Multi-device cooperative dynamic path planning algorithm in complex environment of edge computing

Abstract

This invention discloses a multi-device collaborative dynamic path planning algorithm for complex environments in edge computing, specifically relating to the field of neural network technology. The algorithm includes collaborative perception data alignment modeling, collaborative spatiotemporal relationship graph construction, collaborative intent tensor generation, candidate path cluster generation and executable path corridor formation, collaborative adjudication and conflict resolution, edge rolling complex planning, and model self-calibration. Edge coordination nodes uniformly model the status, tasks, and environmental information of multiple devices. They extract collaborative relationships between devices through graph neural networks and combine this with collaborative intent tensors to predict future occupancy relationships, generating candidate path clusters and executable path corridors. Based on collaborative adjudication scores, they allocate actions such as priority passage, delayed entry, detour switching, or time-sharing passage. Rolling complex planning and incremental model updates are implemented based on complex planning trigger values, thereby improving the continuity, adaptability, and execution stability of multi-device collaborative path planning in complex environments.

Description

Technical Field

[0001] This invention relates to the field of neural network technology, and more specifically, to a multi-device collaborative dynamic path planning algorithm for complex environments in edge computing. Background Technology

[0002] Existing edge computing technologies for multi-device collaborative path planning in complex environments are widely used in warehousing and logistics, park delivery, inspection operations, and the increasingly developing low-altitude economy, such as multiple drones performing inspection, delivery, or monitoring tasks in urban low-altitude corridors. This type of technology typically involves each target device uploading its local sensing data, positioning information, and task information to an edge node. The edge node then synchronizes time and unifies spatial coordinates, constructs a planning space based on the fused environmental information, and generates execution paths for each target device through multi-agent path search, priority ranking, or time slot allocation. After the paths are distributed, the target devices execute according to the planned paths. When changes in obstacles, device offsets, or environmental conditions are detected, the edge node re-triggers path planning to achieve collaborative operation of multiple devices in a shared space.

[0003] The shortcomings of existing technologies are as follows: First, existing methods mostly rely on spatial proximity or fixed priorities for collaborative judgment, lacking a unified way to express the temporal occupancy relationships, task overlap relationships, and communication status changes among multiple devices in low-altitude channels, narrow corridors, or complex intersection areas, resulting in insufficient basis for collaborative decision-making. Second, existing methods usually directly output a single path during the path generation stage. In dynamic environments, when temporary obstacles or changes in channel congestion occur, the path lacks adjustment space, requiring frequent overall path recalculation, which affects the continuity of execution. Third, in the process of conflict resolution and execution adjustment, existing methods mainly rely on simple rules or local obstacle avoidance strategies, with insufficient comprehensive consideration of task time limits, device status, and path quality. At the same time, they have limited utilization of feedback information during execution, making it difficult to continuously optimize the collaborative effect in complex environments or scenarios with dense operation of multiple devices in low-altitude environments. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, this invention provides a multi-device collaborative dynamic path planning algorithm for complex edge computing environments, which solves the problems mentioned in the background art through the following scheme.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a multi-device collaborative dynamic path planning algorithm for complex edge computing environments, comprising:

[0006] The edge coordination node receives local status packets sent by each target device in the target device set, performs timestamp alignment, coordinate unification and missing status item filling on the local status packets, constructs the node feature vector corresponding to each target device, and establishes a sliding time-series buffer of fixed length.

[0007] The edge coordination node constructs a collaborative spatiotemporal relationship graph based on the node feature vectors corresponding to each target device. It establishes graph edge connections based on the predicted motion envelope intersection, the possibility of meeting in the channel where the target point is located, the cross-occupancy of the task area, and the link switching association. It then performs graph neural network relationship encoding based on the node feature vectors and edge feature vectors to obtain the graph encoding state vector corresponding to each target device.

[0008] The edge coordination node performs temporal occupancy prediction for future time windows based on graph-encoded state vectors and spatial segment sets, generates a coordination intent tensor, and outputs the coordination occupancy prediction results for each target device. The coordination occupancy prediction results include a list of conflict spatial segments, priority entry ranking results, and recommended action modes.

[0009] The edge coordination node generates candidate path clusters for each target device based on the collaborative occupancy prediction results. It performs time parameterization processing and path evaluation on each candidate path in the candidate path cluster, retains the top few candidate paths with the best evaluation values ​​as the main candidate path set, and extracts common passable areas based on the main candidate path set to form the executable path corridors corresponding to each target device.

[0010] The edge coordination node performs spatiotemporal intersection analysis on the executable path corridors corresponding to each target device, generates a set of conflict segments, and calculates the collaborative decision score of each target device based on the current task priority, expected conflict duration, optimal candidate path evaluation value, remaining target arrival time limit, remaining power value, link stability, and priority entry strength. Based on the collaborative decision score and recommended action mode, it assigns priority passage, delayed entry, detour switching, or time-sharing passage actions, and outputs the main execution path, alternative path set, and conflict zone action label corresponding to each target device.

[0011] Edge coordination nodes generate short-term execution trajectory segments based on the main execution path and distribute them to each target device. During execution, they collect actual trajectory segments, actual speed segments, information on newly appearing obstacles, records of actual conflicts, records of task changes, and records of link changes. They calculate the complex planning trigger amount. When the complex planning trigger amount exceeds the complex planning threshold, they re-execute the construction of the collaborative spatiotemporal relationship graph, the generation of collaborative intent tensors, the generation of candidate path clusters and the formation of executable path corridors, as well as collaborative adjudication and conflict resolution. When the complex planning trigger amount does not exceed the complex planning threshold, they maintain the current executable path corridor and continue to distribute short-term execution trajectory segments. At the same time, they use the prediction results and the actual execution results to perform model self-calibration.

[0012] The technical effects and advantages of this invention are as follows:

[0013] 1. This solution constructs a collaborative spatiotemporal relationship graph, which integrates the position, motion, task, and link states of each target device in the target device set, as well as the relative relationships between devices, into a graph structure. This graph structure is then combined with graph neural network encoding to form a graph-encoded state vector. This allows edge coordination nodes to obtain the spatial competition, task intersection, and link association relationships between multiple devices before path generation. Furthermore, this solution uses a collaborative intent tensor to uniformly analyze the probability of shared occupancy, priority entry, yielding, and parallel passage of each spatial segment within a future time window. This expands the multi-device collaborative relationship from static proximity determination to a structured expression oriented towards future temporal occupancy, improving the completeness and coherence of the basic information for collaborative decision-making in complex environments.

[0014] 2. In the path generation stage, this scheme does not directly use a single discrete path as the final output. Instead, it first generates a cluster of candidate paths and then extracts a common passable area from the main candidate path set to form an executable path corridor. Since the executable path corridor includes the corridor centerline, left and right boundaries, time window, maximum allowable speed and upper limit of turning change, when the target device encounters local disturbances during execution, it can complete short-term correction within the current corridor without immediately abandoning the entire path planning result. Combined with the collaborative adjudication and conflict resolution processing in step S5, the edge coordination node can also perform targeted action allocation between priority passage, detour switching, delayed entry and time-sharing passage.

[0015] 3. In the execution phase, this solution further introduces collaborative adjudication scores and replanning triggers. Task urgency, remaining energy and motion sustainability, optimal candidate path evaluation value, priority entry intensity, link stability and local autonomy are jointly incorporated into the conflict adjudication process. During execution, actual trajectory segments, actual speed segments, information on newly appearing obstacles, actual conflict occurrence records, task change records and link change records are continuously collected to determine whether rolling replanning is triggered. At the same time, the edge coordination node uses the prediction results and the actual execution results to form training samples and performs small-batch incremental updates on the model used in steps S2 and S3, so that the model parameters can gradually adapt to the traffic patterns, congestion patterns and equipment motion differences in the current deployment environment. Attached Figure Description

[0016] Figure 1 This is the overall flowchart of the multi-device collaborative dynamic path planning of the present invention.

