A multi-path transmission system of unmanned aerial vehicle ad hoc network cross-layer coupling prediction

Abstract

The application discloses a kind of unmanned aerial vehicle self-organizing network cross-layer coupling prediction multipath transmission system.The system includes: sensor module and main control board card, sensor module gathers wind field three-dimensional vector data and unmanned aerial vehicle three-axis acceleration and angular velocity, the multi-core processor of main control board card calculates wind field disturbance compensation acceleration and node future three-dimensional coordinate prediction vector, and link retention time is output through space-time graph convolution network model, and programmable logic gate array activates multipath distribution mechanism when main path link retention time is lower than early warning threshold, and redundant encoding fragments are sent to standby path according to exponential rise ratio.The application can overcome network prediction lag defects, reduce data loss rate under strong wind disturbance, and protect uninterrupted stable transmission of key video stream.

Description

Technical Field

[0001] This application relates to the field of mobile ad hoc networks and emergency communication technology, and specifically to a multipath transmission system for cross-layer coupling prediction in UAV ad hoc networks. Background Technology

[0002] In highly challenging operational scenarios such as earthquake emergency rescue in canyon areas, UAV ad hoc networks undertake the core tasks of real-time video transmission and command and dispatch control. The unique geographical environment of canyons presents multiple physical spatial constraints and electromagnetic interference to the communication network of UAV formations. Traditional UAV ad hoc network routing protocols typically assess link connectivity based on received signal strength indicators or the packet loss rate of heartbeat messages between nodes. When UAVs fly into areas obscured by canyon walls or complex terrain, radio frequency signals experience unpredictable and rapid fading. Due to the strict layered independence of traditional open systems interconnection protocol stacks, the network layer often experiences a significant delay after receiving continuous packet loss feedback from the physical layer before passively triggering route rediscovery and topology reconstruction mechanisms. By this time, the physical transmission link is often completely broken, directly causing transmission interruptions of high-frequency, high-definition video streams for tens to hundreds of milliseconds, severely impacting the continuity of footage at the emergency rescue site.

[0003] Strong, sudden wind shear and irregular airflow are common in deep mountain valleys. Conventional techniques typically rely on coordinate sequences provided by airborne GPS devices for predicting future network topology. However, under strong airflow disturbances, drones can experience severe yaw and unexpected spatial displacement. GPS devices typically have a refresh rate of only 1 Hz to 5 Hz, and their low sampling rate and centimeter- to meter-level positioning accuracy are simply insufficient to accurately capture the instantaneous, highly dynamic physical drift caused by wind shear in drones. Topology prediction models based on linear trajectory extrapolation using this low-frequency and lagging coordinate data have spatiotemporal errors. Ultimately, this can lead to a dangerous situation where the primary and backup communication links, pre-calculated by multipath transmission algorithms, may simultaneously fail during actual high-dynamic flight, failing to guarantee the smooth migration of underlying data streams in extreme environments. Summary of the Invention

[0004] This application provides a multipath transmission system for cross-layer coupling prediction in UAV ad hoc networks, which solves the problem of transmission interruption caused by the lag in topology change prediction and slow convergence of multipath routes in existing UAV ad hoc networks under strong wind and complex obstruction environments.

[0005] This application provides a multipath transmission system for cross-layer coupling prediction in an unmanned aerial vehicle (UAV) ad hoc network, deployed in a flight formation consisting of multiple UAV nodes. Each UAV node carries communication and computing payloads, including: The sensor module, including a three-axis pitot tube and an inertial measurement unit, is configured to acquire three-dimensional vector data of the environmental wind field and high-frequency three-axis acceleration and angular velocity of the UAV; The main control board includes a multi-core processor and a programmable gate array, which are connected to and receive data collected by the sensor module via a bus. The multi-core processor is configured to calculate wind field disturbance compensation acceleration based on the three-dimensional vector data, determine the UAV's reference dynamic state based on the UAV's high-frequency three-axis acceleration and angular velocity, and calculate the node's three-dimensional coordinate prediction vector in the future time window based on the wind field disturbance compensation acceleration and the UAV's reference dynamic state. The multi-core processor is also configured to broadcast the three-dimensional coordinate prediction vector through the radio frequency transceiver front end and receive the corresponding coordinate prediction vector of the neighboring nodes, input the integrated whole network sequence into the spatiotemporal graph convolutional network model, and output the link hold-up time of each edge; The programmable gate array is configured to activate a multipath allocation mechanism when the link hold time on the current main transmission path is lower than a preset warning threshold. Before the physical signal of the main path is disconnected, redundant encoded fragments of the data stream are simultaneously sent to the main path and a non-intersecting backup path. The proportion of data packets allocated to the backup path increases exponentially with the decay of the link hold time.

