Highway bridge unmanned aerial vehicle cluster cooperative detection planning method based on edge computing

Abstract

The present application relates to the field of edge computing, and discloses a highway bridge unmanned aerial vehicle cluster cooperative detection planning method based on edge computing, comprising the following steps: generating a bridge three-dimensional structure shadow matrix and injecting a link state database; dynamically upgrading a resource-rich unmanned aerial vehicle to a micro-edge node to construct a hybrid computing power network; based on a computing power service intention, combining a channel state and a computing power load to calculate a joint measurement cost, routing detection data to a first target node; predicting a trajectory cutting into a physical shielding area to trigger cross-layer linkage, and proactively migrating computing network context data units to a second target node; degrading feature extraction in the physical shielding area, and sending to the second target node after driving out to complete fusion. The present application solves the communication disconnection and computing power interruption problem caused by bridge entity shielding, avoids on-board buffer overflow and downtime, and realizes seamless cross-layer pre-takeover and spatial continuous identification of defect detection tasks.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, specifically to a collaborative detection and planning method for highway bridge unmanned aerial vehicle (UAV) swarms based on edge computing. Background Technology

[0002] Using drone swarms for routine inspections and defect detection of highway bridges is a common practice. Bridge inspection typically requires high-frequency acquisition of high-resolution images and 3D point cloud data, resulting in massive amounts of data. The onboard computing power and battery capacity of a single drone are insufficient to meet the demands of long-term local processing. Therefore, the current standard solution is to introduce an edge computing architecture, transmitting the inspection data collected by the drones via wireless networks to edge computing nodes deployed around the bridge or on the ground for collaborative processing, thereby alleviating the resource pressure on the drones.

[0003] However, bridges are large and complex three-dimensional structures, and their massive solid components can physically block radio signals. When drones conduct close-range reconnaissance flights under bridges or inside complex components, they often enter areas of signal obstruction, causing sudden interruptions in the wireless communication link between the drone and external edge computing nodes.

[0004] Existing computational offloading schemes and network routing mechanisms are primarily designed for open airspace and lack the ability to anticipate physical environmental obstructions. In such scenarios, once a communication link is lost, the original computational offloading process will directly fail, forcing the interruption of detection tasks being executed on edge nodes and rendering previously consumed computational resources useless. Simultaneously, the drone continues to perform data collection tasks while in the blind zone. Since it cannot transmit data outwards, a large amount of raw, high-dimensional data accumulates in the drone's local storage. This not only easily leads to onboard buffer overflows causing system failures but also results in a backlog of processing detection data within the blind zone. When the drone follows its flight path out of the obstruction zone and restores communication, the system typically needs to reallocate computational tasks and start transmitting and processing from scratch. This repeated interruption and restart causes significant time delays, making it impossible to guarantee the spatial continuity of bridge defect detection results. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a collaborative detection and planning method for highway bridge drone swarms based on edge computing. This method solves the technical problems of easy congestion of computing power of a single edge node and lack of environmental awareness in conventional routes when drones inspect complex bridges, as well as wireless communication interruption caused by physical components, which in turn leads to the failure of unloading tasks and discontinuous data processing.

[0006] To achieve the above objectives, the present invention provides a collaborative detection and planning method for highway bridge drone swarms based on edge computing, applicable to a system comprising multiple data acquisition drones and fixed edge nodes. The method includes: A 3D structural shadow matrix of the bridge is generated based on the model of the bridge to be detected, and this matrix is ​​injected into the link-state database of the underlying network protocol of the system. Specifically, a 3D architectural information model of the highway bridge to be detected is obtained as the model of the bridge. Combining the frequency diffraction characteristics of the wireless communication equipment, the detection spatial domain is discretized and voxel-mapped using the ray tracing principle to generate a 3D structural shadow matrix of the bridge that maps the physical occlusion attributes of the bridge entity. This 3D structural shadow matrix is ​​injected into the link-state database of the underlying network protocol as a spatial routing table, enabling nodes in the hybrid computing network to maintain spatial prior data containing 3D coordinate indices and corresponding physical occlusion Boolean values.

[0007] Based on the resource status of each data acquisition drone, at least one data acquisition drone is dynamically upgraded to a micro-edge node and jointly constructs a hybrid computing network with fixed edge nodes. The system monitors the remaining battery energy and available computing resources of each data acquisition drone in real time as its resource status. If the remaining battery energy and available computing resources of any data acquisition drone are both greater than the corresponding preset threshold, the corresponding data acquisition drone broadcasts its computing resource identifier to other network nodes in the system through the underlying control plane signaling, automatically upgrading its network role to a micro-edge node.