[0017] Figure 2 This is a flowchart of the collaborative spatiotemporal relationship diagram and collaborative intent tensor generation process of the present invention.

[0018] Figure 3 This is a flowchart of the candidate path cluster decision and rolling complex planning process of the present invention. Detailed Implementation

[0019] The technical solutions of 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.

[0020] refer to Figures 1-3 The multi-device collaborative dynamic path planning algorithm shown in the image, applicable to complex edge computing environments, includes:

[0021] Step S1: Collaborative Sensing Data Alignment Modeling: This step unifies the local sensing data, positioning data, task data, and environmental data aggregated by edge coordination nodes generated by target devices within the same planning cycle into standardized data units that can be input into the neural network. First, timestamp unification, coordinate system unification, device state completion, and target semantic mapping are completed, forming the node foundation data required for the subsequent collaborative spatiotemporal relationship graph. The output data directly serves as the graph construction input in Step S2, therefore, it is necessary to ensure that the state expression dimension, semantic meaning, and time sampling window of each target device are fixed.

[0022] At the start of planning period t, the target equipment set Each target device in The local computing unit generates a corresponding local status packet and sends it to the edge coordination node e. The local status packet encapsulates the current position coordinates, current attitude angle, current velocity vector, current acceleration vector, body shape envelope parameters, remaining battery power, current task identifier, task priority, target point coordinates, local obstacle grid segments, and local link quality indicators in a fixed field order. The edge coordination node receives and registers the local status packets reported by each target device, and writes the target device number, reception time, and planning period number into each local status packet, ensuring that status data from different target devices within the same planning period enters a unified processing queue.

[0023] The edge coordination node first performs a timestamp alignment operation in the unified processing queue, using the unified planning reference time corresponding to the planning cycle as the alignment reference. For local state packets earlier than the unified planning reference time, it performs state extrapolation; for local state packets later than the unified planning reference time, it stops writing to the current cycle and transfers them to the processing queue for the next planning cycle. The state extrapolation is based on the target device. The most recent effective position, velocity vector, and acceleration vector calculations are used to determine the current position and velocity at the unified planning reference time. The extrapolated results replace the original timestamp's corresponding position and velocity fields, ensuring that all target devices... It has a unified state expression at a given moment within the planning period.

[0024] After completing timestamp alignment, the edge coordination node performs a coordinate unification operation. This operation uses the global reference coordinate system maintained by the edge coordination node as the unique coordinate benchmark for each target device. The current position coordinates, target point coordinates, local obstacle grid segments, and body shape envelope parameters are uniformly mapped.

[0025] Given the current location coordinates and the target point coordinates, the edge coordination node determines the corresponding target device... The installation pose calibration parameters and the transformation relationship from the local coordinate system to the global reference coordinate system are used to complete the coordinate transformation; for local obstacle grid segments, the edge coordination node maps each occupied cell in the grid segment to the corresponding position in the global reference coordinate system, and performs coordinate transformation on adjacent target devices. The reported overlapping obstacle regions are merged for consistency; for the body shape envelope parameters, the edge coordination node maps the front boundary, rear boundary, left boundary, and right boundary of the body to the envelope expression in the global reference coordinate system to ensure that different target devices are considered in the subsequent path planning stage. The passage boundaries adopt a unified geometric benchmark.

[0026] After the obstacle region consistency merging is completed, the edge coordination node assigns a unique obstacle number to each independent obstacle target in the global reference coordinate system and writes the obstacle number back to each target device. The corresponding local obstacle grid segment records enable subsequent planning cycles to track changes in the spatial location of obstacles based on the same obstacle number.

[0027] After coordinate unification is completed, the edge coordination node performs fill-in on missing state items in the local state packet. The edge coordination node is for each target device. Maintain the valid state sequence of the three most recent planning periods. When any fixed field is missing within a planning period, retrieve the data from the target device. Extract historical values ​​of the same field, associated values ​​of adjacent fields, and changing trends during the planning period from the effective state sequences of the last three planning cycles. Generate the filler value for the current planning cycle according to the fixed interpolation rules, and write the generated filler value into the corresponding field position.

[0028] If the current location coordinates are missing, the edge coordination node calculates the current location coordinates based on the most recent valid location and the most recent valid velocity vector; if the current velocity vector is missing, the edge coordination node calculates the current velocity vector based on the displacement changes of the last two planning cycles; if the task priority is missing, the edge coordination node writes a fixed priority value into the task configuration table corresponding to the current task identifier; if the local link quality indicators are missing, the edge coordination node reads the target device. The most recent valid link quality metric is recorded and written to the current cycle record; in the case of missing local obstacle grid segments, the edge coordination node retrieves information from adjacent target devices. The unified obstacle grid results within the corresponding area are used to form replacement fragments, and the target device is retained in the replacement fragment record. Numbering and alternative source markers. The local state packet after completion maintains a fixed field order, fixed data dimensions, and fixed semantic meaning; it does not rewrite field names or rearrange field order.

[0029] After all fields are completed, the edge coordination node provides each target device with... Constructing node feature vectors The node feature vector is formed by sequentially concatenating position features, velocity features, heading features, mission features, aircraft size features, energy features, link features, environmental congestion features, and local passable direction features. Specifically, the position feature consists of the current position coordinates and the target point coordinates; the velocity feature consists of the current velocity vector and the current acceleration vector; the heading feature consists of the current attitude angle and its corresponding direction encoding; the mission feature consists of the current mission identifier and mission priority; the aircraft size feature consists of the aircraft's shape envelope parameters; the energy feature consists of the remaining battery power and the estimated range marker; the link feature consists of the local link quality index and its stability encoding; and the environmental congestion feature consists of the obstacle density and target device within a predetermined range around the current position. The density statistics are composed of local passable direction features, which are composed of angular interval encoding of passable directions in local obstacle grid segments.

[0030] The edge coordination node performs field order verification, dimension verification, and number binding on the node feature vector to target the device. Number i and node feature vector Establish a one-to-one correspondence and assign the feature vectors of the nodes that pass the verification. Write the node feature set for the planning period t.

[0031] In order to enable the subsequent construction of the collaborative spatiotemporal relationship graph to utilize the state evolution information within the continuous planning cycle, the edge coordination node e is for each target device. A sliding time-series buffer of length L is established. The sliding time-series buffer stores the target device sequentially according to the planned cycle order. In the node feature vectors of the most recent L planning periods, when a new node feature vector is written, the node feature vector corresponding to the earliest planning period is removed from the sliding time-series cache, thus ensuring that the sliding time-series cache always maintains a fixed length.

[0032] When writing to the sliding time-series cache, the edge coordination node simultaneously records the planning cycle number, original local state packet number, completion mark, and coordinate unification mark corresponding to each node's feature vector, so as to ensure that subsequent steps can trace the source of the node's feature vector generation.

[0033] Step S1 is used to complete the target device. gather The collaborative sensing data alignment and modeling processing within the planning period t outputs a model that can be directly used in step S2 to construct a collaborative spatiotemporal relationship graph. Node feature vectors used And its corresponding sliding time-series cache records.