[0006] In a preferred embodiment, the configuration of the three-dimensional coordinate prediction vector of the multi-core processor computing node in the future time window specifically includes: calculating the three-axis wind pressure disturbance force of the external wind field in the body coordinate system based on the three-dimensional vector data collected by the three-axis Pitot tube, combined with the preset aerodynamic drag coefficient and the windward area parameter of the UAV; dividing the three-axis wind pressure disturbance force by the current takeoff weight mass parameter of the UAV to obtain the wind field disturbance compensation acceleration; superimposing the wind field disturbance compensation acceleration on the three-dimensional translational acceleration of the UAV's reference dynamic state, and updating the cumulative spatial displacement based on the attitude angle obtained by the inertial measurement unit to output the three-dimensional coordinate prediction vector.

[0007] In a preferred embodiment, the process of updating the cumulative spatial displacement based on the attitude angles obtained by the inertial measurement unit and outputting the three-dimensional coordinate prediction vector is configured to perform the following operations: extracting the attitude angle parameters of the current epoch, constructing a cosine rotation matrix from the body coordinate system to the navigation coordinate system; using the cosine rotation matrix to perform coordinate system transformation on the three-dimensional total acceleration vector after superimposing the wind field disturbance compensation acceleration, obtaining the actual acceleration vector in the navigation coordinate system; performing a difference operation between the actual acceleration vector and the gravity acceleration vector, filtering out the static gravity component, and extracting the pure motion acceleration component; within the width of the set future time window, using the second-order Runge-Kutta integration algorithm to perform double numerical integration on the pure motion acceleration component, with the integration step size configured as a microsecond-level constant matching the high-frequency sampling period of the inertial measurement unit, and appending the calculated three-dimensional displacement offset to the current absolute spatial coordinates to generate the three-dimensional coordinate prediction vector.

[0008] In a preferred embodiment, the spatiotemporal graph convolutional network model includes a spatiotemporal feature fusion module and a hold-time regression module connected in sequence. The spatiotemporal feature fusion module is configured to construct a graph adjacency matrix based on relative Euclidean distance, extract the spatial topological correlation features between each UAV node in the entire network sequence through multi-layer graph convolution operations, and process the evolution trajectory of node coordinates along the time dimension using a one-dimensional dilated convolution kernel. The hold-time regression module is configured to perform nonlinear mapping through a fully connected feedforward network based on the extracted spatiotemporal joint feature vector, and output the remaining available time between the current connected node pairs until the radio frequency signal strength decays to the disconnection threshold as the link hold-time.

[0009] In a preferred embodiment, the spatiotemporal feature fusion module constructs a weight update mechanism for a graph adjacency matrix based on relative Euclidean distance by performing the following operations: calculating the dynamic spatial physical distance and relative approach speed between any two communicating nodes based on the received coordinate prediction vectors of the neighboring nodes; defining an initial connected Boolean value based on a preset maximum communication coverage radius, and determining the connected Boolean value as true if the dynamic spatial physical distance is less than the maximum communication coverage radius; calculating a spatial weight coefficient for edges determined to be true, wherein the spatial weight coefficient is configured as the product of a distance decay term and a speed penalty term, wherein the distance decay term is calculated based on the ratio of the difference between the maximum communication coverage radius and the dynamic spatial physical distance to the maximum communication coverage radius, and the speed penalty term is determined by using a natural logarithmic decay function with a negative relative approach speed as the exponent; and generating the graph adjacency matrix by filling the spatial weight coefficient into the corresponding elements of the node relationship graph.

[0010] In a preferred embodiment, the multipath allocation mechanism configured by the programmable gate array specifically includes: maintaining a single primary path data transmission mode when the link hold time is greater than or equal to the preset warning threshold; when the link hold time falls below the preset warning threshold, triggering the underlying hardware split channel, reading the node sequence of the backup path according to the set source routing protocol table, starting the forward error correction coding module to redundantly expand the original video data stream to generate the redundant coding fragments, and distributing the redundant coding fragments to the primary path and the backup path at a proportional line rate before the physical link actually drops.

[0011] In a preferred embodiment, the configuration of the proportion of data packets allocated to the backup path increasing exponentially with the decay of the link hold time specifically includes: configuring a basic traffic splitting start ratio and a maximum traffic splitting upper limit ratio; calculating the difference between the preset warning threshold and the currently updated link hold time to obtain the warning time depth; constructing a natural base exponential function term based on the warning time depth, wherein the coefficient is a pre-calibrated traffic splitting aggressive coefficient constant; multiplying the basic traffic splitting start ratio by the natural base exponential function term to obtain a dynamic allocation ratio target value; when the dynamic allocation ratio target value exceeds the maximum traffic splitting upper limit ratio, performing hardware-level truncation processing on the dynamic allocation ratio target value, forcibly setting it to the maximum traffic splitting upper limit ratio, and controlling the hardware flow table to distribute the redundant coded fragments accordingly.