[0008] Based on the computing power service intent of detection data collected by the target data acquisition drone among multiple data acquisition drones, a joint metric cost is calculated by combining channel state and computing power load. This joint metric cost is then used to route the detection data to the first target node in the hybrid computing power network for processing. The detection data is encapsulated into data packets, with the computing power service intent encapsulated in the header. This intent includes the computing power type identifier required for the detection task and the estimated computing power demand. Candidate target nodes in the hybrid computing power network are determined. For any candidate target node, its corresponding joint metric cost is obtained by weighted summation of the channel cost function and the computing power load cost function calculated to that candidate target node. The formula for calculating the joint metric cost is as follows:

[0009] in, Indicates arrival at the candidate target node The joint metric cost; and The weighting coefficients are assigned by the system based on the sensitivity of the current detection task to transmission delay and computing power. This represents the channel cost function, the value of which is related to the arrival time at the candidate target node. The packet loss rate of the communication link is positively correlated with the received signal strength indication and negatively correlated with the received signal strength indication. The computational load cost function is represented by the candidate target nodes. The saturation of computing resources is determined after accepting the estimated computing power demand. The calculation of computing power resource saturation is based on the ratio of the sum of the currently occupied computing power of the candidate target node and the estimated computing power demand to its total available computing power. Based on the calculated joint metric cost of each candidate target node in the hybrid computing power network, the underlying network protocol executes a dynamic anycast addressing mechanism to directly route the detection data to the candidate target node with the lowest current joint metric cost, which is then selected as the first target node.

[0010] When the flight control system of the target data acquisition UAV, combined with the 3D shadow matrix of the bridge structure, predicts that the UAV's flight trajectory will enter the physical obstruction zone, cross-layer linkage is triggered. The detection task being executed in the first target node is encapsulated as a network context data unit and migrated to the second target node in the hybrid computing network, which is in a non-physical obstruction state. The flight control system acquires the current spatial coordinates and instantaneous velocity vector of the target data acquisition UAV in real time, and generates a predicted flight trajectory based on these coordinates. The formula for calculating the spatial coordinate integral of the predicted flight trajectory is as follows:

[0011] in, This indicates the estimated flight path at time spatial coordinates, Indicates the current spatial coordinates. This represents the instantaneous velocity vector. Within a preset prediction time window, integral collision detection is performed on the estimated flight trajectory, comparing it with the bridge's 3D structure shadow matrix in the link state database of the underlying network protocol. If the spatial coordinates on the estimated flight trajectory correspond to an occlusion state in the bridge's 3D structure shadow matrix, it is determined that the target data acquisition UAV will enter the physical occlusion zone.

[0012] When it is determined that the UAV will enter a physically obstructed area, the flight control system sends a spatial entry trigger primitive directly to the underlying network protocol via a cross-layer communication interface. Upon receiving the spatial entry trigger primitive, the underlying network protocol switches its routing mode for the detected data to a proactive pre-routing mode. The first target node suspends its processing and packages the incomplete intermediate data packets and the algorithm state machine progress into a computational network context data unit. Before the network connection is lost when the target data acquisition UAV enters the physically obstructed area, the underlying network protocol selects a second target node from among the nodes in the hybrid computing network that are not physically obstructed, based on the recalculated joint metric cost, and sends the computational network context data unit to the second target node via control plane signaling to achieve task pre-takeover.

[0013] The target data acquisition drone performs local degradation feature extraction within the physical obstruction area to obtain incremental feature data. After leaving the physical obstruction area, it sends the incremental feature data to a second target node, which then fuses the incremental feature data with the computing network context data unit to output the detection and identification results. Specifically, after the target data acquisition drone enters the physical obstruction area and disconnects from the hybrid computing network, it automatically activates the local computing power to perform degradation processing. By reducing the data sampling rate or truncating the number of layers in the deep feature extraction network, the acquired detection data is compressed into a feature vector sequence, which is then used as cached incremental feature data. When the target data acquisition drone leaves the physical obstruction area according to its flight trajectory and restores network connectivity, it sends the cached incremental feature data to the second target node that has taken over the computing network context data unit. The second target node activates the merging operator and, based on the algorithm state machine progress recorded in the computing network context data unit, performs channel-dimensional concatenation and incremental feature mapping calculations on the computing network context data unit and the received incremental feature data to output the bridge defect detection and identification results.

