the transpose symbol in linear algebra.

[0093] Subsequently, the cosine similarity of three consecutive frames is subjected to a moving average filter, and the filtered result is input into the hysteresis comparator 404 for determination. The determination logic expression is as follows: .

[0094] In the above formula, Indicates a Boolean-type characteristic mutation event signal; This represents the moving average filter value indicating the similarity of three consecutive frames. This indicates a preset low threshold condition for judgment; This represents the preset high-threshold judgment condition. The cosine similarity after moving average filtering satisfies the low-threshold judgment condition. When the second trigger condition is met, a Boolean-type characteristic mutation event signal with a value of 1 is output.

[0095] S200: When the information entropy does not reach the first triggering condition and the feature mutation measurement does not reach the second triggering condition, the communication silence of the wireless communication network 300 is maintained, and each heterogeneous intelligent agent independently runs the action strategy module 401.

[0096] The heterogeneous behavior cooperation strategy model maintains communication silence. When the hysteresis comparator 404 in the heterogeneous behavior cooperation strategy model determines the Shannon information entropy... If the first triggering condition is not met, and the cosine similarity after moving average filtering does not meet the low threshold judgment condition of the hysteresis comparator 404 (the second triggering condition is not met), the heterogeneous behavior cooperation strategy model maintains the communication silence of the wireless communication network 300. During the period of maintaining the communication silence, the first airborne computing unit 101 and the second airborne computing unit 201 independently run the action strategy module 401 to reduce bandwidth occupation and energy consumption.

[0097] During the communication silence period, the heterogeneous behavior cooperation strategy model performs local dimensionality reduction and local high-frequency control. While maintaining the communication silence, the heterogeneous behavior cooperation strategy model extracts locally dimensionality-reduced observation features. The formula for calculating these locally dimensionality-reduced observation features is as follows: .

[0098] In the above formula, Represents the locally reduced-dimensional observation feature vector; Represents a local dimension reduction mapping matrix; This represents the original observation encoding vector of the i-th agent.

[0099] The heterogeneous behavior collaboration strategy model obtains pre-defined heterogeneous agent role codes as hard constraint bias terms. These codes are embedded into the action distribution generation formula, and the formula for calculating the mean action value generated by the action distribution is as follows: .

[0100] The above formula, This represents the mean of the actions generated by the action distribution; Represents the action mapping function; Indicates the role constraint weight; This represents the pre-defined heterogeneous agent role encoding.

[0101] Based on local dimensionality reduction observation features and hard constraint bias terms, the heterogeneous behavior cooperative strategy model independently drives the action strategy module 401 to generate action distributions. The action strategy module 401 outputs two-dimensional driving commands or three-dimensional flight commands to the control chassis 104 or the flight control system 203. Under the constraints of the hard constraint bias terms, the heterogeneous behavior cooperative strategy model ensures that the physical obstacle avoidance actions do not deviate from the intended action.

[0102] The heterogeneous behavior collaboration strategy model reconstructs the value function and updates parameters to address the timing deviations in asynchronous execution. It constructs a continuous-time discounted value function for a semi-Markov decision process model using a variable step-size discount coefficient. The model then reconstructs the fixed discount factor into an exponentially decaying discount function based on the actual communication time interval. The formula for calculating the exponentially decaying discount function is as follows: .

[0103] In the above formula, This represents a variable step size discount factor based on the actual communication time interval; This represents the actual communication time interval between two event triggers; η represents the preset base decay rate, and satisfies η>0.

[0104] Based on this variable step size discount coefficient, the state value function (continuous-time cumulative reward + variable step size discount) is: .

[0105] The character-state Q-value function is: .

[0106] The time difference (TD) target value is: .

[0107] In the above formula, Indicates the current state. Indicates the actual communication time interval. The next state after; Indicates the current role. Indicates the target role in the next time step; This represents the action performed at time t. Indicates the action to be taken in the next moment; This represents the immediate reward function generated by a state-action pair in the continuous time domain. The expression represents the expected value operation; `max` represents the maximum value operation. This represents the overall action value function calculated based on the target network.