[0012] This application also provides a multipath transmission method for UAV ad hoc networks based on multi-source sensor cross-layer coupling prediction, used to configure the communication and computing payloads in a flight formation consisting of multiple UAV nodes, including: Three-dimensional vector data of the environmental wind field and high-frequency three-axis acceleration and angular velocity of the UAV were collected, using a three-axis Pitot tube and an inertial measurement unit. The wind field disturbance compensation acceleration is calculated based on the three-dimensional vector data. The reference dynamic state of the UAV is determined based on the high-frequency three-axis acceleration and angular velocity of the UAV. Based on the wind field disturbance compensation acceleration and the reference dynamic state of the UAV, the three-dimensional coordinate prediction vector of the node in the future time window is calculated. The radio frequency transceiver front end is controlled to broadcast the three-dimensional coordinate prediction vector and receive the corresponding coordinate prediction vector of neighboring nodes. The integrated whole network sequence is input into the spatiotemporal graph convolutional network model, and the link hold-up time of each edge is output. When the link hold time on the current main transmission path is lower than a preset warning threshold, the multipath allocation mechanism is activated. Before the physical signal of the main path is disconnected, the redundant encoded fragments of the data stream are sent to the main path and the non-intersecting backup path at the same time. The proportion of data packets allocated to the backup path increases exponentially as the link hold time decreases.

[0013] This application also provides a communication and computing payload device, including a memory and a multi-core processor. The memory stores computer execution instructions, and the multi-core processor is configured to execute the above-described UAV ad hoc network multipath transmission method based on multi-source sensor cross-layer coupling prediction when running the computer execution instructions.

[0014] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the above-described UAV self-organizing network multipath transmission method based on multi-source sensor cross-layer coupling prediction.

[0015] This application overcomes the inherent defects of traditional network mechanisms, such as prediction lag and slow recovery after disconnection, by directly transmitting airborne physical sensor data to the network layer in real time for high-frequency dynamic compensation and performing a smooth traffic transfer with a disconnection-then-reconnection mechanism at the programmable hardware layer. This reduces the data loss rate under strong crosswind disturbances and ensures uninterrupted and stable transmission of video streams for critical missions. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall structure of the system provided in the embodiments of this application.

[0017] Figure 2 This is a logical structure block diagram of the main control board and internal modules provided in the embodiments of this application.

[0018] Figure 3 This is a flowchart of the method provided in the embodiments of this application.

[0019] Figure 4 This is a structural block diagram of the spatiotemporal graph convolutional network model provided in the embodiments of this application.

[0020] Figure 5 This is a schematic diagram of the structure of the communication and computing payload device provided in the embodiments of this application.

[0021] Explanation of reference numerals in the attached figures: In the diagram: 100-System, 101-UAV node, 102-Main control board, 103-Sensor module, 104-Triaxial Pitot tube, 105-Inertial measurement unit, 106-RF transceiver front end, 107-Rigid probe, 108-Rigid support base, 109-Nonlinear silicone vibration damping layer, 201-Multi-core processor, 202-Programmable gate array, 203-Ultra-high-speed internal bus, 204-Forward error correction coding module, 401-Spatiotemporal feature fusion module, 402-Hold time regression module, 500-Communication and computing payload equipment, 501-Memory. Detailed Implementation

[0022] The following will describe in detail the UAV self-organizing network multipath transmission system based on multi-source sensor cross-layer coupling prediction provided in the embodiments of this application, with reference to the accompanying drawings and examples.

[0023] refer to Figure 1 , Figure 1 This is a schematic diagram of the overall structure of the system provided in this embodiment. The system 100 provided in this embodiment is deployed in a flight formation consisting of multiple UAV nodes. Each UAV node 101 carries an independent communication and computing payload hardware system. This hardware system serves as the foundation for supporting high-frequency topology prediction and includes a main control board 102 located at the core of the fuselage, sensor modules 103 arranged on the periphery, and a radio frequency transceiver front-end 106 for radio communication. The main control board 102 is the computing hub of the entire payload, responsible for handling all cross-layer interactions of physical data and network routing. The system 100 also includes a power management module that provides stable operating voltage for each component, a clock circuit module for supporting the high-frequency clock of the on-chip system, and peripheral storage circuitry to ensure large-scale data caching. Through this complete electronic support architecture, the system 100 constitutes an independent unit capable of performing high-load computing.