[0014] This invention provides a collaborative detection and planning method for highway bridges using unmanned aerial vehicle (UAV) swarms based on edge computing. It offers the following advantages: 1. This invention injects the three-dimensional structural shadow matrix of a bridge into the underlying network protocol, enabling network nodes to possess spatial prior perception capabilities. This mechanism can proactively avoid sudden physical obstructions caused by complex bridge structures, effectively solving the problems of data loss and network paralysis that are prone to occur when drones are in transit, and significantly improving communication reliability.

[0015] 2. This invention dynamically upgrades resource-rich UAVs into micro-edge nodes and employs a combined cost of channel quality and computing load for routing. This breaks through the limitations of traditional routing that relies solely on signal strength, accurately prevents congestion of detection data on a single node, achieves global balancing of heterogeneous computing power and bandwidth, and significantly reduces system processing latency.

[0016] 3. To address the network outage blind zone, this invention utilizes trajectory prediction to trigger cross-layer linkage. Before the network outage, the task context is migrated; during the outage, local degradation feature extraction is performed; and after connection is restored, only incremental features are sent for result stitching. This design achieves seamless task relay both inside and outside the blind zone, ensuring the continuity and integrity of bridge detection results. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figure 2 This diagram illustrates the application scenarios and network topology of the cluster collaborative detection system of the present invention. Figure 3 This is a schematic diagram illustrating the generation and mapping principle of the three-dimensional bridge structure shadow matrix of the present invention. Figure 4 This is a diagram illustrating the service intent-based joint cost routing logic and packet structure of the present invention. Figure 5 This is a timing diagram of the cross-layer linkage and detection task takeover migration of the present invention; Figure 6 This is a network structure diagram of the feature downgrading extraction and channel fusion inside and outside the blind zone in this invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see the appendix Figure 1 -Appendix Figure 6 This invention provides a collaborative detection and planning method for highway bridge drone swarms based on edge computing, applicable to systems containing multiple data acquisition drones and fixed edge nodes.

[0020] In this system, a physical execution and computing environment is constructed by multiple drones performing high-frequency image or point cloud acquisition and fixed edge servers deployed on infrastructure surrounding the bridge. To accurately quantify the system topology, the set of physical nodes of the data acquisition drones in the system is defined as... ,in The total number of nodes for the data acquisition drone; the fixed edge node set is defined as... ,in This represents the total number of fixed edge nodes connected to the system. All nodes within the system are equipped with wireless radio frequency transceiver front-ends and airborne / edge processing units, relying on wireless communication links for service data interaction and underlying signaling synchronization.

[0021] The edge computing-based UAV swarm collaborative detection and planning method for highway bridges may include the following steps: A 3D structural shadow matrix of the bridge is generated based on the model of the bridge to be detected, and this matrix is ​​injected into the link state database of the underlying network protocol of the system. The system extracts the spatial coordinates of physical components such as piers and box girders by parsing the 3D architectural information model of the highway bridge to be detected, and maps them into spatial prior topology data readable by the network layer. Based on this, the underlying network protocol maintains a global 3D spatial occlusion matrix, enabling the communication network plane to directly obtain obstacle attributes from the physical environment, breaking down the physical isolation between the spatial and communication layers.

[0022] Based on the resource status of each data acquisition drone, at least one of the data acquisition drones is dynamically upgraded to a micro-edge node, and together with the fixed edge node, a hybrid computing network is constructed. The system continuously monitors the set. The remaining battery power and idle processor cycles of each drone are monitored. When a specific data acquisition drone meets the set upgrade threshold, a network role transition is triggered via control plane broadcast, shifting it from a simple computing power consumer to a computing power supplier. At this point, the data is collected... The fixed edge nodes and the micro edge nodes that have completed their role upgrades are combined to form a heterogeneous hybrid computing network that supports task offloading.