[0108] The heterogeneous behavior collaboration strategy model, based on a variable step-size discount coefficient, performs a time-decay-based weighted integral operation on the single-step execution reward within the communication quiescent interval to calculate the core term of the cumulative single-step reward used for updating the target Q-value. The formula for synthesizing the cumulative single-step reward is as follows: .

[0109] In the above formula, This indicates the core item for cumulative single-step rewards; This indicates the total number of high-frequency single-step actions within the current communication silence interval; Indicates the current single-step action step number, and satisfies ; Indicates the first The single-step time interval; Indicates the first The single-step execution reward.

[0110] The heterogeneous behavior collaborative strategy model addresses the multi-timescale credit allocation problem by using a cumulative single-step reward synthesis formula that assigns smaller weights to micro-actions further from the decision-making time. To balance physical motion safety and collaborative task benefits, the heterogeneous behavior collaborative strategy model adjusts the single-step execution reward... The multi-objective weighted synthesis formula is as follows: .

[0111] In the above formula, where, This represents the macro-level task reward, used to characterize the positive feedback of heterogeneous intelligent agents advancing towards the target and completing search and rescue missions. It represents physical safety penalties, used to characterize negative feedback when heterogeneous intelligent agents experience collisions, overspeeding, or abnormal postures; Indicates the weight of macro-level tasks; This represents the physical security weight. To ensure the absolute execution security of the system, the heterogeneous behavior collaboration strategy model pre-sets the physical security weight to be strictly greater than the macro-task weight, i.e., it satisfies... .

[0112] During the model training phase, the heterogeneous behavior collaboration strategy model expands and reconstructs the data structure of the experience replay pool. Specifically, the heterogeneous behavior collaboration strategy model transforms the traditional quintuple... Upgraded to a six-tuple that includes the dimension of actual communication time intervals. This includes adding the actual time interval between the current decision and the previous decision. During training, the network extracts and reads this actual time interval. And calculate the corresponding variable step size discount factor in real time. This ensures that the discount factor absolutely matches the actual time sequence.

[0113] During model training sampling, the heterogeneous behavior collaboration strategy model calculates and sets sampling weights that decay over time based on real communication time intervals. The formula for calculating the sampling weights is: .

[0114] in, This represents the sampling weight that decays over time; This represents the preset sampling attenuation coefficient, and satisfies... The heterogeneous behavior cooperative strategy model sets sampling weights that decay over time, so that samples with longer decision intervals and more network noise have lower sampling weights. This allows for weighted random sampling during replay pooling, thereby physically suppressing the interference of samples with large temporal deviations on training and ensuring algorithm convergence.

[0115] S300: When the information entropy reaches the first triggering condition or the feature mutation metric reaches the second triggering condition, the communication silence is broken, distributed autonomous coordination is carried out through the wireless communication network 300, and physical control commands are output to the physical actuator to complete the coordinated action.

[0116] The hysteresis comparator 404 in the heterogeneous behavior cooperation strategy model generates a hardware interrupt and a global broadcast. When the first trigger condition or the second trigger condition is met, the hysteresis comparator 404 in the heterogeneous behavior cooperation strategy model generates a hardware interrupt signal to break the communication silence, and connects to the wireless communication network 300 through the communication interface 301 to perform a global broadcast, thereby enabling the first heterogeneous intelligent agent 100 and the second heterogeneous intelligent agent 200 to enter a distributed autonomous cooperative state.