[0034] Step S2: Construction of Cooperative Spatiotemporal Relationship Graph: Based on the node basic data output in Step S1, a cooperative spatiotemporal relationship graph is constructed that can simultaneously express the relative positional relationships, relative motion relationships, target conflict relationships, and link cooperative relationships of devices. The cooperative spatiotemporal relationship graph is a dynamic graph structure oriented towards cooperative path planning, where nodes correspond to target devices, and edges correspond to the coupling relationships between devices during the planning cycle. After being encoded by a graph neural network, the output cooperative spatiotemporal relationship graph generates a cooperative intent tensor for Step S3. Provides basic representation.

[0035] During the planning period, the edge coordination node reads the node feature vectors output in step S1 and the corresponding sliding temporal cache records, and constructs a collaborative spatiotemporal relationship graph using the target device set as the node set. .

[0036] The edge coordination node first determines each target device A node index table is established based on the corresponding current position coordinates, current velocity vector, current attitude angle, target point coordinates, task features, link features, and environmental congestion features. Each record in the node index table is fixedly bound to the target device number, node feature vector, planning cycle number, and sliding time-series cache entry address to ensure that each node in the collaborative spatiotemporal relationship graph corresponds to a unique target device. .

[0037] After the edge coordination node completes the node index table, it uses the target device. With target equipment Based on pairwise combinations, the coupling relationship between devices is filtered, and for any two target devices... and If any of the following conditions are met: the predicted motion envelope has spatial intersection within the future window, the target point's channel may have passing traffic, the task area has overlapping occupancy, or the link switching may cause incomplete information and require enhanced collaborative association, the edge coordination node will establish a graph edge connection from node i to node j in the collaborative spatiotemporal relationship graph, and simultaneously establish a graph edge connection from node j to node i, so that the collaborative spatiotemporal relationship graph maintains a bidirectional expression of the relationship between devices.

[0038] When performing the predicted motion envelope crossover determination, the edge coordination node calls the target device. and target equipment Using the current position coordinates, current velocity vector, current acceleration vector, and body shape envelope parameters within the planning period, the spatial occupancy area within the future fixed prediction window is discretized. If the body shape envelope areas corresponding to the two overlap at any discrete moment, the target device is marked as having a predicted motion envelope intersection. When performing the target point's channel meeting determination, the edge coordination node maps the coordinates of the two target points to the corresponding channel numbers based on the channel segment index in the global reference coordinate system. If the two target devices... and If the target points are located in the same width-constrained channel segment and move in opposite directions, then the target device pair is marked as potentially meeting; when performing task area cross-occupancy determination, the edge coordination node maps the current task identifier to the corresponding task area set. If two target devices... and If the task regions of a target device pair intersect, the pair is marked as having overlapping task regions. During link switching association determination, the edge coordination node compares the local link quality indicators and stability codes of the two target devices. If either target device experiences a link quality degradation within the current planning period and its region is adjacent to the planned region of the other target device, the pair is marked as requiring a link association edge. The edge coordination node writes all target device pairs that meet any of the determination conditions into an edge candidate set, and then generates the edge set of the collaborative spatiotemporal relationship graph from the edge candidate set.

[0039] After the set of edges in the collaborative spatiotemporal graph is determined, the edge coordination nodes construct an edge feature vector for each graph edge. Edge feature vectors The relative distance, relative speed, relative heading difference, predicted earliest conflict time, predicted shared area, task priority difference, link quality difference, and channel width constraint are encapsulated in a fixed field order.

[0040] The relative distance is determined by the target device. With target equipment The current position coordinate difference in the global reference coordinate system is calculated; the relative speed is calculated from the difference in the current speed vectors of the two; the relative heading difference is composed of the current attitude angle difference of the two and its direction interval encoding; the earliest predicted conflict time is determined by the discrete time when the motion envelopes first intersect within the future fixed prediction window; the predicted shared area is determined by the area of ​​the overlapping region of the body shape envelopes of the two at the corresponding discrete time; the task priority difference is directly calculated from the task priorities of the two target devices; the link quality difference is composed of the difference in the local link quality index and the difference in stability encoding of the two target devices; the channel width constraint encoding is formed by the width level of the channel segment where the future passage area of ​​the two is located and the meeting capacity status encoding.

[0041] The edge coordination node generates edge feature vectors. At the same time, the edge number, the starting node number i, the ending node number j, the planning period number t, and the edge generation reason mark are recorded so that each graph edge can be traced back to its specific creation conditions.

[0042] The edge coordination node e in the collaborative spatiotemporal relationship graph After construction is complete, the node feature vectors Edge feature vectors and each target device Neighbor set The input graph neural network encodes relations. The graph neural network employs a graph attention structure with edge-gated features, and within a planning period t, it encodes the target device... The node state is updated, and the updated graph-encoded state vector is generated. Calculate according to the following formula:

[0043]

[0044] in, Indicates the target device The graph-encoded state vector in the planning period t. This represents the activation function. Represents the node feature mapping matrix. Represents the edge feature mapping matrix. Indicates the target device A set of connected neighbors. Indicates the target device For target equipment Attention weights. Edge coordination node e traverses its neighbor set according to the graph edge connections. First, examine the edge feature vectors. Perform mapping, then apply attention weights. Aggregate neighbor influences to target devices. Graph-encoded state vector It also includes the status information of this device and the coupling impact information of adjacent target devices. This is to ensure that each target device... All node updates are completed within a unified semantic space, and the edge coordination node e shares the same set of node feature mapping matrices with all target devices. Edge feature mapping matrix and activation function .

[0045] The attention weight From the target device Node feature vectors Target equipment Node feature vectors and the edge feature vector between the two The formula is determined together as follows:

[0046]

[0047] in, Represents the attention projection vector. , , This represents the attention mapping matrix. The edge coordination node e first converts the node feature vector... , and edge eigenvectors Projected onto the same attention computation space, and then processed through hyperbolic tangent transformation and exponential normalization, the target device is obtained. For target equipment Attention weights This distinguishes the impact intensity of different neighboring target devices on the current target device. For graph edges that predict the earliest conflict time, predict a larger shared area, have stronger task area cross-occupancy, or have more obvious link switching correlation, their edge feature vectors are... This results in a greater weight contribution in attention calculations, thus giving the corresponding neighboring target device a higher proportion of influence in node updates.

[0048] During the execution of the graph neural network, the edge coordination nodes further utilize the sliding time-series cache records output from step S1 to continuously enhance the graph construction results of the current planning cycle. For the target device... The edge coordination node reads the historical sequence of its node feature vectors from the most recent L planning periods, performs temporal concatenation before encoding the graph in the current period, and appends the temporal concatenation result to the input of the current node's feature vector, so that the current node carries historical state evolution information during relation encoding. If the edge coordination node detects the target device in the current planning period... If the local link quality index drops to the preset switching condition, resulting in the current local status packet of the target device being incomplete or not arriving in time, the edge coordination node will enable local graph completion processing.

[0049] In the local graph completion process, the edge coordination node is located from the target device. Read the most recent valid node feature vector from the sliding time-series cache record. The most recent valid graph-encoded state vector, combined with the motion extrapolation results of the current planning cycle, generates a temporary node feature vector, which is then used to replace the graph encoding of the missing input in the collaborative spatiotemporal relationship graph. This temporary node feature vector is then used in relation to the target device. For connected graph edges, edge coordination nodes update their corresponding edge feature vectors based on the most recent valid edge feature vector and the current state of neighboring target devices, maintaining the integrity of the edge set and neighbor set. After local graph completion, the collaborative spatiotemporal relationship graph can still complete the unified relationship encoding of all target devices within the current planning cycle, avoiding the interruption of input for subsequent collaborative intent analysis due to link fluctuations of individual target devices.