[0023] Based on the computing power service intent of the detection data collected by the target data acquisition drone among the multiple data acquisition drones, a joint metric cost is calculated by combining channel state and computing power load. This joint metric cost is then used to route the detection data to the first target node in the hybrid computing power network for processing. The target data acquisition drone generates raw sensor data while performing its detection task. The network protocol stack parses the computing power requirement parameters encapsulated in the header of this data packet and detects the real-time RF signal quality and computing queue length of each candidate node in the entire network. By calculating the two-dimensional joint metric cost, the underlying routing mechanism directly addresses the data packet to the first target node corresponding to the minimum cost, executing defect identification or feature extraction algorithms.

[0024] When the flight control system of the target data acquisition UAV, in conjunction with the 3D shadow matrix of the bridge structure, predicts that the flight trajectory of the target data acquisition UAV will enter the physical obstruction zone, cross-layer linkage is triggered. The detection task being executed in the first target node is encapsulated into a computing network context data unit and migrated to the second target node in the hybrid computing network that is in a non-physical obstruction state. The flight control system acquires its own velocity and position vectors in real time and compares them with the shadow matrix in the link state database. Before determining that a physical obstruction disconnection is about to occur, the flight control system sends a cross-layer primitive to the underlying network protocol. After receiving the instruction, the first target node suspends its current computing process, packages the intermediate state data and algorithm state machine progress in memory, and relays them to the second target node in the open airspace.

[0025] The target data acquisition UAV performs local downgraded feature extraction within the physical obstruction area to obtain incremental feature data. After leaving the physical obstruction area, it sends the incremental feature data to the second target node, which then fuses the incremental feature data with the computing network context data unit to output the detection and identification result. During the period when the external network connection is cut off within the physical obstruction area, the UAV uses its local onboard chip to perform dimensionality reduction processing and cache the compressed incremental features. After the network connection is restored, the target data acquisition UAV sends the cached data to the second target node, activates the fusion operator to splice it with the migrated context data unit, completes the computing power continuity, and finally outputs a seamlessly spliced ​​bridge defect detection and identification result.

[0026] In this embodiment, the steps of generating a 3D structural shadow matrix of the bridge based on the model of the bridge to be detected and injecting the 3D structural shadow matrix of the bridge into the link state database of the underlying network protocol of the system are described in detail. The system first acquires the 3D architectural information model of the highway bridge to be detected and extracts the inherent geometric structure and spatial distribution parameters of the physical components of the bridge. Based on the set spatial resolution, the system divides the target detection spatial domain of the bridge into a discrete 3D voxel mesh. For any voxel unit within the spatial domain, the system assigns it an independent 3D spatial coordinate system index. .

[0027] Furthermore, the system combines the carrier frequency diffraction characteristics of the internal wireless communication equipment and uses the ray tracing principle to calculate the signal transmission reachability of each voxel unit. By simulating the emission of electromagnetic rays and calculating their penetration loss and diffraction attenuation when passing through solid bridge piers or box girders, the system generates a three-dimensional structural shadow matrix of the bridge that maps the physical occlusion properties of the bridge entity. .

[0028] In this matrix, the occlusion state of each discrete voxel is precisely calibrated using a Boolean function. The specific mapping relationship is expressed as follows:

[0029] in, Indicates spatial coordinate index as The boolean value for physical occlusion of a voxel unit; This represents the predicted comprehensive path loss of the radio frequency signal reaching the voxel unit, calculated based on ray tracing. This represents the critical path loss threshold that enables the system's wireless radio frequency front-end to maintain an effective communication connection. When the calculated Boolean value is 1, the voxel coordinate region is determined to be in a physically blocked state; when the Boolean value is 0, it is determined to be in an open and unblocked state.

[0030] The system uses the generated 3D bridge structure shadow matrix as a spatial routing table and statically injects it into the link-state database of the underlying network protocol. This operation ensures that each node in the hybrid computing network maintains spatial prior data containing 3D coordinate indices and corresponding physical occlusion Boolean values ​​before initiating routing, thus laying the foundation for subsequent communication blind spots caused by spatial physical occlusion.

[0031] This invention provides a detailed explanation of the mechanism for dynamically upgrading at least one data acquisition drone to a micro-edge node based on the resource status of each data acquisition drone. During the lifecycle of the collaborative detection operation, the system monitors the resource status of each data acquisition drone in real time through an onboard diagnostic interface. This resource status is specifically quantified as the current remaining battery energy and available computing power.

[0032] definition For the first A data acquisition drone is in constant motion The remaining battery energy state value, For the drone at any time Available computing resources that are idle and ready for external unloading. Simultaneously, the system is configured with a preset threshold for basic energy to maintain a safe return. The minimum available computing power threshold for undertaking effective computing tasks. .