[0117] In the heterogeneous behavior cooperative strategy model, the action strategy module 401 performs role reselection and action reconstruction. In the distributed autonomous cooperative state, the processors 302 of the first airborne computing unit 101 and the second airborne computing unit 201 perform calculations and access globally synchronized observation data. The action strategy module 401 in the heterogeneous behavior cooperative strategy model performs an inner product matrix operation on its own observation encoding and the pre-trained heterogeneous agent role encoding, calculating and selecting the target upper-level role that maximizes the current action value function. Subsequently, the action strategy module 401 in the heterogeneous behavior cooperative strategy model concatenates its own observation encoding with the confidence-corrected attention weight matrix, dynamically reconstructing and selecting the corresponding lower-level role under the constraint of the target upper-level role. Based on the determined lower-level role, the action strategy module 401 in the heterogeneous behavior cooperative strategy model outputs two-dimensional driving commands including acceleration, deceleration, and steering to the control chassis 104, or three-dimensional flight commands including three-dimensional spatial displacement and yaw rotation to the flight control system 203. By reselecting roles and updating commands, the physical pauses caused by the difference in control frequencies between the control chassis 104 and the flight control system 203 are improved, thereby enhancing the smoothness of operation and the robustness of control of the air-ground cooperative system.

[0118] The system's control robustness was verified in a simulation environment. In a simulation environment with a multipath fading model and line-of-sight occlusion, delay parameters and random packet loss rate parameters for 300 nodes in the wireless communication network were configured. In a comparative scheme running a centralized map distribution strategy, the first heterogeneous agent 100 failed to acquire the global grid map within the control cycle, and the local path planner received outdated environmental perception data. The Long Short-Term Memory network processed this outdated data, causing a shift in the attention weight matrix elements. The first heterogeneous agent 100 output a steering angle command based on the shifted weights, causing the vehicle's outline to enter the obstacle expansion region. The second heterogeneous agent 200, due to not receiving the state feedback frame from the first heterogeneous agent 100 within the control cycle, experienced an increased calculation cycle in the flight control algorithm, resulting in discontinuous rotor speed output. When running this system, the airborne cooperative strategy model calculated that the current communication delay exceeded the set adaptive time threshold, and the finite state machine switched the system state to a communication silence mode. The first heterogeneous agent 100 activates the feature mapping bridging module 405, substitutes the equations of the dual-wheel differential kinematic model to deduce the vehicle's pose, and outputs a virtual observation vector. In the multi-objective reward function, the coefficient of the safety obstacle avoidance penalty term is set to be greater than the coefficient of the target-oriented reward term. The reinforcement learning algorithm synthesizes weights based on this reward function and outputs braking control quantities, causing the first heterogeneous agent 100 to stop driving and maintain a parking state. The second heterogeneous agent 200 calls local inertial navigation data to maintain the calculation of the multi-rotor hovering control law.

[0119] Preferably, the performance evaluation data of this application shows the system's technical indicators under communication-constrained conditions. When the information entropy of the intent probability distribution and the cosine similarity of local observation features do not reach a set threshold, the radio frequency module is in a dormant state. Compared with the fixed-period state broadcast mechanism, the system reduces the average channel occupancy rate of the wireless communication network 300. The system suspends network communication threads, reduces the tensor calculation load of the airborne graphics processor, reduces the overall inference power consumption of the device, and increases battery power supply time. In the Monte Carlo simulation experiment with set network packet loss rate parameters, the system executes a write protection instruction to lock the memory pages of the data buffer 303, the feature mapping bridging module 405 outputs the extrapolated virtual observation tensor, and the collision avoidance test indicators show that the system maintains a safe distance between the first heterogeneous intelligent agent 100 and the obstacle. The angular velocity data collected by the airborne inertial measurement unit shows that within the control cycle after the network link is re-established, the system substitutes preset coefficients to perform observation tensor fusion calculation. The measured value of the angular acceleration variance in the yaw direction of the second heterogeneous intelligent agent 200 decreased, the system reduced the abrupt change amplitude of the rotor control signal, and controlled the mechanical vibration amplitude of the second heterogeneous intelligent agent 200 during the network recovery phase.