[0024] The sensor module 103 comprises two core sensing devices: a three-axis pitot tube 104 physically connected to the exterior of the UAV fuselage and a microelectromechanical system (MEMS) inertial measurement unit 105 directly mounted near the main control board 102. The three-axis pitot tube 104 is configured to acquire real-time wind speed and direction vector data within the three-dimensional spatial environment wind field, used to sense sudden wind shear within the canyon. To avoid interference from the downwash airflow generated by the high-speed rotation of the UAV rotor on the natural wind field parameters, the three-axis pitot tube 104 is extended and fixed to the undisturbed flow field region at the far front end of the UAV node fuselage via a high-strength rigid probe 107. The inertial measurement unit 105 is configured to continuously output the UAV's current three-axis acceleration and angular velocity at a high-frequency sampling rate greater than 200 Hz. This configuration, exceeding the conventional refresh rate, allows instantaneous maneuvers to be captured immediately. To reduce the impact of mechanical resonance in the fuselage structure on high-precision components, the inertial measurement unit 105 and the main control board 102 are conformally integrated via a rigid support base 108, with a nonlinear silicone damping layer 109 with broadband absorption capability sandwiched between the rigid contact surfaces. The nonlinear silicone damping layer 109 can simultaneously suppress low-frequency wind swaying and high-frequency motor chatter. By employing this physical assembly method combining mechanical isolation and rigid extension, the signal-to-noise ratio of the source sensor data can be improved, thereby providing cleaner and more reliable dynamic data for the subsequent prediction algorithms of the network layers.

[0025] See Figure 2 , Figure 2 This is a block diagram of the logical structure of the main control board and internal modules provided in this application embodiment. The main control board 102 adopts a high-performance system-on-a-chip architecture, such as using an industrial-grade chip. It deeply integrates a multi-core processor 201 and a programmable gate array 202, which are directly connected and exchange data through an ultra-high-speed internal bus 203. The multi-core processor 201 is typically configured to run high-level topology prediction algorithms and routing algorithms, and is responsible for complex floating-point calculations; while the programmable gate array 202 is configured to execute high-frequency scheduling of the media access control layer protocol at the lower level, including data packet encapsulation and high-speed physical distribution of multi-path data frames. The multi-core processor 201 also has a built-in or coupled forward error correction coding module 204, which is used to generate redundant data in advance to resist packet loss when the channel quality deteriorates. The sensor module 103 directly transmits the collected hard real-time data to the main control board 102 through hardwired bus interfaces such as integrated circuit buses and serial peripheral interfaces. Meanwhile, the RF transceiver front-end 106 includes an orthogonal frequency division multiplexing transceiver operating in the 5.8 GHz band and a high-gain phased array antenna, configured to perform directional transmission and omnidirectional reception of physical layer wireless RF signals between nodes. By adopting a heterogeneous architecture of system-on-chip and a bus-level direct data transmission channel, the system breaks through the latency barrier caused by the traditional bottom-up, layer-by-layer delivery of protocol stacks, making millisecond-level routing switching response possible.

[0026] During the operation of system 100, the multi-core processor 201 performs node-level dynamic state estimation. The processor performs position offset deduction based on the high-frequency acceleration vector obtained by the inertial measurement unit 105 and the wind pressure disturbance vector provided by the three-axis pitot tube 104. Specifically, the system performs a three-dimensional coordinate update calculation including the influence of wind disturbance, that is, it solves for the predicted three-dimensional coordinate vector of the node in the future time window based on the time step, base velocity, flight control drive force, and wind disturbance. In a preferred embodiment, this core calculation logic is implemented through the following dynamic state equation formula:

[0027] Among them, physical variables Indicates the moment of the predicted future time window; variable Indicates at time Node 3D coordinate prediction vector; variables This represents the coordinate vector of the UAV's reference position at the current time t, which is initially provided by a slowly corrected reference from the onboard GPS module; variable Represents the spatial velocity vector of the UAV at the current moment; variable Indicates the discrete-time constant or integration step size for performing the derivation; variables This represents the three-dimensional translational acceleration in the UAV's baseline dynamic state, specifically the actual driving acceleration output by the flight control system after eliminating the influence of static gravity; variables This represents the wind field disturbance compensation acceleration. It should be noted that the refresh rate of GPS devices is typically only 1 Hz to 5 Hz, far lower than the 100 Hz level required for prediction. The integration process in this embodiment effectively fills the positioning blind spot and filters out wind drift error.