[0033] The system performs real-time condition determination on the above parameters through the onboard processing unit. If the monitoring indicators of any data acquisition UAV meet the following constraints: and

[0034] The system then determines that the corresponding data acquisition drone currently has sufficient surplus resources.

[0035] Once the above conditions are met, the corresponding data acquisition drone immediately initiates the network role transition process. This node broadcasts an announcement message carrying a computing resource identifier to all other network nodes in the system via lower-level control plane signaling, declaring its computing power availability. After completing the network layer announcement, the data acquisition drone automatically upgrades its network topology role to a micro-edge node. Its idle computing power is then registered in the system's global resource pool, establishing a logical peer relationship with the original fixed edge nodes to jointly construct a hybrid computing network with distributed processing capabilities.

[0036] This embodiment elaborates on the routing addressing mechanism that combines the computing power service intent based on detection data with the joint metric cost calculation of channel state and computing power load. After the target data acquisition UAV acquires high-resolution images or point cloud detection data of the bridge surface, the protocol stack slices and encapsulates them into standardized data packets. Subsequently, the system explicitly encapsulates the corresponding computing power service intent in the header field of the data packet. This computing power service intent includes an identifier of the computing power type required for the current detection task and the estimated computing power requirement to complete the detection task.

[0037] After receiving a data packet to be sent, the underlying network protocol selects computing units with matching processing capabilities from the hybrid computing power network based on the computing power type identifier, constructing a set of candidate target nodes for the current route. For any candidate target node in this set, the system needs to evaluate the combined link cost and queuing cost of routing the data packet to that node and completing the computation. To this end, the system introduces a joint metric cost for quantification, which is obtained by weighted summation of the channel cost function and the computing power load cost function calculated to the candidate target node.

[0038] Specifically, reaching the The formula for calculating the joint metric cost of candidate target nodes is as follows:

[0039] in, Indicates arrival at the candidate target node The joint metric cost; Represents the channel cost function; Represents the computing power load cost function; and These represent the weight coefficients for the corresponding evaluation dimensions.

[0040] In this invention, the channel cost function This accurately characterizes the negative impact of space communication link transmission quality on data packet delivery delay. The function is determined by the arrival time at candidate target nodes. The current packet loss rate of the wireless communication link is jointly determined by the received signal strength indicator. The specific calculation logic is constructed to be strictly positively correlated with the packet loss rate and negatively correlated with the received signal strength indicator. High packet loss rate and low received signal strength will trigger the retransmission mechanism of the underlying link, resulting in a significant increase in the channel cost function value, which will then be rejected by the system in the metric routing.

[0041] Computing load cost function This characterizes the processor queuing latency when the candidate target node executes the unloading task. This function is determined by the candidate target node. The saturation of computing resources after receiving data packets and estimating computing power demand directly determines the system's ability to acquire candidate target nodes. The amount of computing resources currently occupied in the queue is summed with the estimated computing power requirement in the header of the aforementioned data packet, and then divided by the total physical available computing power of the candidate target node to calculate the computing power resource saturation. As this saturation approaches the computing performance bottleneck, the computing power load cost function value will increase non-linearly.

[0042] Before making routing decisions, the system adaptively allocates weighting coefficients for the channel cost function and the computing load cost function based on the service characteristics of the current detection data. When a specific defect detection task is highly sensitive to air interface transmission delay, the system adjusts the weighting coefficients accordingly. The value is chosen to favor finding high-quality channels; when the detection task has the characteristics of high concurrency and large-scale matrix operations, and is highly sensitive to the throughput of the computing resource pool, the system will increase the value. The value of is determined by this dynamic configuration mechanism, which ensures accurate load balancing of detection data for different service intentions across heterogeneous resources in the cluster.

[0043] The joint metric cost of each candidate target node in the hybrid computing network is calculated. The routing control plane of the system's underlying network protocol activates and executes a dynamic anycast addressing mechanism. The underlying network protocol performs numerical sorting and comparison within the candidate target node set, directly routing the detection data to the candidate target node with the lowest calculated joint metric cost. The system then designates this node with the lowest cost as the first target node to undertake this detection task, receiving the data and initiating subsequent processing.