[0120] Based on the technical defects in CN109000664A, where the aerial robot standby wake-up mechanism triggered by a preset fixed-time command may lead to communication congestion and high-frequency kinematic state loss of synchronization, and the objective problem in the comparative document CN117519171A, where the system relying on continuous high-bandwidth communication (SAM algorithm and AMCL algorithm) is prone to control decision distortion under weak communication conditions such as instantaneous packet loss, the embodiments of the present invention fully extend the multi-dimensional correlation technology features of timing correction and dynamic silent decision-making based on intent recognition and feature mutation.

[0121] To address the high latency issue in weak communication networks, this embodiment constructs a high-precision global timing alignment and dynamic confidence assessment mechanism. The first airborne computing unit 101 and the second airborne computing unit 201 receive PPS (Pulse Per Second) signals through built-in BeiDou or GPS receiver modules, achieving precise synchronization of the global clock reference. The processor 302 extracts the timestamp from the MAC layer hardware register of the Ethernet or wireless network card 301 to accurately calculate the actual communication latency difference. During intent recognition, the system dynamically calculates a monotonically decreasing exponential penalty factor based on this actual latency difference and directly multiplies it by the original attention score matrix of the deep neural network for weight reduction and renormalization. This hardware-software coupling mechanism effectively suppresses the interference of lagging historical data on intent prediction, ensuring the robustness of intent recognition.

[0122] Specifically, the processors 302 of the first airborne computing unit 101 and the second airborne computing unit 201 independently run the heterogeneous behavior cooperation strategy model. Under relatively stable spatial physical conditions, when the Shannon information entropy of the intent recognition probability distribution does not reach a preset entropy threshold (first trigger condition), and the cosine similarity of adjacent time steps output by the timing coding module 402, after filtering, does not meet the low-threshold judgment condition of the hysteresis comparator 404 (second trigger condition), the heterogeneous behavior cooperation strategy model maintains communication silence in the wireless communication network 300, and each agent independently runs the action strategy module 401 to output continuous dynamic commands. This mechanism avoids ineffective global high-frequency broadcasting, effectively saving channel bandwidth and reducing the energy consumption of the airborne computing units. However, in the event of a sudden change in the global situation, as long as any of the above trigger conditions are met, the system will generate a hardware interrupt signal to break the network communication silence, broadcast the global status via Ethernet or wireless network card 301, and quickly restore the distributed autonomous cooperative state.

[0123] To address the technical contradictions of open-loop deadlock and control divergence caused by the failure of physical state extrapolation in weak communication conditions, this embodiment reconstructs the fault-tolerant degradation and safety recovery mechanism with dynamics. When instantaneous packet loss occurs in network transmission and the arrival delay is greater than or equal to the adaptive judgment threshold, the finite state machine 403 intercepts and suspends the original input path, triggering the feature mapping bridging module 405 to perform state extrapolation based on kinematic differential equations and generate virtual observation feature vectors. During communication disconnection, the boundary protection and recovery module 406 performs an absolute safety judgment on the effectiveness of state extrapolation based on preset time window limits and feature deviation thresholds. Under extreme constraints exceeding the effective extrapolation range (e.g., the cumulative extrapolation time reaches the maximum allowable extrapolation time window, or the feature deviation reaches the preset deviation threshold), the action strategy module 401 receives the degradation command from the boundary protection and recovery module 406 and controls the control chassis 104 with low-frequency characteristics or the flight control system 203 with high-frequency characteristics to perform on-the-spot parking, hovering, or local obstacle avoidance actions, thereby effectively avoiding the physical collision hazards caused by timing discrepancies. When the wireless communication network 300 recovers and the transmission delay converges back to within the adaptive judgment threshold, the feature mapping bridging module 405 adopts a linear weighted fusion mechanism to gradually and seamlessly replace the virtual observation feature vector with the newly arrived real observation feature vector through a transition smoothing function, ensuring that the control commands sent to the physical actuator are continuous and smooth.