[0050] After the graph neural network completes the relation encoding, each target device at the edge coordination node... Output the corresponding graph-coded state vector And the target device number i, planning period number t, and node feature vector are... Neighbor set Related edge feature vectors and graph-coded state vector A binding relationship is established and written into the graph coding result set of the current planning cycle. Each record in the graph coding result set retains a unique target device number, a unique planning cycle number, and a unique graph coding state vector, enabling step S3 to directly read the graph coding state vector. And continue to generate collaborative intent tensors Thus, step S2 completes the collaborative spatiotemporal relationship diagram. The construction and graph neural network relation encoding process output a graph encoded state vector that can be used in step S3. Records of their corresponding relationships.

[0051] Step S3: Cooperative Intent Tensor Generation: This step transforms the graph-encoded state vector obtained in Step S2 into a structured occupancy prediction result that can directly participate in cooperative decision-making. This step further analyzes the entry order, dwelling trend, yielding tendency, and parallel traversal tendency of each target device in multiple spatial segments within a future time window, forming a cooperative intent tensor. This tensor is a core intermediate variable for subsequent steps S4 to generate candidate path clusters and S5 to perform collaborative decision-making.

[0052] During the planning period, the edge coordination node reads the graph-coded state vector and its corresponding records output in step S2, and records all target devices. The graph-encoded state vectors are written into the collaborative intent analysis input set according to the target device number order. The edge coordination node first performs spatial segmentation processing on the planning area, dividing the global reference coordinate system coverage area into a set of fixed-numbered spatial segments. Each spatial segment corresponds to a unique segment number z, and each spatial segment records the channel type, area boundary, passage width, and allowed passage direction, enabling subsequent occupancy prediction and collaborative relationship analysis to be performed under a unified spatial indexing system. The current location coordinates and target point coordinates are mapped to the corresponding spatial segment numbers, and an index relationship is established between the target device number and the spatial segment number for subsequent occupancy prediction result write-back.

[0053] After the spatial segmentation set is constructed, the edge coordination node takes the graph-coded state vector as input and performs a process for each target device. Perform time-series occupancy prediction for future time windows. The future time window consists of a set of discrete moments of fixed length H, where each discrete moment is denoted as H. The edge coordination node invokes the time-series prediction network deployed on the edge side for each target device. At each discrete time The occupancy status of the space segments is predicted to obtain the occupancy probability, entry order probability, and speed change trend. Here, the occupancy probability represents the target device... At discrete time The probability of entering spatial segment z, and the probability of entering the target device. Compared to the order of entry of other target devices in the same spatial segment z, the velocity change trend indicates the target device's... The direction and magnitude of velocity changes before and after entering the spatial segment z.

[0054] During the prediction process, the edge coordination node inputs the graph-encoded state vector together with the historical node feature vector in the sliding time-series cache record in step S1, so that the prediction result can reflect the continuous changes in the state of the target device.

[0055] After obtaining the occupancy prediction results of all target devices, the edge coordination node e, for any two target devices... and In each spatial segment z and at each discrete time... The probability of shared occupancy, priority entry, yielding, and parallel passage are calculated.

[0056] The probability of co-occupation From the target device and target equipment In spatial segmentation z and discrete time... The occupancy probability is obtained by joint calculation; the priority entry probability is obtained by... Indicates the target device Prioritized over the target device in this spatial segmentation and discrete time. Probability of entering; probability of yielding Indicates the target device Delayed entry at this spatial segmentation and discrete time to allow the target device to pass. The probability of parallel passage; Indicates the target device With target equipment The probability of simultaneous passage within this spatial segment is determined by combining the output of the time-series prediction network with the channel width of the spatial segment, the envelope of the target device's shape, and the predicted velocity change trend. Edge coordination nodes impose consistency constraints on each probability value, ensuring consistency within the same spatial segment z and at discrete times. The probabilities corresponding to different action patterns maintain a uniform distribution range.

[0057] After completing the calculation of each probability quantity, the edge coordination node e generates a coordination intent tensor according to a unified definition. Any element in the collaborative intent tensor Calculate using the following formula:

[0058]

[0059] in, Indicates the target device With target equipment In spatial segmentation z and discrete time... The strength of the cooperative intent under the following conditions Indicates the probability of shared occupancy. Indicates the target device Priority entry probability, Indicates the target device Probability of yielding Indicates the probability of parallel passage. , , This represents the weighting coefficients of the three intent components and is fixed during system deployment.

[0060] The edge coordination node connects all target devices. , All spatial segments z and all discrete time points Calculate in sequence The calculation results are then written into the collaborative intent tensor according to the target device number, spatial segment number, and discrete time index. The corresponding positions are used to form a complete multidimensional structured expression of collaborative relationships.

[0061] After the edge coordination node completes the construction of the coordination intent tensor, it performs a traversal analysis of the tensor to extract each target device. A list of conflict space segments within a future time window. This list consists of all space segments that exceed a set conflict threshold, and records the corresponding discrete time range and the target device number j involved. The edge coordination node further, within each conflict space segment z, according to... The size of the target device is used to sort the multiple target devices involved, generate a priority sorting result, and bind the sorting result with the corresponding discrete time range for use in subsequent path generation and conflict resolution stages.

[0062] The edge coordination node for each target device Comparing the probabilities of three types of action patterns in the conflict space segmentation, when When the target device's recommended operating mode in that spatial segment remains above the entry threshold within the corresponding spatial segment and discrete time range, it is marked as having priority passage; when When the distance remains consistently above the yield threshold within the corresponding range, the recommended action mode will be marked as delayed entry; when When the corresponding spatial segment satisfies the channel width constraint and remains above the parallel threshold, the recommended action mode is marked as parallel passage; when none of the above three probabilities meet the corresponding threshold and there is an alternative path, the recommended action mode is marked as detour switching.

[0063] After generating the conflict space segmentation list, priority entry ranking results, and recommended action mode tags, the edge coordination node reorganizes these results according to the target device number into a single-device collaborative occupancy prediction result. Each target device... The cooperative occupancy prediction result includes a fixed set of conflict space segment numbers, corresponding discrete time ranges, priority entry sorting positions, and recommended action mode markers. The edge coordination node binds the cooperative occupancy prediction result to the target device number i and writes it into the cooperative intent analysis result set for the planning cycle, while retaining the original data structure of the cooperative intent tensor for subsequent steps. Step S3 completes the process of generating the cooperative intent tensor from the graph-encoded state vector and outputs the cooperative occupancy prediction result, which can be directly used in step S4 for candidate path cluster generation and executable path corridor formation.

[0064] Step S4: Candidate Path Cluster Generation and Executable Path Corridor Formation: Based on the collaborative occupancy prediction results generated in Step S3, this step generates candidate path clusters and executable path corridors for each target device. Multiple candidate paths satisfying kinematic constraints, obstacle avoidance constraints, and cooperative constraints are generated, and stable spatial envelopes are extracted from the candidate paths to form an executable path corridor that can be directly deployed for execution. This step expands the path planning output from a single discrete trajectory to a corridor structure that can be continuously executed under local perturbations, thus leaving room for adjustment in the conflict resolution of step S5.

[0065] During the planning period, the edge coordination node reads the cooperative occupancy prediction results and cooperative intent tensor output in step S3, and then, for each target device in the target device set... The process involves generating candidate path clusters and forming executable path corridors. The edge coordination node first targets the device. The path planning inputs are determined, including the current location coordinates, target point coordinates, aircraft shape envelope parameters, local obstacle grid segments, environmental congestion characteristics, and the set of conflict space segment numbers and corresponding discrete time ranges recorded in the cooperative occupancy prediction results. Based on these inputs, the edge coordination node constructs a search space in the global reference coordinate system, encoding the channel boundaries, obstacle-occupied areas, and impassable areas in the spatial segment set as path search constraints, and mapping the time windows corresponding to the conflict space segments in the cooperative occupancy prediction results as temporal constraint regions, so that both spatial and temporal constraints can be satisfied during the path search process.