[0044] This embodiment details the mechanism by which the target data acquisition UAV predicts its flight trajectory and triggers cross-layer linkage for detection task migration. When performing close-range detection operations, the flight control system of the target data acquisition UAV acquires its current absolute spatial coordinates and instantaneous three-dimensional velocity vector in real time through its inertial navigation and positioning module. Using these physical kinematic parameters, the system continuously integrates the coordinates of the future spatial travel path within a preset prediction time window. The formula for calculating the spatial coordinate integral of the predicted flight trajectory is as follows:

[0045] in, This indicates the estimated flight trajectory at the integral future time. Three-dimensional spatial coordinates; Indicates the current spatial coordinates of the target data acquisition drone; Indicates that drones are in The instantaneous velocity vector at time t.

[0046] After acquiring the continuous predicted flight trajectory, the flight control system executes a proactive integral collision detection algorithm on the underlying computational plane. The system extracts discretized spatial coordinates on the predicted flight trajectory and compares them with the 3D bridge structure shadow matrix residing in the link state database of the underlying network protocol, mapping them spatially. If the Boolean value obtained from querying the shadow matrix corresponds to an occlusion state, the system determines that the target data acquisition UAV is about to enter the physical obstruction zone that would cause a communication link interruption.

[0047] In this invention, when the system determines that it is about to enter a physically obstructed area, the flight control system immediately invokes the cross-layer communication interface, bypassing the traditional layer-by-layer interaction logic of the network protocol stack, and directly sends a spatial entry trigger primitive to the control plane of the underlying network protocol. After receiving the hardware-level trigger primitive, the underlying network protocol forcibly blocks the conventional on-demand routing addressing process and switches the routing mode of the detection data produced by the UAV for the target data acquisition to a forward-looking pre-routing mode.

[0048] In proactive pre-routing mode, upon receiving the cooperative signaling, the first target node, which was originally tasked with defect identification and processing, immediately suspends the current processing on its computing core. The first target node extracts the intermediate data packets that have not yet completed forward propagation computation, the network layer cache queue, and the precise execution progress of the current feature extraction algorithm state machine from memory, and rigorously packages them into a highly cohesive computing network context data unit.

[0049] Before the target data acquisition UAV actually enters the physical obstruction zone due to physical inertia and the radio frequency network connection is lost, the underlying network protocol filters out safe nodes in the global hybrid computing network whose current location attributes are in a non-physical obstruction state. Based on the real-time channel and load status of these safe nodes, the system recalculates the joint metric cost, thereby establishing the second target node corresponding to the minimum cost. The first target node then sends the encapsulated computing network context data unit to the second target node via a high-priority control plane signaling channel, achieving seamless pre-takeover of the detection task before the communication link is lost.

[0050] This embodiment elaborates on the local degradation processing mechanism of the target data acquisition UAV within the physical obstruction zone and the fusion calculation process after it leaves the obstruction zone. When the target data acquisition UAV actually enters the physical obstruction zone and loses its external radio frequency connection with the hybrid computing network, the original cooperative offloading link is physically blocked. To prevent buffer overflow of the high-frequency acquired raw image or point cloud data in the onboard memory and cause system crash, the underlying control logic of the target data acquisition UAV automatically wakes up the local onboard processing unit to perform local degradation feature extraction operations on the continuously generated detection data.

[0051] During local degradation feature extraction, the system proactively reduces the sensor spatial sampling rate and dynamically truncates the number of deep feature extraction network layers running on the airborne side, based on the limited computational overhead and energy consumption constraints of the airborne processing unit. Assume the total number of layers in the complete detection network model originally intended to be executed at the first target node is... The target data acquisition drone only retains and activates the data in the blind zone. A shallow network structure of layers, in which Through this truncated network, the input high-dimensional raw detection data is compressed and mapped into low-dimensional incremental feature data. The specific dimensionality reduction mapping calculation relationship is expressed as follows:

[0052] in, This represents the extracted incremental feature data sequence; This represents the raw, high-dimensional detection data collected by the target data acquisition drone within the physically obstructed area; Indicates from the previous Degradation feature extraction function composed of layered networks; Indicates the previous The pre-trained weight parameter set corresponding to the layer network. The incremental feature data sequence generated after mapping processing is statically stored in the airborne security cache queue, waiting for the restoration of the external communication connection status.