[0028] To obtain the wind field disturbance compensation acceleration in the formula The system divides the three-axis wind pressure disturbance force of the external wind field in the body coordinate system by the current takeoff weight of the UAV. The three-axis wind pressure disturbance force is derived from three-dimensional vector data collected by a three-axis Pitot tube, combined with preset aerodynamic drag coefficients and UAV frontal area parameters. After obtaining the wind field disturbance compensation acceleration, a cosine rotation matrix is ​​constructed to transform from the body coordinate system to the navigation coordinate system by extracting the attitude angle parameters of the current epoch. This rotation matrix is ​​used to transform the total three-dimensional acceleration superimposed with wind pressure effects to the navigation system, and then the gravity constant is filtered out by difference operation to extract the pure motion acceleration component. Subsequently, within the set future time window width T, a second-order Runge-Kutta integration algorithm is used for double numerical integration. The integration step size of this process is strictly configured as a microsecond-level constant matching the sampling rate of the inertial measurement unit. By employing microsecond-level fine-grained dual integral iteration, the system can accurately capture instantaneous crosswind impact displacement, correcting flight trajectory prediction drift caused by relying solely on a low-frequency GPS system.

[0029] In addition to internal node computation, the multi-core processor 201 encapsulates and broadcasts its own 3D coordinate prediction vector in periodic control frames via the radio frequency transceiver front-end 106, and synchronously receives corresponding coordinate prediction vectors from neighboring nodes. The integrated network-wide coordinate sequence is then fed as the input source to... Figure 4 The spatiotemporal graph convolutional network model shown is as follows. Figure 4 As shown, the spatiotemporal graph convolutional network model includes a spatiotemporal feature fusion module 401 and a time-preserving regression module 402, which are internally cascaded. In the core graph structure of the spatiotemporal graph convolutional network model, graph nodes correspond to each UAV, and graph edges correspond to each physical communication link. The spatiotemporal feature fusion module 401 extracts the network topology correlation along the spatial dimension through multi-layer graph convolution operations, and processes the evolution trajectory of node coordinates along the temporal dimension using a one-dimensional dilated convolution kernel to capture the changing trends of longer historical sequences without increasing the computational load.

[0030] Before extracting features, the spatiotemporal feature fusion module 401 constructs a graph adjacency matrix based on relative Euclidean distance. The weight update mechanism of the graph adjacency matrix is ​​configured as follows: calculate the dynamic spatial physical distance d and the relative approach speed v between any two communication nodes. Based on the preset maximum communication coverage radius R, when the dynamic spatial physical distance d is less than the maximum communication coverage radius R, the Boolean value of the two nodes is determined to be true. For real connection edges, calculate their spatial weight coefficients, which are obtained by multiplying the distance attenuation term and the speed penalty term. The distance attenuation term is determined based on the ratio of (Rd) to R; while the speed penalty term is determined by using the natural logarithmic attenuation function with the negative relative approach speed v as the exponent. These weight elements are filled to generate a complete graph adjacency matrix. Subsequently, the hold-time regression module 402 performs nonlinear mapping through a fully connected feedforward network and outputs the link hold-time of each edge. This link hold-time is defined as the remaining available time based on the current wind field disturbance and flight trajectory until the radio frequency signal strength attenuates to the disconnection threshold. By embedding an exponential decay penalty mechanism containing dynamic approach velocity v and physical boundary R into the adjacency matrix of the graph convolutional network, the system enables the network weights to detect drastic changes in the imminent departure of two drones from the communication range, thereby improving the accuracy of link breakage time warning.

[0031] The Programmable Gate Array (PGA) 202 relies on the link hold-up time parameters of each side output by the spatiotemporal graph convolutional network to perform a multipath allocation mechanism. Specifically, when the link hold-up time on the current main transmission path is greater than or equal to a preset warning threshold, the PGA 202 maintains a high-bandwidth, low-overhead single main path data transmission mode. When the link hold-up time is detected to fall below the warning threshold, the hardware scheduler of the PGA 202 immediately triggers a split channel at the underlying level. The scheduler queries the set source routing protocol table to select a backup path that is completely disjoint from the main path node. Before a physical link actually drops, the forward error correction coding module 204 is started to redundantly expand the original video data stream and distributes the generated redundant coded fragments to the main and backup dual paths at a dynamic proportional line rate.