[0066] The edge coordination node searches for the target device within the search space. Perform multi-branch path search and generate candidate path clusters. Each candidate path in the candidate path cluster consists of a sequence of path points arranged in order, and each path point contains spatial coordinates and a corresponding estimated arrival time.

[0067] During the path search process, the edge coordination node generates three types of paths according to preset rules: the first type is a direct path, which connects the current position and the target point with the shortest distance under the conditions of satisfying obstacle avoidance constraints and body shape envelope constraints; the second type is a detour path, which introduces alternative passage segments to achieve path offset when encountering conflict space segments or highly congested areas; the third type is a buffer waiting path, which introduces a time delay segment at the entrance of the conflict space segment to achieve conflict avoidance by delaying the entry time.

[0068] When generating each candidate path, the edge coordination node performs time parameterization processing on the path point sequence, so that each node in the path point sequence is bound to a unique discrete time, thereby ensuring that the path can be matched with the coordination intention tensor at a unified time scale.

[0069] In candidate path cluster Once generated, the edge coordination node processes each candidate path. Calculate path evaluation value In calculating path length cost At that time, the edge coordination node accumulates the spatial distances between adjacent path points in the path point sequence to obtain the total path length; in calculating energy consumption cost At that time, the edge coordination node estimates the energy consumption required to execute the path based on the velocity changes corresponding to the path points and the body dynamics model; in calculating the cost of trajectory curvature changes... At that time, the edge coordination node accumulates the angle changes formed by three adjacent points in the path point sequence to obtain a path smoothness index; in calculating the conflict risk cost At that time, the edge coordination node maps the spatial segment number and discrete time index corresponding to the path point to the coordination intention tensor. For all path points shared with other target devices, the corresponding cooperative intent strength is accumulated; the execution cost of link instability is calculated. At that time, the edge coordination node determines the link quality index distribution within the path coverage area and the target device. The risk of communication interruption that may occur during the link characteristic calculation process. Path evaluation value. Calculate using the following formula:

[0070]

[0071] in, Indicates the target device The comprehensive evaluation value of the candidate path p, Indicates the path length cost. Indicates the cost of energy consumption. This represents the cost of changes in trajectory curvature. Indicates the risks and costs of conflict. This indicates the execution cost due to link instability. , , , , This represents a uniformly set evaluation coefficient.

[0072] The edge coordination node for candidate path clusters Perform the above evaluation calculation on all candidate paths and rank them according to their evaluation values. Sort the candidate paths from smallest to largest and select the top K candidate paths as the main candidate path set.

[0073] After the set of primary candidate paths is determined, the edge coordination nodes perform spatial consistency analysis on the selected paths, extract the overlapping areas of each path on the same spatial segment from the path point sequence, and perform boundary expansion processing on these overlapping areas to form a continuous passable spatial zone.

[0074] The boundary expansion process is aimed at the target device. Based on the envelope parameters of the machine's shape, the left and right boundaries of the corridor segment are obtained by extending the path centerline to both sides by a fixed safety distance. Edge coordination nodes connect the corridor segments in adjacent spatial segments according to the path sequence, forming an executable path corridor. Each corridor segment records the corridor centerline, left boundary, right boundary, corresponding time window, maximum allowable speed, and upper limit of turning change. The time window is determined by the discrete time range of the corresponding spatial segment in the path point sequence. The maximum allowable speed is determined by the speed change trend of the path point and the kinematic constraints of the body. The upper limit of turning change is determined by the cost constraint of trajectory curvature change.

[0075] During the formation of the executable path corridor, the edge coordination node writes the conflict space segment constraints in the collaborative occupancy prediction results into the corridor segment record. For the space segments with conflicts, the conflict identifier is marked in the corresponding corridor segment and the recommended action mode mark is recorded, so that the subsequent step S5 can directly read the corridor-level conflict information when making collaborative decisions.

[0076] The edge coordination node completes the task of all target devices. After constructing the executable path corridor, a unified binding relationship is established between the main candidate path set, executable path corridor, and path evaluation value ranking results for each target device, and written into the path planning result set of the planning cycle. Step S4 completes the generation of candidate path clusters and the formation of executable path corridors, and outputs path-level input data for use in step S5 to perform collaborative adjudication and conflict resolution.

[0077] Step S5: Collaborative Decision-Making and Conflict Resolution: This step is used to determine and resolve conflicts between executable path corridors of multiple target devices. It calculates the collaborative decision-making score by combining the collaborative intent tensor, candidate path scores, timing occupancy windows, task urgency, and link continuity. Then, based on the ruling, give way, detour, or time-sharing actions are assigned to each target device. The output of this step is the device-level execution decision that has completed conflict resolution, which is directly fed into step S6 for rolling replanning.

[0078] During the planning period, the edge coordination node reads the path planning result set output in step S4 and performs a path planning process on all target devices in the target device set. Executable path corridor Perform spatiotemporal intersection analysis. The edge coordination node traverses each executable path corridor in the order of planning cycle number, target device number, and corridor segment, mapping the corridor centerline, left boundary, right boundary, and corresponding time window to a unified spatiotemporal occupancy table.

[0079] For any two target devices and If its corresponding executable path corridor and If there is boundary overlap within the same spatial segment and an intersection in the corresponding time window, the edge coordination node records the overlapping area as a conflict segment and writes a conflict segment number, a set of involved target device numbers, a spatial segment number, a time window range, and a corridor segment number for that conflict segment. After the edge coordination node has traversed all executable path corridors, it forms a set of conflict segments for the current planning period and records executable path corridors that have not entered any conflict segment set. Directly mark it as a conflict-free corridor, so that the corresponding target device Keep the current main execution path unchanged.

[0080] For target devices entering the conflict fragment set The edge coordination node further summarizes the decision inputs for the target device under the current planning cycle. These decision inputs are fixed and include the current task priority, estimated conflict duration, optimal candidate path evaluation value, remaining target arrival time, remaining battery power, link stability, and priority entry strength. The current task priority is directly taken from the task characteristics in step S1; the estimated conflict duration is determined by the executable path corridor. The overlap length of the time window between the executable path corridors of other target devices and the time window within the conflict segment is determined; the evaluation value of the optimal candidate path is taken from the minimum path evaluation value corresponding to the main candidate path set in step S4; the remaining amount of the target arrival time limit is obtained by subtracting the corresponding time of the current planning cycle from the specified completion time limit in the task configuration table; the remaining power value is taken from the energy characteristics in step S1; the link stability is determined by the local link quality index and its stability encoding in step S1; the priority entry strength is determined by the cooperation intent tensor output in step S3 that is related to the target device. The corresponding priority entry probability components are accumulated.

[0081] The edge coordination node establishes a decision input record according to the target device number i, based on the aforementioned decision input, for subsequent collaborative decision scoring. Computation provides a unified data foundation.

[0082] After the adjudication input record is formed, the edge coordination node e addresses each target device entering the conflict fragment set. Calculate the collaborative adjudication score The collaborative adjudication score Calculate using the following formula:

[0083]

[0084] in, Indicates the target device The score of the collaborative ruling, Indicators representing the urgency of the task. This represents a joint indicator of surplus energy and sports sustainability. Indicates the target device The optimal candidate path evaluation value, Represented by the cooperative intent tensor The priority entry intensity of the extracted data. This represents a joint indicator of link stability and local autonomy. , , , , This represents the decision coefficient obtained from unified training.