[0053] In this invention, as the target data acquisition UAV continues to fly out of the physical obstruction zone due to physical inertia, its onboard radio frequency front-end re-captures the broadcast signals of the system network nodes, and the topology of the underlying physical link is restored to a connected state. After the underlying network protocol stack detects the restoration of the network heartbeat, it immediately queries the suspended migration task table to extract the network address of the second target node that has pre-taken over the computation task. Subsequently, the target data acquisition UAV actively pulls the incremental feature data sequence temporarily stored in the onboard cache queue. And then send it to the second target node.

[0054] Upon receiving the incremental feature data, the second target node immediately invokes the merging operator process resident in the system kernel. The merging operator first parses the network context data units migrated from the previous cross-layer migration, extracting the algorithm state machine execution progress and intermediate state data feature map of the first target node at the time of network interruption. Subsequently, according to the state machine progress instructions, the system performs rigorous spatial alignment and splicing mapping of the received incremental feature data sequence and the unpacked intermediate state data feature map along the channel dimension of the feature tensor. The fused feature matrix after channel splicing is sequentially injected into the subsequent unexecuted computation graph nodes of the algorithm state machine to continue the deep forward propagation of higher-order features. Through this cross-node and cross-temporal incremental fusion mechanism, the system outputs complete and spatially continuous bridge defect detection and identification results, completely avoiding computational power interruptions and detection gaps caused by spatial occlusion.

[0055] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A collaborative detection and planning method for highway bridge UAV swarms based on edge computing, applied to a system containing multiple data acquisition UAVs and fixed edge nodes, characterized in that... include: A three-dimensional structural shadow matrix of the bridge is generated based on the model of the bridge to be detected, and the three-dimensional structural shadow matrix of the bridge is injected into the link state database of the underlying network protocol of the system. Based on the resource status of each of the data acquisition drones, at least one of the data acquisition drones is dynamically upgraded to a micro-edge node, and together with the fixed edge node, a hybrid computing network is constructed. Based on the computing power service intent of the detection data collected by the target data acquisition drone among the multiple data acquisition drones, the joint metric cost is calculated by combining the channel state and computing power load. The detection data is then routed to the first target node in the hybrid computing power network for processing using the joint metric cost. When the flight control system of the target data acquisition drone, in conjunction with the three-dimensional structure shadow matrix of the bridge, predicts that the flight trajectory of the target data acquisition drone will enter the physical occlusion area, cross-layer linkage is triggered. The detection task being executed in the first target node is encapsulated as a computing network context data unit and migrated to the second target node in the hybrid computing network that is in a non-physical occlusion state. The target data acquisition drone performs local degradation feature extraction within the physical obstruction area to obtain incremental feature data. After leaving the physical obstruction area, it sends the incremental feature data to the second target node, which then fuses the incremental feature data with the computing network context data unit to output the detection and recognition result.

2. The method for collaborative detection and planning of UAV swarms for highway bridges based on edge computing as described in claim 1, characterized in that, The step of generating a three-dimensional structural shadow matrix of the bridge based on the model of the bridge to be detected, and injecting the three-dimensional structural shadow matrix of the bridge into the link state database of the underlying network protocol of the system, specifically includes: A three-dimensional architectural information model of the highway bridge to be detected is obtained as the model of the bridge to be detected. Combining the frequency diffraction characteristics of the wireless communication device, the detection spatial domain is discretized and voxel-mapped using the ray tracing principle to generate the three-dimensional structural shadow matrix of the bridge that maps the physical occlusion attributes of the bridge entity. The three-dimensional structure shadow matrix of the bridge is injected as a spatial routing table into the link state database of the underlying network protocol, so that the nodes in the hybrid computing network maintain spatial prior data containing three-dimensional coordinate indexes and corresponding physical occlusion Boolean values.

3. The method for collaborative detection and planning of UAV swarms for highway bridges based on edge computing as described in claim 1, characterized in that, The step of dynamically upgrading at least one of the data acquisition drones to a micro-edge node based on the resource status of each of the data acquisition drones specifically includes: The remaining battery energy and available computing power of each of the data acquisition drones are monitored in real time as the resource status. If the remaining battery energy and available computing power of any of the data acquisition drones are both greater than the corresponding preset threshold, the corresponding data acquisition drone broadcasts a computing resource identifier to other network nodes in the system through the underlying control plane signaling, and automatically upgrades its network role to the micro-edge node.