[0124] It should be noted that the specific embodiments described above are exemplary. Those skilled in the art can devise various solutions inspired by the disclosure of this invention, and these solutions all fall within the scope of this invention and its protection. Those skilled in the art should understand that this specification and its accompanying drawings are illustrative and not intended to limit the scope of the claims. The scope of protection of this invention is defined by the claims and their equivalents. This specification contains multiple inventive concepts; terms such as "preferredly," "according to a preferred embodiment," or "optionally" indicate that the corresponding paragraph discloses an independent concept. The applicant reserves the right to file divisional applications based on each inventive concept.

Claims

1. An air-ground collaborative rescue system based on intent recognition, comprising at least one first heterogeneous intelligent agent (100), at least one second heterogeneous intelligent agent (200), and a wireless communication network (300); the first heterogeneous intelligent agent (100) and the second heterogeneous intelligent agent (200) are connected through the wireless communication network (300), and both operate a heterogeneous behavior cooperation strategy model; characterized in that, The heterogeneous behavior collaboration strategy model is configured as follows: Acquire observational data, calculate the information entropy representing the deterministic nature of the intended prediction, and the characteristic mutation metric representing the risk of drastic changes in the local environment; When the information entropy does not reach the first triggering condition and the feature mutation metric does not reach the second triggering condition, communication silence is maintained, and each heterogeneous intelligent agent independently generates action distribution based on local dimensionality reduction observation features and outputs control commands to the physical actuator. When a transient packet loss is detected that causes a temporal input discontinuity, each heterogeneous agent extrapolates the virtual observation feature vector based on the historical hidden layer state and the physical kinematic equations. When the information entropy reaches the first triggering condition, or the feature mutation metric reaches the second triggering condition, the communication silence is broken, a global state broadcast is performed through the network to restore distributed collaboration, and physical control commands are output to the physical actuators to complete the collaborative action.

2. The system according to claim 1, characterized in that, The heterogeneous behavior collaboration strategy model acquires observation data in the following ways: Sliding window statistics are performed based on round-trip time delay to calculate an adaptive decision threshold; When the arrival delay of a data packet is detected to be greater than or equal to the judgment threshold, the original input path is intercepted and suspended using a finite state machine (403); The historical hidden layer states extracted from the heterogeneous intelligent agent are substituted into the physical kinematic equations to calculate the virtual observation feature vector; The virtual observation feature vector is input into the time-series coding module (402).

3. The system according to claim 1 or 2, characterized in that, The heterogeneous behavior collaboration strategy model performs intent recognition in the following ways: Extract hardware timestamps to calculate communication latency time differences; Based on the communication delay time difference, an exponential penalty factor is determined; Calculate the original attention score matrix, multiply the exponential penalty factor into the original attention score matrix, and then perform renormalization calculation through Softmax to output the attention weight matrix.

4. The system according to any one of claims 1 to 3, characterized in that, The information entropy is Shannon information entropy; the methods for determining whether the triggering condition has been met and for breaking the communication silence include: The Shannon information entropy is logically compared with a preset entropy threshold to determine whether the first triggering condition has been met. The cosine similarity of the observation coding vectors of adjacent time steps is filtered by moving average; the filtered cosine similarity is input into the hysteresis comparator (404) for judgment. When the cosine similarity after moving average filtering meets the low threshold judgment condition, the judgment reaches the second trigger condition. When the first or second trigger condition is met, a hardware interrupt signal is generated to break the communication silence and a global broadcast is performed.

5. The system according to any one of claims 1 to 4, characterized in that, During the period of communication silence maintained by the wireless communication network (300), each heterogeneous intelligent agent independently runs the action strategy module (401), specifically including: Extracting local dimensionality-reduced observation features; Obtain the pre-defined heterogeneous agent role codes as hard constraint bias terms; Based on the local dimensionality reduction observation features and the hard constraint bias term, the action strategy module (401) is independently driven to generate the action distribution.