[0085] When calculating the task urgency index, the edge coordination node combines the current task priority with the remaining time to reach the target to generate a time-limit pressure representation for the target device. When calculating the joint index of remaining energy and motion sustainability, it combines the remaining power value with the expected energy consumption corresponding to the current main execution path to generate a sustainable execution representation. When calculating the priority entry strength, it extracts the priority entry probability contribution of the target device within the spatial segment and time window corresponding to the conflict segment from the coordination intent tensor. When calculating the joint index of link stability and local autonomy, it writes the link stability of the target device and the target device's ability to continuously execute the current corridor segment locally. The edge coordination node sequentially calculates the coordination adjudication score for all target devices involved in the same conflict segment and writes the results into the conflict segment adjudication table.

[0086] After obtaining the conflict segment adjudication table, the edge coordination node sorts the multiple target devices within each conflict segment in descending order according to their collaborative adjudication scores, and performs action assignment based on the recommended action mode markers output in step S3. For the target device ranked first and whose priority entry strength exceeds the entry threshold... The edge coordination node assigns a priority passage action to the target device and maintains the time window of the target device's current main execution path within the conflict segment unchanged. For target devices that do not receive a priority passage action... The edge coordination node continues to check whether there is a detour path in its primary candidate path set. If there is a detour path and the path evaluation value corresponding to the detour path is... If the switching threshold is met, a detour switching action is assigned to the target device, and this detour path is set as the new main execution path. At the same time, the executable path corridor corresponding to the target device is reconstructed. The sequence of corridor segments following the conflict segment. If the target device... If the detour switching conditions are not met, the edge coordination node continues to determine whether it and other target devices have parallel accommodation space within the conflict segment. The determination of parallel accommodation space is based on the fact that the envelope of the two target devices' body shapes simultaneously satisfy the left and right boundary separation constraints within the spatial segment where the conflict segment is located, and the parallel passage probability corresponding to step S3 meets the threshold. When parallel accommodation space is established, the edge coordination node assigns time-sharing passage actions to the relevant target devices and divides independent time windows for each target device within the same conflict segment, so that multiple target devices occupy the same spatial segment sequentially according to the assigned time windows. For target devices that do not meet the conditions for priority passage, detour switching, and time-sharing passage, the edge coordination node assigns a delayed entry action to them and inserts a waiting period before the entry of the conflict segment into the main execution path of the target device, while adjusting the expected arrival time of the subsequent path points.

[0087] After the action assignment is completed, the edge coordination node coordinates with each target device. An execution decision record is generated after conflict resolution. This record includes a primary execution path, a set of alternative paths, conflict zone action labels, and an adjusted time window. The primary execution path is a sequence of path points adjusted through priority access maintenance, detour switching, time-sharing passage allocation, or delayed entry time. The alternative path set retains other primary candidate paths not selected as the current primary execution path in step S4. The conflict zone action label is fixed as one of four types: priority access, detour switching, time-sharing passage, or delayed entry. The adjusted time window corresponds to the actual execution sequence of the conflict segment and subsequent corridor segments. The edge coordination node binds the execution decision record to the target device number and synchronously updates the action allocation and time window fields in the conflict segment adjudication table, ensuring that the execution decision of any target device can be traced back to the corresponding conflict segment adjudication result.

[0088] After completing the action allocation for all conflict segments, the edge coordination node re-verifies all updated executable path corridors. This verification includes checking for boundary overlaps in the conflict segments, whether new time intersections occur within the adjusted time windows, and whether the main execution path after the detour still satisfies spatial segmentation constraints. If the verification results indicate that a conflict segment still has spatiotemporal overlap after action allocation, the edge coordination node recalculates the collaborative adjudication score for the corresponding target device based on the updated execution decision records and re-executes the sorting and action allocation until the conflict segment meets the non-overlapping execution condition. If the verification results indicate that the conflict segment has been resolved, the edge coordination node marks the conflict segment as adjudicated and writes all relevant target device execution decision records into the collaborative adjudication result set for the current planning cycle.

[0089] Step S5 completes the spatiotemporal conflict analysis, collaborative adjudication score calculation, action allocation and conflict resolution processing of the executable path corridor output in step S4, and outputs the main execution path, alternative path set and conflict zone action label for each target device, and uses the result as the input for step S6 to perform edge rolling complex planning and model self-calibration.

[0090] Step S6: Edge Rolling Complex Planning and Model Self-calibration: This step continuously monitors environmental changes, trajectory deviations, obstacle insertions, link fluctuations, and task changes during device execution, and adjusts the parameters based on the complex planning trigger amount. Determine whether it is necessary to re-enter steps S2 to S5 to perform rolling complex planning. This step also uses the actual execution results to perform edge-side mini-batch self-calibration of the graph neural network encoder and the collaborative intent prediction network, so that the model gradually adapts to the channel habits, device movement differences and local congestion patterns in the current deployment environment.

[0091] During the planning period, the edge coordination node reads the collaborative decision result set output in step S5 and writes the main execution path, alternative path set, conflict zone action label, and adjusted time window corresponding to each target device into the execution scheduling queue. The edge coordination node generates a short-time execution trajectory segment for each target device according to the target device number and time window order. This short-time execution trajectory segment consists of a sequence of continuous path points in the main execution path and includes the spatial coordinates, estimated arrival time, maximum permissible speed, and upper limit for steering change corresponding to each path point.

[0092] The edge coordination node sends short-time execution trajectory segments to the motion control module of the corresponding target device, enabling the target device to move according to the execution constraints issued in the current planning cycle. For target devices assigned priority passage actions, the edge coordination node issues short-time execution trajectory segments according to the time window of the original main execution path; for target devices assigned delayed entry actions, the edge coordination node writes a waiting period before issuing short-time execution trajectory segments before the entry of conflict segments; for target devices assigned detour switching actions, the edge coordination node issues short-time execution trajectory segments corresponding to the switched main execution path to the target device; for target devices assigned time-sharing passage actions, the edge coordination node writes the divided independent time windows into the timing field of the short-time execution trajectory segments, enabling the target device to enter the corresponding conflict space segment within the specified time window.

[0093] During the execution of a short-term trajectory segment by the target device, the edge coordination node continuously collects the actual trajectory segment, actual speed segment, information on newly appearing obstacles, actual conflict records, task change records, and link change records of the target device, and writes the above execution data into the execution monitoring queue according to the target device number and the planning cycle number.

[0094] The actual trajectory segments consist of a continuous sequence of position coordinates reported by the target device within the current execution cycle. The actual velocity segments consist of velocity vectors on the corresponding position coordinate sequences. New obstacle information consists of the location, boundary, and occupied space segment number of newly detected obstacles by the target device or edge coordination node within the current execution cycle. Actual conflict records consist of spatial proximity events and temporal overlap events that occur between the target device and other target devices during execution. Task change records consist of target point modification, task priority adjustment, or task status change information sent by the task allocation system. Link change records consist of the changes in local link quality indicators and their stability codes within the execution cycle. The edge coordination node performs number verification, time verification, and spatial segment mapping verification on the data in the execution monitoring queue to ensure that the calculation input of subsequent complex planning trigger quantities remains consistent with the execution status of the current planning cycle.

[0095] After the edge coordination node e completes the monitoring data verification, it performs checks on each target device. Calculate complex planning trigger quantity Complex planning trigger quantity Used to characterize the target device Under the current planning period t, is it necessary to re-enter steps S2 to S5 to perform rolling replanning? (The replanning trigger amount is missing.) Calculate using the following formula:

[0096]

[0097] in, Indicates the target device The complex planning trigger quantity in the planning period t This indicates the deviation from the trajectory. Indicates the insertion strength of the new obstacle. Indicates the intensity of link fluctuations. Indicates the intensity of task change. , , , This indicates a uniformly set trigger coefficient.