[0032] The proportion of packets allocated to backup paths exhibits a strictly exponential increase as the link hold time approaches zero. The system sets the initial base traffic splitting ratio during operation to... P 0 Meanwhile, to prevent congestion on backup links, a maximum traffic offloading limit is set at [percentage missing]. P max The system obtains the difference between the preset warning threshold and the link hold time in real time as the warning time depth. Based on this time depth, a dynamic allocation calculation logic is constructed. In a preferred embodiment, this calculation logic is implemented through a decay formula that includes a natural base exponential function term:

[0033] Among them, physical variables P target This represents the target value for the dynamic allocation ratio that the underlying hardware should execute; a constant. P 0 This represents the initial proportion of traffic splitting when the link hold time just reaches the warning threshold; variable This indicates the warning time depth; the constant k is a pre-calibrated traffic diversion aggressiveness coefficient, and its larger value represents a faster rate of traffic transfer to the alternative path. When the dynamic allocation ratio target value obtained after scaling to the natural base exceeds the preset maximum diversion limit ratio... P max At that time, the programmable gate array 202 performs mandatory hardware-level truncation processing on it, locking it to... P max This controls the underlying high-speed flow table to distribute redundant fragments according to the corresponding bandwidth usage. The table below shows the transfer relationship of backup path quotas over time under a set of typical operating conditions:

[0034] By adopting an exponential underlying traffic migration mechanism controlled by a maximum upper limit, the majority of core data streams can be smoothly transitioned to a safe backup path in the last second before the link is predicted to break, thereby fundamentally eliminating the video stuttering and black screen phenomenon caused by disconnecting the link first and then addressing in the traditional system.

[0035] At another level, embodiments of this application also relate to Figure 3 The illustrated UAV ad hoc network multipath transmission method based on multi-source sensor cross-layer coupling prediction is used to configure the communication and computing payloads in a flight formation consisting of multiple UAV nodes. This method performs the following workflow steps: For step S301, three-dimensional vector data of the environmental wind field and high-frequency three-axis acceleration and angular velocity of the UAV are acquired. This process is completed by the three-axis Pitot tube and the inertial measurement unit working together to capture the high-frequency physical environment.

[0036] For step S302, the wind field disturbance compensation acceleration is calculated based on the three-dimensional vector data, the UAV's reference dynamic state is determined based on the UAV's high-frequency three-axis acceleration and angular velocity, and the three-dimensional coordinate prediction vector of the node in the future time window is calculated based on the wind field disturbance compensation acceleration and the UAV's reference dynamic state. This process includes drag conversion and dual Runge-Kutta integration processes to complete the calculation and accumulation of microsecond-level attitude shifts.

[0037] In step S303, the RF transceiver front-end is controlled to broadcast the three-dimensional coordinate prediction vector and receive the corresponding coordinate prediction vectors from neighboring nodes. The integrated network sequence is then input into the spatiotemporal graph convolutional network model, outputting the link uptime for each edge. This process regresses the remaining reliable uptime of the link by analyzing topological correlation weights such as relative proximity speed and coverage radius.

[0038] For step S304, when the link hold time on the current main transmission path is lower than a preset warning threshold, a multipath allocation mechanism is activated. Before the physical signal of the main path is disconnected, redundant encoded fragments of the data stream are simultaneously sent to the main path and a non-intersecting backup path. The proportion of data packets allocated to the backup path increases exponentially with the decay of the link hold time. This process allows control to be transferred to the underlying hardware flow table, completing rapid error correction and seamless breakpoint routing handover.

[0039] By employing coordinated scheduling steps across the hardware, algorithm, and protocol layers, the system constructs a network self-healing defense.

[0040] refer to Figure 5 , Figure 5This is a schematic diagram of the communication and computing payload device provided in an embodiment of this application. The communication and computing payload device 500 includes at least one memory 501 and a multiplexed multi-core processor 201. The memory 501 stores a rich set of computer execution instructions or firmware code. The multi-core processor 201 is deeply interconnected with programmable logic circuits and is configured to execute various algorithmic logics when running the computer execution instructions, thereby completing the prediction of future node displacement and adaptive path switching decisions under highly dynamic wind deflection disturbance resistance conditions.

[0041] In summary, the working principle provided in this application exhibits a distinct synergistic effect across technical levels: when facing narrow canyon environments filled with sudden strong crosswinds, the system completely breaks down the barriers between traditional physical electromechanical control and digital network routing. The underlying hard-connected sensor modules provide high-precision wind speed and acceleration information. Multi-core processors, relying on these parameters, establish dynamic compensation equations with wind shear resistance capabilities, calculating future trajectory points with accuracy far exceeding that of satellite positioning. The spatiotemporal graph convolutional network further thoroughly understands the entire network topology data, accurately defining the time points when each communication link is disconnected. Before the connection is lost, the high-speed programmable gate array (FPGA) preemptively transmits redundant data to backup safe routes at an exponentially increasing allocation rate. These core actions are interconnected, eliminating packet loss points caused by topology perception lag, and ensuring the continuity and high availability of high-definition video streams in the emergency rescue network under extremely challenging environments.