6. The system according to any one of claims 1 to 5, characterized in that, The heterogeneous behavior cooperative strategy model is constructed based on multi-agent reinforcement learning. The value function calculation and training methods of the heterogeneous behavior cooperative strategy model include: A continuous-time discount value function for a semi-Markov decision process model is constructed using a variable step-size discount coefficient. Based on the variable step size discount coefficient, the single-step execution reward within the communication silence interval is weighted by a time decay to calculate the target Q value; During the model training phase, the data structure of the experience replay pool is expanded to include a six-tuple containing the dimension of the actual communication time interval. During sampling, sampling weights that decay over time are calculated and set based on the actual communication time interval.

7. The system according to any one of claims 1 to 6, characterized in that, When the heterogeneous behavior collaboration strategy model calculates the communication delay time difference, the calculation method for the communication delay time difference includes: Synchronization of the global clock reference is achieved by using the synchronization pulse signal output by the global satellite navigation system; Extract the hardware timestamp from the MAC layer register of the communication interface (301), the hardware timestamp including the receive timestamp and the send timestamp; The hardware timestamp is linearly corrected based on the clock frequency offset to calculate the actual communication delay time difference.

8. The system according to any one of claims 1 to 7, characterized in that, The heterogeneous behavior collaboration strategy model also includes a feature mapping bridging module (405) for calculating virtual observation feature vectors. The feature mapping bridging module (405) includes: The physical state extraction unit is used to obtain the true kinematic state; Cross-domain mapping units are used to align high-dimensional, implicit mathematical features with low-dimensional, explicit physical models driven by mechanisms, using linear projection matrices and biases. The joint inference unit is used to fuse the mapped features and combine them with kinematic differential equations to predict the physical state; An observation generation unit is used to reverse-map the predicted physical state back to the observation feature space to generate the virtual observation feature vector.

9. The system according to any one of claims 1 to 8, characterized in that, The heterogeneous behavior collaboration strategy model is also configured with a boundary protection and recovery module (406), which is configured as follows: The validity of the virtual observation feature vector is determined based on a preset time window limit and feature deviation threshold. When the effective range is exceeded, the current heterogeneous intelligent agent role is frozen and only local obstacle avoidance and attitude maintenance actions are performed, or the cooperative task is terminated and the heterogeneous intelligent agent is controlled to park or hover in place. Once the wireless communication network (300) is restored, the virtual observation feature vector is gradually replaced by linear weighted fusion, and the continuity of instructions is ensured by a transition smoothing function.

10. A method for air-ground coordinated rescue based on intent recognition, characterized in that, The method is applied to a system comprising at least one first heterogeneous intelligent agent (100), at least one second heterogeneous intelligent agent (200), and a wireless communication network (300), wherein the first heterogeneous intelligent agent (100) and the second heterogeneous intelligent agent (200) are connected through the wireless communication network (300) and both run a heterogeneous behavior cooperation strategy model; the method is executed by the processors of the first heterogeneous intelligent agent (100) and the second heterogeneous intelligent agent (200), and includes: Acquire observational data, calculate the information entropy representing the deterministic nature of the intended prediction, and the characteristic mutation metric representing the risk of drastic changes in the local environment; When the information entropy does not reach the first triggering condition and the feature mutation metric does not reach the second triggering condition, communication silence is maintained, and each heterogeneous intelligent agent independently generates action distribution based on local dimensionality reduction observation features and outputs control commands to the physical actuator. When a transient packet loss is detected that causes a temporal input discontinuity, each heterogeneous agent generates a virtual observation feature vector based on the historical hidden layer state and physical kinematic equations extrapolated from the physical kinematic equations. When the information entropy reaches the first triggering condition, or the feature mutation metric reaches the second triggering condition, the communication silence is broken, a global state broadcast is performed through the network to restore distributed collaboration, and physical control commands are output to the physical actuators to complete the collaborative action.

Images