[0098] When calculating trajectory deviation, the edge coordination node aligns the actual trajectory segment of the target device with the short-term execution trajectory segment corresponding to the current main execution path point by point, and calculates the spatial offset between the two at the same expected arrival time. When calculating the insertion intensity of new obstacles, it maps the information of newly appearing obstacles to the spatial segments and time windows covered by the current main execution path, and calculates the impact of the new obstacles on the occupancy of the current executable path corridor. When calculating the link fluctuation intensity, it compares the link change record with the baseline value of the local link quality index in step S1 to obtain the degree of link fluctuation within the current execution cycle. When calculating the task change intensity, it uniformly encodes the target point modification amount, task priority adjustment amount, and task state change amount in the task change record as a task change representation. After the edge coordination node completes the calculation of each component, it substitutes them into the above formula to obtain the complex planning trigger amount of the target device.

[0099] The edge coordination node compares the replanning trigger amount with a preset replanning threshold. When the replanning trigger amount of the target device exceeds the replanning threshold, the edge coordination node marks the target device as a replanning target device and immediately starts the rolling replanning process. During the rolling replanning process, the edge coordination node retains the current execution state of the target device as the new starting state, writes the latest actual trajectory segment endpoint, actual velocity segment end velocity vector, newly appeared obstacle information, link change record, and task change record of the target device into the new node feature vector construction input, and then re-sends the updated node feature vector into step S2 to perform collaborative spatiotemporal relationship graph reconstruction, step S3 to perform collaborative intent tensor recalculation, step S4 to perform candidate path cluster regeneration and executable path corridor re-formation, and step S5 to perform collaborative adjudication and conflict resolution.

[0100] After the edge coordination node regenerates the main execution path, alternative path set, and conflict zone action label for the replanned target device, it overwrites the old record of the corresponding target device in the original execution scheduling queue and issues a new short-term execution trajectory fragment, so that the target device can switch to the updated path execution state without interrupting the current execution process.

[0101] When the replanning trigger amount of the target device does not exceed the replanning threshold, the edge coordination node marks the target device as a persistent execution target device and keeps the current executable path corridor unchanged. For persistent execution target devices, the edge coordination node does not re-execute steps S2 to S5, but instead extracts a new short-term execution trajectory segment within the subsequent time window based on the current main execution path, and sends the new short-term execution trajectory segment to the motion control module of the target device, so that the target device continues to execute along the current executable path corridor. When sending a new short-term execution trajectory segment, the edge coordination node synchronously updates the executed path point markers, remaining path point markers, and next execution cycle number in the execution scheduling queue to ensure that the time base for subsequent execution monitoring and replanning trigger amount calculation remains continuous.

[0102] After completing the complex planning judgment and execution scheduling update, the edge coordination node will form a training sample with the prediction results and the actual execution results in the current planning cycle. The training sample will include node feature vectors, edge feature vectors, graph encoding state vectors, collaborative intent tensors, main execution paths, actual trajectory segments, actual speed segments, actual conflict occurrence records, and execution result labels corresponding to the complex planning triggering quantities.

[0103] The edge coordination node writes training samples into the edge-side sample buffer and establishes an index relationship according to the target device number, planning cycle number, and execution result label. When the computational load of the edge coordination node meets the preset conditions and there is no target device in the target device set that is undergoing complex planning, the edge coordination node reads a fixed batch of training samples from the edge-side sample buffer and performs small-batch incremental updates on the graph neural network used in step S2 and the time-series prediction network used in step S3.

[0104] During the incremental update process, the edge coordination node keeps the input field definition, node feature vector structure, edge feature vector structure, and collaborative intent tensor structure unchanged, and only updates the network parameters in the unified training settings. The updated parameter version number is written into the model version record table to ensure that the model called in the next planning cycle is consistent with the congestion mode, channel passage mode, and target device movement differences in the current edge environment.

[0105] Once the edge coordination node has completed all the execution, monitoring, replanning determination, and model incremental update of the current planning cycle, it marks the execution status of the planning cycle as the end state and starts the processing flow of the next planning cycle.

[0106] When starting the next planning cycle, the edge coordination node inherits the latest actual state of the current target device as the initial state input for the next planning cycle, and saves the sample index, model version number, and execution scheduling queue status formed in the current planning cycle in the cycle switching record. Step S6 completes the edge rolling complex planning and model self-calibration processing, enabling the target device to respond to environmental changes, trajectory deviations, obstacle insertions, link fluctuations, and task changes based on the complex planning trigger values ​​during execution, and continuously adjusts the model used in steps S2 and S3 on the edge coordination node based on the actual execution results.

[0107] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0108] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-device collaborative dynamic path planning algorithm for complex edge computing environments, characterized in that, include: The edge coordination node receives local status packets sent by each target device in the target device set, performs timestamp alignment, coordinate unification and missing status item filling on the local status packets, constructs the node feature vector corresponding to each target device, and establishes a sliding time-series buffer of fixed length. The edge coordination node constructs a collaborative spatiotemporal relationship graph based on the node feature vectors corresponding to each target device. It establishes graph edge connections based on the predicted motion envelope intersection, the possibility of meeting in the channel where the target point is located, the cross-occupancy of the task area, and the link switching association. It then performs graph neural network relationship encoding based on the node feature vectors and edge feature vectors to obtain the graph encoding state vector corresponding to each target device. The edge coordination node performs temporal occupancy prediction for future time windows based on graph-encoded state vectors and spatial segment sets, generates a coordination intent tensor, and outputs the coordination occupancy prediction results for each target device. The coordination occupancy prediction results include a list of conflict spatial segments, priority entry ranking results, and recommended action modes. The edge coordination node generates candidate path clusters for each target device based on the collaborative occupancy prediction results. It performs time parameterization processing and path evaluation on each candidate path in the candidate path cluster, retains the top few candidate paths with the best evaluation values ​​as the main candidate path set, and extracts common passable areas based on the main candidate path set to form the executable path corridors corresponding to each target device. The edge coordination node performs spatiotemporal intersection analysis on the executable path corridors corresponding to each target device, generates a set of conflict segments, and calculates the collaborative decision score of each target device based on the current task priority, expected conflict duration, optimal candidate path evaluation value, remaining target arrival time limit, remaining power value, link stability, and priority entry strength. Based on the collaborative decision score and recommended action mode, it assigns priority passage, delayed entry, detour switching, or time-sharing passage actions, and outputs the main execution path, alternative path set, and conflict zone action label corresponding to each target device. Edge coordination nodes generate short-term execution trajectory segments based on the main execution path and distribute them to each target device. During execution, they collect actual trajectory segments, actual speed segments, information on newly appearing obstacles, records of actual conflicts, records of task changes, and records of link changes. They calculate the complex planning trigger amount. When the complex planning trigger amount exceeds the complex planning threshold, they re-execute the construction of the collaborative spatiotemporal relationship graph, the generation of collaborative intent tensors, the generation of candidate path clusters and the formation of executable path corridors, as well as collaborative adjudication and conflict resolution. When the complex planning trigger amount does not exceed the complex planning threshold, they maintain the current executable path corridor and continue to distribute short-term execution trajectory segments. At the same time, they use the prediction results and the actual execution results to perform model self-calibration.