[0042] It should be noted that the specific values ​​or platform names regarding system-on-a-chip models, frequency bands, sampling rates, and warning thresholds in the embodiments are merely examples to explain the preferred embodiments of this application. Those skilled in the art can achieve the design objective of preventing sudden channel fading by using other heterogeneous computing platforms with similar architectures, adjusting specific environmental monitoring frequencies or time windows, and such alternative solutions should all be covered within the scope of protection of this application.

Claims

1. A multipath transmission system for cross-layer coupling prediction in an unmanned aerial vehicle (UAV) ad hoc network, deployed in a flight formation consisting of multiple UAV nodes, each UAV node carrying communication and computing payloads, characterized in that... include: The sensor module, including a three-axis pitot tube and an inertial measurement unit, is configured to acquire three-dimensional vector data of the environmental wind field and high-frequency three-axis acceleration and angular velocity of the UAV; The main control board includes a multi-core processor and a programmable gate array, which are connected to and receive data collected by the sensor module via a bus. The multi-core processor is configured to calculate wind field disturbance compensation acceleration based on the three-dimensional vector data, determine the reference dynamic state of the UAV based on the high-frequency three-axis acceleration and angular velocity of the UAV, and calculate the three-dimensional coordinate prediction vector of the node in the future time window based on the wind field disturbance compensation acceleration and the reference dynamic state of the UAV. The multi-core processor is also configured to broadcast the three-dimensional coordinate prediction vector through the radio frequency transceiver front end and receive the corresponding coordinate prediction vector of the neighboring nodes, input the integrated whole network sequence into the spatiotemporal graph convolutional network model, and output the link hold-up time of each edge; The programmable gate array is configured to activate a multipath allocation mechanism when the link hold time on the current main transmission path is lower than a preset warning threshold. Before the physical signal of the main path is disconnected, redundant encoded fragments of the data stream are simultaneously sent to the main path and a non-intersecting backup path. The proportion of data packets allocated to the backup path increases exponentially with the decay of the link hold time.

2. The UAV self-organizing network multipath transmission system based on multi-source sensor cross-layer coupling prediction as described in claim 1, characterized in that, The configuration of the three-dimensional coordinate prediction vector for the multi-core processor computing node in the future time window specifically includes: Based on the three-dimensional vector data collected by the three-axis Pitot tube, and combined with the preset aerodynamic drag coefficient and the UAV's frontal area parameters, the three-axis wind pressure disturbance force of the external wind field in the body coordinate system is calculated. Divide the three-axis wind pressure disturbance force by the current takeoff weight mass parameter of the UAV to obtain the wind field disturbance compensation acceleration; The wind field disturbance compensation acceleration is superimposed on the three-dimensional translational acceleration of the UAV's baseline dynamic state, and the accumulated spatial displacement is updated based on the attitude angle obtained by the inertial measurement unit to output the three-dimensional coordinate prediction vector.

3. The UAV self-organizing network multipath transmission system based on multi-source sensor cross-layer coupling prediction as described in claim 2, characterized in that, The process of updating the accumulated spatial displacement based on the attitude angle obtained by the inertial measurement unit and outputting the predicted three-dimensional coordinate vector is configured to perform the following operations: Extract the attitude angle parameters of the current epoch and construct a cosine rotation matrix from the body coordinate system to the navigation coordinate system; The coordinate system is transformed by the cosine rotation matrix to obtain the actual acceleration vector in the navigation coordinate system after superimposing the wind field disturbance compensation acceleration; The difference between the actual acceleration vector and the gravitational acceleration vector is calculated, and the static gravity component is filtered out to extract the pure motion acceleration component. Within the set width T of the future time window, the pure motion acceleration component is double-integrated using a second-order Runge-Kutta integration algorithm. The integration step size is configured as a microsecond-level constant matching the high-frequency sampling period of the inertial measurement unit. The calculated three-dimensional displacement offset is added to the current absolute spatial coordinates to generate the three-dimensional coordinate prediction vector.

4. The UAV self-organizing network multipath transmission system based on multi-source sensor cross-layer coupling prediction as described in claim 1, characterized in that, The spatiotemporal graph convolutional network model includes a spatiotemporal feature fusion module and a time-preserving regression module connected in sequence. The spatiotemporal feature fusion module is configured to construct a graph adjacency matrix based on relative Euclidean distance, extract the spatial topological correlation features between each UAV node in the entire network sequence through multi-layer graph convolution operation, and process the evolution trajectory of node coordinates along the time dimension using a one-dimensional dilated convolution kernel. The hold-time regression module is configured to perform nonlinear mapping through a fully connected feedforward network based on the extracted spatiotemporal joint feature vector, and output the remaining available time between the current connected node pairs until the radio frequency signal strength decays to the disconnection threshold, which is then used as the link hold-time.