4. The method for collaborative detection and planning of UAV swarms for highway bridges based on edge computing according to claim 1, characterized in that, The computing power service intent based on the detection data collected by the target data acquisition drone, combined with channel state and computing power load to calculate the joint metric cost, specifically includes: The detection data is encapsulated into a data packet, and the computing power service intent is encapsulated in the header of the data packet. The computing power service intent includes the computing power type identifier required by the detection task and the estimated computing power requirement. Candidate target nodes in the hybrid computing network are determined; for any candidate target node, the corresponding joint metric cost is obtained by weighted summation of the channel cost function and the computing load cost function calculated to that candidate target node; The channel cost function is determined by the packet loss rate and received signal strength indication of the communication link to the candidate target node; the computing power load cost function is determined by the computing power resource saturation of the candidate target node after it undertakes the estimated computing power demand; and the weight coefficients of the channel cost function and the computing power load cost function are adaptively allocated according to the sensitivity of the current detection task to transmission delay and computing power.

5. The method for collaborative detection and planning of UAV swarms for highway bridges based on edge computing according to claim 4, characterized in that, The step of routing the detection data to the first target node in the hybrid computing network for processing using the joint metric cost specifically includes: Based on the calculated joint metric cost of each candidate target node in the hybrid computing network, the underlying network protocol executes a dynamic anycast addressing mechanism to directly route the detection data to the candidate target node with the lowest current joint metric cost, which will then be used as the first target node.

6. The method for collaborative detection and planning of UAV swarms for highway bridges based on edge computing according to claim 1, characterized in that, The predicted flight trajectory of the target data acquisition drone will enter the physical obstruction area, specifically including: The flight control system acquires the current spatial coordinates and instantaneous velocity vector of the target data acquisition UAV in real time; Based on the current spatial coordinates and instantaneous velocity vector, a predicted flight trajectory is generated, and an integral collision detection is performed on the predicted flight trajectory within a preset prediction time window. The predicted flight trajectory is then compared and calculated with the bridge three-dimensional structure shadow matrix in the link state database of the underlying network protocol. If the spatial coordinates on the estimated flight trajectory correspond to an occlusion state in the three-dimensional shadow matrix of the bridge structure, then it is determined that the target data acquisition UAV will enter the physical occlusion area.

7. The method for collaborative detection and planning of UAV swarms for highway bridges based on edge computing according to claim 6, characterized in that, The triggering of cross-layer linkage specifically includes: When it is determined that the flight control system will enter the physical obstruction area, the flight control system directly sends a space entry trigger primitive to the underlying network protocol through the cross-layer communication interface. After receiving the spatial cut-in trigger primitive, the underlying network protocol switches the routing mode for the detection data to the forward pre-routing mode.

8. The method for collaborative detection and planning of UAV swarms for highway bridges based on edge computing according to claim 1, characterized in that, The step of encapsulating the detection task being executed in the first target node into a network context data unit and migrating it to the second target node in the hybrid computing network that is in a non-physically obscured state specifically includes: The first target node pauses its processing and packages the incomplete intermediate data packets and the algorithm state machine progress into the computing network context data unit. Before the network connection is lost when the target data acquisition UAV enters the physical obstruction area, the underlying network protocol selects the second target node from among the nodes in the hybrid computing network that are in a non-physical obstruction state, based on the recalculated joint metric cost, and sends the computing network context data unit to the second target node through control plane signaling to achieve task pre-takeover.

9. The method for collaborative detection and planning of UAV swarms for highway bridges based on edge computing according to claim 1, characterized in that, The local degradation feature extraction specifically includes: After the target data acquisition drone enters the physical obstruction area and disconnects from the hybrid computing network, it automatically activates the local computing power to perform a degradation processing mechanism, compresses the acquired detection data into a feature vector sequence, and uses the feature vector sequence as the cached incremental feature data.

10. The method for collaborative detection and planning of UAV swarms for highway bridges based on edge computing according to claim 9, characterized in that, The step of sending the acquired incremental feature data to the second target node, whereby the second target node fuses the incremental feature data with the computing network context data unit and outputs the detection and recognition result, specifically includes: After the target data acquisition drone leaves the physical obstruction area according to the flight trajectory and restores the network connection, it will send the cached incremental feature data to the second target node that has taken over the computing network context data unit. The second target node activates the merging operator, concatenates the computational network context data unit with the received incremental feature data and performs incremental calculations, and outputs the bridge defect detection and identification results.

Images