2. The multi-device collaborative dynamic path planning algorithm for complex environments in edge computing according to claim 1, characterized in that: When the edge coordination node performs timestamp alignment on the local state packet, it uses the unified planning reference time as the alignment reference. For local state packets earlier than the unified planning reference time, it performs state extrapolation based on the most recent valid position, the most recent valid velocity vector, and the current acceleration vector. For local state packets later than the unified planning reference time, it transfers them to the next planning cycle for processing. When the edge coordination node performs coordinate unification, it uses the global reference coordinate system as the only coordinate reference to perform unified mapping of the current position coordinates, target point coordinates, local obstacle grid segments, and body shape envelope parameters, and assigns a unique obstacle number to each independent obstacle target. When the edge coordination node performs missing state item completion, it calls the valid state sequence of the last three planning cycles to interpolate the missing current position coordinates, current velocity vector, task priority, local link quality index and local obstacle grid fragments. Then, it constructs the node feature vector according to the fixed field order and writes the node feature vector into the sliding time-series cache according to the planning cycle order.

3. The multi-device collaborative dynamic path planning algorithm for complex environments in edge computing according to claim 1, characterized in that: The collaborative spatiotemporal relationship graph uses each target device in the target device set as a node set. When any two target devices meet any of the following conditions: the predicted motion envelope has spatial intersection in the future window, there is a possibility of passing each other in the channel where the target point is located, there is cross-occupancy in the task area, or the link switching may lead to incomplete information, a bidirectional graph edge connection is established between the corresponding nodes. The edge feature vector includes at least the relative distance, relative speed, relative heading difference, predicted earliest conflict time, predicted shared area, task priority difference, link quality difference, and channel width constraint encoding. When the edge coordination node detects link fluctuations that cause the current local state packet to be incomplete or not to arrive in time, it reads the most recent valid node feature vector and the most recent valid graph encoding state vector from the sliding time-series buffer, combines them with the motion extrapolation results of the current planning cycle to generate a temporary node feature vector, and uses the temporary node feature vector to participate in the graph neural network relationship encoding of the collaborative spatiotemporal relationship graph.

4. The multi-device collaborative dynamic path planning algorithm for complex environments in edge computing according to claim 1, characterized in that: Before generating the collaborative intent tensor, the edge coordination node first divides the planning area into a set of spatial segments. Each spatial segment corresponds to a unique segment number and records the channel type, area boundary, passage width, and allowed passage direction. The edge coordination node uses the graph-coded state vector as input to perform temporal occupancy prediction on the spatial segment occupancy of each target device at each discrete moment in the future time window, and obtains the occupancy probability, entry order probability and speed change trend. The edge coordination node calculates the probability of shared occupancy, priority entry, yielding, and parallel passage for any two target devices at any spatial segment and at any discrete time, and generates a coordination intent tensor accordingly. The edge coordination node extracts a list of conflict space segments, priority entry ranking results, and recommended action modes for each target device based on the collaborative intent tensor. The recommended action modes are fixed as priority passage, delayed entry, detour switching, and parallel passage.

5. The multi-device collaborative dynamic path planning algorithm for complex environments in edge computing according to claim 1, characterized in that: When generating candidate path clusters, the edge coordination node uses the current location coordinates of the target device as the starting point and the target point coordinates as the ending point to construct a search space in the global reference coordinate system that includes both spatial and temporal constraints. The spatial constraints include at least the channel boundaries, obstacle-occupied areas, and impassable areas in the spatial segment set, and the temporal constraints include at least the time windows corresponding to the conflict spatial segments in the collaborative occupancy prediction results. The candidate path cluster includes at least direct paths, detour paths, and buffer waiting paths. Each candidate path consists of a sequence of path points arranged in order, and each path point is bound to spatial coordinates and an estimated arrival time. The edge coordination node calculates the path length cost, energy consumption cost, trajectory curvature change cost, conflict risk cost, and link instability execution cost for each candidate path, and retains the top few candidate paths with the best evaluation values ​​as the main candidate path set according to the comprehensive evaluation results.

6. The multi-device collaborative dynamic path planning algorithm for complex environments in edge computing as described in claim 5, characterized in that: The edge coordination node extracts a common passable area based on the overlapping area of ​​each candidate path in the same spatial segment in the main candidate path set, and expands the boundary of the common passable area based on the body shape envelope parameters of the target device to form an executable path corridor. The executable path corridor is composed of continuous corridor segments. Each corridor segment includes a corridor centerline, left boundary, right boundary, corresponding time window, maximum allowable speed, and upper limit of steering change. For spatial segments with conflicts, the edge coordination node writes a conflict identifier and a recommended action mode marker into the corresponding corridor segment, and then outputs the executable path corridor after establishing a binding relationship with the main candidate path set.

7. The multi-device collaborative dynamic path planning algorithm for complex environments in edge computing according to claim 1, characterized in that: When the edge coordination node performs spatiotemporal intersection analysis on the executable path corridor, it maps the corridor centerline, left boundary, right boundary and corresponding time window of each target device to a unified spatiotemporal occupancy table. When the executable path corridors corresponding to any two target devices have overlapping boundaries in the same spatial segment and the corresponding time windows intersect, the overlapping area is recorded as a conflict segment, and the conflict segment is written with conflict segment number, set of target device numbers involved, spatial segment number, time window range and corridor segment number. The edge coordination node summarizes the current task priority, expected conflict duration, optimal candidate path evaluation value, remaining target arrival time limit, remaining power value, link stability, and priority entry strength for each target device entering the conflict segment, generates a decision input record for each target device, and calculates the collaborative decision score based on the decision input record.

8. The multi-device collaborative dynamic path planning algorithm for complex environments in edge computing according to claim 7, characterized in that: When the edge coordination node assigns actions based on the collaborative decision score and the recommended action pattern, the target devices in the same conflict segment are sorted according to the collaborative decision score. For target devices ranked first and whose priority entry strength exceeds the entry threshold, a priority passage action is assigned; for target devices that do not receive a priority passage action but have a detour path in the main candidate path set and meet the switching threshold, a detour switching action is assigned, and the corresponding detour path is set as the new main execution path; for target devices that meet the parallel capacity conditions and whose parallel passage probability meets the threshold, a time-sharing passage action is assigned, and an independent time window is divided within the same conflict segment; for target devices that do not meet the aforementioned action assignment conditions, a delayed entry action is assigned, and a waiting period is inserted before the entry of the conflict segment. After the action allocation is completed, the edge coordination node generates an execution decision record containing the main execution path, the set of alternative paths, the action labels of the conflict area, and the adjusted time window. It then performs a verification on the updated executable path corridor again until the conflict fragments meet the non-overlapping execution condition.

9. The multi-device collaborative dynamic path planning algorithm for complex environments in edge computing according to claim 1, characterized in that: During the execution of short-term trajectory segments by each target device, the edge coordination node continuously collects actual trajectory segments, actual speed segments, information on newly appearing obstacles, records of actual conflicts, records of task changes, and records of link changes, and calculates the complex planning trigger amount based on the trajectory deviation, the strength of new obstacle insertion, the strength of link fluctuation, and the strength of task change. When the number of complex planning triggers exceeds the complex planning threshold, the current execution state of the target device is retained as the new starting state. The latest actual trajectory segment endpoint, the actual speed segment end velocity vector, newly appeared obstacle information, link change record and task change record are written into the new node feature vector construction input. The collaborative spatiotemporal relationship graph construction, collaborative intent tensor generation, candidate path cluster generation and executable path corridor formation, as well as collaborative adjudication and conflict resolution are re-executed. When the number of complex planning triggers does not exceed the complex planning threshold, the current executable path corridor remains unchanged and new short-term execution trajectory fragments continue to be issued; The edge coordination node also forms training samples from node feature vectors, edge feature vectors, graph encoding state vectors, collaborative intent tensors, main execution paths, actual trajectory segments, actual speed segments, actual conflict occurrence records, and execution result labels. When the computational load meets preset conditions, it performs small-batch incremental updates on the graph neural network and the temporal prediction network to complete model self-calibration.

Images