5. The UAV self-organizing network multipath transmission system based on multi-source sensor cross-layer coupling prediction as described in claim 4, characterized in that, The spatiotemporal feature fusion module constructs a weight update mechanism for a graph adjacency matrix based on relative Euclidean distance, configured to perform the following operations: Based on the received coordinate prediction vectors of the neighboring nodes, calculate the dynamic spatial physical distance d and the relative approach speed v between any two communication nodes; Based on the preset maximum communication coverage radius R, an initial connectivity Boolean value is defined. Under the condition that the dynamic spatial physical distance d is less than the maximum communication coverage radius R, the connectivity Boolean value is determined to be true. For edges that are determined to be true, a spatial weight coefficient is calculated. The spatial weight coefficient is configured as the product of a distance decay term and a velocity penalty term. The distance decay term is determined based on the ratio of (Rd) to R, and the velocity penalty term is determined by operating on a natural logarithmic decay function with a negative relative proximity velocity v as the exponent. The graph adjacency matrix is ​​generated by filling the corresponding elements of the node relationship graph with the spatial weight coefficient.

6. The UAV self-organizing network multipath transmission system based on multi-source sensor cross-layer coupling prediction as described in claim 1, characterized in that, The multipath allocation mechanism configured in the programmable gate array specifically includes: When the link hold time is greater than or equal to the preset warning threshold, the data transmission mode of a single main path is maintained; When the link hold time falls below the preset warning threshold, the underlying hardware diversion channel is triggered. According to the set source routing protocol table, the node sequence of the backup path is read, the forward error correction coding module is started to redundantly expand the original video data stream to generate the redundant coding fragment, and the redundant coding fragment is distributed to the main path and the backup path at a proportional line rate before the physical link actually drops.

7. The UAV self-organizing network multipath transmission system based on multi-source sensor cross-layer coupling prediction as described in claim 6, characterized in that, The configuration where the proportion of data packets allocated to the backup path increases exponentially with the decay of the link hold time specifically includes: Configure the basic traffic splitting start ratio as follows: P 0 And the maximum diversion limit ratio is P max ; Calculate the difference between the preset warning threshold and the currently updated link hold time to obtain the warning time depth. ; Based on the warning time depth Constructing the natural base exponential function term , where k is a pre-calibrated radical coefficient constant for the diversion; The basic splitting start ratio P 0 Multiply by the natural base exponential function term to obtain the target value of the dynamic allocation ratio; The target value of the dynamic allocation ratio exceeds the maximum diversion limit ratio. P max At that time, hardware-level truncation processing is performed on the target value of the dynamic allocation ratio, forcibly setting it to the maximum upper limit ratio of the current distribution. P max Based on this, the hardware flow table is controlled to distribute the redundant coded fragments.

8. A multipath transmission method for UAV ad hoc networks based on multi-source sensor cross-layer coupling prediction, used to configure the communication and computing payloads in a flight formation consisting of multiple UAV nodes, characterized in that... include: Three-dimensional vector data of the environmental wind field and high-frequency three-axis acceleration and angular velocity of the UAV were collected, using a three-axis Pitot tube and an inertial measurement unit. The wind field disturbance compensation acceleration is calculated based on the three-dimensional vector data. The reference dynamic state of the UAV is determined based on the high-frequency three-axis acceleration and angular velocity of the UAV. Based on the wind field disturbance compensation acceleration and the reference dynamic state of the UAV, the three-dimensional coordinate prediction vector of the node in the future time window is calculated. The radio frequency transceiver front end is controlled to broadcast the three-dimensional coordinate prediction vector and receive the corresponding coordinate prediction vector of neighboring nodes. The integrated whole network sequence is input into the spatiotemporal graph convolutional network model, and the link hold-up time of each edge is output. When the link hold time on the current main transmission path is lower than a preset warning threshold, the multipath allocation mechanism is activated. Before the physical signal of the main path is disconnected, the redundant encoded fragments of the data stream are sent to the main path and the non-intersecting backup path at the same time. The proportion of data packets allocated to the backup path increases exponentially as the link hold time decreases.

9. A communication and computing payload device, characterized in that, It includes a memory and a multi-core processor, wherein the memory stores computer execution instructions, and the multi-core processor is configured to execute, when running the computer execution instructions, the UAV self-organizing network multipath transmission method based on multi-source sensor cross-layer coupling prediction as described in claim 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the UAV self-organizing network multipath transmission method based on multi-source sensor cross-layer coupling prediction as described in claim 8.

Images